id
stringlengths 33
33
| doc_url
stringlengths 33
33
| title
stringlengths 13
211
| publication_date
timestamp[s] | update_date
timestamp[s] | authors
stringlengths 6
16.6k
| category_primary
stringclasses 60
values | category_all
stringlengths 5
132
| abstract
stringlengths 123
1.99k
| text
stringlengths 190
2.08k
| embeddings
sequencelengths 768
768
|
|---|---|---|---|---|---|---|---|---|---|---|
http://arxiv.org/abs/2402.02622v2 | http://arxiv.org/pdf/2402.02622v2 | DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging | 2024-02-04T21:44:09 | 2024-03-21T10:57:40 | Matteo Pagliardini, Amirkeivan Mohtashami, Francois Fleuret, Martin Jaggi | cs.CL | cs.CL, cs.LG | The transformer architecture by Vaswani et al. (2017) is now ubiquitous
across application domains, from natural language processing to speech
processing and image understanding. We propose DenseFormer, a simple
modification to the standard architecture that improves the perplexity of the
model without increasing its size -- adding a few thousand parameters for
large-scale models in the 100B parameters range. Our approach relies on an
additional averaging step after each transformer block, which computes a
weighted average of current and past representations -- we refer to this
operation as Depth-Weighted-Average (DWA). The learned DWA weights exhibit
coherent patterns of information flow, revealing the strong and structured
reuse of activations from distant layers. Experiments demonstrate that
DenseFormer is more data efficient, reaching the same perplexity of much deeper
transformer models, and that for the same perplexity, these new models
outperform transformer baselines in terms of memory efficiency and inference
time. | DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging: The transformer architecture by Vaswani et al. (2017) is now ubiquitous
across application domains, from natural language processing to speech
processing and image understanding. We propose DenseFormer, a simple
modification to the standard architecture that improves the perplexity of the
model without increasing its size -- adding a few thousand parameters for
large-scale models in the 100B parameters range. Our approach relies on an
additional averaging step after each transformer block, which computes a
weighted average of current and past representations -- we refer to this
operation as Depth-Weighted-Average (DWA). The learned DWA weights exhibit
coherent patterns of information flow, revealing the strong and structured
reuse of activations from distant layers. Experiments demonstrate that
DenseFormer is more data efficient, reaching the same perplexity of much deeper
transformer models, and that for the same perplexity, these new models
outperform transformer baselines in terms of memory efficiency and inference
time. | [
0.0064409859,
0.0846855715,
0.0159086194,
0.0423200354,
-0.0217095446,
0.0039961119,
0.0174108706,
-0.0486415103,
-0.046843145,
0.0255290195,
-0.0164813213,
-0.0470363088,
0.0354340486,
0.0440237187,
-0.0175618269,
-0.0686082914,
0.0027087675,
-0.0009514086,
-0.0654203966,
-0.0140959891,
-0.0149059854,
-0.008251165,
-0.0040769191,
0.036566209,
0.0017944451,
-0.0082274089,
-0.0432975255,
0.0066139302,
0.0434898101,
-0.0134561881,
0.0137706483,
0.0553514473,
0.0260288008,
0.0606830642,
0.0000021348,
0.0004798515,
-0.0114482371,
-0.0119850263,
-0.0496090688,
0.0223826598,
-0.0150487861,
0.0035936588,
0.0042840051,
-0.0169411153,
-0.0184114967,
-0.0595206842,
0.076414451,
0.0947446898,
0.0729140267,
0.0329812057,
-0.0073448955,
-0.0600263439,
0.055582922,
-0.0507057011,
-0.0079802414,
-0.0542169698,
0.0033348636,
-0.0253050551,
-0.0616402514,
-0.0149513679,
-0.0331472196,
0.0424558111,
0.0592911541,
0.0001668366,
0.0131072709,
0.0398995019,
0.0137340967,
-0.0082835732,
0.0035070432,
0.0830295384,
-0.0533038154,
0.0114817871,
-0.0132675366,
-0.0069864402,
-0.026652867,
-0.0195043385,
-0.0360652283,
0.0008453915,
0.0245430153,
0.0042377235,
-0.0513821654,
0.0408078171,
0.0641799867,
0.0012666595,
-0.0608836859,
0.0080335187,
0.0108334869,
0.0094913812,
0.00941406,
-0.0059196772,
-0.0079660835,
-0.0465058088,
0.0040054331,
0.0187070724,
-0.029606225,
-0.0031912024,
-0.0024359988,
-0.0305093583,
-0.0166975074,
-0.0067508654,
0.0610782951,
0.0412194319,
-0.0380229913,
0.0474995039,
-0.0174340624,
0.0513262488,
-0.0625239834,
0.0163260326,
-0.0164670758,
-0.0086951982,
-0.0217734445,
-0.0373093262,
-0.0295753684,
0.0567792207,
-0.0137039209,
-0.0359424874,
-0.0415159613,
0.0204522498,
-0.0327044353,
-0.0406118557,
0.0107143074,
0.0386686586,
-0.033093363,
-0.0177017841,
0.0130116325,
0.0613686219,
0.0190937407,
0.009551472,
0.0396254137,
-0.0254735332,
-0.0390877612,
-0.0204625558,
0.0158417933,
-0.0175455566,
0.0250225235,
0.0258096699,
-0.0201360732,
-0.0375467055,
-0.0307392459,
0.0159575585,
-0.0016586031,
-0.0476280563,
0.0297587253,
-0.0489392914,
0.036104437,
-0.0174751766,
0.0240444057,
-0.0125550916,
0.0079754638,
0.020044351,
-0.0178645812,
0.0548246391,
0.0163599905,
-0.0282377377,
0.0090475157,
0.0317724161,
0.016994277,
0.039579168,
0.0252430532,
0.0502331369,
-0.0293803383,
-0.0208463985,
-0.0070394981,
0.0214613695,
0.0094259158,
0.0109413117,
-0.0195101872,
-0.0414072424,
-0.0743834004,
0.0014557475,
0.0317032225,
0.0374623127,
-0.0717321709,
-0.0195734482,
0.0398520008,
0.0169229843,
0.0238798503,
0.0565291606,
0.021392459,
-0.0198562089,
-0.0091662426,
0.0283287987,
-0.0579934902,
0.0875466391,
0.0229191463,
-0.0325831361,
0.017425891,
0.0593145825,
-0.0555467308,
-0.0258136131,
0.0413718633,
-0.0116952136,
0.0277719013,
0.0033039232,
0.0147162024,
0.0442072526,
-0.0181145053,
0.1274750829,
0.0242549572,
-0.0445235893,
-0.0308339689,
-0.0619595647,
0.0079262247,
0.0174841192,
0.0254167207,
0.0027653608,
-0.0121918758,
0.0287824925,
0.0180345476,
0.0316259563,
0.0010670974,
0.0104437899,
0.078493841,
-0.033987198,
-0.0589382648,
-0.0084089991,
0.0221808385,
0.0219026431,
0.0165709648,
-0.0360928103,
0.0355473161,
-0.0276983865,
0.013459634,
-0.0034555939,
0.0007083818,
0.0037199918,
0.0594329908,
0.0081364298,
0.0175065231,
0.0069522127,
-0.0010925351,
-0.00997042,
0.0312894583,
-0.0419797227,
-0.053040415,
0.0176663026,
-0.0747185796,
0.0076829842,
0.0160132386,
-0.012371541,
-0.0008620371,
0.0920306072,
-0.0468562245,
-0.0460859127,
0.0550371297,
-0.0408106111,
0.0567545518,
-0.025854921,
0.0101242457,
-0.0122115901,
0.0406320877,
-0.0231116656,
0.0578513555,
-0.0314992368,
-0.0150978062,
-0.0056905439,
-0.034859281,
0.0038732756,
0.0022335695,
-0.0143796494,
-0.0184453372,
-0.0365263857,
0.0020022173,
0.0061126854,
-0.028501831,
0.0047376882,
-0.0153457206,
-0.0258455444,
-0.0174668022,
-0.0260059852,
-0.0087936688,
0.0051314076,
0.0843431503,
0.0541655049,
-0.0109588737,
-0.021923041,
0.0250181835,
-0.0033777098,
0.0051874178,
-0.0139800142,
-0.00615972,
-0.0365028754,
-0.0252584778,
-0.0498727262,
0.0149611393,
0.0682285205,
0.0063364096,
-0.0319186337,
-0.0323386416,
-0.0273297448,
-0.0070747202,
0.0244426038,
-0.0162285604,
-0.045213528,
-0.0172834098,
-0.0084956484,
0.0201477241,
0.070401594,
0.0102469707,
0.0160326418,
0.0259421561,
-0.0075679924,
-0.0378947556,
0.0266959555,
-0.0543625765,
0.0208926052,
-0.0581731535,
0.0181753952,
0.0115303285,
0.0542867184,
-0.0148832612,
-0.0031430849,
-0.0203459691,
-0.037954811,
0.0216167923,
-0.0320673101,
0.0254946183,
-0.0258536655,
0.0047783474,
0.0130662937,
-0.0170240942,
-0.0425798297,
0.0593745485,
-0.0091216918,
-0.0103128217,
0.005474058,
0.0094835265,
-0.0421075299,
0.0483539999,
-0.0015963544,
0.0439489,
-0.0357200168,
-0.0507980846,
-0.0341345966,
0.0061261375,
-0.0027806733,
-0.0006554094,
0.0717894882,
-0.0471023768,
0.0049402555,
0.0007230068,
-0.0654735118,
-0.0582175329,
-0.0006708251,
-0.0414739698,
0.0202145558,
0.0220279079,
-0.0515265167,
-0.1134183705,
-0.00597231,
0.0618005432,
0.0582911558,
0.0105094025,
-0.0230176263,
-0.0004923582,
0.0216980074,
0.009969363,
0.0611578859,
0.0555437803,
-0.0168372393,
0.0603929684,
0.0784331113,
-0.0274566561,
0.0038582333,
-0.0952169001,
0.029613331,
0.0253859013,
0.012305689,
-0.0104977693,
-0.0072406782,
0.025526965,
-0.0494634695,
-0.0125243478,
0.0670917109,
-0.0126096699,
-0.0837504342,
0.0110647185,
0.0100779869,
0.0363173001,
0.0074464325,
0.0193014368,
-0.0496869162,
-0.0456216112,
0.0021054712,
-0.0038411352,
0.0400330573,
-0.0062885527,
0.0517400987,
-0.0735195354,
0.0432112068,
-0.0076243058,
0.0150898136,
-0.022305252,
-0.0395321622,
-0.0085577443,
-0.0328922756,
-0.0054428349,
-0.0155458087,
0.0083420873,
-0.0591012053,
0.024499137,
-0.0038889188,
-0.0268596597,
0.0491839908,
-0.0205222871,
-0.0319456942,
0.0646457747,
-0.0148603218,
-0.004292794,
-0.0736644939,
0.0225573573,
0.0094606448,
0.0013264337,
0.0044832109,
0.0886931941,
0.0373339802,
-0.0277762506,
-0.0355103239,
0.0202774145,
-0.0281645972,
-0.034184657,
-0.0676396489,
0.0434552804,
-0.0187853537,
0.0068462924,
-0.0105234114,
-0.0015172447,
-0.018896237,
0.0363801159,
0.0016240876,
-0.0023090991,
-0.0183259435,
-0.0086186389,
0.0335312821,
-0.0458199829,
-0.0500425249,
0.0397544391,
0.0336703099,
-0.0581831858,
-0.0186322574,
0.0453446321,
-0.0191489309,
0.0382801294,
0.006746158,
-0.0403412804,
0.0376367979,
-0.0094424672,
-0.0548168793,
0.0086538726,
-0.0206565727,
0.0394667387,
-0.0429025851,
-0.0071354494,
0.0305113997,
-0.0064858231,
-0.0280735847,
0.0269653033,
-0.1253154278,
-0.0518713742,
0.0378748663,
0.02605992,
0.0203754157,
0.0473076589,
-0.0072502005,
-0.012422855,
0.0655496493,
0.0018061463,
-0.0026408399,
0.0345015563,
0.0400023311,
-0.0521787107,
-0.0153616043,
0.0431592464,
-0.0475970134,
0.0293482654,
0.0570787303,
0.0039597801,
-0.035418991,
-0.0033442492,
0.0200122818,
-0.0383362509,
0.0534879304,
0.032430578,
0.0608513616,
0.0007144786,
0.0503045507,
0.0156593118,
-0.0036790494,
-0.0677122027,
-0.0083395019,
0.0396063179,
0.0208593383,
0.0338629186,
0.0267822631,
0.0487333387,
0.0190267917,
-0.0215383936,
-0.0119602922,
-0.0302575734,
-0.0484104343,
-0.0065969662,
0.0121600963,
-0.0046960674,
-0.0042252112,
0.0281906091,
0.0287130289,
0.0173675232,
0.0129882218,
0.0192770194,
-0.0057157441,
-0.080215767,
0.0236917716,
-0.1240792796,
-0.0589869805,
-0.0074969893,
-0.0457757786,
0.048806157,
-0.0095344996,
0.0237561688,
0.0053288471,
0.0024644292,
-0.0094052712,
0.0029921487,
-0.0183805153,
-0.0249957256,
0.020930795,
0.0245731734,
0.0220071245,
-0.0084240204,
0.0706167668,
-0.0159227084,
0.0382129215,
-0.0046797437,
0.0301251523,
-0.0224209409,
-0.0538413711,
0.0218530539,
0.0461526103,
-0.0148560507,
-0.0089564063,
-0.0192620475,
-0.0626933947,
0.0470890887,
-0.0500892475,
0.0391786508,
-0.0247502811,
0.0284674782,
0.0308676101,
0.0230994206,
-0.0492436886,
0.0117383916,
-0.0297856852,
-0.0108080357,
-0.0064551327,
-0.0042418167,
-6.777549248e-33,
0.0269929431,
0.0387572236,
0.0142351473,
0.0108700115,
-0.0427342393,
0.0421996787,
-0.0423470698,
-0.0135392165,
-0.0242726561,
0.0543294698,
-0.0265123062,
0.0317065604,
0.0120242033,
0.0385808423,
0.0180986281,
-0.0318350531,
0.0630374402,
0.0282765906,
-0.0091914777,
0.0557003617,
0.0103754783,
0.0173781905,
0.071587421,
-0.0210858714,
-0.063554652,
0.029916143,
-0.0241974145,
0.0158671569,
-0.0279348716,
0.0407334194,
-0.0369094349,
-0.032541994,
0.0022067216,
-0.0418747887,
-0.0397339277,
-0.007305339,
-0.0671039745,
0.015299188,
-0.0290140528,
0.0163472556,
-0.0494923592,
0.0066581196,
0.0262956657,
0.0173641555,
0.0004737595,
-0.0887704715,
0.0155951101,
0.0026397617,
-0.0293821245,
-0.0314853862,
-0.0154352924,
0.0234219786,
-0.0351190455,
-0.029080525,
0.000368378,
0.0143719641,
-0.0189384352,
0.0319215655,
-0.085289456,
0.0050270222,
0.0097808903,
0.0256500486,
0.0283873323,
0.0414737388,
0.0033470711,
0.0494613238,
0.0355162583,
-0.0306002256,
-0.0436977558,
0.0185192935,
0.0276829228,
0.0712901726,
0.0322920419,
0.0060672918,
0.0640834495,
0.0078465035,
-0.0696333274,
0.0086021349,
-0.0545160472,
0.0878840387,
0.0185239557,
0.0079831248,
-0.0061364495,
-0.0216580965,
-0.0176723935,
0.0069191544,
-0.0345599204,
0.0084621841,
-0.0141563145,
0.0396900773,
-0.0324545205,
-0.010288002,
0.0355889387,
-0.0325885527,
0.0259672888,
0.008180852,
0.0273177419,
0.0310803838,
0.0232461616,
-0.0042545418,
0.0048742704,
-0.0155230127,
0.014882056,
0.0127138887,
0.0084586656,
0.0069644847,
0.0304150935,
0.0697189495,
-0.0464989208,
-0.0361840539,
-0.0500961132,
-0.0038806235,
-0.0001613345,
0.0569252558,
0.0030828777,
0.0035731622,
-0.0219128411,
0.0160860699,
-0.0336563215,
0.0337460227,
-0.0211420953,
0.0203526933,
0.0296925846,
0.0535470061,
-0.0264573358,
-0.0533397198,
-0.0636103675,
-0.0422143489,
0.0551431701,
-0.0887714997,
-0.0432292894,
0.0428115018,
2.925e-7,
0.0068545747,
0.0026160311,
0.0185972992,
-0.0055114236,
0.0131727718,
0.0161553081,
0.0000168305,
0.0135953482,
0.0151859922,
-0.0006247499,
0.0027368893,
0.0516600125,
0.0397020765,
0.014053341,
-0.0811985806,
-0.0232012346,
-0.0733281225,
-0.0243331175,
-0.0196024738,
-0.0441339649,
-0.0062196609,
0.1192936897,
0.0200613569,
-0.0386465304,
0.0589062572,
0.0107114818,
0.0542332381,
0.028490629,
0.0463766418,
0.007912999,
0.0342937969,
-0.0172936264,
0.0111544542,
0.0063023088,
-0.0372856967,
0.0298749208,
0.0002524514,
0.0584821552,
-0.0052393014,
-0.0613640398,
-0.0001719059,
0.0244168565,
0.025003111,
-0.0319865979,
0.044722829,
-0.0597455129,
-0.0354951136,
-0.0160529483,
-0.0374524742,
-0.0033952738,
0.0214763638,
0.0280731022,
-0.0041668043,
0.0114050778,
0.0002262203,
-0.0245661885,
0.0039946809,
-0.0796257034,
-0.009277693,
-0.0498004816,
-0.0521833301,
-0.0139652099,
-0.0472102389,
-0.0875638127,
0.0350133106,
0.0172848906,
-0.0064679906,
2.857786331e-34,
0.0405021235,
-0.0062828204,
0.012350847,
0.0282301474,
0.0130769024,
-0.0268622991,
0.0299375951,
-0.0132035725,
-0.0002841815,
-0.1034528464,
-0.0226631351
] |
http://arxiv.org/abs/2402.02619v7 | http://arxiv.org/pdf/2402.02619v7 | Increasing Trust in Language Models through the Reuse of Verified Circuits | 2024-02-04T21:33:18 | 2024-06-16T22:09:57 | Philip Quirke, Clement Neo, Fazl Barez | cs.LG | cs.LG, cs.CL | Language Models (LMs) are increasingly used for a wide range of prediction
tasks, but their training can often neglect rare edge cases, reducing their
reliability. Here, we define a stringent standard of trustworthiness whereby
the task algorithm and circuit implementation must be verified, accounting for
edge cases, with no known failure modes. We show that a model can be trained to
meet this standard if built using mathematically and logically specified
frameworks. In this paper, we fully verify an auto-regressive transformer model
for n-digit integer addition. To exhibit the reusability of verified modules,
we insert the trained integer addition model into a larger untrained model and
train the combined model to perform both addition and subtraction. We find
extensive reuse of the addition circuits for both tasks, easing verification of
the more complex subtractor model. We discuss how inserting verified task
modules into LMs can leverage model reuse to improve verifiability and
trustworthiness of language models built using them. The reuse of verified
circuits reduces the effort to verify more complex composite models which we
believe to be a significant step towards safety of language models. | Increasing Trust in Language Models through the Reuse of Verified Circuits: Language Models (LMs) are increasingly used for a wide range of prediction
tasks, but their training can often neglect rare edge cases, reducing their
reliability. Here, we define a stringent standard of trustworthiness whereby
the task algorithm and circuit implementation must be verified, accounting for
edge cases, with no known failure modes. We show that a model can be trained to
meet this standard if built using mathematically and logically specified
frameworks. In this paper, we fully verify an auto-regressive transformer model
for n-digit integer addition. To exhibit the reusability of verified modules,
we insert the trained integer addition model into a larger untrained model and
train the combined model to perform both addition and subtraction. We find
extensive reuse of the addition circuits for both tasks, easing verification of
the more complex subtractor model. We discuss how inserting verified task
modules into LMs can leverage model reuse to improve verifiability and
trustworthiness of language models built using them. The reuse of verified
circuits reduces the effort to verify more complex composite models which we
believe to be a significant step towards safety of language models. | [
-0.0061851325,
0.0982872769,
0.0169238094,
0.034222573,
-0.0111556165,
-0.0250175092,
0.0255429447,
-0.0486590452,
-0.0088767111,
-0.0426569581,
-0.0086099878,
-0.0042754407,
0.0337617062,
0.0527060069,
0.0723649487,
0.0123369973,
0.0516326018,
-0.0312808342,
-0.0433860682,
0.0012932098,
-0.0210979003,
-0.0210378226,
0.0036812734,
0.0240058228,
-0.0114110364,
-0.0091216974,
0.0045747068,
-0.0612617098,
0.0052924985,
0.0080990363,
-0.0051593324,
0.0616044104,
-0.0148216831,
0.0564526096,
0.0000022714,
-0.008322713,
-0.0427248552,
0.0384110659,
-0.0864174813,
0.0324936993,
-0.0089034587,
0.0058665485,
-0.0410956703,
0.0364842489,
-0.0205096696,
-0.0117101455,
0.0640784651,
0.0303994901,
0.0789040998,
0.0705473721,
0.0034226335,
0.0016907388,
0.0745084137,
-0.0198710114,
-0.0012617842,
-0.012536821,
0.01906606,
0.0146056209,
-0.0103404876,
-0.0182107966,
0.0098078391,
-0.0135088237,
0.0497668311,
-0.0142121855,
-0.014831434,
0.009963898,
-0.0244827829,
0.0046062036,
-0.0603256412,
0.0151501521,
-0.0449829437,
-0.0806276873,
-0.0064335954,
-0.0055097924,
-0.0602042638,
-0.0106665595,
-0.0248429552,
0.0255620982,
0.0095482012,
-0.0089970874,
-0.0080044549,
-0.0238287691,
0.0637543574,
-0.0096222106,
0.0031565516,
0.0949382186,
0.0267641917,
0.0058972943,
-0.0337040164,
-0.0103128655,
-0.0324146263,
-0.0633015111,
-0.0066481414,
0.0295251738,
0.0269459132,
0.018849263,
-0.0432796478,
-0.0096547622,
0.0063244682,
-0.0625881478,
0.0180539209,
0.0475449786,
0.0028644591,
-0.0100490004,
-0.0484578349,
0.0649699122,
-0.0250168387,
0.0248490907,
0.0042219651,
0.0525972284,
0.0121489419,
-0.0191517398,
-0.02742544,
0.0441751368,
0.0125218835,
-0.0016031942,
-0.0553846508,
0.05305776,
-0.0422126018,
0.0566695444,
0.0176890958,
-0.0343186073,
0.0334568173,
-0.0443426818,
0.0085680662,
0.0058661816,
0.028870197,
0.0099560516,
-0.0090285968,
-0.0789875239,
-0.0064589945,
0.0045521762,
-0.0444580093,
0.0060313987,
-0.0025282842,
0.0771695152,
0.0059247403,
-0.0364706255,
-0.0520065315,
-0.0427809134,
-0.0336122066,
-0.0472461544,
0.001967103,
-0.101053521,
-0.0183294732,
-0.0046389066,
0.0068810279,
-0.008076244,
0.0276678298,
0.0352546535,
0.0029856935,
0.0585940517,
0.0142546976,
-0.0153237162,
-0.0394152962,
0.0011107064,
-0.0061703403,
0.0162219647,
0.0487558953,
-0.0191986412,
0.0460308827,
-0.0118177217,
0.0464123003,
0.0492301695,
0.0151264518,
-0.0104304776,
-0.02456172,
-0.0237189997,
0.0165433511,
-0.0608256608,
-0.0463698097,
0.0886143148,
-0.0453767851,
-0.0131773558,
0.0460673869,
0.012067752,
0.0001325688,
0.0579383858,
-0.0029476793,
-0.0320883729,
0.0242348816,
0.0010941055,
-0.0975498334,
0.0468067266,
-0.0636726692,
0.0244418103,
0.0464753211,
0.0198811963,
-0.0179473329,
-0.0446518697,
-0.0349600352,
-0.0541659594,
0.0760906264,
-0.0022060547,
0.0376076959,
-0.0082784882,
-0.0406253189,
0.073684074,
0.005028334,
-0.104784891,
-0.0404789783,
-0.1068327278,
0.041498743,
0.0315459669,
0.06236545,
0.0123672001,
0.0549360998,
-0.0264119525,
-0.0233678557,
-0.0016122492,
0.0339291506,
-0.0023664599,
-0.0064153643,
-0.0220518354,
-0.0187540185,
0.022799788,
0.0509179309,
0.0149584264,
0.0478795543,
-0.0186881404,
0.0202231798,
-0.0100893555,
-0.0282181017,
0.0056965835,
-0.0187766701,
-0.0351410322,
0.0032819372,
0.0242891833,
-0.0297224559,
0.0048870505,
-0.0091386605,
0.0565510765,
0.0232724771,
0.0230767839,
-0.0200470071,
-0.0264888275,
-0.0455857739,
0.0374840498,
0.06724637,
-0.0144421672,
0.0029372338,
0.0506243743,
0.01209863,
-0.0378630795,
0.0507638641,
-0.0751807168,
0.030521255,
-0.0342883505,
0.0427704453,
-0.0171111785,
-0.0237746034,
-0.0002922131,
0.0345251262,
-0.0070791659,
0.0050127204,
-0.0351735577,
-0.0263242088,
0.020331949,
0.0069644307,
-0.0634349957,
-0.0101968022,
-0.0029462893,
0.0654584542,
0.0347218737,
-0.0196678136,
0.0031639116,
-0.0347900838,
-0.0454414636,
-0.0348734297,
-0.0320629999,
0.0677346662,
-0.0090757096,
-0.013464639,
0.0285106283,
-0.0078337258,
-0.011787666,
-0.0321941301,
0.0168878119,
-0.0020232315,
-0.0032057576,
-0.0340198129,
-0.0438964851,
-0.0040585799,
-0.0405436046,
-0.0272712037,
0.0789851844,
-0.0230085514,
0.0147568714,
-0.048239015,
-0.0019370737,
0.0011395522,
0.0429130048,
-0.0127841262,
-0.014087487,
0.0293865465,
0.0229494013,
-0.009142736,
0.0035716188,
-0.0134848757,
-0.0315262862,
0.0479471721,
0.0003269341,
-0.0169376079,
-0.0190427955,
0.0061643249,
-0.0556560531,
0.0245564058,
-0.0083876029,
-0.038937483,
0.0787002891,
-0.0234103184,
-0.010266006,
-0.0274839383,
-0.0129182115,
0.033791896,
0.0094116526,
0.0258546993,
-0.0727438331,
-0.0161788855,
0.0212446507,
-0.0182437003,
-0.0799994394,
0.0611129552,
-0.0588369183,
-0.0175424665,
-0.0209613089,
0.0009006598,
-0.0020893952,
0.0306280442,
0.0186377224,
0.0249293875,
0.0514931455,
0.0275953691,
-0.0127270492,
-0.0104247099,
0.0422228314,
-0.0662831739,
0.0933730751,
-0.0208619703,
0.019456055,
-0.0081225084,
-0.0319005363,
-0.0615685023,
0.0181442089,
-0.0282981172,
0.0284816455,
-0.0145025272,
-0.0407533795,
-0.0155866984,
0.0370962061,
0.0611568913,
0.045880679,
-0.0165902358,
-0.0150677105,
-0.0143659841,
-0.0032404419,
-0.0033842511,
-0.0363130979,
0.0644245893,
-0.0585370287,
-0.0039456827,
0.0126552908,
0.0524952523,
0.0311760362,
-0.0713089854,
-0.0160287712,
0.0238237139,
-0.0132773146,
0.0031711569,
0.0578589849,
-0.0117326202,
-0.0278384723,
0.0040731756,
0.0729170144,
0.0076965471,
-0.0356594883,
-0.0027308089,
-0.0229758862,
0.0042528999,
0.0206469893,
0.0422873907,
-0.1034652889,
0.0221416745,
0.0206828155,
0.0037689148,
-0.0253939815,
0.0021917752,
-0.0267753731,
-0.1060528979,
0.0219073426,
0.043573454,
-0.0239307694,
-0.0307367481,
-0.0564703979,
-0.028230438,
-0.0148668392,
0.030657772,
0.0274052043,
-0.0013052052,
-0.0313887075,
0.0155653954,
-0.0648812056,
-0.0092408946,
0.0210911855,
0.0078329248,
0.0144326445,
0.0727932006,
-0.032587789,
-0.0189207923,
-0.0710988194,
0.0126959514,
0.0491288565,
-0.0161915086,
0.0447104797,
0.008192773,
0.0267535821,
-0.0174047705,
0.0027444933,
0.0045354082,
-0.0274950787,
-0.0302766804,
-0.0377263539,
0.0789734051,
-0.0040090773,
-0.0329984725,
0.008629255,
-0.0198707078,
0.0070537501,
0.031394802,
-0.0133429179,
0.0201068837,
-0.0224083401,
-0.0522570349,
-0.0086818328,
-0.0212335195,
-0.0566111654,
-0.0036803922,
0.0368786529,
-0.0439975336,
-0.0167070534,
0.0355951898,
-0.0522457249,
0.0324459672,
0.0046405974,
-0.0707703456,
0.0796926171,
-0.02438768,
-0.0695741773,
0.0017832782,
-0.0694127902,
0.0694714487,
-0.0703587234,
0.0377648026,
-0.0060228328,
-0.0306224339,
-0.0131888483,
0.0203255173,
-0.0922643095,
-0.0841619745,
-0.0101915132,
0.0796569437,
0.0372003727,
-0.0054836385,
-0.0031314592,
-0.026836304,
0.022339033,
0.0004061466,
-0.0506870151,
0.0313135311,
0.0642871559,
-0.0425481088,
-0.0035342509,
-0.0275628306,
-0.0095764948,
0.0116154905,
0.0375521593,
-0.0185922813,
-0.0500623323,
-0.0204221066,
0.0399574041,
0.0258506406,
0.0442761146,
0.0083407378,
0.0768191442,
0.0007350204,
0.0440115742,
-0.0231471937,
0.011349028,
-0.0306397062,
0.0265694633,
0.0080208266,
0.00627496,
0.001925374,
-0.0572933815,
0.0264845546,
0.0363257006,
-0.0003456743,
0.0440249741,
-0.0249766093,
0.045947887,
0.0102441246,
-0.034687046,
-0.0243667942,
-0.006008259,
0.0143494494,
0.0216156598,
0.0046793101,
-0.0211363211,
0.004305481,
0.0114745917,
-0.0203635097,
0.0166995227,
-0.0442531966,
-0.0787679553,
-0.0421379544,
0.0024973322,
0.0098705404,
0.0353449322,
0.0083788829,
-0.0448723845,
0.0062675737,
-0.0555394255,
-0.0027363247,
0.0026891031,
-0.0214751083,
0.0068859034,
0.0356598571,
0.009376266,
-0.0408532061,
0.0400857776,
-0.0162057523,
0.0273290537,
-0.0306174606,
-0.0158725604,
-0.0259011798,
-0.0651736781,
-0.014890952,
0.0280598216,
-0.0552454516,
-0.0424105786,
-0.0321916379,
-0.0814690217,
0.0103058955,
0.0220170151,
0.0478960127,
0.0008251127,
0.0242015626,
0.0491037667,
-0.0096067339,
0.0302272364,
-0.0098592713,
-0.0340384282,
0.0416134484,
0.0154307047,
-0.0254345909,
-6.360724829e-33,
-0.0068522799,
-0.0167602971,
0.0335193202,
0.0106537025,
0.0237151049,
0.0423321724,
-0.0552798137,
0.0557468347,
-0.0097953873,
-0.0175433792,
0.0179150533,
0.0071028559,
0.0318524763,
0.0085686538,
-0.0006619594,
0.0361374542,
0.0571407713,
-0.00929124,
-0.0437579192,
-0.0017972698,
0.0216010641,
-0.0343869962,
0.0720723644,
-0.058536116,
0.0540303849,
-0.0059889378,
-0.0499433242,
-0.0144879604,
-0.0406965874,
0.0771455318,
0.0107150469,
0.0261776149,
-0.0153935757,
-0.0554645993,
-0.0061599533,
0.0345942639,
-0.0378456898,
-0.0068380088,
-0.0050352025,
-0.0043513719,
-0.0062759765,
-0.0238926373,
0.0464606062,
-0.0176816136,
-0.0482333042,
0.0147695765,
-0.020690782,
-0.021101689,
0.0014659252,
0.0122375181,
-0.0214550328,
0.0385835096,
0.0290139094,
-0.0168586299,
0.0239343382,
0.0326514393,
-0.0364836045,
-0.0639822483,
-0.0523862876,
0.0655269101,
0.0621992163,
0.0248683486,
0.0589199476,
-0.0140533186,
-0.0070216917,
0.0413777493,
0.0340032689,
-0.0091520259,
-0.0287381131,
-0.0003005484,
0.040751338,
0.0553551912,
0.0978325084,
0.0270094126,
0.0477216318,
-0.020715598,
-0.0364168547,
0.0207100529,
0.0394642241,
0.0338219963,
0.0425123014,
0.0056700488,
-0.0608944669,
-0.0122602209,
-0.0198187996,
0.026150519,
-0.0579850711,
-0.0333528221,
-0.0284539238,
0.0466603674,
-0.0078774532,
0.0285613127,
0.0054945899,
-0.0613592938,
-0.0028728587,
0.0347193964,
0.0397671051,
0.0046506212,
-0.0244765729,
0.0202918854,
0.0028136666,
-0.007866879,
-0.0215747636,
-0.0301102567,
0.0086959321,
0.0197948962,
0.0113280332,
0.0723114163,
-0.0785421506,
0.0068674367,
-0.0330073833,
-0.0092809359,
0.0544057302,
0.0500487797,
0.0268090218,
0.003617658,
-0.0044370811,
-0.0066489205,
-0.05801934,
-0.0486262403,
-0.0030130243,
0.0698407739,
0.0015198584,
0.0628740937,
0.0168182123,
-0.0239037443,
-0.0326575339,
0.0280429628,
0.0513582714,
-0.0081755351,
-0.0067406604,
-0.0164625794,
3.103e-7,
0.0007041043,
0.0047222944,
0.0276628919,
0.0101191169,
0.0241434276,
0.0332839899,
-0.0141366683,
0.0230166521,
-0.0287841577,
-0.0171075854,
-0.0349601842,
0.0014420357,
0.0358872414,
0.0188679621,
0.0226244163,
-0.0172745716,
-0.0548391677,
0.0151840029,
0.0269310102,
-0.0170439072,
0.0041531082,
0.0745205209,
0.0357981622,
0.0084126256,
-0.008715976,
-0.0397926271,
0.0440731905,
0.0031024809,
-0.003450369,
-0.0016818178,
0.027203463,
0.0247989055,
0.005906601,
0.0374965146,
-0.0121409008,
0.0278341714,
-0.0027805921,
0.0685809851,
0.0084251398,
-0.0034007658,
-0.0061606602,
-0.058391124,
-0.0335844383,
-0.0194636285,
0.0251213592,
-0.0189638231,
-0.0025760597,
-0.0775667578,
-0.0192568023,
-0.0109539842,
0.0296840295,
0.0121413115,
0.0173634477,
0.000984917,
0.0171233416,
0.0113975117,
0.0315680653,
-0.0439579487,
-0.0713493824,
0.0051583629,
-0.0255801454,
-0.0552544184,
-0.0133081879,
-0.0026042063,
0.0127816927,
-0.0330136605,
0.0130825723,
3.078051013e-34,
0.0331999734,
-0.0355747156,
-0.0067369863,
0.0477485955,
0.0585514642,
-0.0127210198,
-0.01966298,
-0.0000532825,
0.0151287727,
-0.0136853065,
-0.0440589115
] |
http://arxiv.org/abs/2402.02617v1 | http://arxiv.org/pdf/2402.02617v1 | Layer-Wise Analysis of Self-Supervised Acoustic Word Embeddings: A Study on Speech Emotion Recognition | 2024-02-04T21:24:54 | 2024-02-04T21:24:54 | Alexandra Saliba, Yuanchao Li, Ramon Sanabria, Catherine Lai | cs.CL | cs.CL, cs.SD, eess.AS | The efficacy of self-supervised speech models has been validated, yet the
optimal utilization of their representations remains challenging across diverse
tasks. In this study, we delve into Acoustic Word Embeddings (AWEs), a
fixed-length feature derived from continuous representations, to explore their
advantages in specific tasks. AWEs have previously shown utility in capturing
acoustic discriminability. In light of this, we propose measuring layer-wise
similarity between AWEs and word embeddings, aiming to further investigate the
inherent context within AWEs. Moreover, we evaluate the contribution of AWEs,
in comparison to other types of speech features, in the context of Speech
Emotion Recognition (SER). Through a comparative experiment and a layer-wise
accuracy analysis on two distinct corpora, IEMOCAP and ESD, we explore
differences between AWEs and raw self-supervised representations, as well as
the proper utilization of AWEs alone and in combination with word embeddings.
Our findings underscore the acoustic context conveyed by AWEs and showcase the
highly competitive SER accuracies by appropriately employing AWEs. | Layer-Wise Analysis of Self-Supervised Acoustic Word Embeddings: A Study on Speech Emotion Recognition: The efficacy of self-supervised speech models has been validated, yet the
optimal utilization of their representations remains challenging across diverse
tasks. In this study, we delve into Acoustic Word Embeddings (AWEs), a
fixed-length feature derived from continuous representations, to explore their
advantages in specific tasks. AWEs have previously shown utility in capturing
acoustic discriminability. In light of this, we propose measuring layer-wise
similarity between AWEs and word embeddings, aiming to further investigate the
inherent context within AWEs. Moreover, we evaluate the contribution of AWEs,
in comparison to other types of speech features, in the context of Speech
Emotion Recognition (SER). Through a comparative experiment and a layer-wise
accuracy analysis on two distinct corpora, IEMOCAP and ESD, we explore
differences between AWEs and raw self-supervised representations, as well as
the proper utilization of AWEs alone and in combination with word embeddings.
Our findings underscore the acoustic context conveyed by AWEs and showcase the
highly competitive SER accuracies by appropriately employing AWEs. | [
-0.0471056886,
0.0506452061,
0.0186362229,
0.0427313559,
0.0018321748,
-0.0258468837,
0.0298996437,
-0.0528551973,
0.0030177992,
-0.0379513651,
-0.0934022442,
-0.0174348727,
0.007561022,
0.0012932318,
-0.0192337297,
-0.0253883749,
0.0147145847,
0.0109298946,
0.0254949797,
0.0006492967,
0.0188353118,
-0.0078206006,
-0.0376835465,
0.0573822744,
-0.0363574103,
-0.0159649048,
-0.009449047,
0.0155820427,
0.0110337129,
-0.0247587152,
0.0116908643,
0.0360614136,
0.0767882764,
0.0260952804,
0.0000018551,
-0.0114467768,
-0.0020948735,
-0.0091243805,
-0.0610597022,
0.041559279,
0.0463340692,
0.0244775657,
-0.0434860885,
0.0117301112,
0.0044585369,
-0.0477362536,
0.0376002826,
0.0140508953,
0.0262057204,
0.0330156796,
-0.0092920754,
-0.0155760162,
-0.0003635793,
-0.0208790246,
-0.0187173672,
0.0013121938,
-0.0050945841,
-0.033090584,
-0.0089071477,
-0.0449269153,
-0.0551722609,
0.0231722388,
-0.0286909863,
0.0146917691,
0.058734823,
0.017364433,
-0.0403008498,
-0.0242991056,
-0.0143468576,
0.0729005411,
0.0111618442,
0.0131426612,
-0.0226683728,
0.010272068,
-0.0201338213,
-0.0201059598,
-0.013156577,
-0.011912792,
0.0540327281,
0.0242102686,
-0.0177642833,
0.1081467718,
0.028435586,
0.0117852334,
-0.0266370382,
0.0310710557,
-0.0356119946,
-0.0376084819,
0.0094298646,
-0.0446507297,
0.0094619943,
0.0289599895,
0.0200939234,
-0.0238827541,
0.023691386,
-0.0129373195,
-0.0325561315,
-0.0128578395,
-0.0224433709,
-0.0649327561,
-0.0173666421,
0.0208131094,
-0.067259945,
0.052073516,
-0.0183866732,
0.0586860403,
0.0438156985,
-0.0218463223,
-0.0292790737,
0.020641569,
-0.0243110564,
0.0349872969,
-0.0084831361,
0.024553651,
-0.0064165085,
-0.0833240524,
-0.0617045052,
0.0038325868,
-0.0281233788,
0.0328943431,
-0.0054925429,
0.0221280083,
0.0183976945,
0.0125921359,
0.0273984037,
0.0120684085,
-0.0196094736,
0.0478043891,
-0.0246945824,
-0.0241907015,
-0.0004459079,
0.0161860753,
0.0291623473,
0.0012751562,
0.0338425897,
0.0547129586,
0.0217617769,
-0.0446548872,
-0.0150387296,
0.0110761225,
-0.0517108999,
-0.0205677953,
-0.0119143827,
-0.023324931,
0.0634983182,
-0.0313238986,
0.0092416983,
-0.0128844064,
-0.0003531565,
-0.0448869951,
0.0022715328,
0.0977400616,
0.025884226,
-0.0172239412,
0.0329600759,
0.0242581423,
0.0083896536,
0.014084179,
0.0645401329,
0.0148883807,
0.0672183484,
-0.0624596328,
0.0049329214,
0.010566663,
-0.0106745921,
-0.0232027061,
-0.0416220129,
-0.0155101661,
-0.0225043911,
0.054348696,
0.0607854836,
-0.0010276792,
-0.0357314795,
-0.0429723822,
0.060240645,
0.1158777177,
0.0052351933,
0.0259514749,
-0.0213996395,
0.0149714611,
0.0304498728,
-0.0158004835,
-0.0150609287,
-0.0061578485,
-0.0033901555,
0.0011509943,
0.043214906,
0.0021997325,
-0.0379453003,
-0.0296197552,
-0.0321953036,
-0.0341069251,
-0.0220759064,
0.0121119739,
0.0270104613,
-0.0453358553,
-0.0259924661,
-0.0099139269,
-0.0132831251,
-0.0807026699,
0.0015674555,
-0.0019767699,
-0.0015823082,
0.0567504987,
0.0183525477,
-0.0378328115,
-0.0353842862,
-0.0308951922,
-0.0217285659,
-0.0089591611,
0.0242949706,
-0.0100353118,
0.0721695423,
-0.0770118609,
-0.0196163543,
-0.0350596681,
0.0205390416,
0.0261724815,
-0.0236732811,
0.0565802231,
0.0183077455,
-0.0713912621,
0.0253708437,
0.0074395747,
-0.0196698606,
-0.0756590068,
0.0086362865,
0.0477688834,
0.0475327559,
0.0344514064,
-0.0519716702,
-0.0024862697,
0.0548722893,
-0.0074202125,
-0.0290786866,
0.03924568,
-0.0537864864,
-0.0169907045,
0.0301878229,
-0.0326196961,
0.0476166457,
0.0128295878,
-0.014020293,
-0.0355510898,
0.0550448969,
-0.0048125363,
0.0488977134,
-0.0485966876,
0.0172345582,
-0.0246171169,
0.0712558255,
0.0225217771,
-0.0008292389,
0.0213767812,
-0.0223640539,
-0.0227051433,
0.0206900779,
0.0335739963,
-0.0186813604,
-0.0619642809,
0.0159492977,
0.0180143639,
0.0244627185,
-0.0062175794,
-0.0340959318,
-0.0175337773,
-0.0230411496,
-0.109786965,
-0.0070022386,
-0.0147172492,
-0.0191528853,
-0.032230664,
-0.0027088388,
-0.0375201963,
-0.011889223,
-0.0046068858,
-0.0255909376,
-0.0297116786,
-0.015741989,
-0.0259133354,
-0.0256516803,
-0.0240600053,
-0.0865392685,
0.0203987155,
0.0079037212,
0.0013404152,
-0.0246293508,
0.0044831112,
0.0079959417,
-0.0144851925,
-0.0100269811,
0.0296449065,
-0.0017569436,
-0.0447323248,
0.0146180103,
-0.0121990293,
-0.0087559912,
0.0504563451,
0.0114614721,
0.089046739,
0.0441000275,
-0.0035977892,
-0.0360804163,
0.027939057,
-0.0006990871,
0.0688048825,
-0.0001210557,
0.048637379,
-0.0406028554,
0.0412006192,
0.0358169973,
0.0205987561,
-0.0189655218,
-0.0195016563,
0.013666844,
-0.0110464385,
-0.0205229167,
-0.0526989251,
-0.0326665342,
-0.0291268583,
0.0230799206,
-0.0533503704,
0.020929426,
-0.056898836,
-0.03520808,
-0.0694132745,
-0.0695654228,
-0.0163827892,
0.0673724487,
-0.0706888214,
-0.0079909861,
-0.016921781,
-0.0003943369,
0.0293721557,
0.0162494984,
-0.0013783387,
-0.0646890253,
-0.0099595664,
-0.0012784243,
-0.0239665881,
0.0456301011,
-0.0395593494,
-0.0290858969,
-0.0773119852,
-0.0496276319,
-0.0178191941,
0.003082877,
-0.0214493647,
-0.0211606864,
0.0073669837,
0.0258065611,
0.0063899294,
0.0264885575,
0.0033664776,
0.0054895654,
0.0362427197,
0.007462204,
0.0423650406,
0.0000803625,
0.0540188476,
0.0717841759,
-0.0122548733,
0.0777321532,
-0.0049435436,
-0.0632540733,
0.0349797197,
0.0203153379,
0.0158157013,
-0.0209293403,
-0.0237736534,
0.0247873385,
-0.0292677432,
0.0480159856,
0.1334838271,
0.0391126052,
-0.0236135814,
0.0306657217,
0.0020180719,
-0.0196840093,
0.0194196254,
-0.0120124575,
-0.0061170897,
-0.0266445559,
0.0213179179,
0.0368478522,
0.0371230952,
0.0163715649,
-0.0080605811,
-0.0522926711,
0.0474026911,
0.0204480588,
-0.004005983,
-0.0240327679,
-0.0414457023,
0.0315066986,
-0.0463042334,
-0.0019596126,
-0.0785767213,
0.0087618185,
-0.0327756368,
0.0276257508,
-0.0717917308,
-0.0585656166,
0.0429286361,
-0.0399747118,
0.0749876574,
0.0277291238,
0.0654087439,
-0.0270718988,
-0.0624790713,
0.0386226363,
0.0143202515,
-0.0247238725,
-0.0162416827,
0.0389687829,
-0.0366057083,
-0.0174108148,
0.0431369878,
0.0293264445,
-0.0240238272,
0.0038624085,
0.0084042912,
0.0909470618,
-0.0251058172,
-0.009961009,
-0.0310461335,
0.0118186697,
0.0331476368,
0.0799810514,
0.0113068363,
-0.0315369889,
-0.0248185359,
0.0207473617,
0.0382712558,
0.035850659,
-0.0197862498,
-0.0004372347,
0.001937826,
-0.0000848065,
-0.0312162563,
-0.0701597184,
0.0150756426,
0.0308308527,
0.0076272991,
-0.0115599195,
-0.0263645146,
0.019003639,
-0.0424481668,
0.0472833812,
0.0462805033,
0.0275903773,
-0.0240726043,
0.0231630076,
0.0542057529,
-0.0025605669,
-0.0302013606,
0.0133157149,
-0.0093363123,
-0.0599376932,
0.0299300607,
-0.0524811372,
0.0053085587,
0.0213973857,
-0.0192122106,
-0.0103903124,
0.0966688544,
0.0121685676,
-0.0452317744,
-0.0015770269,
0.0682605878,
0.0193805918,
-0.0313350409,
0.0248991828,
-0.0324652866,
0.0394630097,
-0.0355159044,
0.0092256879,
0.0450973473,
-0.0397559181,
-0.0483874045,
-0.0279602781,
0.0671472698,
0.0011440325,
-0.0030925833,
-0.0560602881,
0.014725144,
0.0050681555,
0.0477628484,
-0.0541494451,
-0.03922984,
0.0151421139,
-0.0060175033,
-0.0534888022,
-0.0087572299,
-0.015054754,
-0.013710903,
-0.0103958203,
0.0796888545,
0.0273289904,
-0.0280421525,
-0.0386069715,
0.0283604395,
-0.0112088975,
0.02146817,
-0.0041847131,
-0.0257460102,
-0.0010137978,
0.0041460055,
-0.0277938023,
-0.0173330959,
-0.0304184221,
0.0344241187,
-0.0272880532,
0.008369755,
-0.0198515188,
0.0266474988,
0.0268306378,
-0.0823068321,
0.0029950319,
-0.0177776013,
0.04011143,
-0.025274666,
-0.0152209057,
-0.0632005334,
0.0196158756,
0.0267587695,
0.0002433562,
-0.0578463785,
0.0318280011,
0.0403177366,
0.0028255959,
0.049015075,
-0.0293352194,
0.0190025214,
-0.0530523211,
-0.014309519,
0.0561010018,
0.024061149,
-0.0158775989,
0.0205583069,
-0.0461923294,
-0.0610886812,
0.0318845399,
-0.0453908294,
-0.0171374343,
0.0073872344,
0.0109300911,
0.0530490242,
-0.0337232426,
0.0704877973,
0.0172544345,
-0.0207589753,
-0.0350466669,
-0.0479598977,
0.0159740131,
-5.330047708e-33,
0.0094535798,
0.0110443402,
-0.0268813614,
-0.0159670301,
-0.0405301489,
0.0567025952,
0.0139726652,
0.028120121,
0.0080247214,
0.017975267,
-0.0086963568,
0.0080375848,
0.030011246,
-0.0102885189,
0.0014459912,
-0.0197587367,
0.0347753093,
-0.0529629663,
0.0017149129,
-0.0304977428,
0.0165647008,
0.0474732183,
0.0356241725,
-0.0804779306,
0.023401171,
0.0100527089,
0.0142573398,
-0.0440329574,
0.0902675092,
0.0597708151,
-0.0354602933,
0.0336395577,
-0.0274674203,
0.021108849,
-0.0179915559,
0.0093712267,
-0.0625675023,
-0.0161789712,
-0.0040546595,
0.0370329544,
-0.0311934203,
0.0043608034,
0.0447382145,
0.0268039703,
-0.0379964635,
-0.0464784019,
0.0049645635,
0.0005282899,
-0.0157621428,
-0.0363454893,
-0.0504540019,
0.0388204642,
-0.0360871851,
-0.0232119467,
0.0947636589,
0.022304548,
0.0169338509,
-0.0012546144,
-0.0652426258,
0.0043658866,
-0.0253231283,
-0.0096531147,
0.0336572677,
-0.0162077714,
-0.0043685208,
0.0804023668,
0.0493498035,
-0.0863876119,
-0.0130763389,
-0.0066641858,
0.0654632002,
0.0317908451,
0.0131920762,
-0.024187088,
0.0143239656,
-0.0086048357,
-0.0083924597,
0.0020613375,
0.0451902263,
0.010493841,
0.0729648918,
0.0466786884,
-0.054606501,
0.0110477172,
-0.0640131533,
-0.023458669,
-0.0105743194,
-0.0087561421,
0.0583748408,
0.0044022142,
0.0190944634,
-0.0016040294,
0.0225268919,
-0.0369818807,
0.0040791435,
0.027568575,
-0.0272072535,
0.021489691,
0.0460014157,
0.0340613872,
0.0160550456,
0.0189439729,
0.0463822559,
0.0559700653,
0.0600837693,
-0.048704993,
-0.0210126452,
0.0731403679,
-0.0262728967,
-0.0228612162,
-0.0198527221,
-0.0361903794,
-0.0261512026,
0.0014570522,
0.0030911106,
0.0312063862,
-0.0090544717,
-0.0119169727,
-0.0231202301,
0.0444253534,
0.0191784818,
0.0030907872,
-0.0101294657,
0.0355987214,
-0.0208526049,
-0.0039005051,
0.0375933349,
-0.0153755192,
0.0103980694,
-0.0437823646,
0.0111884214,
-0.0089811468,
2.679e-7,
-0.0533500053,
0.0214871615,
0.0092042144,
-0.0684340671,
0.0400475413,
0.0225234218,
-0.0490745679,
0.009204031,
-0.0304658394,
-0.0049734879,
0.0563845485,
0.0068534371,
0.0878261626,
0.0292695072,
-0.0618320554,
-0.0105446968,
-0.0391723923,
0.0406849347,
-0.0325972848,
0.006968569,
0.0366874784,
0.1039197296,
0.0463517234,
0.0001739409,
-0.0161690861,
-0.013006446,
0.0151905883,
-0.0454487354,
0.0362076089,
-0.0285990294,
-0.0175764244,
-0.0090656336,
-0.0184432175,
-0.0348896682,
-0.0320771597,
0.0603529252,
0.0639561862,
0.0350020081,
0.0192938372,
-0.066039741,
0.0085821208,
0.0406923965,
-0.0257794037,
-0.0155160502,
0.0519462675,
-0.0090398649,
-0.0200898759,
-0.0210437104,
0.0005243611,
0.005700646,
0.0948728323,
-0.0324564911,
-0.0031527407,
-0.0066569746,
0.0127712116,
-0.0337373875,
-0.0328441858,
-0.0927183479,
0.0067099012,
0.0248131491,
-0.0623037554,
0.0256946888,
-0.029031001,
0.013314575,
0.0762627646,
-0.016427841,
-0.0077724503,
1.941805547e-34,
0.0029530455,
-0.0205132011,
-0.0311877429,
0.0240313131,
0.0128194336,
-0.0150527386,
0.0244940408,
-0.029382633,
0.0578556992,
-0.0263736472,
-0.0025379069
] |
http://arxiv.org/abs/2402.02611v2 | http://arxiv.org/pdf/2402.02611v2 | PuzzleBench: Can LLMs Solve Challenging First-Order Combinatorial Reasoning Problems? | 2024-02-04T20:56:09 | 2024-02-22T14:42:45 | Chinmay Mittal, Krishna Kartik, Mausam, Parag Singla | cs.AI | cs.AI, cs.CL, cs.LG | Recent works show that the largest of the large language models (LLMs) can
solve many simple reasoning tasks expressed in natural language, without
any/much supervision. But, can they also solve challenging first-order
combinatorial reasoning problems, such as graph coloring, knapsack and
cryptarithmetic? To answer this question, we present PuzzleBench, a dataset of
31 such challenging problems along with a few solved instances for each
problem. These problems are all first order, i.e., they can be instantiated
with problem instances of varying sizes, and most of them are NP-hard,
requiring several reasoning steps to reach the solution. We first observe that
LLMs, even when aided by symbolic solvers, perform rather poorly on our
dataset. In response, we propose a new approach, Puzzle-LM, which combines LLMs
with both symbolic solvers and program interpreters, along with feedback from
solved examples, to achieve huge performance gains. Our extensive
experimentation and analyses offer new insights into the reasoning abilities
and limitations of present-day LLMs. | PuzzleBench: Can LLMs Solve Challenging First-Order Combinatorial Reasoning Problems?: Recent works show that the largest of the large language models (LLMs) can
solve many simple reasoning tasks expressed in natural language, without
any/much supervision. But, can they also solve challenging first-order
combinatorial reasoning problems, such as graph coloring, knapsack and
cryptarithmetic? To answer this question, we present PuzzleBench, a dataset of
31 such challenging problems along with a few solved instances for each
problem. These problems are all first order, i.e., they can be instantiated
with problem instances of varying sizes, and most of them are NP-hard,
requiring several reasoning steps to reach the solution. We first observe that
LLMs, even when aided by symbolic solvers, perform rather poorly on our
dataset. In response, we propose a new approach, Puzzle-LM, which combines LLMs
with both symbolic solvers and program interpreters, along with feedback from
solved examples, to achieve huge performance gains. Our extensive
experimentation and analyses offer new insights into the reasoning abilities
and limitations of present-day LLMs. | [
0.0337994546,
0.0930320919,
-0.0139691588,
0.0707200319,
-0.0720084533,
0.0065334486,
0.0059367763,
0.0138933538,
0.0162297338,
-0.0161391161,
0.0308998115,
0.0209447555,
-0.0143092014,
0.0292067509,
0.0116528831,
-0.0106455432,
-0.0055034454,
-0.091261439,
0.0170096233,
-0.0227715932,
-0.0081239268,
0.0320624746,
-0.0159248654,
0.0033855177,
-0.0024101799,
-0.0260812752,
-0.0399204753,
0.0001718145,
-0.0229485985,
-0.0145344939,
-0.0345794111,
0.0423487388,
0.0062721339,
0.0085132373,
0.0000023612,
-0.0449211895,
-0.0393239595,
-0.0178172532,
-0.0805812925,
-0.0133292917,
0.0074416492,
-0.0158129875,
0.0144799938,
0.0351519547,
-0.0445371903,
-0.001339562,
0.0528877936,
0.0391181111,
0.0873054117,
0.0510267131,
0.0142079871,
-0.0010417704,
-0.0000626212,
0.0117250932,
0.0548252352,
-0.0571204871,
0.0141671775,
0.0423941202,
0.0346331894,
-0.0187650174,
-0.0285132527,
0.0143938586,
0.0411274098,
-0.0227499157,
-0.0202770848,
-0.0137201436,
-0.017797742,
0.0140252532,
-0.0236076266,
0.0631594583,
-0.0406032167,
-0.0110808918,
-0.0053114095,
-0.0064168181,
-0.0096896645,
0.0199519042,
-0.0142163038,
0.0820012763,
-0.0312983207,
0.0052517499,
0.0083261309,
0.0036574288,
0.0301914047,
0.0289006922,
-0.0458470732,
0.1208699718,
-0.0146852005,
-0.0007550216,
0.0072056381,
-0.0473785624,
-0.0161789544,
-0.065227434,
0.02477834,
0.0403129049,
-0.0309718829,
0.0024682609,
0.0137935439,
-0.0054383501,
-0.0032596486,
0.0191308651,
0.0637818873,
0.0369435847,
0.0105092097,
0.0645491555,
0.0266978461,
0.0823934898,
-0.0214043539,
0.016882794,
-0.0564194061,
-0.0028364032,
-0.0280181095,
-0.0128209228,
-0.0459648483,
0.032782685,
-0.040610265,
0.0097314883,
0.0121088624,
0.0193966664,
-0.0112199876,
0.0519611277,
0.0303606279,
0.0008361579,
0.0175852682,
-0.055908151,
-0.0328460112,
-0.0364191569,
-0.028571574,
-0.0087931585,
-0.0230189878,
-0.0283387657,
-0.0179289468,
0.0145914508,
0.0165998675,
-0.0579154752,
0.0439127311,
0.0865382627,
0.0394946299,
-0.0089736274,
-0.0261579808,
-0.0213615987,
0.0074773757,
-0.0873681232,
-0.006344548,
-0.0979827568,
0.0449801646,
0.009258071,
0.0119681237,
-0.0450783148,
0.012444769,
-0.0169701204,
-0.0097850868,
0.0843739435,
0.0437228233,
0.0223754942,
-0.0164798759,
-0.0218576007,
-0.0113206301,
0.0493956432,
0.0393528603,
0.0407202691,
0.0035924874,
0.0011227927,
0.0150029371,
0.0412905924,
-0.0065184101,
0.0187682081,
-0.0274412967,
0.0002578232,
0.0082979873,
0.008955122,
-0.0329414196,
0.0663310289,
-0.0040904754,
-0.0513840802,
0.0053025647,
0.0525692962,
-0.0169377625,
-0.0224312581,
-0.0114884116,
0.0195240267,
0.0180086717,
-0.0268418621,
-0.0493849665,
0.0610402301,
-0.0606666543,
-0.0097079305,
-0.0183489211,
0.035012234,
-0.0035081641,
-0.0487973168,
-0.0299006384,
-0.0049234168,
0.0394411571,
0.0426075347,
0.0599690601,
-0.0042288327,
-0.0352555551,
0.0105499541,
0.0045191809,
-0.0236393251,
-0.0429272354,
-0.0361765213,
0.0321329683,
0.084964931,
0.0710157007,
-0.041912239,
-0.0310013741,
0.0083089629,
0.0017458623,
0.0176514443,
0.0010216006,
0.0110290367,
0.014736495,
-0.0497394875,
-0.0135160424,
-0.0319528691,
0.0124948425,
0.0164818149,
0.0279863589,
-0.0450723544,
-0.0191418789,
-0.0428106971,
-0.016853651,
0.0078301029,
-0.0146738961,
-0.0000467693,
0.0109182429,
0.0062400722,
0.0201099254,
-0.0068688998,
-0.0061970591,
0.0334754214,
-0.0086831916,
-0.0532014146,
-0.0280689728,
-0.0343050212,
0.002844224,
-0.0404010378,
0.0365255587,
-0.0404693671,
-0.0406966917,
0.0567044728,
-0.0837798715,
-0.0160564519,
0.0501344986,
-0.0234906115,
-0.0082389861,
-0.0047549992,
0.0062157693,
0.0124857062,
-0.0153939733,
-0.0184577405,
0.0487966724,
-0.0367975347,
-0.0141700134,
-0.0228696447,
-0.0139792468,
0.0126782348,
-0.0525210164,
-0.0098781735,
0.0343837105,
0.0175067298,
0.0165910516,
-0.0080877505,
0.0092965309,
-0.0149597814,
-0.035352245,
-0.0295684338,
0.0326587632,
0.0063745603,
0.0217928663,
-0.0137498388,
0.0117455106,
0.0036438578,
-0.010911148,
-0.0578384995,
0.0023648469,
0.0202995259,
0.0180714838,
0.0201867502,
0.004713519,
-0.0302251875,
-0.0252312664,
-0.0422853157,
-0.022626752,
0.0316849053,
-0.0074864472,
-0.0299154501,
-0.0144953961,
0.0564892367,
-0.0332083441,
0.0188448206,
-0.0152863124,
0.0071619544,
-0.0117387446,
-0.0756775737,
0.0066716247,
0.0730658099,
-0.0307414941,
0.0291704424,
0.0531718694,
-0.0768671855,
-0.0548932329,
0.0192083549,
0.0096892584,
-0.0066485354,
-0.0042136358,
-0.0723859295,
-0.012537309,
0.0697789714,
-0.0024379613,
-0.008013593,
-0.0322631001,
0.0543823913,
0.0233311448,
0.0268388782,
-0.0721414313,
-0.0158217158,
0.0485092774,
0.0485306419,
-0.0295395385,
-0.0356147662,
0.0580881909,
-0.0188603289,
-0.0134427669,
0.0024262455,
-0.0224721115,
-0.1003443077,
0.1018066034,
0.0452181399,
0.0261783879,
0.0072703036,
-0.0184086058,
-0.0376627855,
-0.028304657,
0.0409319624,
-0.0514895506,
0.0175133515,
-0.0644521862,
0.0187507365,
-0.0040875021,
-0.0523104556,
-0.0190887842,
0.0027311861,
-0.0593495481,
-0.0222626645,
0.0220672674,
-0.0299645644,
-0.0229229666,
0.0464213304,
0.0007985654,
-0.0083011584,
-0.0082640396,
0.0087623708,
-0.0138594015,
0.006288182,
0.0160882454,
0.0262454879,
0.0217808504,
-0.0335065201,
0.0229434036,
0.0463602804,
0.0430897996,
0.0092915148,
-0.0258252062,
0.0581252202,
0.0280151684,
-0.0007503969,
-0.0083975317,
0.0229729079,
-0.0043657981,
-0.0636484548,
0.0025255529,
0.0783047676,
0.0480355248,
0.0102958884,
-0.0055200499,
-0.0233773012,
-0.0298594441,
0.0264658965,
0.0557586774,
-0.0964071304,
-0.0367277712,
0.0350383706,
-0.009198158,
0.0052194083,
-0.0745518431,
-0.0450460315,
-0.1069226712,
0.0118249366,
0.0458957441,
-0.0249920934,
-0.0373843089,
-0.0213055704,
-0.0382668711,
-0.0134163136,
-0.019965997,
0.0204216614,
0.0157097802,
-0.0025077269,
-0.0131353186,
-0.0347072668,
0.0187229179,
0.0448898003,
0.0139731038,
0.0256639794,
0.0542517044,
-0.0039469907,
-0.007189793,
-0.0355375595,
0.0614095181,
0.0784162208,
-0.0049087266,
-0.0203740858,
0.0053476347,
0.0295728873,
-0.0199279971,
-0.0154572017,
0.0171816573,
-0.0602862835,
-0.0260130074,
0.0297524594,
0.0802073404,
-0.0102235787,
-0.0236185566,
0.0022952203,
-0.0248046089,
-0.0204103291,
0.0533626489,
0.0175144412,
-0.0100766318,
-0.058708705,
-0.1087552384,
-0.0203594528,
-0.047848396,
-0.0380896106,
0.0516662747,
0.0368605033,
0.027445009,
-0.0282649603,
0.033403933,
0.0636420473,
0.0195034612,
-0.0068715266,
-0.0386322699,
0.0017820987,
0.0027170572,
-0.0311511904,
0.0334442481,
-0.0566884056,
0.0446531549,
-0.0348587818,
-0.0322369598,
-0.0097838361,
-0.0485641062,
-0.0081089772,
-0.0114667676,
-0.0946002901,
-0.0284228288,
-0.0195017532,
0.0371556431,
0.0303222016,
0.006175593,
-0.002601342,
-0.0263519157,
-0.0114775244,
0.0677211359,
-0.0258639008,
-0.003109009,
0.0107097132,
-0.0032985613,
0.0136636049,
0.0173469801,
-0.0313589685,
0.0337948538,
0.0279570464,
0.0696047768,
-0.0256535672,
0.030181665,
0.0552018993,
0.0769001693,
-0.025194915,
0.0507185347,
0.0061721466,
-0.0098799746,
0.0204060953,
-0.023572281,
-0.0796970278,
-0.0491356403,
-0.0031227749,
0.0228097923,
-0.0142760072,
0.0356809683,
-0.0211048163,
0.0457734652,
0.0362811759,
0.0242556185,
0.0109990975,
0.0289048776,
0.0043318076,
0.085816294,
-0.0146285053,
0.0054330775,
0.062908344,
-0.0054889964,
0.038128376,
0.0258378945,
0.0329857245,
0.022464944,
0.0077012316,
-0.0965640172,
-0.0091619547,
-0.0231144726,
-0.1101079658,
0.0338827632,
0.0009783953,
0.0737472251,
0.0025009655,
-0.0367228389,
-0.0207196251,
0.0066292011,
-0.0128338141,
0.0084397811,
0.0680986643,
-0.0428472385,
0.0326902494,
0.0548400283,
0.0215669479,
-0.0243612919,
0.0357671231,
0.029886594,
0.0435623862,
0.0190656055,
-0.0008155315,
-0.0406848006,
-0.0267800558,
-0.0131258946,
0.0140525065,
0.013835636,
0.0244466756,
-0.0644944757,
-0.0722253621,
0.0305007789,
0.0214135405,
0.0259361956,
0.0102388635,
0.050562501,
0.0261227731,
0.0279041044,
-0.0106995981,
-0.0198814459,
0.0346525125,
0.0115192169,
-0.0317857116,
0.0110082626,
-6.547181008e-33,
-0.0295832735,
-0.0614302829,
0.020728264,
0.0230980311,
0.0535243489,
-0.0066048992,
-0.0513839573,
-0.0036571654,
0.0180141032,
-0.0212165713,
-0.0087758722,
0.0222039334,
-0.0005764051,
0.0172291528,
0.0009313789,
-0.0077915248,
0.007140907,
0.0311896745,
-0.0386910364,
-0.0382270813,
0.017001899,
-0.0377702676,
-0.0092093106,
-0.0327961668,
0.0009544222,
0.0182209536,
-0.0092547378,
0.0532736965,
-0.0615585893,
0.0133488597,
-0.0127672413,
0.007334149,
0.0073329317,
-0.0433422923,
-0.0050523062,
0.0009248793,
-0.0238585919,
-0.0349118821,
0.0301011316,
-0.0019990387,
-0.0539265424,
-0.045030646,
0.0176564455,
0.01226144,
0.0033886095,
0.0104191843,
-0.0070802048,
-0.051028531,
0.0035602762,
0.0023766772,
0.0406223908,
0.0192062464,
-0.0438953266,
0.0073794774,
0.044767756,
-0.0286259074,
-0.0255205482,
0.0241498444,
0.0176294297,
0.0511889793,
0.0413486734,
-0.0153809292,
0.039911259,
-0.0201216619,
0.0102724386,
0.0253708381,
0.0101378951,
-0.0223045554,
-0.0374813899,
0.0540373661,
-0.0437368415,
0.0731022507,
0.0551090725,
0.0136374738,
0.0223356616,
0.0319194645,
-0.0503345393,
-0.0095796399,
0.0174233243,
0.0341945253,
0.0550116524,
0.0185825583,
-0.0348858647,
-0.0553833246,
-0.0500024296,
0.0262027271,
-0.0561474264,
0.0036982126,
0.0181867555,
-0.0242521502,
-0.0381375253,
-0.0784322619,
0.0065427604,
-0.0253178645,
-0.0316518582,
0.0652593002,
0.0368189663,
0.0017443099,
-0.0035233994,
0.0422911495,
-0.0343580768,
-0.0231510196,
-0.0215716772,
-0.0559115298,
0.0291477945,
0.0242199339,
-0.0026493142,
0.042752441,
-0.0735636279,
-0.004663853,
-0.0703261644,
0.0146576222,
0.0163461193,
-0.0058561247,
0.0318149887,
-0.0040840926,
-0.0066997944,
0.0265309885,
-0.0151002174,
-0.055282101,
-0.0044692317,
0.0170680303,
-0.0259666648,
0.000253042,
-0.0075756712,
-0.0126409214,
-0.0506735556,
-0.0309742801,
0.0254054144,
-0.0127572082,
-0.0100574885,
0.003081765,
3.063e-7,
0.0096502444,
0.0459740497,
0.020967029,
0.0550244339,
0.036941912,
0.0191333406,
-0.050373964,
-0.029417973,
-0.0009355365,
-0.0298747923,
-0.0143605564,
0.0963905379,
0.0394483767,
0.0448584147,
-0.0538509488,
0.0747817382,
-0.0041021132,
-0.0162968244,
-0.0334382169,
0.0154905329,
0.0378321595,
0.0640023127,
0.0354983732,
0.0119133973,
-0.0212994926,
-0.0762306303,
-0.0099439137,
-0.0198541898,
0.0015676718,
-0.0174897071,
-0.0332968198,
-0.0039865691,
-0.015236726,
0.0391080976,
0.0293426365,
-0.0196681656,
0.0003325655,
0.0597024783,
0.0101373466,
-0.0059846337,
-0.0300114453,
-0.043359302,
-0.0157325529,
0.0033543603,
0.0920317173,
-0.0130716516,
-0.0594383292,
-0.0261449311,
0.0254768021,
-0.0518169105,
-0.0255683381,
0.0061289244,
0.0022141414,
0.0146441841,
-0.0181355663,
-0.0181235038,
0.0173850805,
-0.0511310622,
-0.0171647556,
-0.0231088772,
-0.0072733588,
-0.0356020667,
-0.0247699358,
0.0658437014,
0.0235929862,
-0.026436707,
-0.0082179755,
3.085603335e-34,
0.0561751649,
-0.002373887,
0.0190503374,
0.0360597745,
0.0611449666,
-0.0038875176,
-0.0845231637,
0.0195244756,
0.0393680893,
-0.0942158923,
-0.0200928394
] |
http://arxiv.org/abs/2402.14594v1 | http://arxiv.org/pdf/2402.14594v1 | Improving Assessment of Tutoring Practices using Retrieval-Augmented Generation | 2024-02-04T20:42:30 | 2024-02-04T20:42:30 | Zifei FeiFei Han, Jionghao Lin, Ashish Gurung, Danielle R. Thomas, Eason Chen, Conrad Borchers, Shivang Gupta, Kenneth R. Koedinger | cs.CY | cs.CY, cs.AI, cs.CL, cs.HC, cs.IR | One-on-one tutoring is an effective instructional method for enhancing
learning, yet its efficacy hinges on tutor competencies. Novice math tutors
often prioritize content-specific guidance, neglecting aspects such as
social-emotional learning. Social-emotional learning promotes equity and
inclusion and nurturing relationships with students, which is crucial for
holistic student development. Assessing the competencies of tutors accurately
and efficiently can drive the development of tailored tutor training programs.
However, evaluating novice tutor ability during real-time tutoring remains
challenging as it typically requires experts-in-the-loop. To address this
challenge, this preliminary study aims to harness Generative Pre-trained
Transformers (GPT), such as GPT-3.5 and GPT-4 models, to automatically assess
tutors' ability of using social-emotional tutoring strategies. Moreover, this
study also reports on the financial dimensions and considerations of employing
these models in real-time and at scale for automated assessment. The current
study examined four prompting strategies: two basic Zero-shot prompt
strategies, Tree of Thought prompt, and Retrieval-Augmented Generator (RAG)
based prompt. The results indicate that the RAG prompt demonstrated more
accurate performance (assessed by the level of hallucination and correctness in
the generated assessment texts) and lower financial costs than the other
strategies evaluated. These findings inform the development of personalized
tutor training interventions to enhance the the educational effectiveness of
tutored learning. | Improving Assessment of Tutoring Practices using Retrieval-Augmented Generation: One-on-one tutoring is an effective instructional method for enhancing
learning, yet its efficacy hinges on tutor competencies. Novice math tutors
often prioritize content-specific guidance, neglecting aspects such as
social-emotional learning. Social-emotional learning promotes equity and
inclusion and nurturing relationships with students, which is crucial for
holistic student development. Assessing the competencies of tutors accurately
and efficiently can drive the development of tailored tutor training programs.
However, evaluating novice tutor ability during real-time tutoring remains
challenging as it typically requires experts-in-the-loop. To address this
challenge, this preliminary study aims to harness Generative Pre-trained
Transformers (GPT), such as GPT-3.5 and GPT-4 models, to automatically assess
tutors' ability of using social-emotional tutoring strategies. Moreover, this
study also reports on the financial dimensions and considerations of employing
these models in real-time and at scale for automated assessment. The current
study examined four prompting strategies: two basic Zero-shot prompt
strategies, Tree of Thought prompt, and Retrieval-Augmented Generator (RAG)
based prompt. The results indicate that the RAG prompt demonstrated more
accurate performance (assessed by the level of hallucination and correctness in
the generated assessment texts) and lower financial costs than the other
strategies evaluated. These findings inform the development of personalized
tutor training interventions to enhance the the educational effectiveness of
tutored learning. | [
0.034530703,
-0.019022299,
-0.0454783626,
-0.0018691034,
-0.014829237,
0.0029749181,
0.0701559708,
-0.0406179577,
-0.0261926185,
-0.027803082,
-0.0364302695,
-0.0096888607,
-0.02992066,
-0.0274725985,
0.032564193,
-0.0894793943,
-0.0031938178,
0.0394206531,
0.0136557948,
0.025225006,
0.0142455865,
-0.0032637117,
-0.0453874283,
0.0603335612,
-0.0378530845,
-0.0510062389,
-0.0073045301,
-0.0269067287,
0.0241503529,
-0.0277394522,
0.0410257205,
0.0663535669,
-0.0132810995,
0.0246880818,
0.0000021967,
-0.0482912995,
-0.0259924699,
-0.0231733527,
-0.0122748408,
-0.0421542302,
0.0473299697,
0.087528035,
0.0208211914,
0.0216380414,
-0.0821264535,
0.0098231919,
0.030012032,
0.0169056095,
0.0904363617,
0.037597768,
0.0006944019,
-0.0093204146,
0.0230981112,
-0.0148267858,
0.0477118567,
0.0048453221,
0.0037846467,
0.0039977017,
-0.0042724721,
0.0102915298,
-0.0457350798,
0.0175211243,
0.0309966113,
-0.0386515185,
0.0452519916,
-0.0060466328,
-0.0259373821,
-0.0099585755,
-0.0166727565,
0.0391480252,
0.0545186326,
0.037375059,
-0.0217104144,
-0.0088604502,
0.0616757944,
0.0933656991,
-0.0495162606,
0.0229746178,
-0.0397938862,
-0.0190519523,
-0.0095325662,
0.0085654948,
-0.0488398485,
-0.0071924031,
0.0487567931,
0.0038029174,
-0.0144644082,
-0.0393613949,
-0.0079842452,
0.0410539508,
0.009259264,
-0.0283026397,
0.0452666618,
0.0557415597,
-0.0051811407,
0.0165732205,
-0.0091661559,
-0.0840788111,
-0.0223845709,
0.0648322403,
0.0676648915,
0.0599632487,
-0.0697429925,
-0.0369730219,
-0.0054175104,
0.0672024935,
0.045054961,
0.0263837334,
-0.0210798252,
0.1036448255,
-0.0549514927,
-0.0336106457,
-0.0578774177,
0.0469211228,
0.0013527335,
0.0223022047,
-0.0728803948,
-0.0197124537,
-0.0557091422,
0.0086155664,
-0.0319636241,
0.0078193145,
-0.0208277758,
0.0411375947,
-0.0263171922,
0.0177880004,
-0.02635918,
0.0057348856,
0.0384508297,
-0.0910949707,
-0.0005004061,
0.0048889113,
0.0206960887,
0.0289774891,
0.0168284364,
0.0068044416,
-0.0339757837,
-0.0528758317,
-0.0458368771,
-0.0549112558,
0.0074782227,
-0.012150716,
0.013844844,
-0.0294551756,
-0.0135221668,
-0.0038840359,
-0.0115643116,
-0.0078610713,
0.0139442813,
0.0322992764,
0.0508223884,
0.0203938857,
0.000214418,
0.0002938619,
0.0346555226,
0.0197657272,
-0.0611148961,
0.0077134762,
0.0169729441,
0.0167727433,
0.0111097815,
-0.0371543206,
0.0890413746,
-0.0200587176,
0.0131065054,
-0.0118124411,
0.0368360914,
-0.0287236646,
0.0004119074,
0.018276846,
-0.0127152735,
0.0083862077,
-0.0457821488,
0.0482524261,
0.0103999125,
0.0467210039,
-0.0045605432,
0.0710178167,
0.0538600944,
0.0330338404,
0.0484616794,
0.0191722326,
-0.0654861182,
0.0142233847,
-0.0495756604,
0.0080035115,
0.0050829165,
0.0107461968,
-0.0173995532,
-0.0706570372,
-0.0022358822,
0.0137856053,
-0.038062416,
0.0179424006,
0.0375742279,
0.0043555251,
-0.062426988,
0.0140219983,
-0.0258360542,
-0.0932362452,
0.01090049,
-0.0715779439,
0.0312913768,
0.0411842689,
0.064104788,
-0.0378036276,
-0.0296170171,
-0.0254539438,
0.0083801849,
-0.0017340537,
-0.0040966799,
-0.0012336618,
-0.0206977576,
-0.0232377872,
-0.0111019341,
-0.018589776,
0.0230555441,
0.039748881,
0.021041384,
0.0242076572,
0.0348410718,
0.0195951872,
-0.0191031005,
0.0186562836,
-0.0233708173,
-0.0844012871,
-0.0415430553,
0.0134235024,
-0.000891977,
0.0352993011,
-0.0106554702,
0.0172614343,
-0.0056895204,
0.033734832,
-0.0106044849,
0.0017672104,
-0.0923313349,
-0.0207436588,
0.0089896182,
-0.0281815268,
-0.0242144447,
0.015163268,
-0.0028999706,
-0.0143521326,
0.0534020886,
-0.0248171557,
0.0273339879,
-0.040540684,
0.0238000602,
-0.0381841324,
-0.0144296251,
0.0028437711,
0.0063827992,
0.0439818427,
0.0415087715,
0.0286262073,
0.0114810336,
-0.0010301027,
-0.0590089969,
-0.0133739701,
-0.0480057038,
-0.0564367808,
0.0087483767,
-0.0087310942,
0.0287819374,
-0.0781321153,
-0.0049693529,
-0.0405692533,
-0.0088690743,
0.0052067069,
0.029040629,
-0.0079230387,
-0.0596586727,
-0.0416138247,
0.0040163896,
-0.049504105,
0.0059332117,
0.0138548333,
-0.0225831512,
-0.0131524783,
0.0081423065,
-0.0018397275,
-0.0562135391,
-0.0138596362,
-0.0324310437,
0.0170736741,
0.0997176617,
0.0415993109,
-0.0181472152,
0.0525839627,
-0.013877023,
-0.0065130754,
-0.0150799854,
-0.0420406498,
0.0388499126,
-0.0173303429,
0.0907368436,
0.0578646883,
0.0183556844,
0.0606448129,
-0.0269213878,
-0.0632857382,
-0.0494266413,
0.0227492433,
-0.0108485408,
-0.0590538792,
0.0349983275,
-0.0388948545,
-0.0032972598,
0.0314158686,
0.0147771966,
-0.0367547944,
-0.0369282737,
0.0327440724,
0.009458825,
-0.0066319304,
-0.0290486757,
-0.0336422995,
0.0112171648,
-0.0170029756,
0.0104946382,
-0.0447344556,
0.0444608331,
0.0214842223,
-0.0338552333,
-0.0255853105,
0.0524403863,
0.002197382,
0.0306020789,
-0.0221564285,
-0.0165021438,
-0.0774504542,
-0.0318207592,
-0.0422942191,
-0.0027703037,
0.0482378155,
-0.060921602,
-0.018334372,
-0.0237778593,
-0.0311998706,
-0.0140029062,
0.0355491191,
-0.0689045563,
0.0165710021,
0.0139863566,
-0.0159212761,
0.0324174203,
-0.0027481252,
-0.0315548442,
-0.0300040171,
0.0472958535,
0.0221472867,
-0.0457566269,
-0.0477202795,
0.0066812071,
-0.0005857805,
-0.0062918589,
-0.0123685692,
0.0133700315,
-0.0482260659,
0.0671316907,
0.0378286727,
0.0606861897,
-0.0382978991,
-0.0084816972,
0.003700302,
0.0244780257,
-0.0136886314,
0.0093676262,
0.0459022075,
-0.0079942998,
-0.04880457,
0.0071455245,
0.0815730691,
0.0255979672,
-0.042409461,
-0.0307038054,
-0.0471874997,
-0.0385935195,
-0.023230901,
0.0168788303,
-0.0733977929,
-0.0021579317,
-0.0151351904,
0.0653467551,
-0.0223390982,
-0.0435480773,
0.0094542243,
-0.0934277028,
0.0568880923,
0.002054513,
-0.0183593724,
-0.0159890577,
0.0133722331,
-0.005015783,
0.0374692492,
-0.0071548568,
0.0028608388,
-0.0107718939,
0.0604658797,
0.0548500158,
-0.0000767489,
-0.026176583,
-0.0050145905,
-0.0011177476,
-0.0057567339,
0.0354199074,
0.0051158876,
-0.0049347677,
-0.0668746829,
-0.0532738939,
0.08404544,
0.0039589116,
-0.0049968804,
0.0483915135,
-0.0311398152,
0.0376848802,
-0.0167887099,
0.0510826372,
-0.0315880291,
-0.0062515591,
0.0605274737,
0.0512373745,
-0.008647358,
0.0080602141,
-0.0185787771,
0.0324093364,
0.0562547967,
0.0372113883,
0.0195881203,
-0.0038022839,
0.0215707123,
-0.0029271077,
0.0713883415,
-0.0511636324,
-0.0326806009,
-0.0003253186,
0.0171890743,
0.0341748111,
-0.0127139287,
-0.0474332683,
0.0248546861,
0.0530911908,
-0.0140835727,
-0.0624624416,
-0.0015153077,
0.0270056389,
-0.0490512773,
0.0471835136,
-0.0025879168,
0.0047133085,
-0.0624964163,
-0.0422971174,
0.0073880213,
-0.0518902093,
-0.0230876561,
0.0006658003,
-0.0200034194,
-0.0201117843,
0.0318510123,
0.0758770332,
0.02231065,
-0.0237711035,
0.0578607582,
0.0208675712,
0.0013386061,
0.0277229119,
-0.0174055938,
0.0076980037,
0.0202634688,
0.0330926031,
0.0126682604,
-0.024250336,
-0.0007840147,
0.0287967268,
0.0081631681,
0.0072333789,
0.0375628136,
-0.000907197,
0.0125749074,
0.039079424,
0.049842231,
-0.0153727951,
0.0196174216,
-0.0505386628,
-0.0090618283,
-0.0416252986,
0.0057312045,
-0.0477699898,
-0.0135477474,
0.0043543866,
0.0190983322,
0.0043686307,
-0.0246273261,
-0.0439711213,
0.0304595269,
-0.0377239808,
0.0384215452,
-0.0398706608,
0.0098005608,
-0.0088047944,
-0.0008338251,
-0.0003066045,
-0.0402597561,
0.0272217449,
0.0418874137,
0.0207765307,
0.1255677342,
0.0200787596,
0.0226906799,
0.043867711,
0.0214877557,
-0.0133074932,
-0.0662935749,
-0.0519167222,
0.0369963162,
-0.032797575,
0.0037363837,
-0.0051883291,
-0.0200636256,
0.0386819765,
0.0221279394,
0.0202973504,
-0.0075878697,
-0.0677829683,
0.0152509436,
-0.0121445656,
0.0407517701,
-0.0211171284,
0.029164305,
-0.0720068365,
0.0033368156,
-0.0515572466,
-0.0281242989,
-0.0185112655,
-0.0003928098,
0.0353767946,
0.004096047,
0.0405298285,
-0.0059597506,
-0.0793117285,
-0.0132216746,
-0.0340420566,
0.0210299194,
-0.0186378453,
-0.0404296294,
0.0528389215,
0.0638803169,
-0.0561679266,
-0.032537628,
-0.0081295688,
0.0753632486,
-0.0680263564,
-0.015568532,
-0.0252412912,
-5.711115794e-33,
-0.0195467677,
-0.0509771779,
0.0540790148,
0.0409671254,
-0.0157960169,
-0.0029384613,
-0.0416971371,
0.0406125709,
0.0291466564,
0.0038743292,
-0.0088120131,
0.024129713,
-0.0028129723,
0.040995881,
0.0191631615,
0.0077239461,
0.0361285768,
0.0104685146,
-0.0127060292,
0.0395026058,
-0.0297335871,
0.009775673,
0.0717857704,
-0.0372874662,
0.05794001,
-0.0392067395,
0.0071856491,
-0.0311502218,
0.0142467376,
-0.0103806481,
0.0007829723,
0.050920099,
-0.014039861,
-0.061020609,
0.0135068661,
0.0345058292,
-0.0164816696,
-0.0872825086,
-0.0386090167,
0.0081763035,
-0.1027798355,
-0.0272141322,
-0.0191780198,
-0.0009452056,
-0.0503287055,
-0.0020026797,
0.055476401,
0.0237662885,
0.0434679799,
-0.0114643211,
-0.0356289558,
-0.0032437702,
-0.0473696068,
-0.0714624971,
-0.0236451626,
0.0850894079,
0.0037142462,
-0.0067167943,
-0.0393368155,
0.0445704572,
0.0716171935,
-0.002053075,
0.0330862738,
-0.0005038233,
-0.0159980599,
0.0090162754,
0.0339838751,
0.0092763938,
-0.0233165342,
-0.0205014795,
0.0451123677,
0.0140343569,
0.058460623,
0.0274659749,
0.0269643702,
0.0046655713,
-0.0090326536,
0.0883188024,
0.051974453,
0.0175828412,
-0.0061634309,
-0.04424797,
-0.0233102031,
-0.0054228762,
-0.0267953463,
-0.03472168,
0.0073733791,
-0.0417239293,
0.0615214743,
-0.0205417611,
0.0182573721,
-0.0070119733,
0.0018555003,
-0.0039163269,
0.0368952341,
-0.0016455054,
-0.0645692497,
0.0299973004,
-0.0347752385,
-0.0188395064,
-0.0269446466,
-0.0252618771,
0.0366385207,
-0.0043619224,
0.0125678573,
-0.0062120217,
-0.0216701869,
0.0310124308,
-0.014847937,
-0.0082220118,
-0.0257146787,
0.0514076762,
0.0508756116,
-0.0462676771,
0.0073903347,
0.0341528915,
-0.0194687285,
0.0403304771,
0.002786983,
-0.0017425163,
-0.0437776223,
-0.0081981029,
-0.0222033374,
-0.0229208246,
-0.0110072084,
0.0098762317,
-0.0123515669,
-0.0180599894,
-0.013835921,
-0.0329451151,
-0.0035590243,
-0.0333218239,
2.98e-7,
-0.0147848213,
0.0372047648,
0.0283091832,
-0.0446358472,
-0.028002223,
0.0176185686,
0.0278366283,
-0.0002635894,
0.0841394514,
0.0069674468,
0.0349016339,
0.0150286639,
0.0371337645,
0.0243003275,
-0.0489483625,
-0.0463959314,
-0.0339195058,
0.0054888404,
-0.0063944347,
0.0396014117,
0.1141766086,
-0.0357176885,
0.0056100208,
-0.0019653169,
-0.0054649385,
-0.0053253137,
0.033595603,
0.0169415362,
0.0116200596,
0.0150666311,
0.0498700999,
-0.0115946261,
0.0322622545,
0.036184065,
0.0067321262,
0.0457532555,
0.0009442449,
0.0158443488,
0.0050561223,
0.0042997822,
0.0225286465,
0.0425317101,
-0.016590355,
-0.0191322565,
-0.0106102377,
-0.0812548175,
0.0131025044,
0.0718236417,
-0.0823245123,
-0.0234918278,
0.0096403621,
-0.0147698196,
-0.021422755,
0.034313567,
0.0015280928,
-0.0196448322,
0.0147639103,
-0.0050442098,
0.0074569671,
0.0145496307,
-0.0477883592,
-0.0285974164,
-0.0473058224,
-0.0055250754,
-0.0254473947,
-0.0299940314,
-0.0043685292,
2.289773915e-34,
0.0001079574,
0.0043877321,
0.0605375767,
0.0054100771,
0.013632617,
0.0331488512,
-0.0455691181,
0.0470245741,
0.0466571935,
-0.0241383389,
-0.0480633117
] |
http://arxiv.org/abs/2402.02591v1 | http://arxiv.org/pdf/2402.02591v1 | On the performance of phonetic algorithms in microtext normalization | 2024-02-04T19:54:44 | 2024-02-04T19:54:44 | Yerai Doval, Manuel Vilares, Jesús Vilares | cs.CL | cs.CL | User-generated content published on microblogging social networks constitutes
a priceless source of information. However, microtexts usually deviate from the
standard lexical and grammatical rules of the language, thus making its
processing by traditional intelligent systems very difficult. As an answer,
microtext normalization consists in transforming those non-standard microtexts
into standard well-written texts as a preprocessing step, allowing traditional
approaches to continue with their usual processing. Given the importance of
phonetic phenomena in non-standard text formation, an essential element of the
knowledge base of a normalizer would be the phonetic rules that encode these
phenomena, which can be found in the so-called phonetic algorithms.
In this work we experiment with a wide range of phonetic algorithms for the
English language. The aim of this study is to determine the best phonetic
algorithms within the context of candidate generation for microtext
normalization. In other words, we intend to find those algorithms that taking
as input non-standard terms to be normalized allow us to obtain as output the
smallest possible sets of normalization candidates which still contain the
corresponding target standard words. As it will be stated, the choice of the
phonetic algorithm will depend heavily on the capabilities of the candidate
selection mechanism which we usually find at the end of a microtext
normalization pipeline. The faster it can make the right choices among big
enough sets of candidates, the more we can sacrifice on the precision of the
phonetic algorithms in favour of coverage in order to increase the overall
performance of the normalization system.
KEYWORDS: microtext normalization; phonetic algorithm; fuzzy matching;
Twitter; texting | On the performance of phonetic algorithms in microtext normalization: User-generated content published on microblogging social networks constitutes
a priceless source of information. However, microtexts usually deviate from the
standard lexical and grammatical rules of the language, thus making its
processing by traditional intelligent systems very difficult. As an answer,
microtext normalization consists in transforming those non-standard microtexts
into standard well-written texts as a preprocessing step, allowing traditional
approaches to continue with their usual processing. Given the importance of
phonetic phenomena in non-standard text formation, an essential element of the
knowledge base of a normalizer would be the phonetic rules that encode these
phenomena, which can be found in the so-called phonetic algorithms.
In this work we experiment with a wide range of phonetic algorithms for the
English language. The aim of this study is to determine the best phonetic
algorithms within the context of candidate generation for microtext
normalization. In other words, we intend to find those algorithms that taking
as input non-standard terms to be normalized allow us to obtain as output the
smallest possible sets of normalization candidates which still contain the
corresponding target standard words. As it will be stated, the choice of the
phonetic algorithm will depend heavily on the capabilities of the candidate
selection mechanism which we usually find at the end of a microtext
normalization pipeline. The faster it can make the right choices among big
enough sets of candidates, the more we can sacrifice on the precision of the
phonetic algorithms in favour of coverage in order to increase the overall
performance of the normalization system.
KEYWORDS: microtext normalization; phonetic algorithm; fuzzy matching;
Twitter; texting | [
0.0303240027,
-0.0190093219,
-0.0038439257,
-0.0049597062,
-0.0783244744,
-0.0081770495,
0.0209145863,
0.0316195115,
-0.0337431356,
-0.0134983491,
0.0448015854,
-0.0361956395,
0.0633344427,
-0.0092877727,
0.014102431,
-0.0414986424,
0.0076615475,
-0.0154785272,
0.00678324,
-0.0436493196,
0.0791299865,
0.014861457,
0.0110668233,
0.0318717025,
-0.014661653,
0.0352329686,
0.0319767967,
-0.0137874354,
-0.0027055759,
0.0299506783,
-0.0280609354,
0.0236941874,
-0.0509106629,
0.0413939282,
0.0000015507,
-0.0354996137,
-0.0531352013,
-0.026982924,
0.0523701906,
-0.013335228,
0.0599593371,
-0.018123515,
-0.0476405583,
0.0791786611,
0.01138735,
0.0001172294,
0.1019893214,
-0.0019885374,
0.0156765915,
0.052643206,
-0.0405718498,
-0.0808925405,
-0.0066502136,
0.0206577871,
-0.0081811026,
-0.006977479,
0.0132221123,
-0.0244468898,
0.0008522078,
-0.0195324142,
0.032234557,
0.0422293171,
-0.0434262455,
-0.02732265,
0.0099874623,
-0.0120071433,
0.0092527131,
-0.0321142189,
0.009247194,
0.0238642879,
-0.0598113388,
-0.052588243,
0.0319319814,
0.025284104,
-0.0230511352,
0.0085445018,
0.0080895768,
-0.0235957,
-0.0369508266,
0.0304537974,
0.0202333219,
0.0143907042,
0.0089281974,
0.0245753136,
-0.0184266511,
0.0378005728,
0.0176895969,
-0.0412862934,
-0.0252450965,
-0.0102886669,
0.023019189,
-0.0098604122,
0.0948216915,
0.0307128802,
0.0226039346,
-0.0081927422,
-0.0184838437,
-0.0322871432,
0.0124933179,
-0.0129539715,
-0.0229880922,
0.0173508674,
0.0766785145,
-0.0211737007,
-0.011428738,
-0.0199031141,
0.004523078,
-0.0299202409,
-0.0480372235,
-0.031669464,
-0.0351756625,
0.0143314302,
-0.091265738,
0.0351205319,
0.0171115585,
0.0013464383,
-0.053491544,
-0.0156331602,
0.011224228,
-0.0114123197,
-0.0045401612,
0.0027568543,
-0.0044095484,
-0.0209334567,
0.0386749022,
0.0424765833,
-0.0507857986,
0.0163672101,
-0.059103895,
0.0141793955,
0.0487814061,
0.02201901,
0.0913421884,
0.0256838929,
-0.0055815438,
0.0140337609,
-0.0011135004,
-0.0131971976,
-0.0536013171,
-0.0280305222,
-0.0301380605,
0.0175355002,
-0.0540845506,
-0.0642678663,
-0.0243500732,
-0.0160649922,
0.0071292827,
0.007303467,
-0.0164939724,
0.0127399061,
-0.0115383752,
0.0532137863,
-0.0560500249,
0.0250402503,
0.065331161,
-0.0164516196,
0.0344525836,
0.0416225642,
0.0642354414,
0.0240548663,
0.0476726294,
-0.0514165834,
0.0081918109,
0.062199492,
0.0075414386,
-0.0133973705,
-0.0017171251,
-0.0637382269,
-0.0530811734,
0.046847783,
0.0010250555,
0.1127894521,
0.0031283915,
-0.0048166355,
0.023194721,
0.0924061015,
0.0728846863,
0.0203236956,
0.0208187662,
0.0166640822,
0.026918292,
0.0138811907,
0.0416932255,
-0.0373619199,
-0.0494411439,
0.0008749944,
0.0298942979,
0.0414939076,
-0.0366023928,
0.0080565214,
0.011795288,
-0.0136269517,
-0.0081962226,
-0.0468493514,
0.0244908109,
-0.0048606684,
0.0126165058,
0.0250767395,
-0.0205758493,
-0.0278766192,
0.0442170948,
-0.0141550032,
-0.0029513324,
0.0700955763,
0.0538523495,
-0.0061608911,
0.0128315343,
-0.0106089059,
-0.0758705139,
0.0375049524,
-0.0297095831,
-0.0414302908,
-0.024347892,
-0.0684822276,
0.0159326419,
0.1156160608,
-0.0125513049,
-0.0077052591,
-0.0006690609,
0.0395396501,
0.0227436945,
-0.0065615023,
-0.0686990768,
0.0024565842,
0.0242266823,
-0.0337304063,
0.0207115896,
0.0176095478,
-0.0110180918,
0.0828026384,
0.0117974877,
0.0001178778,
-0.0121582532,
-0.0304039456,
0.0255683791,
0.0170706175,
-0.0031429771,
-0.030722674,
0.0731731951,
-0.0585419312,
0.0186909325,
0.0240729433,
0.0138574904,
-0.0665988773,
0.0167409182,
-0.0428327471,
0.008082293,
-0.0480750538,
0.0573464036,
-0.059005294,
-0.0192446951,
-0.0023938487,
-0.0359262377,
0.0150859216,
0.0163842794,
-0.0086564077,
-0.0048790644,
0.0340232924,
-0.0322578698,
-0.0374494977,
0.0332089067,
-0.0298913158,
0.063965641,
0.0248826537,
-0.0181745626,
-0.0259774644,
-0.0274928845,
-0.0430521443,
-0.0230411664,
0.0052271285,
0.0405528545,
0.0198978093,
0.0745978653,
0.039355021,
-0.0084608626,
-0.0363167115,
0.0120984586,
0.0026407638,
0.0148088457,
0.0524105988,
-0.0310207363,
-0.0017827853,
-0.0249502081,
-0.0139725422,
-0.0050185742,
0.0088607194,
0.0186465494,
-0.0235582832,
-0.0407196023,
0.0101674218,
-0.0266883206,
0.0110785076,
0.0618523546,
-0.0753685236,
-0.0089226421,
-0.0041448232,
0.0407007448,
-0.0183599759,
0.0072786077,
-0.0379442126,
-0.0075488654,
0.019001374,
0.0046074786,
-0.0544713922,
0.0143459784,
0.0323945768,
-0.0275254231,
-0.045584131,
-0.0166006219,
0.0597827882,
0.002825733,
0.0145609332,
-0.0647967532,
0.0241956357,
-0.0512232557,
-0.0142739229,
-0.0151718697,
-0.0072663031,
-0.0456631258,
0.0214656573,
-0.0217752047,
-0.0415005721,
0.0439709015,
-0.0266467743,
0.0021699734,
-0.070572786,
-0.040405564,
-0.0063723018,
-0.0048843976,
-0.0405034311,
0.0613101684,
-0.0190116558,
-0.0444504581,
0.02464341,
0.0032238734,
0.0536108948,
-0.0449623317,
0.0008881823,
0.0413404591,
0.020648621,
-0.0106039243,
-0.0186650157,
0.0210924167,
-0.0521235093,
-0.0352843851,
0.0286213513,
0.08059562,
-0.0375311524,
-0.0008593537,
0.0934379324,
0.0066384277,
0.0143654132,
0.0047211181,
0.026992036,
0.0171603933,
0.0878775492,
-0.0151913548,
-0.0086554913,
-0.0455704927,
0.0484853163,
0.0461515225,
-0.0143338833,
0.0869309381,
-0.0262572691,
-0.0175419394,
-0.0074534942,
0.0107029621,
-0.0540653951,
0.0120970123,
-0.0144030722,
0.0305969827,
0.011383554,
0.0335647389,
0.1203389689,
0.0383987203,
-0.0206821803,
-0.0154037736,
0.029539166,
-0.0022002955,
-0.0429187454,
-0.0145941572,
-0.0865752175,
0.0077942759,
0.0096025737,
0.0453063957,
-0.0196596384,
-0.03232437,
0.0026532551,
-0.0567225441,
-0.0279890634,
0.0439963117,
-0.0451508425,
-0.0251498576,
-0.0066300095,
0.0110620419,
-0.0200783033,
0.0226470567,
0.0342079103,
-0.0063159894,
-0.0021652144,
0.0260062404,
-0.0397843719,
-0.0573788546,
-0.0319627002,
-0.0750240088,
0.0623445883,
-0.0051084668,
0.0477811694,
-0.0245208517,
-0.0001507421,
-0.0286890082,
-0.0275129471,
-0.0156529583,
-0.0275309496,
0.0199233647,
0.021467166,
-0.0118296528,
-0.0597500093,
-0.0136238746,
0.0053268368,
0.0244908556,
-0.0071890401,
0.0312425196,
0.0141816419,
-0.020478636,
0.014731355,
0.0278174505,
0.0089439275,
0.0380655564,
0.0006440644,
-0.0585204773,
-0.0368384235,
-0.0402138308,
-0.0090269921,
-0.0664791316,
-0.0408547744,
-0.0104873041,
0.0079751266,
-0.106197007,
-0.0747172683,
-0.0392044112,
0.0291035101,
0.0701884255,
0.0221374426,
0.0101315053,
0.0799826756,
-0.0094326399,
-0.0084532583,
-0.0657222569,
-0.0034912522,
0.0207628962,
-0.0333713479,
0.0343361795,
0.0325685553,
-0.0196344908,
0.027885085,
0.0008498644,
-0.0265227333,
-0.0693221465,
0.0445093028,
-0.0524352379,
0.0290156584,
0.0005255091,
-0.0644019991,
-0.098751232,
0.0519036986,
0.0437496565,
-0.0191987474,
-0.0080569396,
0.043107193,
0.0519587994,
-0.0102568381,
0.0348323323,
-0.0034266636,
0.0285617691,
-0.0441123322,
-0.0119954199,
-0.0251295771,
-0.0494198725,
0.0074839313,
0.047561992,
-0.0259545203,
-0.0519262999,
-0.0034243546,
-0.016795164,
-0.0376096666,
0.0217559747,
0.013419684,
0.0100798663,
0.0075728148,
0.0580373034,
0.0319249146,
-0.0201311298,
-0.0560356602,
-0.0235543754,
-0.0410038605,
-0.0123752728,
0.0381286964,
-0.0308049805,
0.065103218,
0.0022899592,
0.0058161411,
0.0082811937,
0.053375259,
-0.0124413222,
-0.0097914767,
-0.0297593735,
-0.0743006915,
-0.0031882664,
0.0177161153,
-0.0310580786,
0.0319934078,
-0.0164306071,
0.0446175896,
-0.0182438381,
0.0030724551,
0.0940735564,
-0.0095383981,
0.0661014766,
0.0098267021,
0.0223696064,
-0.0043482766,
-0.0307776686,
-0.0343252569,
0.0351896174,
-0.0332695805,
-0.0077298796,
-0.0175994001,
0.0141466279,
0.0500723384,
-0.0512072369,
0.0170110688,
0.0148684327,
-0.0285162348,
-0.0160790142,
0.0294768382,
-0.0586084947,
0.0567335226,
-0.040608108,
-0.0344520248,
-0.0432920344,
-0.0399977304,
0.0466539413,
-0.024379449,
0.0047164327,
0.0086630695,
0.0010868915,
0.0050676134,
0.014085535,
0.0605991408,
-0.0075938101,
-0.0189903751,
-0.006115383,
-0.0399015062,
0.0181159489,
-5.811616887e-33,
0.0287798718,
-0.0433986709,
-0.0322450139,
0.0021844038,
0.0084934048,
-0.0210719518,
-0.0097628068,
-0.0015127175,
-0.0162719376,
-0.0092502879,
-0.0098914634,
-0.0023019242,
0.0312473159,
-0.0196208097,
0.0040408871,
0.0339922532,
0.0255173892,
-0.0440277979,
-0.0007361182,
0.0041588834,
0.0239159558,
0.0246344805,
0.0157300849,
-0.0701795444,
-0.0138842911,
-0.0022318016,
-0.0485310741,
0.0207068622,
0.065357171,
0.0333687998,
-0.0151344538,
0.0782011971,
0.0266265869,
0.0180406943,
0.0103024812,
0.0446455106,
-0.0729940683,
-0.0370261855,
0.0377715416,
0.0202873871,
-0.0089684445,
-0.0354673229,
0.0415029787,
-0.0250022523,
-0.0543041043,
0.0511398092,
-0.0110811479,
-0.0046567554,
0.0450063646,
-0.0106792785,
-0.0342937782,
0.054491654,
0.0116819385,
0.0019698294,
0.1080912501,
-0.022466287,
-0.0140449014,
-0.0192587767,
-0.0121901678,
0.0571393445,
0.0305005237,
0.0511735864,
0.0165567193,
-0.0006544297,
-0.0327240601,
-0.0324388221,
0.0792687982,
-0.0023081077,
0.0353052989,
-0.0184439141,
0.0188816413,
-0.0151769631,
0.0424175635,
0.0547983237,
0.035692662,
-0.0434529148,
0.0063176393,
0.0257323459,
0.0242571346,
0.0033977106,
-0.0427554995,
-0.0010817924,
-0.0320275314,
-0.0309989117,
-0.015075624,
-0.0522567332,
-0.0384819023,
-0.0047825235,
-0.025286749,
-0.0083202533,
-0.0221909471,
0.0625574216,
0.0155001422,
-0.0298195053,
0.0162892751,
-0.0155611802,
0.025527956,
0.0360177308,
-0.0202451311,
0.0266966168,
-0.0392384902,
-0.0574405454,
0.0485751815,
0.0104324939,
-0.0237032752,
0.0069008567,
-0.0393940657,
0.0254242066,
-0.0588709749,
0.0028289831,
-0.0024589945,
-0.0003313541,
-0.0027173075,
-0.0152842822,
-0.0064498172,
-0.0186197869,
0.0171888378,
-0.0523448326,
0.080633983,
-0.0150477858,
0.005037724,
-0.0084310742,
0.0076581803,
-0.0099638673,
-0.0274188239,
0.0381211005,
0.0056215823,
0.0207365993,
-0.0010985001,
0.0140002584,
-0.0204953756,
-0.0099858772,
2.455e-7,
0.0429051034,
0.0601993836,
0.0748562291,
-0.0188847445,
0.032192599,
-0.0124935471,
0.013565301,
-0.0009790327,
-0.0336962007,
0.0567389429,
0.0623522028,
-0.0541710369,
0.0613616928,
0.0259874407,
-0.0678635463,
0.0098233195,
-0.034167029,
-0.017812144,
-0.0175915714,
0.0760167614,
0.0091839451,
0.036380101,
-0.0047620921,
-0.0414233468,
0.0012810361,
-0.0976523086,
-0.018413648,
0.0054613557,
0.0283989627,
0.0002212002,
0.0538144559,
-0.0010986207,
-0.001628887,
-0.0040771309,
-0.0055227866,
0.0298210029,
0.0268487725,
0.0231138598,
-0.0388901196,
0.0426645912,
0.0047897152,
-0.0560826287,
-0.0074691223,
-0.0687064305,
0.0412500836,
0.0317271464,
-0.0417714529,
-0.0200769566,
0.0082339402,
-0.0130079677,
-0.0197375007,
-0.0153762214,
-0.0571416952,
-0.0003175233,
0.0128878811,
0.0231214855,
-0.0304926299,
-0.0122111058,
-0.0202195924,
0.0217998791,
0.0063551916,
-0.0547560677,
-0.011615512,
-0.002646293,
0.0343293585,
-0.0292142313,
0.0612782687,
1.785470997e-34,
0.0320781656,
-0.0571223982,
-0.0130679738,
0.0135690505,
-0.0374005027,
0.0079358555,
-0.0035605226,
0.0184304733,
0.0041686012,
-0.0566158183,
-0.014038506
] |
http://arxiv.org/abs/2402.02572v1 | http://arxiv.org/pdf/2402.02572v1 | A Quantitative Discourse Analysis of Asian Workers in the US Historical Newspapers | 2024-02-04T17:32:52 | 2024-02-04T17:32:52 | Jaihyun Park, Ryan Cordell | cs.CL | cs.CL | Warning: This paper contains examples of offensive language targetting
marginalized population. The digitization of historical texts invites
researchers to explore the large-scale corpus of historical texts with
computational methods. In this study, we present computational text analysis on
a relatively understudied topic of how Asian workers are represented in
historical newspapers in the United States. We found that the word "coolie" was
semantically different in some States (e.g., Massachusetts, Rhode Island,
Wyoming, Oklahoma, and Arkansas) with the different discourses around coolie.
We also found that then-Confederate newspapers and then-Union newspapers formed
distinctive discourses by measuring over-represented words. Newspapers from
then-Confederate States associated coolie with slavery-related words. In
addition, we found Asians were perceived to be inferior to European immigrants
and subjected to the target of racism. This study contributes to supplementing
the qualitative analysis of racism in the United States with quantitative
discourse analysis. | A Quantitative Discourse Analysis of Asian Workers in the US Historical Newspapers: Warning: This paper contains examples of offensive language targetting
marginalized population. The digitization of historical texts invites
researchers to explore the large-scale corpus of historical texts with
computational methods. In this study, we present computational text analysis on
a relatively understudied topic of how Asian workers are represented in
historical newspapers in the United States. We found that the word "coolie" was
semantically different in some States (e.g., Massachusetts, Rhode Island,
Wyoming, Oklahoma, and Arkansas) with the different discourses around coolie.
We also found that then-Confederate newspapers and then-Union newspapers formed
distinctive discourses by measuring over-represented words. Newspapers from
then-Confederate States associated coolie with slavery-related words. In
addition, we found Asians were perceived to be inferior to European immigrants
and subjected to the target of racism. This study contributes to supplementing
the qualitative analysis of racism in the United States with quantitative
discourse analysis. | [
0.0206610728,
0.0975067914,
-0.0149505166,
-0.0033213494,
0.0057536699,
0.0513871871,
0.0583534576,
0.0081111128,
-0.0103656622,
-0.0732498094,
0.0087301023,
0.0173571967,
0.0671453178,
-0.0807560533,
-0.0330812372,
-0.0068514063,
0.0039457376,
0.0828412026,
0.043065019,
-0.0178858265,
-0.0025717583,
0.0318496004,
0.006216608,
0.0410379358,
-0.0538394786,
-0.0060746372,
0.0451295152,
-0.0227859281,
0.0053519392,
-0.0350638777,
0.0028670281,
0.0192121994,
0.0053658918,
0.0200298019,
0.0000016901,
-0.0175559651,
0.0069052484,
-0.0131068351,
0.0308854841,
-0.0059675192,
0.0837780908,
-0.001738208,
-0.0155019835,
0.0145449191,
-0.0309659615,
-0.0144125903,
0.0326084942,
0.0335821398,
0.0353192911,
0.0096660098,
-0.0035471597,
-0.0187504701,
-0.0343284085,
0.021828549,
-0.0357850343,
0.0398402885,
0.0459646583,
0.0028522238,
-0.0306500793,
-0.0713965222,
-0.0231622588,
0.0203216951,
0.0011561614,
0.027414212,
-0.0270054191,
-0.0057307626,
-0.0317820385,
0.0430200845,
0.0172111951,
-0.0040821549,
0.0444755964,
0.0096202968,
0.006436062,
0.0504891612,
-0.0456425361,
0.0077782804,
0.0161846615,
0.0149635514,
-0.0274261031,
-0.0002610399,
-0.0056673726,
-0.0353045762,
0.0105400775,
0.0551348217,
0.0195525438,
0.1018534154,
-0.0090058194,
-0.0017337493,
-0.0219199657,
-0.0125908852,
-0.0106309531,
-0.0301420875,
0.0523399413,
0.0309494771,
-0.0352575257,
0.041859407,
0.0278651398,
-0.0665811524,
-0.0034314217,
-0.0758237168,
0.0062659779,
0.0193202998,
-0.0251952969,
-0.0339068696,
0.0598303489,
0.0201372001,
0.0123892073,
-0.0075891274,
0.0138430148,
-0.0029762264,
-0.0237160679,
-0.022670893,
-0.002956514,
-0.0034936895,
0.0121674975,
0.0162148532,
-0.0448416099,
-0.0384831987,
-0.0294740964,
-0.0050459164,
-0.0262368396,
0.0059564062,
-0.0415926948,
0.0301127136,
-0.0412596427,
-0.0082082208,
-0.0766522959,
0.0273611657,
-0.063876316,
-0.0162161496,
0.0513095334,
0.0168010984,
-0.0219372287,
-0.0281357542,
-0.0185178444,
-0.0223308373,
-0.0131603377,
0.0180950556,
-0.0381199941,
-0.007653906,
-0.0264048371,
-0.0299237501,
-0.0325004086,
-0.0535912476,
-0.0213680044,
0.0180793758,
0.0198408756,
-0.0764085427,
0.0135317883,
0.0151511664,
-0.0302846283,
0.03904837,
0.0277567711,
0.0055476399,
0.0924752504,
0.0236864872,
0.0169996861,
0.039891649,
0.069124721,
-0.0015208265,
0.0423521847,
-0.00887419,
-0.0109724673,
0.0075458521,
-0.0331398137,
0.0108179506,
-0.0662445873,
0.0157715343,
-0.0372286178,
-0.0143316472,
-0.0004228869,
0.0546775535,
0.0227956679,
-0.0107595241,
-0.0140500143,
0.1321869493,
-0.0740730986,
0.0118315369,
0.0419752747,
0.0498905815,
-0.0523078106,
-0.0023884657,
-0.0151370913,
-0.0617051497,
-0.0012034331,
-0.0750801116,
0.0122626312,
0.0337471738,
-0.0256247688,
-0.0225784369,
-0.0277603958,
0.0141025418,
0.057086233,
-0.0462568775,
0.031521339,
0.0048750797,
0.009292203,
0.0086714327,
0.0612983592,
-0.0074939523,
-0.0378392749,
-0.0136755239,
-0.0133295637,
0.060754206,
0.0089573208,
0.0096051339,
0.0931264237,
-0.0212944634,
0.0410932489,
0.0850064009,
0.0096397428,
-0.0035324153,
-0.0161637142,
-0.0158381723,
0.0004017439,
0.0370398574,
0.0370628536,
0.0354395211,
0.0481845848,
0.0580271259,
0.0137123102,
0.0649400428,
-0.0466157906,
0.0102602299,
0.0216129422,
-0.0532636754,
-0.0182735547,
-0.0031498941,
0.0361808278,
0.0167606268,
-0.0316165127,
-0.0347398818,
-0.0139195416,
-0.0006250017,
-0.0414594859,
0.0817332044,
-0.0377457961,
0.0595271997,
-0.0042097489,
-0.0659250394,
-0.0122336382,
0.002674618,
0.0309397075,
0.0080972807,
0.0743155777,
0.0008998719,
-0.0096077975,
-0.0209126696,
0.0098098842,
-0.0508379117,
0.0101571921,
-0.0177748464,
0.0502936915,
-0.0114246141,
0.0232446827,
0.0484815016,
-0.0699205771,
0.0420345701,
0.0178345479,
-0.054757189,
0.0221147481,
-0.0173604377,
0.0434751175,
-0.0280880388,
-0.0561821945,
-0.0049563772,
0.0258434508,
-0.0715182647,
0.0094680171,
0.0042150933,
-0.0431034565,
-0.0028615769,
-0.0036135346,
-0.0240701437,
-0.0007394649,
0.0497796349,
-0.0360393189,
0.0147326561,
0.0342836529,
0.0051620631,
-0.0303439181,
0.0279326178,
-0.0105796037,
-0.0229506511,
-0.0548282042,
0.0381654873,
-0.0131875724,
-0.0328646749,
-0.0114681711,
-0.0104092592,
-0.0444913246,
-0.0046710502,
-0.0047536893,
-0.0395743437,
0.0564119704,
0.0498836376,
-0.01347306,
0.0388659313,
0.0458162352,
0.1059071943,
0.0313723348,
-0.0763622522,
0.0072749569,
-0.0159990285,
-0.0150850294,
0.0245154854,
0.0142163523,
-0.0665567517,
-0.0241483655,
0.0559254102,
0.0533701666,
0.0043993448,
-0.037335515,
0.0397740602,
-0.1103476435,
-0.0329397246,
-0.0117652118,
-0.0494902246,
-0.0562910847,
0.0171614662,
-0.01938463,
-0.0130424891,
-0.0015315611,
0.0261051338,
0.0604703315,
-0.0332064405,
0.0092696296,
0.0080389334,
0.0774575993,
-0.0153995706,
0.0298713837,
-0.0291835256,
-0.0576118641,
0.0006868872,
-0.011558176,
0.0235917773,
-0.0525697917,
0.0355736502,
-0.0243122131,
-0.0061547924,
0.0190978106,
-0.0002998472,
-0.0173179712,
-0.0761398897,
-0.0478614196,
-0.0166253187,
0.0099093905,
-0.0184591394,
-0.0270178802,
0.0447783545,
0.0264003407,
-0.0221882127,
-0.0106554208,
0.0299117416,
0.0070144255,
-0.0158737116,
-0.0070962138,
0.0163991265,
0.0183693618,
0.0412459038,
0.0022946773,
0.0285267886,
-0.0097361226,
0.0225931667,
0.0295286719,
-0.0056484537,
0.0122515196,
-0.025567919,
0.0124483034,
-0.0253706053,
0.0477167889,
-0.0155279534,
0.0398610719,
0.0668880492,
0.0384016149,
-0.0129712466,
0.010914064,
-0.0052032797,
-0.0286615156,
0.0197141524,
0.0068406286,
-0.0921802744,
0.0629863888,
0.009247018,
0.0169291086,
-0.0905591324,
-0.0461208262,
0.0010325059,
-0.0827013701,
-0.0051513142,
-0.0182207841,
0.0233071893,
0.0264720097,
-0.0235425271,
0.029854307,
0.0520978272,
0.0236776732,
-0.0095078116,
0.0389557295,
-0.0066958885,
-0.0041529783,
-0.0353116505,
0.0071129715,
0.0220617279,
0.0163404681,
-0.0404265709,
-0.0262137093,
0.0572608933,
-0.0263210759,
-0.0012945901,
0.0032582597,
-0.0290859956,
-0.0233932938,
-0.0634987652,
0.0226982031,
0.0116633186,
-0.0155010018,
-0.024899194,
0.0077406331,
0.0246128011,
0.0185373332,
0.0462403521,
-0.0303834733,
0.0122436294,
-0.022075681,
-0.0068844799,
0.0019442309,
0.009113837,
0.0299403239,
0.0503679961,
-0.0447304137,
0.0381276794,
-0.0855866373,
0.0396451764,
-0.0081803165,
-0.0792402625,
-0.0128548732,
-0.0230117664,
-0.0798299685,
-0.0396275818,
-0.090150319,
-0.0064402185,
0.0422769934,
0.0671918392,
-0.0160425734,
0.0358649865,
-0.0383230783,
-0.0805540085,
-0.0015474394,
-0.036621578,
0.019504115,
0.0544054732,
-0.0249097943,
-0.00913921,
0.0320354998,
-0.012749943,
-0.0218059868,
-0.0121268844,
-0.0431112088,
0.0912869051,
0.0119491061,
0.0228455346,
-0.0106035247,
-0.0378628671,
-0.0298563242,
0.0347839147,
0.0668836758,
-0.0107001532,
0.0237848815,
-0.0417080522,
-0.0250036139,
0.0079496363,
-0.007020175,
0.0125080021,
0.0479099229,
0.015954705,
-0.0183521863,
-0.0485648997,
0.0299681593,
-0.0076568266,
-0.0222562309,
0.0076461011,
-0.0247716978,
-0.0249607265,
0.009709659,
-0.0440481268,
-0.030509308,
-0.0430238247,
0.0109070251,
-0.0154969795,
-0.0265126117,
-0.0004522101,
-0.0697100386,
-0.0192578137,
-0.0255108196,
0.1028754562,
-0.0488614924,
0.0544397719,
0.0068949601,
0.0814179629,
-0.0012850828,
-0.001912541,
-0.0124009978,
0.0665459111,
-0.0390128493,
0.0308510996,
0.0328234062,
0.0390822925,
0.017075127,
0.0103172269,
-0.0032117665,
-0.0024536289,
-0.029989535,
-0.0189954452,
0.0118422257,
-0.0239769295,
0.0028796946,
-0.0502365865,
0.0147302365,
0.0365901068,
-0.0199297629,
-0.007316187,
0.0072125178,
0.0426226892,
-0.0487261303,
0.0515203997,
0.0167790111,
0.0056970199,
0.0234196745,
-0.0090902289,
-0.0712609291,
0.0319241285,
0.1024779379,
-0.0858251154,
0.0109965587,
-0.0371082276,
0.0257474668,
0.0094114291,
0.0365208872,
-0.0356736705,
-0.0627375618,
-0.0283895433,
0.0143172788,
-0.0089627113,
0.0456010103,
-0.030455118,
-0.0188876111,
0.0259394627,
-0.0006834335,
0.00917341,
-0.0396796428,
0.0390053988,
-0.066273056,
-0.0193200596,
0.035123568,
-5.730997444e-33,
-0.0275598802,
-0.1097101197,
0.0016379224,
-0.0069454731,
-0.0297753122,
-0.0119109508,
-0.045958519,
0.0175781399,
0.0162339006,
0.0063066268,
-0.0395606309,
0.0100145359,
0.0127983093,
-0.009170373,
0.0143596455,
0.011831875,
0.0098643284,
0.0062422007,
-0.0179828946,
-0.0005127197,
0.0323900767,
-0.027866127,
0.0136465058,
-0.0380362198,
0.0530808158,
0.0145283733,
-0.0431313142,
-0.0126937116,
0.0442354567,
0.0222319383,
-0.0264787171,
0.0596854202,
0.0318741165,
0.010668599,
0.0334485918,
0.0883449018,
-0.0368531197,
-0.0107681528,
0.0301318113,
0.0332154669,
0.0762175098,
-0.011755418,
-0.0224047303,
0.0049581034,
-0.0224888138,
0.0162245799,
0.0546548925,
-0.0203073714,
0.0331118815,
0.0020480449,
-0.0347671211,
0.0177286882,
-0.046578344,
0.0405030809,
0.0076930947,
0.0018236149,
0.0181552693,
0.0177965499,
-0.0667012334,
-0.0028165635,
-0.0161423068,
-0.0099115884,
0.0088775465,
0.0106455795,
0.0356976017,
0.0237394795,
0.0189716239,
0.0256586839,
-0.0621136166,
-0.0035101001,
0.0621824265,
-0.0416936912,
-0.0414917767,
0.0470431969,
0.0498039052,
-0.1054773182,
0.0332313292,
-0.0431697033,
0.0546180904,
0.035409078,
0.0166640393,
-0.0085414294,
-0.0272093136,
-0.0588782094,
-0.031173557,
-0.0320835859,
-0.0049195243,
-0.026844345,
0.0118978573,
-0.0147609711,
-0.008773637,
0.0136931539,
0.0261337049,
0.0002681714,
-0.0593366921,
0.014132957,
-0.0503276922,
-0.0138566708,
0.0090830689,
0.0047609764,
-0.0472578034,
0.0343616158,
-0.0333795473,
0.0364383422,
0.0303614475,
0.0123671321,
0.002839661,
-0.0049742302,
-0.0350516848,
0.0120170088,
0.0319887958,
-0.0002900089,
-0.0305357054,
0.0158584267,
0.023912698,
0.0677785501,
0.0050491863,
-0.0242818315,
0.0689752176,
-0.0265627019,
-0.0082010301,
-0.0175829194,
-0.0275140665,
-0.0155034102,
-0.0488402247,
-0.0291791074,
0.041668389,
0.0879055858,
0.0524378531,
-0.0619946718,
-0.0075012287,
-0.0585674755,
2.381e-7,
0.0368517078,
0.035908673,
0.008071986,
0.0441236235,
-0.012928335,
-0.013659752,
0.024887789,
-0.0212060399,
0.0277383253,
-0.0016703233,
0.0104425186,
0.022922134,
0.0245320499,
0.0478668027,
-0.0587367043,
-0.0044244011,
-0.0889581218,
0.0063628103,
-0.0316447541,
0.0456899293,
0.0017188721,
0.0067013851,
0.0083879791,
-0.0040261233,
-0.0712751374,
0.0166849364,
-0.0063339234,
-0.0477197431,
0.0176362861,
0.0079990579,
0.0136179645,
-0.0156676434,
-0.0018229964,
-0.0053377142,
0.0223453641,
0.048896879,
0.0351292379,
0.0151412254,
0.0040062936,
0.063287437,
-0.0120692151,
0.0095359674,
-0.0354120806,
-0.0304526687,
0.004134797,
0.0405672751,
-0.0357900225,
-0.04775621,
-0.0653581098,
-0.0062547559,
0.016523933,
-0.0216872953,
-0.0259958971,
0.0013288329,
0.005747823,
-0.0039481497,
0.032160908,
-0.021688927,
-0.0253474005,
-0.0348956063,
-0.0153692355,
-0.0134347547,
0.0212867428,
0.0035834764,
-0.0261947252,
0.0104846135,
0.0487437472,
1.734987417e-34,
0.0296931807,
-0.0090150265,
-0.0226984248,
-0.0169943403,
0.0508173965,
0.0086320862,
0.0194662903,
-0.0413020663,
0.0221408587,
-0.0138420472,
0.006316239
] |
http://arxiv.org/abs/2402.02564v2 | http://arxiv.org/pdf/2402.02564v2 | A Truly Joint Neural Architecture for Segmentation and Parsing | 2024-02-04T16:56:08 | 2024-03-02T16:33:32 | Danit Yshaayahu Levi, Reut Tsarfaty | cs.CL | cs.CL, cs.AI, cs.LG | Contemporary multilingual dependency parsers can parse a diverse set of
languages, but for Morphologically Rich Languages (MRLs), performance is
attested to be lower than other languages. The key challenge is that, due to
high morphological complexity and ambiguity of the space-delimited input
tokens, the linguistic units that act as nodes in the tree are not known in
advance. Pre-neural dependency parsers for MRLs subscribed to the joint
morpho-syntactic hypothesis, stating that morphological segmentation and
syntactic parsing should be solved jointly, rather than as a pipeline where
segmentation precedes parsing. However, neural state-of-the-art parsers to date
use a strict pipeline. In this paper we introduce a joint neural architecture
where a lattice-based representation preserving all morphological ambiguity of
the input is provided to an arc-factored model, which then solves the
morphological segmentation and syntactic parsing tasks at once. Our experiments
on Hebrew, a rich and highly ambiguous MRL, demonstrate state-of-the-art
performance on parsing, tagging and segmentation of the Hebrew section of UD,
using a single model. This proposed architecture is LLM-based and language
agnostic, providing a solid foundation for MRLs to obtain further performance
improvements and bridge the gap with other languages. | A Truly Joint Neural Architecture for Segmentation and Parsing: Contemporary multilingual dependency parsers can parse a diverse set of
languages, but for Morphologically Rich Languages (MRLs), performance is
attested to be lower than other languages. The key challenge is that, due to
high morphological complexity and ambiguity of the space-delimited input
tokens, the linguistic units that act as nodes in the tree are not known in
advance. Pre-neural dependency parsers for MRLs subscribed to the joint
morpho-syntactic hypothesis, stating that morphological segmentation and
syntactic parsing should be solved jointly, rather than as a pipeline where
segmentation precedes parsing. However, neural state-of-the-art parsers to date
use a strict pipeline. In this paper we introduce a joint neural architecture
where a lattice-based representation preserving all morphological ambiguity of
the input is provided to an arc-factored model, which then solves the
morphological segmentation and syntactic parsing tasks at once. Our experiments
on Hebrew, a rich and highly ambiguous MRL, demonstrate state-of-the-art
performance on parsing, tagging and segmentation of the Hebrew section of UD,
using a single model. This proposed architecture is LLM-based and language
agnostic, providing a solid foundation for MRLs to obtain further performance
improvements and bridge the gap with other languages. | [
-0.0037115321,
0.0640822649,
-0.0119434092,
0.0603952594,
-0.0441589393,
0.0159100275,
0.0127263097,
-0.0088554909,
-0.031541314,
-0.0088971257,
-0.018517185,
-0.0637802482,
0.0180328023,
0.0004798488,
0.0238518436,
-0.0288585518,
0.0244209953,
-0.0581024848,
-0.0145018315,
0.0078118606,
0.0283113755,
0.0209293664,
-0.004314933,
-0.0257777553,
-0.0689845234,
-0.0202315263,
-0.005163088,
-0.0006271334,
-0.0031587323,
-0.0054806615,
-0.0260219388,
0.0063513336,
-0.050658945,
0.022393208,
0.0000017826,
-0.0241393466,
-0.0301208999,
-0.0089233443,
-0.0316203758,
-0.0101441983,
0.0461757481,
-0.0721778572,
-0.0699439794,
-0.0350240618,
-0.0071371403,
-0.0191221591,
0.0400960818,
0.0211819168,
0.0251029618,
0.0745456517,
-0.0200266689,
-0.0854273885,
0.0878422335,
-0.0135144573,
-0.0526347384,
0.0197571237,
-0.0048892121,
-0.0508209541,
-0.0205465928,
0.0058399029,
-0.0071135084,
0.019573031,
0.03532755,
0.0167830847,
-0.0490128957,
0.0124413362,
-0.0234169923,
-0.0394048877,
0.0285991635,
-0.018961018,
-0.0616039559,
-0.0166528579,
0.0268714912,
-0.0106009236,
0.018067712,
-0.0268612765,
-0.0299508385,
0.0699389502,
-0.0294498205,
0.0540906899,
0.0617071204,
-0.0303133801,
0.0228992738,
0.0022499787,
-0.0464454144,
0.0855147317,
0.0344482698,
-0.0505030118,
0.0077156336,
0.0471470319,
-0.0341899022,
-0.0161145069,
0.068748787,
-0.0115237162,
0.0188829843,
0.0153795611,
-0.0021578085,
-0.04808411,
-0.0077873073,
-0.0477141291,
-0.0062769838,
0.0101751518,
-0.0497251339,
0.0000289997,
-0.0581314042,
-0.0098375212,
0.0160362497,
0.0057662954,
-0.0367855728,
-0.0009013751,
0.0012170587,
-0.0011905954,
-0.0354494303,
0.0518569015,
-0.0550673082,
0.0273151863,
-0.0013321793,
0.0167936757,
0.0237413365,
0.0028293806,
0.0217906367,
-0.0289251171,
-0.0204118397,
0.0036157467,
0.0159740821,
-0.0029887457,
-0.0731800646,
-0.0409089513,
-0.016905345,
-0.0653659999,
0.0057727844,
0.0568942167,
-0.005943608,
-0.0477331802,
-0.0024810778,
0.068117097,
-0.0025361611,
-0.0436604917,
-0.0507565103,
0.005773942,
-0.0383899547,
-0.0049078292,
0.0468061864,
-0.0638047829,
0.0378913842,
-0.0025710678,
-0.021134261,
0.0109540494,
-0.0212432407,
0.0273943339,
0.0232435577,
0.0532147177,
0.0137000084,
-0.0558419414,
0.0062947287,
-0.0057763928,
0.0283441748,
0.0383021943,
0.0025208103,
-0.004541093,
-0.0052107614,
0.0170334708,
-0.0317386463,
0.036294885,
-0.0369192623,
-0.0272258129,
-0.0523369461,
-0.0531883612,
-0.0345373824,
0.0170213524,
0.0340956561,
0.0563656762,
-0.0113240657,
-0.0304098818,
0.0040235599,
0.0645419657,
0.0081205824,
0.0594576038,
0.022421822,
0.0071799755,
0.0234304592,
0.0485342108,
-0.0136675071,
-0.0072648865,
0.0147118699,
0.0301803462,
-0.1003584638,
0.067090027,
0.0037531757,
-0.0379443988,
0.0054468652,
-0.0241343193,
-0.0254830215,
-0.0449994877,
0.0527897254,
0.0167722851,
-0.0397117808,
0.0531040765,
0.021963425,
-0.0249786787,
0.017447751,
-0.0203297492,
-0.0038211793,
0.0056620133,
0.0019133773,
-0.0414457284,
-0.0587913468,
0.0326546393,
-0.0357357934,
0.0575061403,
-0.0276415236,
0.0185226351,
0.0183683392,
-0.0447410345,
0.0371160842,
0.0185482856,
0.0247846209,
0.0336604118,
-0.0455530211,
-0.0024325347,
0.0113091022,
0.0341033377,
0.0223449804,
0.0283804368,
0.0177722219,
0.0442996845,
-0.032376349,
-0.0014296906,
0.0114278328,
0.0168490056,
-0.0163636282,
0.0070310766,
0.0265491214,
-0.0679023564,
0.0227463301,
-0.0163205359,
-0.1106498986,
-0.0434172116,
0.0419441722,
0.0095885973,
-0.0372183509,
-0.0279963482,
0.0047875461,
-0.013778829,
0.047050897,
-0.012565691,
0.0821760222,
-0.0369859934,
0.0506926514,
0.0453493744,
0.0424161069,
-0.0393263586,
0.079887569,
0.0030219029,
0.025306873,
-0.0076260217,
-0.0199513473,
-0.0065589906,
-0.0060033598,
-0.0418680608,
0.0149275735,
-0.017798543,
-0.0289086252,
-0.0191257093,
0.0016851308,
-0.0035207516,
-0.0623934716,
-0.0502174459,
-0.0289909616,
-0.0483740605,
0.0711988807,
0.0021461095,
0.0333550498,
0.0222256463,
-0.0064027258,
-0.0283313561,
0.0222278442,
-0.0012384708,
-0.0189646278,
0.019876888,
-0.0142285712,
-0.002358827,
-0.0387511626,
-0.0583229698,
-0.0193902031,
0.0434236676,
0.0342166461,
-0.0424347296,
-0.02237585,
-0.0330309607,
-0.0029131745,
0.0534343198,
0.0218467321,
-0.0136355814,
0.0183730498,
-0.0546204522,
0.0687385947,
0.0457931794,
-0.0293137562,
-0.0382806063,
-0.0148461526,
-0.0655763969,
-0.0001157349,
0.0005366671,
0.0298759118,
0.056072358,
-0.0887006298,
0.0124260075,
-0.0173716396,
0.0779632702,
-0.0369179994,
0.0013848576,
0.0178930592,
-0.0062292283,
-0.0501325838,
0.002904718,
0.0687917024,
-0.0007841281,
-0.0496676825,
0.0308397226,
-0.0410711356,
-0.0089534419,
0.0342589095,
-0.0380700119,
-0.0591192394,
-0.0361506045,
0.0042804801,
-0.0495699197,
0.0187076684,
0.0457465015,
-0.0063924626,
0.0050562872,
-0.0692217499,
-0.042192176,
0.040190693,
-0.0281171259,
0.0109215779,
0.0607810654,
-0.0465152822,
-0.0005091258,
-0.0162185505,
-0.0360170938,
-0.0101589942,
0.0398473889,
-0.0040493845,
0.0334755704,
0.0433181524,
0.0014990026,
0.0443385281,
0.0143538592,
-0.0580882877,
0.0650383532,
0.0270007681,
0.0065978058,
-0.032824412,
0.0549858324,
-0.0253131129,
-0.0046196352,
0.0175894517,
0.0201001484,
0.0238839556,
0.0098754326,
-0.0022885264,
-0.0253111273,
-0.0528514087,
0.0262086205,
0.0417561382,
-0.04581476,
-0.0650935546,
-0.0315941013,
-0.0153858876,
-0.0516678058,
-0.0120329084,
0.0764277056,
0.0437219404,
-0.0251351465,
0.0020436565,
-0.0177345779,
0.0188599825,
0.0020565996,
0.0108899483,
-0.0412208997,
0.0025301659,
0.0336662494,
-0.0222020783,
0.0248519909,
-0.0228391774,
0.0639963001,
-0.0849423409,
-0.0080030877,
0.0060224934,
0.0168003589,
0.0420603417,
0.0199065525,
-0.0325501524,
-0.020132497,
-0.0043679574,
0.0113898301,
-0.0197913852,
-0.0021073588,
0.0379985683,
-0.0088276006,
-0.0027548696,
-0.0193894431,
-0.0408268087,
0.0005955882,
0.048828885,
0.0161105394,
-0.0731492117,
0.0070081246,
0.0846820101,
-0.0321100838,
0.006216262,
-0.0176026709,
-0.0010649263,
0.0416623428,
-0.0418503173,
-0.0294438712,
-0.0217586979,
-0.0217669532,
-0.0098863328,
-0.0175696444,
0.0670418739,
0.0140486658,
0.0535352379,
0.0482877456,
-0.0122875916,
-0.0801506415,
0.0110848509,
-0.0023849311,
-0.006214276,
-0.0780707747,
-0.0454443321,
-0.0065423348,
-0.0663603544,
-0.0049688839,
0.0286180861,
0.031273149,
-0.0125318049,
-0.0861059204,
-0.0682691559,
0.0227867328,
0.0275713671,
-0.0021942081,
-0.0252691265,
0.0087907789,
-0.0409018844,
-0.0107547967,
-0.001842308,
0.041183684,
0.0186955016,
-0.0650159121,
0.0069973599,
0.0242182128,
-0.0032710354,
-0.0092072245,
0.0089422725,
-0.0262735877,
-0.0129824309,
0.0581933968,
0.0312474761,
0.0027523437,
0.0321274623,
-0.0602568276,
-0.024289215,
0.0036647648,
0.0043357718,
-0.0418025181,
0.0317703001,
0.049605038,
0.0354188457,
-0.0167998895,
0.0226959605,
0.0422227308,
0.0395990908,
0.0076963166,
0.019940909,
-0.0522974283,
-0.0113573978,
0.0244025625,
-0.001516621,
0.0303808339,
0.0297176912,
0.0837901011,
-0.0447620563,
0.0155711016,
-0.0434273444,
0.0330274291,
-0.0342381299,
0.0080701318,
0.0401417613,
0.0181458071,
-0.0028345282,
0.017678123,
0.0298924092,
0.0088718869,
-0.0007103909,
0.0503208674,
0.0566215329,
-0.0102083292,
-0.0264587514,
0.0328000374,
0.0186847895,
-0.0201136433,
0.009609648,
0.0337337628,
-0.0767187551,
0.0253889989,
-0.01518708,
-0.0134870363,
-0.035409309,
0.0181649812,
0.0152584314,
0.007062431,
-0.0107498076,
0.0264642425,
0.0390818752,
-0.0334736258,
0.0613765419,
-0.0139372479,
-0.0001053848,
-0.0089084366,
0.0051136645,
0.0005940883,
-0.0118054915,
0.0196914934,
0.0235546622,
-0.0078742318,
-0.0829081237,
0.0227464791,
0.0037996364,
-0.0039717415,
0.010966138,
0.0087658502,
-0.0157692842,
-0.0416810066,
-0.0364870913,
-0.006954852,
-0.0502606407,
0.0085260086,
-0.0507169552,
-0.0566640124,
-0.0230162833,
0.0266267024,
0.0495296568,
-0.0183666199,
0.0378111824,
0.0466291904,
0.0126169594,
-0.0536448397,
-0.007663467,
0.010550661,
0.0410979353,
-0.0331999734,
0.009952968,
-5.454815046e-33,
-0.0272052288,
0.0590776354,
-0.0117069855,
-0.0182767473,
-0.0156959035,
0.0233244896,
-0.0154604297,
0.0133599136,
-0.0246955045,
-0.0117501104,
-0.0275781751,
0.0190002918,
0.0405472629,
0.0303237438,
0.0794162527,
-0.0105742021,
0.0414303318,
-0.0129578765,
0.0457093753,
-0.0397096984,
0.0091439346,
0.0548435189,
0.1035016626,
0.0298282336,
-0.0068530482,
-0.0104681002,
0.0161265563,
0.0285604037,
-0.0205409918,
0.042877946,
0.0019732192,
-0.0061154175,
0.0277804043,
-0.0246680435,
-0.0117252292,
-0.0055280235,
-0.1150353476,
-0.0365104005,
0.0351633839,
0.0560385361,
-0.0494249761,
-0.0080462955,
0.0330665149,
-0.0062176925,
0.0266463328,
0.028646715,
0.0037689921,
-0.0247476287,
0.005620501,
-0.0069163265,
-0.0689429864,
0.0600655861,
0.0184388123,
0.045696497,
0.1058013439,
-0.0066904644,
0.0071281553,
0.0173256174,
-0.030879315,
-0.055411417,
-0.0377363749,
0.0858652443,
0.0192902647,
0.0371658318,
-0.05520989,
0.003073347,
0.0707555041,
-0.0684707835,
0.0482148416,
0.0640381575,
0.0214283336,
0.010462584,
0.0191146806,
-0.0352193825,
0.0294767842,
0.0213944055,
-0.0443219319,
-0.0481213257,
0.0591606312,
0.0577725321,
-0.0390119441,
0.0183356907,
-0.0287390053,
-0.0155674797,
0.0290247239,
-0.0862121806,
-0.0368005484,
-0.0188248437,
0.0006967841,
-0.006409063,
0.007720775,
0.0905298963,
0.0171056818,
0.0136232516,
0.0042351214,
-0.023298353,
-0.0165110603,
0.0423099585,
-0.0185876172,
0.0357393101,
-0.0393099524,
-0.0639155209,
-0.0319869407,
-0.0092779081,
0.0161907952,
0.006618937,
0.0403903089,
0.001132176,
-0.0344480127,
0.0088944593,
-0.051566653,
-0.0017378175,
0.0158385858,
0.094391048,
0.0162345115,
0.0201950334,
-0.0002572322,
0.039119374,
0.043916788,
-0.0084134676,
-0.0247642566,
0.038095966,
-0.0205352064,
0.033485245,
-0.0187373534,
-0.0176200215,
0.0147380922,
0.0061571272,
0.0475574248,
-0.0446677916,
-0.0158978906,
0.0378344022,
2.623e-7,
0.0539514348,
0.0152244894,
0.0635943711,
-0.0218748804,
0.022372786,
-0.001916611,
-0.015448153,
0.0128042707,
-0.0466102511,
0.0071342867,
-0.0317629762,
-0.0027484603,
0.0188459065,
0.0320913158,
-0.0996933356,
-0.0042186235,
-0.0139490571,
-0.0125182057,
-0.0457467698,
0.0099297045,
0.0479038954,
0.1187871769,
-0.0286767706,
-0.0364185534,
-0.032643389,
-0.0139928041,
0.00313188,
0.0179896541,
0.0047355955,
0.0348054692,
0.0167868435,
-0.0442838483,
-0.0185420793,
0.0213200748,
-0.0116417529,
0.0081374506,
0.0387878194,
0.0452168845,
0.0193496477,
0.0226978809,
0.0698182732,
-0.0183545277,
-0.0259770062,
-0.0062180255,
0.0474550426,
-0.0181739964,
0.0019302103,
0.053917475,
-0.0125685362,
0.0299338605,
-0.0439047441,
0.0103777228,
-0.0215061903,
0.0249261744,
0.059290804,
-0.0285763219,
-0.0429327153,
-0.0441104844,
0.00405013,
0.0527129248,
-0.0108899232,
0.0047734464,
-0.0354411043,
0.0135168191,
0.0183285084,
-0.000427972,
-0.0549944974,
2.571130325e-34,
0.0054984018,
-0.0029298081,
0.0100677749,
0.0465709865,
-0.0102979504,
-0.0056613614,
-0.0450416282,
0.0599921532,
-0.0571466684,
-0.0171448775,
-0.0146242678
] |
http://arxiv.org/abs/2402.02563v2 | http://arxiv.org/pdf/2402.02563v2 | Synergy-of-Thoughts: Eliciting Efficient Reasoning in Hybrid Language Models | 2024-02-04T16:45:01 | 2024-05-23T14:20:53 | Yu Shang, Yu Li, Fengli Xu, Yong Li | cs.CL | cs.CL, cs.AI, cs.LG | Large language models (LLMs) have shown impressive emergent abilities in a
wide range of tasks, but still face challenges in handling complex reasoning
problems. Previous works like chain-of-thought (CoT) and tree-of-thoughts (ToT)
have predominately focused on enhancing accuracy, but overlook the rapidly
increasing token cost, which could be particularly problematic for open-ended
real-world tasks with huge solution spaces. Motivated by the dual process
theory of human cognition, we propose "Synergy of Thoughts" (SoT) to unleash
the synergistic potential of hybrid LLMs for efficient reasoning. By default,
SoT uses smaller-scale language models to generate multiple low-cost reasoning
thoughts, which resembles the parallel intuitions produced by System 1. If
these intuitions exhibit conflicts, SoT will invoke the reflective reasoning of
scaled-up language models to emulate the intervention of System 2, which will
override the intuitive thoughts and rectify the reasoning process. This
framework is model-agnostic and training-free, which can be flexibly
implemented with various off-the-shelf LLMs. Experiments on six representative
reasoning tasks show that SoT substantially reduces the token cost by
38.3%-75.1%, and simultaneously achieves state-of-the-art reasoning accuracy
and solution diversity. Notably, the average token cost reduction on open-ended
tasks reaches up to 69.1%. Code repo with all prompts will be released upon
publication. | Synergy-of-Thoughts: Eliciting Efficient Reasoning in Hybrid Language Models: Large language models (LLMs) have shown impressive emergent abilities in a
wide range of tasks, but still face challenges in handling complex reasoning
problems. Previous works like chain-of-thought (CoT) and tree-of-thoughts (ToT)
have predominately focused on enhancing accuracy, but overlook the rapidly
increasing token cost, which could be particularly problematic for open-ended
real-world tasks with huge solution spaces. Motivated by the dual process
theory of human cognition, we propose "Synergy of Thoughts" (SoT) to unleash
the synergistic potential of hybrid LLMs for efficient reasoning. By default,
SoT uses smaller-scale language models to generate multiple low-cost reasoning
thoughts, which resembles the parallel intuitions produced by System 1. If
these intuitions exhibit conflicts, SoT will invoke the reflective reasoning of
scaled-up language models to emulate the intervention of System 2, which will
override the intuitive thoughts and rectify the reasoning process. This
framework is model-agnostic and training-free, which can be flexibly
implemented with various off-the-shelf LLMs. Experiments on six representative
reasoning tasks show that SoT substantially reduces the token cost by
38.3%-75.1%, and simultaneously achieves state-of-the-art reasoning accuracy
and solution diversity. Notably, the average token cost reduction on open-ended
tasks reaches up to 69.1%. Code repo with all prompts will be released upon
publication. | [
0.043856699,
0.0143945981,
-0.0196947586,
0.0565768629,
-0.0698707178,
0.0192275122,
0.0261288173,
-0.0213417355,
0.0167232882,
-0.0402061008,
-0.0361490697,
-0.0294091497,
-0.0075914892,
0.0351861082,
0.0131738149,
-0.0572822206,
0.054540921,
-0.06337253,
0.0261025243,
-0.0162630957,
-0.0033217755,
0.0236260109,
-0.0463692136,
0.0546822399,
-0.014537164,
-0.0278753396,
-0.032781601,
-0.0093137678,
0.0380143113,
-0.0316434242,
-0.0283121578,
0.0265065674,
0.0356061608,
0.0605052039,
0.0000022449,
-0.0649259165,
-0.0031975834,
-0.0124868294,
-0.0170220118,
-0.0194954351,
-0.0067084837,
0.020813629,
0.0333914161,
0.0330062509,
-0.0364142098,
-0.0072492473,
0.0106599201,
0.0495170206,
0.0537405498,
0.044939097,
-0.0035769569,
-0.037217468,
0.000737076,
-0.0016354135,
0.0215863734,
-0.0206951797,
0.0014444957,
-0.0033517776,
0.0485939011,
-0.0121673197,
0.0065014302,
0.0168783944,
0.0434589349,
-0.0046643312,
-0.0060306741,
0.0566888712,
-0.0108487839,
-0.023407178,
-0.0116552943,
0.0382736847,
-0.0293482803,
-0.0218942128,
-0.0098287519,
-0.0137931323,
-0.0099809486,
-0.0169243775,
-0.0473059043,
0.041103147,
-0.0233912095,
0.0291431118,
0.0239933617,
-0.0110723469,
0.0200365372,
0.0236615259,
-0.0624004155,
0.0948524773,
-0.010363413,
-0.0384989902,
0.0417184792,
0.0004651509,
-0.024398759,
-0.0641383976,
0.0000145878,
0.0586510561,
-0.0376937613,
0.0123989191,
0.0006254365,
-0.0441786349,
-0.0179160722,
-0.0166211072,
0.0490071885,
0.0649463832,
0.0047888644,
0.0640303195,
-0.063971445,
0.0142046111,
-0.0221197736,
0.0418833084,
-0.0323213115,
-0.0188333374,
-0.0711898804,
-0.0209401511,
-0.0220241304,
0.0730457306,
-0.0121380826,
-0.0288402736,
-0.0210374426,
0.0333601125,
-0.0224853233,
0.0018176646,
0.0193691701,
0.0157294981,
0.0494633205,
-0.0565151013,
-0.0654675066,
0.0389926322,
-0.0336197466,
0.0436464436,
0.0326583609,
-0.0258122031,
-0.0125946924,
0.0056921206,
0.0000178472,
-0.034007255,
0.01234526,
0.0462574996,
-0.0328606851,
-0.007981292,
-0.0477031767,
-0.0425778478,
0.0116182221,
-0.0768541992,
0.0105891237,
-0.0785402581,
0.0185070261,
0.0169675145,
-0.0272964966,
-0.0027310224,
0.0284767374,
0.0123408344,
-0.0372916199,
0.0889902413,
-0.0202021468,
0.018955078,
0.0014857309,
0.0106879873,
-0.0576746427,
0.070547618,
0.0461393893,
0.0105144652,
-0.0138139008,
0.0363722965,
-0.0106372517,
0.0180999395,
-0.0002828549,
-0.0528513789,
-0.0074340869,
0.0080128741,
-0.010714313,
0.0373977572,
-0.0130412597,
0.0600507446,
-0.0299780201,
0.0352224782,
0.0555499867,
0.0704536289,
-0.0045154924,
0.0207211692,
0.0520804599,
-0.0165689439,
0.0030127785,
-0.0300936159,
-0.0585902222,
0.0616926998,
-0.0545392074,
-0.0040880814,
-0.0067678047,
0.0480122492,
-0.0374475829,
-0.0328045003,
-0.0354240686,
0.0003959828,
0.0050103921,
0.0269558858,
0.0406577699,
-0.0184608661,
-0.0231649056,
0.0608529747,
0.0027674704,
-0.0341628902,
-0.0328607596,
-0.0721053258,
0.0258960817,
0.0590766296,
0.0258129388,
-0.030819213,
-0.0100058746,
-0.0129638091,
-0.0332009904,
0.0357145518,
0.0241723694,
0.0068430011,
-0.0360045843,
-0.0547747463,
-0.0142140975,
-0.0225670636,
-0.0113570336,
-0.0067269057,
-0.0106617324,
0.0008826507,
0.0003872164,
0.0177871641,
0.0242577232,
0.0113972342,
0.0029852078,
0.0354412682,
-0.0233642235,
0.0293333791,
0.0222020485,
0.0071985205,
0.0148932096,
0.0372768156,
0.012899979,
-0.0322867855,
-0.0405868962,
0.042987179,
-0.0439361595,
-0.0566146933,
0.036873091,
-0.0083594797,
-0.0093178889,
0.0243933238,
-0.0140627902,
-0.0488415658,
0.0252183266,
-0.0338246189,
0.0469328798,
-0.0444830209,
0.0251983125,
0.0028426996,
-0.0300708096,
-0.0172563102,
0.106669873,
-0.0180856138,
0.0247221757,
0.0261079296,
-0.0258448943,
0.0283890869,
-0.0196110141,
0.0052635148,
-0.0052201306,
-0.0076973257,
-0.0274899527,
-0.0081790769,
-0.0046237158,
-0.0661663041,
-0.0420969427,
-0.1004441455,
-0.017038133,
-0.0104144998,
0.0122039486,
0.0015171402,
0.0040078051,
0.0082829297,
0.0181842875,
-0.0339394547,
0.0450370386,
0.0423756652,
0.0353820436,
0.0260956157,
-0.0168812051,
-0.0208178628,
0.0086699883,
-0.0008205366,
-0.0516199097,
0.0476413742,
0.0001816683,
-0.0192465149,
-0.0409858525,
0.0392494239,
-0.059217073,
0.0265795011,
0.0000441973,
-0.0098138759,
0.0526445396,
-0.0337157361,
0.0420003384,
0.0777218118,
-0.0061437394,
-0.0240958836,
0.0655068085,
-0.0111340638,
-0.0503486507,
0.0461493768,
0.0002194047,
0.0302283242,
-0.0110206744,
0.0108223064,
-0.0502904393,
0.0583240166,
-0.006617242,
-0.0090011135,
-0.0532826968,
0.0374550112,
0.0138882585,
-0.045172479,
-0.0309775937,
-0.050873965,
-0.0010862622,
0.051700484,
-0.0302769132,
-0.0121613275,
0.0918712467,
-0.0081768604,
-0.003829496,
-0.0429055057,
0.0071697058,
-0.0341260545,
0.0929051489,
-0.0086218538,
0.0092483917,
-0.0047278437,
-0.0044757524,
-0.0655037612,
0.0096810888,
-0.0020221362,
-0.0262916815,
0.054883603,
-0.0470993817,
0.0172105748,
-0.0203988627,
-0.063484326,
-0.0656794161,
0.0181659535,
-0.0639736131,
0.032483153,
-0.0067641414,
-0.0083076879,
-0.0083659776,
-0.0015128319,
0.0599090122,
0.0327540152,
0.0011796267,
-0.0260176398,
-0.0180400871,
0.0132031068,
0.0176201668,
0.0168745201,
0.0563352481,
-0.0040189279,
0.0440450758,
0.0406696163,
0.0725404918,
-0.0257741921,
-0.0235430058,
0.0121458499,
0.0074958107,
-0.0026628592,
-0.0128767919,
0.0154440161,
0.0709108189,
-0.0591635704,
-0.0328544639,
0.1270308793,
0.0617853999,
-0.0393224023,
0.0021828928,
-0.0654143542,
0.0218006168,
0.0393696837,
0.0224909168,
-0.0631254911,
-0.0444925912,
0.0273202006,
-0.0338033922,
-0.007012968,
-0.0583905578,
-0.0103547405,
-0.1160092652,
0.0145680616,
-0.0110554202,
-0.0179828666,
0.0006321918,
0.0060287127,
-0.0106599042,
0.0393238366,
-0.0441912264,
0.032796178,
0.0272417609,
0.0073093432,
0.0065246322,
-0.0035279898,
0.0027573071,
0.0504466631,
-0.0216592718,
0.0115995994,
0.0354091413,
0.0038107058,
0.0405663177,
-0.1071407571,
0.0591805279,
0.0099485088,
0.0269793402,
-0.0295719523,
-0.0271085799,
0.0147756767,
-0.0120070037,
-0.0202151835,
0.0510081351,
-0.048619315,
-0.0360799097,
0.0063422993,
0.1289062351,
-0.0365161598,
-0.0165876206,
-0.0147855999,
0.0062573943,
-0.0272922795,
0.068242006,
0.0132247675,
-0.0285027847,
-0.0456221513,
-0.0101063419,
-0.0613646954,
-0.0310343523,
-0.0297266431,
0.0042529725,
0.0713844597,
0.0130229816,
-0.0324012488,
-0.0373166911,
0.034761522,
0.0576511323,
-0.0136742741,
-0.0018564921,
-0.0028347378,
0.0249009766,
-0.027829295,
0.0145589616,
-0.0418086685,
0.0410906337,
0.0033840796,
0.0135389864,
0.0001720049,
-0.0277273618,
-0.0269044414,
-0.0004713054,
-0.1023423672,
-0.0638491362,
-0.0371520668,
0.0566602759,
0.0298109502,
0.0292685945,
-0.007331884,
-0.0338759758,
0.0090569109,
0.0276181046,
-0.0507658496,
0.0001243566,
0.0221777745,
0.0328844897,
0.0386677347,
0.0530120097,
-0.0281383432,
0.008360724,
0.0345912389,
0.0574337617,
-0.0072204336,
0.0046111504,
0.0194002502,
0.0142398002,
0.0212010052,
-0.0047509125,
0.0121245161,
-0.0402703285,
0.042130135,
-0.0249646902,
-0.0258017816,
-0.047215458,
0.0067254147,
-0.004100787,
-0.0018996961,
0.0162290949,
-0.0368335322,
0.0194935966,
0.015371725,
0.0211099479,
-0.0101588471,
0.0126432013,
0.0466969758,
0.0260143578,
-0.0218648575,
-0.03190808,
-0.009381949,
-0.0187527575,
0.0192323942,
-0.031099176,
0.1114379987,
0.0340123996,
-0.0248330981,
-0.0480463058,
0.0007377868,
-0.0731321424,
-0.0916814581,
0.0389308818,
0.0320190229,
0.0504446402,
0.0358156562,
-0.0021383818,
-0.0582973659,
0.0064129969,
0.0154646114,
0.0235620663,
0.0145037314,
-0.0502954759,
0.0165066738,
0.0246884469,
0.0067798467,
-0.0509110503,
0.0568798073,
-0.0713257492,
0.0153531618,
0.0117172962,
-0.0104292408,
-0.0317500047,
-0.0350185595,
-0.0279757511,
0.049723316,
-0.0055954438,
0.0546862595,
-0.0861353278,
-0.061048653,
-0.0000307446,
0.0202799365,
0.0083730528,
0.0033554141,
0.0494859256,
0.0495171919,
-0.0019939397,
-0.043645665,
0.0127778472,
0.0193663072,
0.0086900471,
-0.0179082453,
-0.0403921567,
-6.427352583e-33,
0.0054160357,
-0.0115337381,
0.063552849,
-0.0104363812,
-0.0333819278,
-0.0246037282,
-0.0388534665,
0.0206562169,
-0.0231616497,
0.0077615478,
-0.0142343314,
0.0049850075,
-0.00237569,
0.0290433858,
0.032393042,
0.0006070927,
0.0057201483,
-0.018549405,
-0.013963744,
-0.0018310092,
0.0556782261,
0.0015066529,
0.0176440012,
-0.0261720028,
-0.010947302,
0.0198668633,
0.0080000153,
0.0586862378,
-0.0372753516,
0.0279264133,
-0.0146652944,
0.0208136048,
0.0483285934,
-0.034449555,
0.0048719221,
0.0156526081,
-0.0454500392,
-0.0381021649,
0.0575968362,
-0.0181628969,
-0.0468227044,
-0.0186877474,
0.0178572498,
0.0027682229,
0.0005977961,
-0.0266362671,
0.0087245107,
-0.029190056,
0.0121382559,
-0.0003106236,
-0.0490561873,
-0.0264856201,
-0.0119665377,
-0.0098614041,
0.0122396443,
0.028645955,
-0.023090696,
0.0129381986,
-0.071871683,
0.0046739471,
-0.0020754521,
0.0113079082,
0.0836328119,
0.0213339403,
0.0042737382,
0.0536432676,
0.0184844341,
0.0019526933,
-0.0744290203,
0.0117484685,
0.0167162642,
0.0507484823,
0.0727396011,
-0.0081722867,
0.0160949435,
0.0223386195,
-0.0875457749,
-0.0295418445,
0.0351944081,
0.0515381098,
0.0587061457,
0.025036959,
0.0063705929,
0.0027364516,
-0.0413314961,
0.0253079273,
-0.0589773543,
-0.0145543655,
0.0252114404,
0.0047326605,
0.0121322684,
-0.0137816435,
0.039614886,
-0.0299802963,
-0.0098329429,
0.0103216758,
0.0090281395,
0.0176271349,
-0.0115632275,
0.0189017989,
-0.0561322235,
-0.0081060966,
0.0042194501,
0.0417375863,
-0.0034162183,
0.0186147038,
0.035201937,
0.0030050329,
-0.0395111479,
-0.010708577,
-0.0965569615,
0.0188146383,
-0.0095638707,
-0.0114655243,
0.0316976011,
0.033817187,
-0.0181841869,
0.0109277507,
-0.0224272013,
-0.0292204078,
-0.0324243084,
0.0226708315,
-0.0136134,
0.0346305259,
-0.010289113,
-0.0044689626,
-0.0448126681,
-0.0182418767,
0.0167569909,
-0.0393765606,
-0.0146469194,
0.0556593239,
2.952e-7,
0.004741692,
0.0214479621,
-0.0223382991,
0.0196486376,
0.0460701399,
0.0436559394,
-0.0114806052,
-0.0176675227,
0.0277457926,
-0.0643264651,
-0.0169825666,
0.0684036836,
0.0375330187,
0.0279673301,
-0.0906007662,
-0.0296967812,
-0.0335480273,
-0.037223924,
-0.0460629463,
0.008179808,
0.0689154863,
0.0664832518,
0.0350062773,
-0.0209404267,
0.0109512359,
-0.0088492297,
0.0595102534,
-0.0058474354,
0.0074200775,
-0.0186634101,
0.0217789914,
-0.0245320126,
0.0170736127,
0.0345900431,
0.0261982102,
0.01606852,
-0.0152371749,
-0.0014640284,
-0.0137079023,
-0.0138988225,
-0.0118563641,
-0.0339612886,
-0.003544528,
-0.0754887536,
0.0847519711,
-0.0619871765,
-0.0208491664,
0.0106645394,
0.021169411,
-0.0185160786,
-0.0013281482,
0.007843297,
-0.0124121951,
0.0234361552,
-0.0051495107,
-0.0064526666,
0.0170326196,
-0.1082533374,
-0.020082375,
0.0110035622,
-0.0376709178,
-0.0324946903,
-0.0347799212,
0.015230516,
0.0420004465,
-0.0591133684,
-0.0192396082,
2.876782963e-34,
0.0153913889,
-0.0103378324,
0.0322791524,
0.0457594767,
0.0271047521,
-0.0150764054,
-0.0243189391,
0.0050776335,
0.029643761,
-0.0939509198,
-0.0414290465
] |
http://arxiv.org/abs/2402.02559v1 | http://arxiv.org/pdf/2402.02559v1 | NavHint: Vision and Language Navigation Agent with a Hint Generator | 2024-02-04T16:23:16 | 2024-02-04T16:23:16 | Yue Zhang, Quan Guo, Parisa Kordjamshidi | cs.CL | cs.CL | Existing work on vision and language navigation mainly relies on
navigation-related losses to establish the connection between vision and
language modalities, neglecting aspects of helping the navigation agent build a
deep understanding of the visual environment. In our work, we provide indirect
supervision to the navigation agent through a hint generator that provides
detailed visual descriptions. The hint generator assists the navigation agent
in developing a global understanding of the visual environment. It directs the
agent's attention toward related navigation details, including the relevant
sub-instruction, potential challenges in recognition and ambiguities in
grounding, and the targeted viewpoint description. To train the hint generator,
we construct a synthetic dataset based on landmarks in the instructions and
visible and distinctive objects in the visual environment. We evaluate our
method on the R2R and R4R datasets and achieve state-of-the-art on several
metrics. The experimental results demonstrate that generating hints not only
enhances the navigation performance but also helps improve the interpretability
of the agent's actions. | NavHint: Vision and Language Navigation Agent with a Hint Generator: Existing work on vision and language navigation mainly relies on
navigation-related losses to establish the connection between vision and
language modalities, neglecting aspects of helping the navigation agent build a
deep understanding of the visual environment. In our work, we provide indirect
supervision to the navigation agent through a hint generator that provides
detailed visual descriptions. The hint generator assists the navigation agent
in developing a global understanding of the visual environment. It directs the
agent's attention toward related navigation details, including the relevant
sub-instruction, potential challenges in recognition and ambiguities in
grounding, and the targeted viewpoint description. To train the hint generator,
we construct a synthetic dataset based on landmarks in the instructions and
visible and distinctive objects in the visual environment. We evaluate our
method on the R2R and R4R datasets and achieve state-of-the-art on several
metrics. The experimental results demonstrate that generating hints not only
enhances the navigation performance but also helps improve the interpretability
of the agent's actions. | [
0.0476604961,
0.0036851815,
-0.004230659,
0.0259079076,
-0.0037061912,
0.0275277104,
-0.0251908302,
-0.0127027091,
0.011348445,
-0.021860892,
0.0087861894,
-0.0001219752,
0.0206222162,
0.0176258292,
0.0224077106,
-0.0727484599,
0.0208773967,
-0.0647845939,
-0.052828256,
0.0417640693,
-0.0013485237,
0.0090077519,
0.0029623085,
0.0554851927,
-0.0365104377,
0.0049388614,
-0.0673787147,
0.0431749076,
0.0495681167,
-0.0696044341,
-0.0001651347,
0.0053072493,
0.0736846104,
-0.000643812,
0.0000019751,
0.0261865184,
0.0332359299,
-0.0312375054,
-0.0092105335,
-0.0651377216,
0.0217037145,
-0.0439131036,
-0.0023134185,
0.0154542755,
-0.0324835666,
-0.0380514562,
0.0635554567,
0.0814883187,
0.0474562831,
0.0553294495,
-0.0171827059,
-0.006140064,
-0.0254286006,
0.0006973026,
0.0454840735,
-0.021744797,
-0.0071509415,
0.0266304724,
-0.0630116165,
0.035114523,
-0.0441572778,
0.0189025588,
0.0392280258,
-0.0153789241,
0.000719887,
0.0655794516,
-0.0037816507,
-0.0226223506,
-0.0184298605,
0.0918936729,
-0.0470466502,
-0.007481541,
-0.0101862187,
-0.0003135613,
-0.0106764669,
-0.0127163166,
0.0023169534,
0.0108658206,
-0.0506328382,
0.0236698873,
0.0197317358,
-0.014246989,
0.0030502384,
-0.0255745836,
0.0345378779,
0.0085782241,
0.005435585,
0.0198140163,
0.0169422347,
0.0163618121,
-0.0152200712,
-0.069433488,
-0.000818767,
0.0378636532,
-0.0290486533,
0.045980759,
-0.045270972,
0.0379347391,
-0.0257032793,
-0.0688299388,
0.0633524656,
0.0400311947,
-0.0845372677,
0.0527162962,
-0.0590665489,
0.0462693386,
-0.0103543755,
0.0565966256,
-0.0258508101,
-0.0204455778,
-0.037379168,
0.0221383348,
-0.0478295237,
-0.0090526948,
-0.0359279513,
-0.0272100009,
0.003364878,
0.0334442966,
0.0127127487,
0.0334161185,
-0.0191237237,
0.0390774757,
-0.0596062317,
0.0085675744,
-0.0238417108,
-0.0077670622,
-0.0430685654,
0.0029301699,
-0.0091348421,
-0.0312047936,
-0.0069236546,
0.0385461003,
-0.0152153661,
-0.0046855607,
-0.0103788991,
0.0963772312,
-0.0240381174,
-0.0424475782,
-0.0914472342,
0.0057285153,
0.0156001123,
-0.05405467,
0.0264940932,
-0.0478711911,
0.0457089245,
-0.0174779315,
-0.0313473567,
-0.0318669975,
0.0431771055,
0.0229602437,
0.0328490324,
0.0711563677,
-0.0210856982,
-0.0245059747,
0.0027447855,
0.0244208612,
-0.0228989478,
0.0504474454,
0.0140074873,
0.0751923248,
-0.0030084848,
0.0129300272,
0.0217601098,
0.0066955555,
-0.0139999846,
0.0347174332,
-0.0112668807,
0.0025723581,
-0.0039078277,
0.0259240568,
0.0228863899,
0.0472320504,
-0.0531030931,
-0.0161054656,
0.0553069413,
0.0290888511,
0.029532928,
0.010125326,
0.0798823088,
-0.0037611972,
-0.0150542194,
-0.009950175,
-0.0641655698,
0.0501074679,
-0.0163546745,
-0.0185492504,
-0.0595737062,
0.0190960858,
-0.0598396547,
-0.0554006509,
0.0282628424,
0.0139664235,
0.0050932257,
-0.0007549939,
0.050955236,
-0.0005333649,
-0.0550104454,
0.003887991,
0.0356814452,
-0.0350495614,
0.0061583123,
-0.0106346831,
0.015806146,
0.0677765235,
0.0171218328,
0.0093395431,
-0.0401972234,
-0.0178439654,
0.0593338273,
0.0346889831,
0.0152204568,
-0.0107389838,
0.0717044249,
-0.0588606298,
-0.0663787872,
0.016112227,
-0.0146745248,
0.0704305395,
0.0168155003,
0.0177457575,
0.038886182,
-0.0445687473,
-0.0223247595,
0.0496843234,
0.0618540794,
-0.0135427602,
-0.0082229832,
0.0173973758,
-0.0170504227,
0.0186250936,
-0.0006934891,
-0.0319508836,
0.0313175283,
-0.0191932414,
0.0300309267,
0.0202689879,
-0.0212990195,
-0.0714489371,
-0.0042186482,
-0.071309194,
-0.0085409833,
0.0331971571,
-0.0226573143,
-0.0271726344,
-0.0131258452,
-0.0410306938,
-0.0093000466,
0.0109502953,
-0.0060690083,
0.0440784879,
0.0255070068,
-0.0005400934,
0.0684550032,
-0.0126567837,
0.0008899166,
0.031268198,
0.0048765666,
0.016064845,
-0.0380899459,
0.0234033316,
-0.0262637828,
0.0365548097,
-0.0091175018,
0.0327258743,
0.0056648804,
-0.0426910892,
-0.0018476249,
-0.0534781516,
-0.0133241368,
-0.0504171364,
-0.0118828891,
0.0120581146,
-0.0254861917,
0.009262532,
-0.0388087071,
-0.019017091,
0.0525625423,
-0.0094105834,
-0.0138252173,
0.0128738685,
0.0155419689,
-0.0399679244,
-0.0399462767,
0.0090205204,
-0.0008179326,
-0.0013438209,
0.0386742651,
-0.0404898152,
-0.1120399684,
0.0187587645,
-0.0004803269,
0.0514236465,
-0.0615803413,
-0.0425549522,
0.0138226068,
-0.0196691211,
0.0034871318,
0.0601898059,
-0.0191950891,
0.0451486893,
0.055585172,
-0.0389588475,
-0.0515440963,
0.0048347516,
-0.016117597,
0.0037687416,
-0.0493489876,
-0.0229048654,
-0.0608197004,
0.0020351564,
-0.002120893,
0.0517816953,
-0.057937488,
-0.0016794656,
-0.0089974217,
-0.0514194109,
-0.0241555963,
0.0578377694,
0.0002779151,
0.0457494222,
-0.0375191122,
-0.0534706078,
0.0494025089,
-0.0188304354,
-0.0260409024,
0.0045922874,
-0.0087751551,
-0.0139309103,
0.0686748996,
0.0377191678,
0.020073954,
0.0167295318,
-0.0383240879,
-0.0353741534,
-0.0241340622,
0.0056790141,
0.0366366208,
0.0506557412,
-0.0894544125,
-0.0154572427,
-0.0339675955,
0.0177455191,
-0.0421482921,
0.0047007087,
0.0141204456,
0.0135062477,
0.0188773684,
-0.0216044132,
-0.0396330059,
0.020518126,
0.0365468785,
0.0396556631,
-0.0023305418,
-0.0353404656,
-0.0146878567,
-0.0138120754,
0.0734576359,
0.001683806,
-0.0210641157,
-0.0497604199,
0.0242372919,
0.061535649,
0.0138201239,
-0.0282868166,
-0.0739818662,
0.032039836,
0.0268109906,
0.029216554,
-0.0129574416,
0.0329656005,
0.0630450621,
-0.0735884681,
-0.0294146165,
0.0603852794,
0.0107320286,
-0.0299905688,
-0.0059396052,
-0.0045782696,
0.0195703749,
0.0326697528,
0.0461365245,
-0.0504490547,
-0.0484862775,
-0.0161748659,
0.0002919824,
0.0306414701,
-0.0031875311,
-0.0267632976,
-0.0538047999,
0.050400462,
0.0110757491,
-0.0575560518,
-0.0267682746,
-0.0079942988,
-0.0049686013,
-0.0159149244,
-0.0161893107,
-0.0095263096,
0.0105085373,
-0.0389820226,
0.0127902972,
0.0180135984,
0.0027298562,
0.0003949854,
-0.0560297929,
0.0284978654,
0.0474344455,
0.0089058317,
0.0115979826,
-0.0406243801,
0.0071791173,
-0.004528977,
0.0392619707,
-0.0328022018,
0.081652455,
-0.0335489996,
-0.0408526435,
0.0181502141,
0.0329971649,
-0.0399895608,
0.0050408891,
0.0011405481,
0.0116055775,
-0.032435853,
0.0277894959,
-0.017506225,
0.0004036352,
0.0293962061,
0.0172944367,
0.0042297263,
0.0058930875,
-0.0536984876,
-0.0353759639,
0.015737148,
-0.0505528003,
-0.010347832,
0.0029105076,
-0.0236707702,
-0.0541450903,
-0.0651409253,
-0.0974879265,
0.0662898198,
0.0727012455,
0.0239250604,
-0.0311755296,
-0.0456465855,
0.012323847,
-0.0241275188,
0.0336281247,
-0.0327489786,
0.038095478,
-0.0406406112,
-0.0325233899,
-0.0082657468,
-0.0568807237,
0.0221337378,
0.0261175074,
-0.0581106097,
-0.0692333877,
-0.0195531361,
0.0749225393,
0.0142379478,
0.0724749044,
0.0053633912,
-0.0069113704,
0.0099820485,
0.0097740423,
0.0245517567,
-0.0096567245,
0.035061352,
-0.0073558008,
0.0000670513,
0.0264174864,
-0.0702446029,
-0.0141507816,
-0.0036720978,
0.0269653667,
-0.0340670496,
0.0183957405,
0.0485925972,
-0.0637254864,
0.0479760654,
-0.0223179441,
0.0699480325,
-0.0240777377,
0.0137318848,
-0.0232706107,
-0.0063433866,
-0.0696404055,
0.0419538282,
0.0293004625,
0.0241992865,
0.047572419,
0.000936961,
-0.0286581218,
-0.013317381,
-0.0399847031,
-0.0047848215,
0.0366531126,
0.0152501129,
-0.0074975635,
0.0445382968,
-0.0163124297,
0.0350850634,
-0.0073908987,
0.0367055088,
0.0149196275,
0.0663221478,
0.0555679016,
-0.006105985,
-0.0312068462,
0.0258683357,
-0.0182409957,
-0.0647917017,
-0.0182425044,
0.0606544316,
0.0786349401,
-0.0050497842,
0.0243562665,
-0.0478762761,
0.0201326925,
0.010996351,
0.020531347,
0.0118174842,
-0.0718683004,
0.0573092364,
0.0278204046,
0.0223384872,
-0.033269383,
0.0574693084,
0.016005788,
0.0575198792,
0.0118783955,
-0.006569108,
-0.0421039611,
0.0250866842,
0.0492016152,
0.0274871606,
-0.0120675052,
0.0455153286,
-0.0185458548,
-0.0357983373,
-0.0199564621,
0.0453859344,
0.004309631,
0.0036420305,
0.0591822825,
0.0088422475,
-0.0032237978,
-0.0780312419,
-0.0411547311,
-0.0619666278,
-0.0275229309,
0.0125120655,
-0.0195870586,
-5.810856857e-33,
-0.0540917888,
-0.0352516733,
0.0551852807,
0.0316997766,
-0.0112045938,
0.0062541203,
-0.0511918478,
-0.0098782731,
-0.0364443846,
0.0380046703,
0.0170412511,
0.0231885016,
0.0230772011,
0.0262049939,
-0.0055542388,
-0.0442943387,
0.0448411331,
0.0348339304,
-0.0323490426,
0.0204083696,
0.0432555638,
-0.0175575279,
0.0646774545,
-0.0689398795,
-0.0003777793,
-0.0165232792,
-0.0170693062,
0.0385922901,
0.0232646,
0.0551601723,
-0.011916928,
0.0314015597,
-0.0059121875,
-0.0198180731,
-0.0428000949,
-0.0221573059,
-0.0614982694,
-0.040301498,
-0.0061340365,
0.0701291412,
-0.087289542,
-0.0131922485,
0.0144247031,
-0.0288855284,
0.0187476426,
-0.0027094008,
0.0553397909,
-0.0216277223,
-0.0348651446,
0.0093286941,
-0.0555783883,
0.0171203911,
-0.0035302117,
-0.0827093869,
0.0513703972,
0.029115038,
0.035803549,
-0.0726557747,
-0.0216876324,
0.0148629546,
0.0446260534,
0.0286802277,
0.0465560071,
0.0764540806,
0.0033098906,
0.0555106923,
0.0898663178,
-0.0129489284,
0.0104099382,
-0.0136963148,
-0.0172547698,
0.0437480584,
0.0192561708,
0.0247909557,
0.0140726538,
0.0201055128,
-0.0256676581,
0.037667349,
-0.0207801517,
0.077260159,
-0.0152327558,
0.0098863347,
-0.0516798347,
-0.0155904749,
-0.0272814333,
-0.0399917997,
-0.0106820334,
-0.0514694676,
0.0558403805,
0.0554010235,
-0.0087196669,
0.0065657799,
0.0467908792,
-0.0136917112,
0.0400262214,
0.0035970621,
-0.0201432128,
0.0113064954,
-0.0047860346,
-0.0003675012,
-0.0587028153,
-0.0379288457,
0.0390298925,
-0.0141522884,
0.0124778291,
0.0046197772,
-0.0022502705,
0.0533791631,
-0.0173234809,
-0.0104535911,
-0.0646876395,
-0.020466201,
0.0180460718,
-0.020795051,
0.0332237706,
-0.0274627283,
-0.0629690513,
0.0581507273,
-0.0126124052,
-0.0353950523,
0.0093069403,
0.0023461969,
0.0078699617,
0.0432858057,
-0.0390906446,
-0.0050674626,
-0.054299593,
-0.001493645,
0.0258195195,
-0.0514153466,
-0.0034623966,
0.0351712257,
2.635e-7,
-0.0187041163,
0.0167583544,
0.0149735389,
0.0243259948,
-0.0263249576,
0.0461424552,
-0.0260827765,
0.0078181773,
-0.024272928,
-0.0472168401,
0.0203951932,
0.0405042619,
-0.0120470095,
0.0228280257,
-0.0536476187,
0.0351955816,
0.047077734,
-0.0184622575,
-0.0348143876,
0.0147819016,
0.0865083188,
0.0754716843,
-0.0177371465,
-0.0122016054,
0.0300573427,
-0.0263392404,
0.0548302904,
-0.0278172847,
0.0020810799,
-0.0339630172,
0.0339015201,
0.0695418417,
-0.0659241825,
0.0274068583,
-0.023372205,
-0.0249484032,
-0.0009369747,
0.0114008803,
-0.045271188,
-0.0226591378,
-0.0131169762,
0.0183683205,
-0.0189900883,
-0.0197753683,
0.0245290026,
-0.0383896269,
-0.0039749835,
0.0317306556,
-0.0241125729,
-0.0063628075,
-0.0154035063,
-0.0104711438,
-0.0312450919,
-0.0242889747,
0.0154608041,
-0.0110759959,
-0.0334952474,
-0.0536392145,
-0.0075854971,
0.0052416618,
0.0204206742,
-0.0216058716,
-0.0341728702,
0.0134277334,
0.0040552486,
-0.0208718702,
-0.056936726,
1.731386777e-34,
0.0318836831,
-0.0296440665,
-0.0205063708,
0.0081766983,
0.0287802368,
0.0002323745,
-0.0072677648,
-0.0025265655,
0.0210352261,
-0.0117298746,
-0.0737295151
] |
http://arxiv.org/abs/2402.02558v1 | http://arxiv.org/pdf/2402.02558v1 | Enhancing Robustness in Biomedical NLI Models: A Probing Approach for Clinical Trials | 2024-02-04T16:18:01 | 2024-02-04T16:18:01 | Ata Mustafa | cs.CL | cs.CL, cs.LG | Large Language Models have revolutionized various fields and industries, such
as Conversational AI, Content Generation, Information Retrieval, Business
Intelligence, and Medical, to name a few. One major application in the field of
medical is to analyze and investigate clinical trials for entailment
tasks.However, It has been observed that Large Language Models are susceptible
to shortcut learning, factual inconsistency, and performance degradation with
little variation in context. Adversarial and robust testing is performed to
ensure the integrity of models output. But, ambiguity still persists. In order
to ensure the integrity of the reasoning performed and investigate the model
has correct syntactic and semantic understanding probing is used. Here, I used
mnestic probing to investigate the Sci-five model, trained on clinical trial. I
investigated the model for feature learnt with respect to natural logic. To
achieve the target, I trained task specific probes. Used these probes to
investigate the final layers of trained model. Then, fine tuned the trained
model using iterative null projection. The results shows that model accuracy
improved. During experimentation, I observed that size of the probe has affect
on the fine tuning process. | Enhancing Robustness in Biomedical NLI Models: A Probing Approach for Clinical Trials: Large Language Models have revolutionized various fields and industries, such
as Conversational AI, Content Generation, Information Retrieval, Business
Intelligence, and Medical, to name a few. One major application in the field of
medical is to analyze and investigate clinical trials for entailment
tasks.However, It has been observed that Large Language Models are susceptible
to shortcut learning, factual inconsistency, and performance degradation with
little variation in context. Adversarial and robust testing is performed to
ensure the integrity of models output. But, ambiguity still persists. In order
to ensure the integrity of the reasoning performed and investigate the model
has correct syntactic and semantic understanding probing is used. Here, I used
mnestic probing to investigate the Sci-five model, trained on clinical trial. I
investigated the model for feature learnt with respect to natural logic. To
achieve the target, I trained task specific probes. Used these probes to
investigate the final layers of trained model. Then, fine tuned the trained
model using iterative null projection. The results shows that model accuracy
improved. During experimentation, I observed that size of the probe has affect
on the fine tuning process. | [
0.0670946538,
0.0403881818,
-0.0182245746,
0.0169418342,
-0.0486917943,
-0.0142848771,
-0.0010393284,
0.0078610582,
0.0104193976,
-0.0350058116,
-0.019571159,
-0.0319087431,
0.0267057698,
0.0645954236,
0.0372606777,
-0.0014761979,
0.0687485561,
-0.0077712219,
0.0205547269,
-0.0147184115,
-0.0339536406,
-0.0146135204,
-0.0156060839,
0.0200021267,
0.0115903607,
-0.0166406669,
0.0013919205,
-0.0160893407,
-0.0067670387,
-0.0547469854,
-0.0100937476,
0.0602562316,
-0.0259480309,
0.0270572063,
0.0000022579,
-0.0458040647,
-0.0082788095,
0.0453071073,
-0.0284130871,
-0.0045288317,
0.0732033327,
-0.0254721921,
0.010643933,
0.0239015911,
-0.0214284491,
0.0126017295,
0.021369338,
0.0719476566,
0.0340359211,
0.0790783837,
-0.0119033521,
0.0015801857,
0.0373420008,
-0.0340460837,
-0.0288359765,
-0.0729139,
0.0515644327,
-0.0112325149,
0.0387784317,
0.0338880792,
0.0191569924,
0.0318586379,
0.0097497394,
0.0249764808,
-0.0124409134,
0.0251765512,
0.0148760984,
-0.0471646376,
-0.0087501314,
0.0340531394,
-0.0390931405,
-0.012689325,
0.0111328363,
-0.0099974079,
-0.0085849212,
0.0118850078,
-0.0319271237,
-0.0103424853,
-0.011373301,
-0.0148910349,
0.0578640178,
-0.0285737049,
0.0449853539,
0.0113547128,
-0.0183139909,
0.0212680958,
0.024489684,
-0.0319198444,
-0.0093424227,
-0.0062320754,
0.0018423118,
-0.030786179,
0.0107590957,
0.0186100695,
-0.0663186982,
0.0112987729,
-0.0267582498,
-0.0528022014,
0.0722199604,
-0.0803281739,
0.0168916881,
-0.0038572038,
-0.0188525263,
0.0183538329,
-0.0267862752,
0.0106681185,
-0.0247738007,
0.0808870867,
-0.0320379622,
0.0083457027,
-0.0442085862,
-0.0214519762,
-0.0153248776,
0.0524570048,
-0.027335614,
-0.0076136724,
-0.0659194589,
-0.0012338447,
-0.0142721217,
0.0246212035,
-0.0440909155,
0.0171721242,
-0.0633829013,
0.0201375671,
-0.0065380125,
0.0206069313,
-0.0429744013,
0.0417739898,
-0.0309697352,
-0.0696303993,
-0.0300832689,
0.0121722696,
-0.0149504542,
-0.0268901996,
0.010981313,
0.068877615,
0.019499965,
-0.0368779898,
-0.0539532043,
-0.060752511,
-0.0628140569,
-0.0755781829,
0.0436698273,
-0.0434577726,
0.007469323,
0.0171814319,
-0.0218983199,
-0.0568842851,
0.0001217863,
-0.0137201557,
-0.0687468424,
0.1087632403,
-0.0182937421,
-0.0366595238,
-0.0121298097,
0.0217448547,
0.0001894133,
0.0541551374,
0.0189416725,
-0.0071831318,
0.0056176949,
-0.0079598874,
-0.0104717026,
-0.0217489768,
0.0455165543,
-0.0185048617,
-0.0388156548,
0.0129703935,
-0.0145192323,
0.0632319301,
-0.0578167401,
0.0771606565,
0.0023714807,
0.0257710926,
0.0514927246,
0.0278572123,
0.0180759281,
0.0193443801,
-0.0315765366,
-0.0619409904,
0.0117306141,
0.0145128826,
-0.0408193357,
0.0089509832,
-0.0428996421,
0.0176038276,
-0.0309669171,
-0.0008747905,
-0.0245089345,
-0.0479131564,
-0.0597432889,
-0.008903157,
0.0081973905,
-0.0128902625,
0.0713432133,
-0.0083077243,
-0.0246347189,
0.0144930882,
0.0412793756,
-0.0086226286,
0.0016411328,
-0.0808292925,
0.0019019077,
0.0056870854,
0.0480635352,
-0.0554900058,
-0.0247515924,
-0.02642598,
-0.0211201273,
0.0143320169,
0.029403612,
-0.012586616,
-0.0009147124,
-0.0151038822,
-0.0171932653,
-0.0211586431,
0.0320129581,
-0.0314752609,
-0.0298646409,
0.0365991704,
0.0003075769,
-0.0586677119,
-0.0108937565,
0.053469412,
-0.0589315891,
0.0074121868,
-0.0130848512,
-0.0036604607,
0.0458321609,
0.0214331597,
-0.021856809,
0.0455157347,
0.0592506044,
-0.0440696143,
-0.0018717564,
-0.0226212889,
-0.0011439055,
-0.0287897233,
0.0303548537,
-0.040660277,
-0.0096379705,
0.0207431689,
-0.0349216685,
-0.0348245427,
-0.0094918041,
-0.0347612798,
-0.0180622526,
-0.0139061017,
0.0222301465,
-0.0196391791,
0.0017976726,
-0.060180407,
0.057273414,
-0.0029415262,
0.0046329666,
0.0015263543,
-0.0361685716,
0.0373053886,
-0.0311039109,
-0.0494511276,
0.0521927737,
-0.0128743611,
-0.0187780056,
-0.0022136343,
-0.024670098,
-0.0320201851,
-0.0399380177,
-0.080421634,
0.0065811616,
0.0000350501,
0.0195921287,
0.000477414,
0.0170287061,
0.0252890978,
-0.0197853167,
-0.0240313839,
-0.0220344029,
0.0175832547,
0.0364385881,
0.0391865708,
0.0414254852,
0.0060753031,
-0.0486648791,
-0.0116614019,
-0.0487417541,
0.0536911674,
0.0176394489,
-0.0307413526,
-0.0344147906,
0.0240560807,
-0.0089153014,
0.004175487,
-0.0034454765,
-0.0462879539,
0.0346250273,
0.0234877523,
-0.0249446798,
0.0983813554,
0.0061212839,
0.0426686145,
0.0449877419,
0.0474212132,
0.0140167549,
0.0101597589,
-0.0208601542,
-0.0061211204,
-0.0142578706,
-0.0127858371,
-0.0058867196,
0.0676352009,
0.0316858478,
-0.0121719101,
-0.0410171375,
0.0561659932,
-0.0308086406,
-0.0178701263,
-0.0038840109,
0.0239855628,
-0.0286362618,
0.0353579968,
-0.0059467284,
-0.0146579258,
0.1308929324,
-0.0038635822,
0.0527441911,
-0.0921171606,
-0.0063499524,
0.010870873,
0.0209571905,
0.0013618274,
0.0293056462,
0.0541581437,
-0.0002832052,
-0.0415334217,
-0.0195663385,
0.0145641789,
-0.03137476,
0.0601402186,
-0.0288868491,
0.0281416476,
0.0211514626,
-0.0497864559,
-0.0542423055,
-0.0136170872,
-0.0358598493,
0.0176701583,
0.0121637154,
0.0263580512,
-0.0419015996,
0.0093775773,
-0.003046599,
0.052336555,
-0.0238862708,
-0.0093259504,
-0.0296624955,
-0.007950197,
0.0583627932,
-0.0091321142,
0.0397267193,
-0.0248682555,
0.016233325,
0.0470926426,
-0.0509911813,
0.0717037022,
-0.0453419536,
0.0080263121,
-0.0275116451,
-0.0485633723,
-0.0149058318,
0.0356347822,
0.030978255,
-0.0530284271,
-0.0373390503,
0.0862573311,
0.01346453,
0.0007097654,
0.0413864031,
-0.0230532419,
-0.0171051435,
-0.0217889715,
0.008545409,
-0.0646914914,
-0.0316375606,
-0.0009048219,
0.0305437706,
-0.0290286392,
-0.0079777185,
0.0189980101,
-0.0695042834,
0.0185173079,
0.0046559628,
-0.0564048178,
-0.0586980283,
-0.0745854601,
0.0185916945,
-0.0311542004,
0.0744998157,
0.0023646373,
0.0218434222,
-0.0115347002,
-0.0022437726,
0.0295046903,
-0.1361916363,
0.0195369702,
-0.0454992503,
0.0750473142,
0.0527448356,
0.0348054059,
-0.0247102994,
0.0077425633,
0.028094355,
0.0120466864,
0.0155988317,
0.0006547461,
0.0817909539,
0.0244104527,
0.0068903286,
-0.0475322492,
0.0430469252,
-0.0337231122,
-0.0092440434,
-0.0388396531,
0.076018326,
-0.0463124923,
-0.0382502154,
-0.0496265292,
-0.0018838543,
0.006683548,
0.0763547719,
0.0091637401,
-0.0226412155,
-0.0424956307,
-0.0004720945,
0.0479079932,
0.009764038,
-0.0354104266,
0.0620098263,
-0.0028542464,
0.0104897125,
-0.0557034463,
-0.0456653535,
0.0284117777,
0.0310926549,
0.0005401163,
0.0382524505,
0.0306597445,
-0.0114566004,
-0.0314078629,
0.043938905,
0.0352627188,
0.0186831262,
-0.0269675963,
0.0093108201,
0.0582407825,
0.0032270825,
-0.0062617511,
0.0167004112,
-0.0585374832,
-0.0509239063,
-0.0099264793,
0.0578541234,
0.028402986,
0.0268373489,
-0.0428459756,
-0.0411992557,
0.0224455874,
0.0311436709,
-0.0250425134,
-0.0208867006,
0.0264019668,
-0.0410157368,
-0.0263186526,
0.0177615639,
-0.0360314026,
0.0115618026,
0.0252860878,
0.0608082265,
0.006900657,
-0.0715516582,
0.0353083499,
0.0625577718,
0.0621507131,
-0.0104843918,
-0.0230754763,
-0.0134744793,
0.0120749911,
-0.0374236442,
0.0131511502,
-0.0091605159,
0.0135666663,
0.0093281856,
0.0358021036,
-0.005979076,
-0.0539633185,
-0.0681125671,
0.0101468377,
-0.0449360125,
0.0122016696,
0.014170249,
-0.0058247875,
0.0507003851,
0.0102764471,
-0.0207739919,
0.0131058507,
-0.0092381556,
0.0281501114,
-0.0323012173,
0.0315964855,
0.0417752787,
0.0170844086,
-0.0557407029,
0.0358542278,
-0.0539173968,
-0.0099791866,
-0.0287526753,
0.0321930237,
0.0239767488,
0.0140702659,
0.0100083621,
-0.0323146358,
0.0031648064,
-0.0138236806,
0.0466645509,
-0.0008749446,
0.0287518278,
0.0547754318,
0.0193400308,
0.0036423851,
-0.0166933555,
0.0673131049,
0.0398707166,
0.0541888215,
-0.030705316,
-0.0182863511,
-0.0511170216,
-0.0125255417,
0.0062607699,
0.0077293031,
-0.0762114376,
0.0227595549,
-0.0513004623,
-0.0640547201,
0.0579844005,
-0.0016103102,
0.0023032869,
-0.0377014466,
0.0377366617,
0.0563053451,
0.0333121903,
-0.0312246624,
0.0008475207,
0.0266027134,
0.0317007266,
-0.0370396003,
-0.018893024,
-7.019729737e-33,
0.0022974142,
-0.0376280285,
-0.0125605296,
-0.0422599614,
0.0060720262,
0.0058048298,
-0.0687242448,
-0.0008577938,
-0.0299956501,
0.0279938076,
-0.0166956894,
0.0193717945,
0.0113519756,
0.0192838535,
-0.0038849702,
-0.0338707119,
0.050878033,
0.0119846622,
-0.0091165295,
-0.0053672339,
-0.0147796534,
0.052102197,
0.0635717213,
-0.0762218535,
0.028033508,
0.0343756191,
-0.0031437455,
0.0244122967,
-0.0653763413,
0.0745667666,
0.0057192799,
0.0277937781,
0.0374010429,
-0.021913195,
0.0108406711,
-0.0419875234,
-0.0611182041,
-0.0369443446,
0.0136385923,
-0.007811876,
-0.0440797172,
-0.0220465306,
0.0736477077,
-0.0458968244,
-0.0080856765,
0.006560287,
0.0171902273,
-0.0211497545,
-0.0344002247,
-0.0135987001,
-0.0597466677,
0.0148739275,
0.012166583,
-0.006749433,
-0.0104933791,
0.0081127677,
-0.0190204345,
0.0284043793,
-0.0393709093,
-0.0186892264,
0.0326401517,
0.024758881,
0.0143524166,
0.0116866613,
-0.0108444355,
0.0674872398,
0.0095425174,
0.0472217724,
-0.0223531574,
-0.0113431374,
0.0327745304,
0.0479345322,
0.0778798983,
0.0367219225,
0.0158208888,
0.0028970756,
-0.1151618734,
-0.0161114018,
-0.0043185698,
0.054793749,
0.0373990126,
-0.0057143383,
-0.0387715586,
-0.0124988649,
-0.051756274,
-0.044128783,
-0.0322709493,
-0.0318699926,
0.0210653041,
-0.0095059089,
-0.014971802,
0.0132271256,
0.0147485156,
-0.0311929062,
0.0361148715,
-0.0094961533,
0.0209730156,
0.0132564008,
-0.030170301,
0.0283488557,
-0.0339168273,
-0.0176544599,
0.0016334894,
-0.008507655,
0.0644819513,
0.0071058068,
-0.0007065719,
-0.0260347556,
-0.0182022545,
-0.0102887414,
-0.0011231595,
-0.0221545175,
0.0299764294,
0.0321916565,
0.0464827865,
0.017556848,
0.0065040635,
0.057287775,
0.0266311727,
0.0368868783,
-0.0443007052,
0.0161163341,
0.0190086532,
0.0053264783,
-0.0352220945,
-0.0324131064,
-0.0343659595,
0.0545340143,
-0.0015151155,
-0.0750515014,
-0.0228974316,
-0.0227877535,
3.208e-7,
0.0664258227,
0.0684155375,
-0.0092048794,
0.0231220704,
0.0213325154,
0.0429757573,
-0.0091553619,
0.0137033211,
-0.0112602031,
-0.0390868746,
0.0181524865,
-0.0103581659,
0.0242192913,
0.0235663541,
-0.0553948507,
-0.0242106095,
-0.0520656928,
0.007091742,
-0.0200935509,
0.0104960445,
0.0346469656,
0.0896866471,
-0.0216349736,
-0.0117024006,
-0.0147576816,
-0.0547902808,
0.0275850669,
0.0246870704,
-0.016171921,
0.061201226,
0.0328621492,
-0.0028220022,
0.0352633931,
0.071124725,
-0.0194846671,
0.0024565924,
-0.0029361739,
0.0594663061,
-0.0521862619,
0.0043316828,
0.0060466561,
-0.0193148348,
0.0369806103,
-0.0393768772,
0.0644772798,
-0.0757628456,
-0.0684228316,
-0.0457481593,
0.0138482079,
-0.0235383678,
0.0683269277,
-0.0181694757,
-0.0018133827,
0.0184823126,
0.0155451652,
0.0190374497,
-0.0178743508,
-0.0449837334,
-0.0051663993,
0.0703969672,
-0.048470933,
-0.0307599604,
0.0339841135,
-0.0394764356,
0.0466077775,
-0.0831411034,
0.0080541931,
3.152822556e-34,
0.0532897711,
-0.0187003426,
0.0149829751,
0.0190274473,
0.0525157601,
-0.0051888358,
-0.0116298646,
0.0075196801,
0.0251600798,
-0.0614191703,
-0.0127139762
] |
http://arxiv.org/abs/2402.02555v1 | http://arxiv.org/pdf/2402.02555v1 | Generalizable Entity Grounding via Assistance of Large Language Model | 2024-02-04T16:06:05 | 2024-02-04T16:06:05 | Lu Qi, Yi-Wen Chen, Lehan Yang, Tiancheng Shen, Xiangtai Li, Weidong Guo, Yu Xu, Ming-Hsuan Yang | cs.CV | cs.CV, cs.CL | In this work, we propose a novel approach to densely ground visual entities
from a long caption. We leverage a large multimodal model (LMM) to extract
semantic nouns, a class-agnostic segmentation model to generate entity-level
segmentation, and the proposed multi-modal feature fusion module to associate
each semantic noun with its corresponding segmentation mask. Additionally, we
introduce a strategy of encoding entity segmentation masks into a colormap,
enabling the preservation of fine-grained predictions from features of
high-resolution masks. This approach allows us to extract visual features from
low-resolution images using the CLIP vision encoder in the LMM, which is more
computationally efficient than existing approaches that use an additional
encoder for high-resolution images. Our comprehensive experiments demonstrate
the superiority of our method, outperforming state-of-the-art techniques on
three tasks, including panoptic narrative grounding, referring expression
segmentation, and panoptic segmentation. | Generalizable Entity Grounding via Assistance of Large Language Model: In this work, we propose a novel approach to densely ground visual entities
from a long caption. We leverage a large multimodal model (LMM) to extract
semantic nouns, a class-agnostic segmentation model to generate entity-level
segmentation, and the proposed multi-modal feature fusion module to associate
each semantic noun with its corresponding segmentation mask. Additionally, we
introduce a strategy of encoding entity segmentation masks into a colormap,
enabling the preservation of fine-grained predictions from features of
high-resolution masks. This approach allows us to extract visual features from
low-resolution images using the CLIP vision encoder in the LMM, which is more
computationally efficient than existing approaches that use an additional
encoder for high-resolution images. Our comprehensive experiments demonstrate
the superiority of our method, outperforming state-of-the-art techniques on
three tasks, including panoptic narrative grounding, referring expression
segmentation, and panoptic segmentation. | [
0.023041226,
0.0655275062,
0.0411142111,
0.0466494523,
0.013813002,
-0.0024628416,
-0.0067638522,
0.0162318926,
-0.0563281067,
-0.0841873363,
0.0070339059,
0.0008702676,
0.0095176212,
0.0320342593,
0.0160662476,
-0.075822182,
-0.0071841329,
-0.0244302638,
-0.0329317451,
0.0134346532,
0.0510904156,
0.0116079468,
-0.0091599962,
0.0289957486,
-0.0687533394,
0.0182974916,
0.0014216229,
-0.0086107524,
0.0207788628,
-0.0381880291,
-0.0222140215,
0.0315223709,
0.0610895976,
-0.0040335413,
0.000001968,
-0.0375218503,
-0.0355665684,
0.0183478668,
0.0128889736,
-0.0089623127,
0.0321961232,
-0.0434289388,
-0.0157417394,
-0.0586472079,
0.0054835952,
0.002671309,
0.0715797693,
0.0944383293,
0.0557639226,
-0.0004452496,
-0.038187813,
-0.0645523593,
0.022610018,
-0.0043249931,
0.0013335038,
-0.0257872734,
0.0208749436,
0.0643859133,
-0.0482639484,
-0.0158372503,
-0.0545501523,
0.0301526655,
-0.022763148,
0.006530141,
0.0302125048,
0.0565879419,
0.0057531307,
-0.035649851,
-0.0130124297,
0.0946131572,
-0.0154378694,
0.0200595241,
-0.0245055016,
0.0225296505,
-0.0146836676,
0.001832886,
-0.0304838307,
0.0412747189,
-0.0052010575,
0.0475694872,
-0.0252094716,
0.0064693457,
-0.0087019065,
0.0083494931,
-0.044220496,
-0.0280462466,
0.0314936675,
0.0124207167,
0.0078577399,
0.0310395714,
-0.0464654528,
-0.046574112,
0.0326393582,
0.0098173795,
0.0206757784,
0.0171007346,
-0.0392111689,
-0.0233800262,
-0.0202731919,
-0.0351954587,
-0.0233523306,
-0.0089600282,
-0.0347247086,
0.0780365914,
-0.0258057285,
0.0093098208,
-0.0324998125,
0.0256443471,
0.0054724119,
-0.0287898425,
-0.0315228589,
0.004335314,
-0.0269027762,
0.0272180121,
-0.0613057576,
-0.0482475273,
-0.0241342429,
-0.0300193373,
-0.0103336461,
0.0190547872,
-0.0238515511,
0.0369132236,
-0.0215332787,
-0.0036810508,
-0.0211925954,
0.0204078089,
-0.0311513804,
0.0030492179,
-0.0126690101,
-0.0306101087,
0.0308744498,
-0.006722468,
0.0244270097,
-0.0219046641,
0.0267676711,
0.0849087089,
-0.0444276705,
-0.064969331,
-0.055922538,
-0.0312867314,
-0.0409945101,
-0.0372943543,
0.0209518448,
-0.0625993535,
0.0102268094,
-0.0219607446,
0.022853123,
-0.0494194366,
0.0344413854,
-0.0007767375,
-0.031118704,
0.1250025183,
-0.0109870713,
-0.0029930216,
0.0250780005,
0.0189976431,
0.0300358161,
0.0736567974,
0.0583713092,
0.0293868389,
0.0549370907,
-0.0006684265,
0.014269365,
0.0550829843,
-0.0137493936,
-0.0124229882,
-0.0523582697,
-0.0378869884,
0.0138129778,
-0.0082453769,
0.0500605516,
0.0381416045,
0.0044461959,
0.0088467784,
0.0840416551,
0.0588049553,
-0.0805414841,
0.0149882268,
0.0215355549,
0.0030871639,
-0.0453810059,
0.0083599649,
-0.0563596748,
0.0151770152,
0.0029690186,
0.0076886946,
0.0071238452,
0.0495171212,
-0.020976752,
0.0089834733,
0.0139992461,
-0.0081442175,
0.020786643,
-0.0065384912,
0.0673327893,
0.0077667902,
-0.066196546,
0.0711887702,
0.0256910902,
-0.0554438904,
0.0003329439,
-0.0026286675,
0.0155757591,
0.0145430416,
0.0399330035,
0.0043724435,
-0.0262038894,
-0.0038937132,
0.0338634439,
0.0686590821,
-0.0433390364,
-0.0083149364,
0.0253926553,
-0.068749845,
-0.0224394873,
0.0099833682,
0.0039970856,
0.0093775075,
0.0242393874,
0.0049870014,
0.0332540348,
-0.0621048585,
-0.0084040128,
0.0706990436,
0.0343630128,
0.0012540554,
-0.0844740793,
0.044078704,
-0.0056459596,
0.0162592698,
-0.0452654473,
-0.02310469,
0.0155602377,
-0.0321842693,
0.0038576366,
0.0229862463,
-0.0448397212,
-0.0109391259,
0.0049716481,
-0.0747836754,
-0.0055870335,
-0.032572899,
-0.0241507404,
-0.0440055169,
-0.0068301149,
-0.0020897968,
0.0387631878,
-0.0333740152,
0.0209355243,
0.0232861899,
0.0521000959,
-0.0351760015,
0.0589279681,
-0.0146574583,
-0.0225521494,
0.0063587008,
-0.0201025382,
-0.0257399436,
-0.016547665,
-0.030325897,
0.0307221729,
0.0172905754,
-0.0141521906,
0.0194677785,
0.0119974585,
-0.0481336825,
-0.0075212475,
-0.114112027,
-0.0224112924,
0.0027315121,
-0.0240494907,
-0.02239866,
0.0049388031,
0.0081500215,
-0.0210136194,
-0.0060613011,
-0.0157331023,
0.0009836553,
0.0050038341,
0.0156263374,
0.032640785,
-0.0342753418,
-0.0159899723,
-0.0027115401,
0.0142280022,
0.0748383924,
0.0182296857,
-0.0812131837,
-0.0279759429,
0.0104751904,
-0.0170160159,
0.0319257155,
-0.0203349218,
-0.0066788048,
-0.0095062787,
-0.0017834792,
0.0457115546,
0.0660886541,
0.0111717144,
0.0293306354,
0.0004571147,
-0.0730914399,
-0.0077903182,
0.014935541,
0.0047531449,
-0.0106745018,
-0.0120976418,
-0.0163544565,
-0.0184501633,
0.0538557544,
-0.0352727324,
0.0506931916,
-0.0312051326,
-0.0013540916,
-0.0013319862,
-0.0436135456,
-0.0443505794,
0.0390155725,
-0.0111066895,
0.000567498,
-0.0413505435,
-0.0419183075,
0.0223952122,
-0.0718660355,
-0.0370601527,
0.0338930339,
0.0061298981,
-0.0186612383,
0.0755374357,
0.0182837453,
0.0254731048,
0.0048821489,
-0.0304986127,
-0.0121991839,
0.0236426685,
-0.0198873375,
-0.0009867144,
0.0964024439,
-0.0520410463,
0.0165929757,
-0.0113485856,
0.0016828461,
-0.0203615073,
-0.0058527738,
-0.0637818575,
-0.0037967681,
-0.0180735383,
-0.0062897298,
-0.0095681511,
0.0601981021,
0.019118892,
0.0022485997,
-0.0009709678,
0.0013285682,
-0.0151047418,
0.0533680394,
0.0410395786,
0.0278744455,
-0.0504095368,
-0.0324856155,
0.0044324165,
0.0273775999,
0.0155675178,
-0.0190095082,
-0.007010608,
0.0134533532,
0.0087680789,
0.0256391335,
-0.0023440612,
0.0041851024,
0.0721556097,
-0.0565834269,
0.0035005233,
0.0835164934,
0.0065592113,
-0.0536181778,
-0.0306411888,
0.0188583173,
0.0109118829,
0.0200888105,
-0.0349278189,
-0.0366436578,
-0.079625614,
0.0179419257,
-0.048705738,
0.0382375605,
-0.024958238,
-0.0173873194,
-0.0806445479,
0.0256291348,
0.003523519,
-0.0576164462,
-0.0362126417,
-0.0241946261,
0.0470437519,
-0.0158087388,
-0.0252884384,
-0.0284064375,
-0.0066448534,
-0.0292097498,
-0.0055455188,
0.0177265909,
-0.0169935916,
0.0146380393,
-0.0663175434,
-0.0092462022,
0.0703607202,
0.0394872501,
-0.0373851322,
-0.0150182908,
0.0050280746,
-0.0361289047,
-0.0090528894,
0.0174657628,
0.0775521621,
-0.080811061,
-0.0223315377,
0.0088927066,
0.0407311618,
0.0213062875,
-0.0069022388,
-0.0079075163,
0.0145276189,
0.0002797169,
0.0169242676,
-0.0291076954,
0.0065076854,
-0.0043478757,
0.0447586551,
0.019847637,
-0.0324437059,
-0.0032297198,
-0.0130043821,
0.0317818783,
-0.0294025335,
-0.0080967331,
0.013914899,
-0.0744367465,
-0.021856688,
-0.0544988699,
-0.0111979414,
0.0767389983,
0.0478783511,
0.0235364698,
-0.0046017841,
0.0295967627,
0.0429971889,
-0.0583067499,
0.0613358431,
-0.0588053688,
0.0235059951,
-0.0543990731,
0.0067709684,
0.0627241582,
-0.0073246569,
-0.0180934202,
0.0520736761,
-0.0457816832,
-0.0158432294,
-0.0003842071,
0.0392303094,
0.0089275958,
0.0307336096,
0.0217723921,
-0.0077566137,
0.0282467324,
0.0060276981,
0.0149375405,
0.0290562026,
-0.0031670567,
-0.0201443452,
-0.0241574347,
0.0096194195,
-0.0487289242,
0.0475194827,
-0.0049245004,
0.0433182269,
-0.0136246439,
-0.0111923451,
0.0030090332,
0.000138903,
0.0543458089,
-0.0472437032,
0.0229312144,
-0.0457767509,
-0.0047975648,
0.0064158197,
-0.0285742283,
-0.0597523712,
0.025423361,
-0.0162160993,
0.0337476768,
0.0240530819,
0.0336452648,
0.0250911172,
0.033616282,
-0.0006785457,
-0.0095258262,
0.0070481095,
-0.0630258545,
-0.0011385323,
0.0244191643,
0.0144834053,
0.0207961015,
-0.0195767581,
0.0204900447,
-0.0290337838,
0.0763392448,
0.0117872963,
-0.0419351608,
-0.0476547629,
0.0414614826,
0.0363487266,
0.0473657511,
-0.0608736649,
0.0813305452,
0.0728325993,
-0.0303372052,
0.0381331779,
-0.0184432548,
-0.0225976184,
0.0217487179,
0.0168690663,
0.0066430867,
-0.0379329585,
0.0646238625,
0.0231382139,
-0.019715298,
-0.0092043895,
0.0303422976,
0.0124245621,
0.0427333377,
0.0202413388,
-0.0227686055,
-0.0510944538,
0.0017088865,
0.0047385977,
-0.0154348556,
0.0108457804,
0.038622383,
-0.0373228528,
-0.0416897275,
0.0187963061,
-0.0201190263,
-0.0302853137,
0.0359813981,
0.0418041088,
0.0333056115,
0.0137253143,
-0.0044867801,
-0.0008985169,
-0.0121961599,
-0.0180966444,
0.0061420118,
0.0050757993,
-5.820061345e-33,
-0.0782317072,
-0.0589078367,
0.014668934,
-0.0057836086,
-0.0506141409,
0.0444065705,
-0.0194875803,
0.0288405549,
0.006733973,
0.0194305554,
0.0111071793,
-0.0361172296,
0.0375912115,
0.0446525328,
-0.0192197151,
-0.0017747841,
0.0540719517,
0.0207836889,
0.0070213531,
-0.0152683081,
-0.0045396988,
0.0398324691,
0.0761877,
-0.0578632019,
0.0068101846,
0.0151484059,
-0.0395948477,
0.0221841056,
0.0661520511,
0.0470173582,
-0.0479590595,
0.0141866365,
0.0127247563,
-0.0437332019,
-0.0161683187,
0.0149679501,
-0.0915871188,
0.0467452668,
0.0240302123,
0.0207622238,
-0.0296450313,
-0.007370512,
0.0279504322,
-0.0221118387,
0.0248683002,
-0.0174573045,
0.0634497926,
-0.0464775823,
0.0181032494,
0.0012335895,
-0.0575753413,
0.0253158007,
0.0046914844,
-0.030059211,
0.0575865693,
-0.045008909,
0.0540339425,
0.003290101,
-0.0247129072,
0.0095640151,
0.015613649,
-0.0073338565,
0.0758867562,
0.0798958391,
0.0072999634,
0.0879075825,
-0.0125389062,
-0.0419353209,
-0.057225816,
0.0120254299,
0.0207989737,
0.0727131963,
0.0107961465,
0.0307130236,
-0.0033323297,
0.0308309998,
-0.0162063781,
0.0124048293,
-0.0696558952,
0.0382219143,
0.0232860595,
0.0222234465,
-0.0106216054,
0.0085003227,
-0.0255531929,
-0.0280927215,
-0.022463996,
-0.0113988463,
0.0237146858,
0.0287938733,
-0.033762373,
0.0793806612,
0.074387446,
0.0214225743,
-0.0052049109,
0.0313381366,
0.0181107912,
0.0576386303,
-0.0219871439,
-0.0400987603,
-0.0269404817,
-0.0191202685,
0.0506537408,
-0.0484275185,
0.0103255762,
0.0093360143,
0.0094824554,
-0.0085301548,
-0.028888138,
-0.0160488747,
-0.0201524608,
-0.0205837637,
-0.0436641499,
0.0459596589,
0.0259332899,
0.008520606,
-0.0693655461,
0.0403075852,
0.0210413206,
0.0310925376,
-0.0023424982,
-0.0421560183,
0.0118263438,
0.043695733,
-0.0422163978,
0.0299379248,
-0.0124830846,
-0.0587635636,
0.039576713,
-0.0504693501,
-0.0160748642,
0.0074210349,
2.572e-7,
0.0134856412,
0.0368803628,
0.037108995,
-0.0260044858,
-0.0255445018,
0.0588779785,
0.0100819478,
0.0193795171,
-0.0153510412,
-0.1053005457,
-0.0090250494,
0.0294445138,
0.0233575534,
0.0173136331,
-0.049263794,
0.0011427889,
0.006668611,
-0.0457716957,
-0.0673586354,
-0.0062786387,
0.0237147622,
0.0942217857,
-0.0019214159,
-0.0294268727,
0.0122833373,
-0.0666531101,
0.059935037,
-0.0306317899,
-0.0292880386,
-0.0021761879,
-0.0205825958,
0.041060064,
-0.053059008,
0.0425295047,
0.0162056349,
-0.0065528643,
0.0892441496,
0.0357214063,
-0.0368175022,
0.0406580344,
-0.0456050932,
0.0261936821,
-0.0378243923,
-0.0039755576,
0.0565587878,
-0.0312488489,
-0.0201692097,
-0.0368695296,
-0.0208925754,
0.013478497,
-0.0178323742,
-0.0160939042,
-0.0482697859,
0.0258408766,
0.0313595459,
-0.0122376755,
-0.0569539778,
-0.1051395983,
-0.0091376016,
0.0112661598,
-0.0191630237,
-0.020189492,
-0.0058300584,
-0.071981065,
0.003646503,
-0.0559342094,
-0.0313115083,
2.478410223e-34,
-0.0065863905,
0.0214186739,
-0.0239478592,
0.033340428,
0.0543638617,
-0.0086651929,
-0.0359708257,
-0.0253060367,
-0.049484279,
-0.057280872,
-0.0032948332
] |
http://arxiv.org/abs/2402.02549v1 | http://arxiv.org/pdf/2402.02549v1 | Are Large Language Models Table-based Fact-Checkers? | 2024-02-04T15:52:59 | 2024-02-04T15:52:59 | Hangwen Zhang, Qingyi Si, Peng Fu, Zheng Lin, Weiping Wang | cs.CL | cs.CL, cs.AI, cs.LG | Table-based Fact Verification (TFV) aims to extract the entailment relation
between statements and structured tables. Existing TFV methods based on
small-scaled models suffer from insufficient labeled data and weak zero-shot
ability. Recently, the appearance of Large Language Models (LLMs) has gained
lots of attraction in research fields. They have shown powerful zero-shot and
in-context learning abilities on several NLP tasks, but their potential on TFV
is still unknown. In this work, we implement a preliminary study about whether
LLMs are table-based fact-checkers. In detail, we design diverse prompts to
explore how the in-context learning can help LLMs in TFV, i.e., zero-shot and
few-shot TFV capability. Besides, we carefully design and construct TFV
instructions to study the performance gain brought by the instruction tuning of
LLMs. Experimental results demonstrate that LLMs can achieve acceptable results
on zero-shot and few-shot TFV with prompt engineering, while instruction-tuning
can stimulate the TFV capability significantly. We also make some valuable
findings about the format of zero-shot prompts and the number of in-context
examples. Finally, we analyze some possible directions to promote the accuracy
of TFV via LLMs, which is beneficial to further research of table reasoning. | Are Large Language Models Table-based Fact-Checkers?: Table-based Fact Verification (TFV) aims to extract the entailment relation
between statements and structured tables. Existing TFV methods based on
small-scaled models suffer from insufficient labeled data and weak zero-shot
ability. Recently, the appearance of Large Language Models (LLMs) has gained
lots of attraction in research fields. They have shown powerful zero-shot and
in-context learning abilities on several NLP tasks, but their potential on TFV
is still unknown. In this work, we implement a preliminary study about whether
LLMs are table-based fact-checkers. In detail, we design diverse prompts to
explore how the in-context learning can help LLMs in TFV, i.e., zero-shot and
few-shot TFV capability. Besides, we carefully design and construct TFV
instructions to study the performance gain brought by the instruction tuning of
LLMs. Experimental results demonstrate that LLMs can achieve acceptable results
on zero-shot and few-shot TFV with prompt engineering, while instruction-tuning
can stimulate the TFV capability significantly. We also make some valuable
findings about the format of zero-shot prompts and the number of in-context
examples. Finally, we analyze some possible directions to promote the accuracy
of TFV via LLMs, which is beneficial to further research of table reasoning. | [
0.0277535301,
0.0376038924,
-0.0008334581,
0.0533922128,
-0.0688166618,
0.036651127,
0.0210718736,
0.0176707264,
0.000945756,
-0.0426834263,
-0.0098000579,
-0.0324519202,
-0.0259282142,
0.0056413426,
-0.0034147766,
0.0611972958,
0.0713110566,
-0.010251388,
-0.0127443699,
0.0144228302,
-0.0165480543,
-0.0205177683,
-0.0419026278,
-0.0141182439,
-0.0165360458,
-0.0057699927,
-0.0198089201,
-0.019574808,
-0.0103979856,
0.0012648989,
-0.0146521963,
0.0700012892,
0.017240461,
0.0339976065,
0.0000020625,
0.0028352519,
-0.0387025476,
-0.0019109648,
-0.0746329129,
-0.0089470027,
0.0004322089,
-0.026543349,
-0.0090374043,
-0.0202607922,
-0.0256985798,
-0.0551151484,
0.0213982742,
0.0766328797,
0.0335037746,
0.0613182969,
0.0065065818,
-0.0256887339,
-0.0003005968,
0.0205041599,
0.0628845692,
-0.0318691209,
0.0262779798,
0.0077837524,
0.0127563011,
0.0074190577,
0.0241107326,
0.0375506468,
0.0389755219,
0.0066868262,
-0.0253104791,
0.0475995243,
-0.03770224,
0.0151855135,
0.0012290819,
0.0771983564,
0.0286168288,
-0.0341143571,
0.0029737356,
0.0037328571,
0.010719425,
0.0010722897,
-0.0272383876,
0.0516288541,
-0.0261740889,
-0.0105994288,
0.0214605853,
-0.0124330437,
0.0094583668,
-0.010446215,
-0.0384435318,
0.0982352197,
-0.0158111472,
-0.0639843121,
0.0094912369,
-0.0124746934,
-0.0270784125,
-0.0587734133,
-0.0264191758,
0.0439841934,
-0.0433829688,
0.0407226682,
-0.0029673211,
-0.0075993915,
0.030870175,
-0.0560255237,
0.0280150827,
0.0602460317,
-0.030766381,
0.0441366173,
-0.0640399605,
0.0320985131,
-0.0254330486,
0.0190226492,
-0.0241551213,
-0.0277016759,
-0.0452612154,
0.0290999655,
-0.0386153236,
0.0406142622,
-0.0373941399,
0.0011132839,
0.0027716863,
0.0028560278,
-0.0334659368,
0.0966750905,
-0.0189751107,
-0.0123513304,
0.0110509209,
0.0089565599,
-0.0610010773,
0.0038045091,
-0.0267220959,
-0.0129841492,
-0.0232184287,
-0.0778565556,
-0.0205218103,
-0.0362800173,
0.0171011277,
-0.0287462678,
0.0648562685,
0.0378067307,
-0.0142672462,
-0.0157191604,
-0.0417926982,
-0.0316292942,
-0.0556028038,
-0.1058137491,
-0.0023906801,
-0.0745569468,
0.0346329473,
0.0268776212,
0.0020392211,
-0.0176160485,
0.0163928531,
0.0613416582,
-0.0111981621,
0.0696391761,
0.0395430252,
0.0110564865,
-0.0194867626,
-0.0208761208,
-0.0074240658,
0.0621236898,
0.0018767648,
0.0026615351,
-0.0102385208,
-0.0015155261,
-0.0319427736,
0.0042026401,
0.0626229495,
0.0230719522,
-0.0038829579,
0.0377554633,
0.0764464661,
0.0039641936,
-0.0421113968,
0.0295272656,
-0.0209224746,
0.0302413143,
0.0408407561,
-0.0091454079,
0.0007112166,
0.0197563842,
0.0352037176,
-0.0221817549,
0.0338735767,
0.0372634567,
-0.0500083379,
0.0337941386,
-0.0410591438,
0.0131105343,
0.0298985969,
-0.003667969,
-0.0202839542,
-0.045368135,
-0.0249199588,
0.0230815131,
-0.0319644883,
0.0200089961,
0.0344685651,
0.0297020189,
-0.051182434,
0.0196252242,
0.0370221138,
-0.0510794781,
-0.0613309741,
-0.0198525544,
0.0454610586,
0.0331072398,
0.0482880101,
-0.0620796233,
-0.0107762162,
-0.0115123875,
0.0362473726,
0.0210836437,
0.0529276431,
-0.0183705762,
-0.0201363042,
0.0221110173,
0.0101996707,
0.0005203603,
0.0426579192,
-0.0356983095,
0.0349468999,
-0.0114451665,
-0.0160883125,
-0.0245586466,
-0.0183263533,
0.0302608479,
0.0419059768,
-0.001236752,
0.0183799714,
0.0269895103,
0.0323812217,
0.0123672141,
-0.0105050905,
0.0244740807,
0.0462080427,
-0.0232308079,
0.0000697307,
-0.0441439785,
-0.0305511598,
-0.0538069271,
0.0491828956,
0.0420772806,
-0.0736992955,
-0.029608164,
-0.0377404802,
-0.0149309868,
-0.02838232,
-0.0090401396,
-0.0248693265,
-0.0203194134,
0.0200807769,
0.0015755685,
0.0131825954,
0.0079718744,
0.1198090836,
0.0052348501,
0.0333181359,
0.0021538895,
-0.0495693907,
0.0180256236,
-0.0208154395,
0.0102940137,
-0.0010459175,
-0.018438708,
0.0061182352,
-0.0002677465,
0.0149273947,
-0.0335095674,
-0.0039154063,
-0.1152808517,
0.0586108416,
-0.0335671566,
-0.0121301515,
0.0213217624,
-0.0123911165,
0.0072170943,
0.0086163934,
-0.0279329438,
-0.0126160197,
0.0165167954,
0.0189302526,
0.0699581727,
-0.0214901064,
-0.0358503684,
-0.0035605263,
-0.0187319871,
-0.05555024,
0.0378048271,
-0.0274200961,
-0.0196394417,
0.0254177321,
0.0053683911,
-0.0101478612,
0.0238212794,
0.0231316164,
0.0294176489,
0.0006424491,
-0.0101684127,
0.0301004797,
0.0577444546,
-0.0374381468,
0.0357320793,
0.0564354286,
-0.0036424431,
-0.0259236377,
0.0511281267,
-0.0068098772,
-0.0062804385,
0.0349592045,
-0.047236342,
0.0192579646,
0.0226779059,
0.0020118833,
-0.0026883204,
-0.0374910645,
0.0173020903,
0.0184707288,
0.0109115494,
-0.0657440871,
0.0497242287,
0.0152351875,
0.0207336862,
-0.0159228425,
0.0142322555,
0.0325782038,
-0.0454879776,
0.0171671994,
-0.038002044,
-0.0199330598,
-0.0473504253,
0.0710330382,
0.0206476767,
-0.0034419938,
0.0101513397,
0.0124464035,
-0.0162037462,
-0.0158561897,
0.0301788133,
-0.0201669056,
0.1068160385,
-0.0483894497,
0.0167684257,
-0.0224888995,
-0.0372202508,
-0.0474651791,
-0.0705775023,
-0.035409186,
0.0372279659,
-0.0370216034,
-0.0130050778,
-0.0327548422,
-0.0334420018,
0.066856645,
0.0395011082,
-0.0236350782,
-0.0149435317,
-0.0419336259,
0.0388492495,
0.0352747142,
-0.0068778037,
-0.0244225711,
-0.0106147798,
0.0455848686,
0.0457638465,
-0.008973157,
0.0044029844,
-0.0158426594,
0.0050952481,
-0.0054426957,
-0.0273119342,
0.0086578904,
0.047611542,
-0.0300096944,
-0.0624900833,
-0.0372331142,
0.0815616101,
0.0415719934,
0.0471601523,
-0.0057463343,
0.0260180067,
0.0042520901,
0.017934734,
0.0472508594,
-0.1158521101,
-0.0431607515,
0.0052357269,
-0.0645259097,
-0.0346768759,
-0.0281416625,
-0.0496972427,
-0.1326921433,
0.0466181338,
-0.0397327691,
-0.0427378379,
-0.0853722245,
-0.0327338614,
0.0155000351,
0.0151862875,
0.0191446822,
0.0627873018,
-0.0156018492,
-0.0084340284,
0.0087672677,
0.0054313424,
-0.0411908813,
0.0474793464,
-0.0479063131,
0.0429096594,
0.0180314966,
0.0061020362,
-0.0245880596,
-0.021283539,
0.0076307035,
0.0045707561,
-0.0064856815,
-0.0107263913,
0.0480918251,
-0.0206845738,
0.0197333135,
0.0001998046,
-0.0123933339,
0.0050173784,
-0.0182627235,
0.0212888457,
0.0747346207,
-0.0324062593,
-0.0156086842,
0.0267874021,
0.0013193985,
-0.0135070393,
0.0975004137,
0.0120774275,
0.0079317456,
-0.0450405404,
-0.0284134597,
-0.0882867873,
-0.0266279988,
0.0106102824,
0.0501313619,
-0.0175604969,
-0.0051270993,
-0.0257426612,
0.0348339826,
0.0641403347,
0.0545303114,
-0.0208009928,
0.0194204189,
-0.0228874013,
0.0054641045,
-0.018555874,
-0.0344445705,
-0.0851046965,
0.0418651141,
-0.0217000619,
0.0065186052,
0.0212235618,
-0.0316580646,
-0.0181163158,
0.0155651607,
-0.06796325,
-0.0160396695,
-0.0146400966,
-0.0099985003,
0.0168515071,
0.0331832133,
-0.0057845362,
0.0102660032,
0.0101654772,
0.030829832,
-0.0611129552,
-0.0029651662,
0.007925245,
-0.0347377919,
0.0221904069,
0.0252054986,
-0.0311175957,
-0.001907394,
0.0243339799,
0.0371762961,
-0.018383041,
0.0245138425,
0.0771877915,
0.0244289879,
0.0310960002,
0.061520312,
-0.0716567114,
-0.0156013854,
0.0489182658,
-0.059870936,
-0.0644435734,
0.0237546191,
0.0018149524,
-0.002609351,
-0.0126106935,
-0.0145309968,
-0.0164002385,
0.0050871968,
0.0606782213,
0.0169646218,
-0.0160564389,
0.0309318248,
-0.0294339545,
0.0198825989,
-0.0465834662,
-0.0454800017,
-0.001057792,
0.0249449238,
0.0278221164,
-0.0155808153,
0.0276198443,
0.0470706001,
-0.0417419411,
-0.0604608692,
-0.0044133267,
-0.0891408026,
0.0020521476,
0.0079493178,
0.0365261994,
0.0556611009,
-0.0134253744,
-0.0135092335,
-0.0442313999,
-0.0202120952,
-0.0005419356,
-0.0004192677,
0.0424986519,
-0.0153676951,
0.0396841578,
0.0126671977,
-0.0138830822,
-0.023754308,
0.0360478945,
0.0256657731,
0.0376733392,
0.041192282,
-0.0133556267,
-0.0578909442,
-0.0036413977,
-0.0287237819,
-0.0080686975,
-0.0280485041,
-0.000648417,
-0.0619821362,
-0.0279783029,
0.040619269,
0.0271218028,
-0.0244369246,
-0.0103835799,
0.0009452011,
0.0424578115,
0.0705947205,
-0.0509733073,
-0.0574500263,
0.0430270135,
0.016955357,
-0.0179485939,
-0.0177881066,
-6.46019663e-33,
-0.0510853715,
-0.0371421836,
0.0241012592,
-0.0317360014,
0.0402544662,
-0.0080828294,
-0.0476431549,
-0.0219367035,
-0.0420357212,
0.0242994297,
-0.0446751826,
0.034406051,
0.0262930039,
-0.0348093584,
-0.0013770111,
-0.0704730079,
0.0374855921,
-0.0008212617,
-0.0161096361,
-0.0142425457,
0.0347258598,
-0.0254887193,
-0.0050882134,
-0.0278365333,
0.0453360043,
-0.0155223729,
0.0000856164,
0.0097695049,
-0.047461994,
0.0323024653,
-0.029612286,
0.0326960012,
0.0431534089,
-0.0840787366,
0.0059281662,
0.0052382546,
-0.1433712691,
-0.0213420819,
0.0227420256,
-0.0116836922,
-0.0032203875,
-0.0115198744,
0.0034604955,
0.0083355494,
-0.0110078426,
-0.021985935,
-0.0033475228,
-0.0372643545,
0.0069004465,
-0.0038007279,
-0.0423809923,
0.0149485236,
0.0182285998,
0.0082083745,
0.036359299,
0.0347068384,
0.0239488706,
-0.0317633972,
-0.049173478,
0.0146953091,
0.0106594022,
0.0184256006,
0.0456002392,
-0.0351359881,
-0.0047126845,
0.0291765854,
-0.0059314868,
0.0249599963,
-0.0555594862,
0.013312377,
-0.0161764976,
0.0624892153,
0.0397559442,
0.0406233259,
0.0562774464,
0.0146155702,
-0.0791861266,
0.0115655074,
0.0447927788,
0.0347691216,
0.0045953863,
-0.039562393,
-0.0522962436,
-0.0137736211,
-0.0312130228,
0.0357503295,
-0.0459223054,
0.002666862,
0.0178818703,
-0.0034413466,
0.0233571585,
-0.0077628884,
0.015798917,
-0.0119206021,
-0.0126309656,
0.0661987215,
-0.0166093744,
-0.0071547837,
0.0068689412,
0.020859573,
-0.0519554876,
-0.0349209979,
0.0022369896,
-0.0233681891,
0.029954277,
0.057167571,
0.0852044374,
0.0168911684,
-0.0687162802,
-0.0130422087,
-0.0583362468,
0.041240681,
-0.0138040194,
-0.0036151658,
0.0172043685,
0.0365276933,
0.0049998336,
0.0418361127,
0.0156147899,
-0.0255020522,
-0.0246211439,
0.0374815203,
-0.0073869079,
0.0081009259,
0.0075609833,
0.0064789336,
-0.0390934907,
0.0411557928,
0.02013666,
-0.0363795869,
0.0189078655,
0.0284199174,
2.71e-7,
0.0210159197,
0.0289421864,
0.021308817,
0.0580454729,
-0.0176054556,
0.0391515754,
-0.0081007695,
-0.0215507578,
-0.0211730637,
-0.0657274649,
0.024152454,
0.0240853075,
0.0512319505,
0.0548096076,
-0.0204771478,
-0.0041189874,
-0.0497368239,
-0.0143481437,
-0.0484230444,
0.0168343652,
0.0444555245,
0.0743229091,
-0.0018795142,
-0.0029541403,
-0.0057688481,
-0.010688467,
0.0276822317,
-0.0093816575,
0.0081996713,
0.0150976116,
0.0324683264,
-0.0247216392,
0.009390736,
0.088431336,
0.0015177559,
0.0113908788,
-0.0335695222,
0.0626729503,
0.0171562936,
0.0517202951,
-0.0354132354,
-0.0628718585,
-0.0015654428,
-0.0563022494,
0.1394878477,
-0.0362870097,
0.0120055117,
0.0220831651,
0.0104952687,
0.0233775992,
-0.0175388381,
0.031452328,
-0.0224380866,
-0.0121546872,
0.0275645405,
0.0046227812,
-0.006796943,
-0.0355215147,
-0.0189643465,
0.036635872,
-0.015642019,
-0.0561147407,
0.0117464168,
-0.016766008,
0.0225945804,
-0.0790813714,
-0.040618062,
2.871928997e-34,
0.0358577818,
-0.0063581145,
0.0388407968,
0.0026266726,
0.0516030379,
-0.0079792868,
-0.0317103229,
-0.0219776761,
0.0242338013,
-0.0721452087,
-0.0129185766
] |
http://arxiv.org/abs/2402.02548v1 | http://arxiv.org/pdf/2402.02548v1 | "What's my model inside of?": Exploring the role of environments for grounded natural language understanding | 2024-02-04T15:52:46 | 2024-02-04T15:52:46 | Ronen Tamari | cs.CL | cs.CL, cs.AI, cs.SI | In contrast to classical cognitive science which studied brains in isolation,
ecological approaches focused on the role of the body and environment in
shaping cognition. Similarly, in this thesis we adopt an ecological approach to
grounded natural language understanding (NLU) research. Grounded language
understanding studies language understanding systems situated in the context of
events, actions and precepts in naturalistic/simulated virtual environments.
Where classic research tends to focus on designing new models and optimization
methods while treating environments as given, we explore the potential of
environment design for improving data collection and model development. We
developed novel training and annotation approaches for procedural text
understanding based on text-based game environments. We also drew upon embodied
cognitive linguistics literature to propose a roadmap for grounded NLP
research, and to inform the development of a new benchmark for measuring the
progress of large language models on challenging commonsense reasoning tasks.
We leveraged the richer supervision provided by text-based game environments to
develop Breakpoint Transformers, a novel approach to modeling intermediate
semantic information in long narrative or procedural texts. Finally, we
integrated theories on the role of environments in collective human
intelligence to propose a design for AI-augmented "social thinking
environments" for knowledge workers like scientists. | "What's my model inside of?": Exploring the role of environments for grounded natural language understanding: In contrast to classical cognitive science which studied brains in isolation,
ecological approaches focused on the role of the body and environment in
shaping cognition. Similarly, in this thesis we adopt an ecological approach to
grounded natural language understanding (NLU) research. Grounded language
understanding studies language understanding systems situated in the context of
events, actions and precepts in naturalistic/simulated virtual environments.
Where classic research tends to focus on designing new models and optimization
methods while treating environments as given, we explore the potential of
environment design for improving data collection and model development. We
developed novel training and annotation approaches for procedural text
understanding based on text-based game environments. We also drew upon embodied
cognitive linguistics literature to propose a roadmap for grounded NLP
research, and to inform the development of a new benchmark for measuring the
progress of large language models on challenging commonsense reasoning tasks.
We leveraged the richer supervision provided by text-based game environments to
develop Breakpoint Transformers, a novel approach to modeling intermediate
semantic information in long narrative or procedural texts. Finally, we
integrated theories on the role of environments in collective human
intelligence to propose a design for AI-augmented "social thinking
environments" for knowledge workers like scientists. | [
0.0645840093,
-0.0017111921,
-0.0492361523,
0.0401921533,
-0.0046517886,
-0.0061895852,
0.0029006393,
-0.0134524256,
-0.0103222448,
-0.0530347079,
-0.0249138251,
0.0145462481,
0.0398874469,
0.0118144443,
0.004810479,
-0.0400995649,
0.0310443453,
-0.0159521364,
-0.0256691165,
0.0257712118,
-0.0191208851,
0.0129664764,
-0.0367597006,
0.0510983095,
0.0089857662,
-0.014500292,
-0.0369642936,
-0.0208634343,
0.014857905,
-0.0463982224,
-0.0203337986,
0.0066830437,
0.0715219006,
0.0103305262,
0.0000021387,
-0.015442959,
0.0306150261,
-0.0221497659,
-0.0248794015,
-0.0382546186,
-0.0006813785,
0.0168885645,
0.0230517611,
-0.0248806477,
-0.0461678132,
-0.0046915594,
0.0598589107,
0.0388903096,
0.0602911673,
0.0129167615,
-0.0068180463,
-0.0303499326,
-0.0339928232,
0.0329236761,
0.0078767929,
-0.0157475024,
-0.0085103242,
0.0712119043,
0.0548019782,
-0.0183039643,
-0.0279591437,
0.0181788187,
0.0383910537,
0.0028577694,
0.0319635496,
0.072517693,
-0.0198979285,
0.0004960545,
-0.0176848993,
0.100701347,
0.0281759873,
-0.0350612216,
0.0085955476,
0.0129290996,
-0.0011651265,
0.0080243722,
-0.045332972,
0.060833741,
-0.020541437,
0.016123835,
0.0164741036,
0.0323241763,
-0.0101961885,
0.021578446,
0.0102647198,
0.085778974,
-0.0344093293,
0.0075273896,
-0.011375213,
0.0323224962,
-0.0211208593,
-0.0384679809,
-0.0029816993,
0.0401499048,
-0.0372791328,
0.0389766246,
-0.0146134729,
0.0117165251,
-0.0090304306,
-0.0095784003,
0.021319896,
0.0343710594,
0.0179880131,
0.0467434786,
0.0234354902,
0.0123601351,
0.0031725746,
-0.0039789719,
-0.0020413566,
0.040815603,
-0.08864858,
-0.0198273845,
-0.0319144018,
0.0353391021,
-0.0721591115,
-0.0308324229,
0.0369390361,
0.0110949362,
-0.0089069745,
-0.0017697574,
-0.013627558,
0.0460252389,
-0.0306525789,
0.0223105103,
0.0017373885,
-0.0348267518,
-0.0876808837,
0.0417837277,
-0.0079028383,
-0.0892353207,
-0.0169641785,
-0.0240270644,
0.019860344,
-0.0043182708,
-0.0155323474,
0.0305763558,
0.0343490131,
-0.0250204075,
-0.0112224044,
-0.042380739,
0.0033571692,
-0.0468752459,
0.0054100412,
-0.0225733425,
0.0099803722,
-0.0083001014,
0.02358355,
-0.0094900578,
0.0198245551,
0.0281911809,
-0.0079601137,
0.0914898887,
-0.0203460548,
-0.013887166,
0.0504450016,
0.0303438306,
0.0431412645,
0.0605857186,
0.0343048722,
-0.0366353691,
-0.0029841915,
0.0111043183,
-0.0104248533,
0.0038562233,
0.0032975359,
-0.0427746996,
-0.0007139195,
-0.0056824302,
-0.0141507573,
0.0579205342,
-0.0149387224,
0.0045943526,
-0.0465550832,
0.0044366871,
0.0531645976,
0.1225133687,
-0.0791443512,
0.0216602273,
0.0249377564,
0.0027385638,
-0.0197692066,
0.0010638197,
-0.0173302349,
0.0388550423,
-0.0124854678,
-0.0180079788,
-0.0574015006,
0.0065059951,
-0.0018373305,
-0.0320020765,
-0.0176351741,
0.0247129221,
0.0591343641,
-0.0196660887,
0.0197034981,
-0.0122870905,
-0.0106988717,
0.0286245495,
0.0172738265,
-0.0730087087,
0.0100889094,
-0.0319793411,
0.0276915412,
0.0446223468,
0.0500280894,
-0.0331139565,
-0.0389264897,
-0.0311087202,
0.0371314436,
0.0770895705,
-0.0174737182,
-0.0293167755,
0.0045481147,
-0.0150896646,
-0.019398516,
-0.0746072009,
-0.0080974242,
0.0216563307,
-0.0139395604,
0.0037742276,
-0.0276942831,
-0.0168502498,
-0.0143717723,
0.0528211482,
-0.0557241142,
0.0089750281,
-0.0870990977,
0.0052422965,
0.0155315734,
0.0205687992,
-0.024353262,
-0.0142393261,
0.0196941141,
-0.0357174985,
-0.0185584184,
0.084996745,
-0.0211840011,
-0.0140543403,
0.0268697012,
-0.091862984,
-0.0134897744,
-0.0029420224,
0.03055875,
-0.0048609241,
0.0263682976,
-0.0611303523,
0.0466359779,
-0.0084244609,
-0.0197484847,
-0.0317994282,
-0.0159493797,
-0.0412711129,
0.0607688949,
-0.0153410072,
0.0000241934,
-0.0020000977,
-0.0518918075,
0.0466712192,
-0.0403109118,
0.0380013771,
0.0170982387,
0.0160768703,
-0.0582858473,
0.0300059803,
0.0017511464,
-0.0947642028,
-0.0046466487,
-0.0949857906,
-0.0213030614,
-0.0242292546,
-0.0248532072,
0.0292036328,
-0.0347170494,
-0.0265478082,
-0.0182232205,
-0.0079914266,
-0.0101832747,
0.0025646649,
0.0023848917,
0.054238271,
0.0245078206,
0.0017335976,
-0.0061509958,
0.0341461487,
-0.045725178,
0.0658132061,
0.0599725805,
-0.0210222062,
-0.0796578899,
0.0163899306,
-0.02717066,
0.0198944081,
0.0002176724,
-0.0284601506,
0.0735691711,
0.0308221355,
0.0455417633,
0.082031548,
-0.0156233227,
0.0747681633,
0.0669858754,
-0.0561237112,
-0.01923595,
0.0126656704,
0.0167793427,
0.0582867414,
-0.0159643069,
-0.0226414856,
0.0136600202,
0.0674263686,
0.012114672,
0.0199463628,
-0.0913126245,
0.0864453539,
-0.0267195404,
-0.0373339057,
-0.0319966935,
-0.0294745676,
-0.0038716411,
0.0322514251,
-0.0501261428,
0.0239402931,
0.0838702545,
-0.0128089506,
0.005490745,
-0.0735606998,
0.0295268632,
-0.0084478259,
0.0835476294,
-0.0229704026,
0.0103593403,
0.0408745036,
-0.0191131812,
-0.0359884948,
0.0052691391,
0.0152892387,
0.0060978802,
-0.0039718389,
-0.0738104656,
0.0057293619,
0.0216313377,
-0.0218708087,
-0.0143818529,
0.0105766142,
-0.0643105134,
-0.0128493914,
0.0053571071,
0.0074957679,
-0.0454262868,
0.0163856689,
0.0436098762,
0.0181163102,
0.0120411972,
-0.0255257916,
-0.0150543461,
0.0138848089,
0.0706668422,
0.051885739,
0.0184666384,
-0.0302369222,
0.0362348109,
0.0388589092,
0.0069420743,
-0.0115786903,
0.0002490562,
0.0175047368,
0.0313033015,
-0.022172926,
-0.0230063722,
0.0273052659,
0.064105846,
-0.0404171757,
-0.0615852699,
0.1022296324,
0.0466689244,
0.0202500131,
0.0339463577,
-0.046477709,
-0.000525148,
0.0365657546,
0.0022081023,
-0.032347586,
-0.0215026606,
0.0044479249,
-0.0322052166,
0.019306602,
-0.088923648,
-0.0048177545,
-0.1010176241,
0.016614005,
-0.0190206021,
-0.1055896133,
-0.0233896784,
-0.0204270724,
0.0105842873,
0.0187668316,
-0.0213250406,
-0.0162356999,
0.0274237674,
0.0328152888,
0.0132809384,
0.0016860262,
-0.0006173346,
0.0141038205,
-0.0523158647,
0.0972918123,
0.017199127,
0.0820707306,
0.0093326066,
-0.0592778847,
-0.0311077423,
0.0026563171,
0.0475539006,
-0.0255077444,
0.0290647279,
-0.0131400991,
-0.0044950219,
-0.0162207913,
0.0395996831,
-0.0442619808,
-0.0475454405,
0.0023020252,
0.0843824074,
-0.0352360159,
0.0198682528,
-0.0061574173,
0.0588237979,
0.0295442529,
0.0601521209,
0.0336934999,
-0.0356078632,
-0.0039303536,
-0.0084822858,
0.0261361413,
-0.0262982957,
-0.063294664,
-0.0297307707,
-0.0001571331,
0.0334401019,
-0.0774052441,
-0.0411196947,
0.0418687649,
0.062581718,
0.025057191,
-0.0139522189,
0.0198112242,
0.0319024809,
-0.026392011,
0.0481310636,
-0.0396214761,
-0.0193460807,
0.0112048062,
-0.0227632988,
0.0241745636,
-0.06263192,
0.0098144962,
0.0150767248,
0.01261428,
-0.0415524766,
0.030122621,
0.0077749859,
0.0236060806,
0.0303642824,
0.0123489629,
-0.0281456802,
-0.0105467793,
-0.0203210451,
-0.0899077952,
0.0281128511,
-0.0225679222,
0.0101710586,
0.0297252592,
-0.00418306,
0.0010800589,
-0.022309633,
-0.0188353211,
0.0475201234,
-0.0545416847,
-0.0204233229,
0.0224560387,
-0.0139418533,
0.0392150357,
-0.0283923056,
-0.0147045385,
-0.0398312695,
-0.0011245387,
-0.0385763235,
-0.0159977265,
-0.0183096714,
-0.0035698614,
0.0029002819,
-0.0033450078,
-0.0023464712,
0.0293144416,
-0.0406551212,
0.0733418539,
0.0159570985,
-0.0017421077,
0.0474654287,
0.0088874782,
-0.0424271747,
0.0056777927,
0.0140611464,
0.0027526838,
-0.0554309748,
0.0683610588,
0.0162675008,
0.0795593932,
0.0425777808,
-0.0685715526,
-0.01010211,
0.022672113,
-0.0382851064,
0.0213314015,
0.0288444497,
0.0560531728,
0.0178772099,
-0.0045361202,
0.0069421395,
-0.0108979838,
0.0032712258,
0.0510398857,
-0.0126273958,
0.0280025844,
-0.0784104541,
0.0035172466,
0.0443809666,
-0.0232400727,
-0.0182960927,
0.0132483561,
-0.0013222515,
0.0633575544,
0.0127852894,
-0.010051514,
-0.0261466503,
0.0193421841,
0.0076361033,
0.0410678275,
0.0039602369,
0.0517137609,
-0.0905206501,
-0.0245636404,
0.0189922545,
0.0644104332,
-0.007681617,
-0.0167985354,
0.0399697609,
0.0271933023,
0.0156232072,
-0.0451131798,
0.0074719605,
0.0552753955,
-0.0572507121,
-0.0019928499,
-0.0572710931,
-5.871456896e-33,
-0.0781110525,
-0.0535732284,
0.0424489565,
0.0144526921,
-0.0437000506,
0.0113935536,
-0.0617980585,
0.0095381038,
0.0098567763,
0.0129425805,
0.0004922331,
0.0261379331,
-0.0116412519,
-0.0269963928,
-0.0180587154,
-0.0398545563,
-0.0042121904,
-0.0086414181,
-0.0339228176,
-0.0144860363,
0.0530098714,
-0.0038367934,
0.0440138392,
-0.0758902878,
0.0091181081,
0.0186414532,
0.0155588323,
0.0204644483,
-0.0192343034,
0.0076743532,
-0.0401308984,
0.0225716867,
0.0299421754,
-0.0348595828,
0.0271399152,
0.0406202376,
-0.0553040057,
-0.0594414845,
0.0327081159,
0.0178247429,
-0.0208009481,
-0.0219265819,
-0.0104712918,
-0.0125815198,
0.0063382941,
0.0191222224,
0.0567922145,
-0.0408301651,
0.0286464076,
-0.0182879977,
-0.0630331188,
-0.0127573702,
-0.0525821932,
-0.0164445825,
-0.017971376,
0.0151998987,
0.0056920182,
-0.0013669513,
-0.0408116989,
0.0301936306,
-0.0092425831,
-0.038183488,
0.0477600619,
-0.0013822011,
0.0432822816,
0.0904307887,
0.0019434347,
0.0468895249,
-0.1024260297,
0.0032516955,
0.0582401901,
0.0120768473,
0.0480588302,
0.0297345202,
-0.0310729649,
0.0321775228,
-0.0603376403,
0.0241354462,
0.017199019,
0.0430193022,
0.0115735801,
0.0244270451,
-0.0319352448,
-0.0073334803,
-0.0408109985,
-0.0114522306,
0.0121382829,
-0.014883006,
0.0081978692,
-0.0459996723,
-0.054589238,
-0.0335588008,
0.0022141216,
-0.0369793326,
0.0097482363,
0.040840596,
0.0110442648,
0.0048579397,
-0.0045331372,
-0.0141120087,
-0.0242869388,
-0.0115170227,
0.0177596007,
0.02647743,
0.0418063812,
-0.0001847035,
-0.0169643443,
0.0303076673,
0.0171840247,
-0.0346578807,
-0.0237854067,
0.0082629006,
0.0176225398,
-0.0030440448,
0.03746612,
0.0005766228,
-0.0300886072,
-0.0071184309,
0.0447467156,
-0.0126424702,
0.0013338652,
0.0114748599,
-0.0356399082,
0.0133398548,
-0.0161114261,
-0.0262492653,
-0.020835869,
0.0160277151,
0.0221421383,
-0.0650975555,
-0.0186827462,
0.0018604577,
2.898e-7,
-0.0014378937,
0.0532280803,
0.038932465,
-0.0481609032,
0.0025092526,
0.0142574841,
0.0091813309,
0.0062780427,
-0.0080478759,
-0.091240406,
0.0029766932,
0.0625261217,
0.0035324392,
0.0021395811,
-0.083326824,
-0.0142933885,
-0.020655958,
0.0000079083,
-0.038950935,
0.0112276776,
0.0285616964,
-0.0132483002,
0.0338104106,
0.0180434696,
-0.0130393505,
0.0010791715,
0.0036191077,
-0.054892607,
0.0437497906,
-0.0026697461,
-0.002409325,
-0.0229785144,
-0.0243480764,
0.0122048575,
0.0346796513,
0.0277630724,
0.019272849,
0.013957873,
-0.0313250721,
0.0226441324,
-0.0154770818,
-0.0570756309,
-0.0159928836,
-0.0472454429,
0.0596056506,
-0.0923493057,
-0.0452428684,
-0.0072773416,
-0.0231337622,
-0.0404579304,
-0.0153290071,
0.0044803158,
-0.0091708442,
-0.0095163221,
-0.0146918967,
-0.0291747414,
0.014040485,
-0.0657905564,
-0.0183009412,
0.002893117,
-0.0248488002,
0.0103427973,
-0.0332600549,
-0.0245690141,
0.0110660102,
-0.0763151497,
-0.0210803207,
2.408107625e-34,
-0.0493364707,
0.0062352666,
0.0127434023,
0.0174655542,
0.0663504899,
-0.0101097571,
0.0189275704,
-0.0009292043,
0.0667698458,
-0.0285228584,
0.0166633129
] |
http://arxiv.org/abs/2402.02547v2 | http://arxiv.org/pdf/2402.02547v2 | Integration of cognitive tasks into artificial general intelligence test for large models | 2024-02-04T15:50:42 | 2024-03-06T02:46:40 | Youzhi Qu, Chen Wei, Penghui Du, Wenxin Che, Chi Zhang, Wanli Ouyang, Yatao Bian, Feiyang Xu, Bin Hu, Kai Du, Haiyan Wu, Jia Liu, Quanying Liu | cs.AI | cs.AI, cs.CL | During the evolution of large models, performance evaluation is necessarily
performed to assess their capabilities and ensure safety before practical
application. However, current model evaluations mainly rely on specific tasks
and datasets, lacking a united framework for assessing the multidimensional
intelligence of large models. In this perspective, we advocate for a
comprehensive framework of cognitive science-inspired artificial general
intelligence (AGI) tests, aimed at fulfilling the testing needs of large models
with enhanced capabilities. The cognitive science-inspired AGI tests encompass
the full spectrum of intelligence facets, including crystallized intelligence,
fluid intelligence, social intelligence, and embodied intelligence. To assess
the multidimensional intelligence of large models, the AGI tests consist of a
battery of well-designed cognitive tests adopted from human intelligence tests,
and then naturally encapsulates into an immersive virtual community. We propose
increasing the complexity of AGI testing tasks commensurate with advancements
in large models and emphasizing the necessity for the interpretation of test
results to avoid false negatives and false positives. We believe that cognitive
science-inspired AGI tests will effectively guide the targeted improvement of
large models in specific dimensions of intelligence and accelerate the
integration of large models into human society. | Integration of cognitive tasks into artificial general intelligence test for large models: During the evolution of large models, performance evaluation is necessarily
performed to assess their capabilities and ensure safety before practical
application. However, current model evaluations mainly rely on specific tasks
and datasets, lacking a united framework for assessing the multidimensional
intelligence of large models. In this perspective, we advocate for a
comprehensive framework of cognitive science-inspired artificial general
intelligence (AGI) tests, aimed at fulfilling the testing needs of large models
with enhanced capabilities. The cognitive science-inspired AGI tests encompass
the full spectrum of intelligence facets, including crystallized intelligence,
fluid intelligence, social intelligence, and embodied intelligence. To assess
the multidimensional intelligence of large models, the AGI tests consist of a
battery of well-designed cognitive tests adopted from human intelligence tests,
and then naturally encapsulates into an immersive virtual community. We propose
increasing the complexity of AGI testing tasks commensurate with advancements
in large models and emphasizing the necessity for the interpretation of test
results to avoid false negatives and false positives. We believe that cognitive
science-inspired AGI tests will effectively guide the targeted improvement of
large models in specific dimensions of intelligence and accelerate the
integration of large models into human society. | [
0.0040525021,
0.0598018467,
-0.057894554,
-0.0133379642,
-0.0191542525,
0.0282770731,
0.0567525886,
0.0220396779,
0.056519568,
-0.0011329514,
-0.0446146205,
-0.0513336547,
-0.0212156866,
0.0389568172,
0.0693603382,
-0.0210085027,
0.0513969399,
-0.0409438051,
0.0256361254,
0.0071307356,
-0.0705983937,
-0.0445805602,
0.0002946754,
0.0072749127,
0.0090404768,
-0.002552167,
0.0141586307,
-0.0187871791,
0.0369648188,
-0.0475414582,
-0.0444336906,
0.0274336357,
0.0273169465,
0.0365087613,
0.000002359,
-0.0487721451,
0.0265555046,
0.0062997132,
-0.0453093164,
0.0006993549,
0.0434248298,
0.0368555859,
0.0126855522,
0.0106047951,
-0.0266743191,
0.012160792,
0.0048689293,
0.0233569369,
0.0100196218,
0.020841442,
-0.0018314389,
0.0155648412,
0.0599569865,
-0.0126180388,
0.0078758178,
0.0066863755,
0.0249683727,
0.0357689261,
0.0303339045,
-0.0438901708,
0.0214921497,
0.0295632239,
0.0483686663,
0.0217076465,
0.0049149194,
0.0276333522,
-0.0113611277,
-0.0218851939,
-0.0140736131,
0.0255099386,
0.0102343978,
0.0087543679,
-0.0070931595,
0.0365525223,
0.0288528875,
0.0311309397,
-0.0553031266,
0.0175108984,
0.0309438836,
-0.0624183305,
0.0148172118,
0.0361045003,
0.0520378761,
-0.0073564192,
-0.0369558781,
0.0614361353,
0.0244704541,
-0.0207846723,
0.0148764513,
-0.0131816305,
-0.0656576455,
-0.0001374657,
-0.0349305309,
-0.0086632818,
-0.0564963892,
-0.0659934655,
0.0117973201,
0.0153067689,
0.0193522703,
-0.0223396625,
0.1214318424,
0.0273672864,
0.047063224,
0.0279632676,
-0.0188109297,
0.100789912,
-0.0159913134,
0.0229268577,
-0.0212547928,
0.0067329989,
-0.0620382428,
-0.0010849652,
0.0239777081,
0.0325652771,
0.0035935841,
-0.0409373865,
-0.0174157917,
0.0121026328,
-0.074452363,
0.0037169079,
-0.0492685661,
0.0414736308,
0.0604079068,
0.0243779477,
-0.0497667976,
-0.0600513034,
-0.0613603145,
0.0192628969,
-0.0005994671,
-0.0033931085,
0.0022970524,
-0.0030003896,
-0.011346695,
-0.0016498785,
0.0143348994,
0.090157859,
0.024637932,
-0.0257631857,
-0.0021688528,
-0.0245687775,
0.0361660011,
-0.0662637725,
0.050081443,
0.0016029411,
0.0392935127,
0.001743408,
0.0152687319,
0.0023906592,
-0.0253624283,
-0.0265358221,
-0.0058929999,
0.0241059046,
0.0150122149,
0.01373812,
0.0398580134,
0.0099470317,
0.0029169088,
0.0562862828,
0.0154525163,
-0.0049874317,
-0.0022311627,
0.0195476469,
0.0223758016,
0.0164557695,
-0.0166033115,
0.0164506268,
0.0044934,
0.0076821554,
0.0055213668,
0.0410216562,
-0.0077741244,
0.0160216838,
-0.0252743233,
0.0071669794,
0.047527615,
0.0704199448,
-0.0398794189,
0.0148187838,
0.0689879134,
-0.0083968844,
0.0127087506,
-0.0186531004,
-0.0797720924,
0.0444351584,
-0.0728563443,
-0.0011306019,
-0.0077236407,
0.0315983407,
-0.0261025392,
-0.0333279185,
0.0007853337,
0.007781676,
0.0537919179,
0.0248312186,
0.0224704202,
-0.0542713217,
0.0030472951,
-0.0333387852,
-0.0009318693,
-0.0212984346,
-0.0504344925,
-0.0800980181,
0.0407693498,
0.0256663598,
0.0245971419,
0.0028728177,
-0.0478648394,
0.0267185904,
-0.0001981457,
0.0249196962,
0.0624023154,
-0.0172481928,
-0.0251127817,
0.0107580572,
0.0023854878,
-0.0051441309,
0.0302886721,
0.0405135714,
0.0004262308,
-0.0065322951,
-0.0059231203,
0.0284359213,
-0.0216737688,
0.0151845971,
-0.045528926,
0.0329054482,
-0.0791636705,
-0.0019653088,
0.083657749,
0.0024005349,
-0.0036359425,
-0.0233589467,
0.0360623822,
-0.0078203538,
-0.040530283,
0.0056536044,
-0.0031683617,
0.0784215108,
-0.0227790866,
-0.023487892,
0.001902074,
0.0490940213,
0.0425911956,
0.002782814,
0.0558160394,
-0.0294337273,
0.0279022306,
0.001307531,
0.053942021,
-0.0629684478,
-0.0086025791,
0.0260338709,
0.053919252,
-0.0022674496,
-0.0031976933,
0.011894661,
-0.0255285669,
0.0529325381,
-0.0587194487,
-0.0002946325,
-0.0100761019,
-0.0214229114,
-0.010493909,
0.0137005672,
-0.0133176493,
-0.0580306314,
0.007904076,
-0.0974702165,
-0.0013304289,
0.0116477935,
0.0061748479,
-0.0084423525,
-0.0473359041,
-0.0109644337,
0.045867689,
0.0117487414,
0.0234785564,
0.0541926809,
0.0052570589,
-0.0112671079,
-0.0427881554,
-0.0106394449,
-0.0292700436,
0.0199321359,
-0.0495224483,
0.0677810684,
0.061157234,
0.0163041726,
-0.0060479697,
-0.0070134504,
-0.0529490896,
-0.0016899942,
-0.003030234,
-0.0664677471,
0.0532490797,
0.0096564572,
0.0136868386,
0.1015704498,
0.0015616596,
0.0109132119,
0.0310375821,
0.0038205099,
0.0074608871,
0.020061424,
0.0102064218,
0.0445834361,
-0.0135114575,
-0.0348389298,
-0.0202361811,
0.0612724386,
-0.0232129786,
0.0336825736,
-0.0199227687,
0.0962037668,
-0.0578251109,
-0.0785152316,
-0.0204933584,
0.0187015906,
0.0034762973,
0.0191538036,
-0.0569647774,
-0.0039873887,
0.0386279449,
0.0159561578,
0.0139152342,
0.0251803529,
0.0106715932,
0.004824108,
0.051149115,
0.0317109376,
0.0031425613,
-0.0243227258,
0.0294702351,
-0.0294852909,
0.0039622998,
-0.0353942402,
-0.020907918,
-0.0316280238,
-0.080571793,
-0.023771394,
0.031391494,
-0.0524425507,
-0.0386008658,
0.0608680546,
-0.0363543779,
0.0495795123,
0.0176878311,
-0.058610227,
-0.0931052417,
-0.0018467525,
0.042640008,
-0.0361335129,
-0.0355002172,
-0.0001627467,
0.0135946786,
-0.0088183684,
0.0279045403,
0.0729988366,
0.0300868601,
-0.0436474718,
0.0436359681,
0.0439334325,
-0.0017250039,
-0.0007019314,
-0.0302220527,
0.0512296148,
0.0104990471,
-0.0042128512,
-0.0054614875,
0.0053578028,
0.0460159145,
-0.0198587999,
-0.0134715913,
0.0998773053,
0.0058971141,
0.0151436813,
0.0173197426,
0.0082863579,
-0.0485874228,
0.044207979,
0.0068610399,
-0.0294414349,
-0.0009063225,
-0.008926576,
-0.0249798,
-0.0448203757,
-0.042447634,
0.0609117858,
-0.0769066587,
0.0480409563,
0.0028620148,
-0.0533747599,
-0.0269841813,
-0.0207112413,
0.0468427613,
-0.0130557884,
-0.0348810293,
-0.010006641,
0.025379438,
-0.026047226,
-0.0387663655,
-0.03397277,
-0.0358818844,
0.0611953586,
-0.0594865456,
0.0477847643,
0.0460169762,
0.042883914,
-0.0395682342,
-0.0716126412,
-0.0122251911,
-0.0045360508,
0.0368450694,
0.012462304,
0.0409005694,
0.0115413284,
0.0411244668,
-0.0297555085,
0.0019544403,
0.0183012597,
-0.0101990271,
0.0204819553,
0.0681577176,
-0.0252006035,
-0.0842087492,
-0.0285912342,
0.0018482086,
-0.011550324,
0.0689878315,
-0.0030782861,
-0.0528247058,
-0.0362916365,
-0.0164776314,
-0.0304330252,
-0.0280359592,
-0.0746257082,
0.0025592186,
0.0553404801,
0.0091977259,
-0.0388409644,
-0.0210785642,
0.031285245,
0.002393527,
-0.0017020658,
-0.0757850483,
-0.0075492142,
-0.0020476615,
-0.0358674899,
0.0138513092,
-0.0038960096,
-0.0091913026,
-0.0453907885,
-0.0261769295,
0.0055293441,
-0.0666574538,
-0.007571823,
0.0383881293,
-0.001283037,
-0.0255410764,
0.0380656905,
0.0255900379,
0.0306165982,
-0.0370088108,
-0.0504818633,
-0.0032592702,
0.0844430998,
-0.0550903566,
-0.0708072633,
0.0463029221,
-0.011844446,
0.0191752743,
-0.0096102729,
0.0022966298,
-0.0176667441,
0.0451269746,
0.0011050933,
-0.0292603187,
0.0029628847,
-0.0194255151,
-0.0148397749,
0.0345067047,
0.0268272236,
0.0309418049,
-0.0083363978,
0.0090669291,
-0.0398268774,
-0.019336449,
-0.0002894097,
-0.0006952088,
-0.0257962532,
0.0071441648,
0.0397769138,
-0.0030737258,
0.0142130358,
0.030249564,
-0.0094624264,
-0.0753455311,
-0.0247901008,
-0.0134593956,
0.0106043285,
-0.0423208363,
0.0202689748,
-0.0085138911,
-0.0152931465,
-0.0832923576,
0.0365173109,
0.0183845349,
0.0812311471,
-0.0308796838,
-0.0138759306,
-0.0634573624,
0.0561741628,
-0.0782877207,
-0.0557015017,
0.0030148441,
0.0880702138,
0.0295558758,
0.0518425219,
-0.0157146733,
-0.0647970662,
-0.0023990006,
-0.0244432855,
-0.0278665144,
0.0067397836,
-0.0338013992,
-0.0172796082,
0.055521328,
-0.0214694403,
-0.0172454175,
0.0026717996,
0.0058483114,
0.0496262722,
-0.0571822226,
-0.0130965868,
-0.0395572037,
-0.0208191276,
-0.0055188751,
0.0346540362,
-0.0125664668,
-0.0018173417,
-0.0586386733,
-0.0506022647,
0.0025388158,
0.0846015885,
0.0113094421,
0.0022924691,
-0.006330092,
0.0236611962,
-0.0354385562,
-0.0706640854,
0.0247639213,
-0.0451255068,
-0.0740271285,
0.0150199896,
-0.031133512,
-6.838493487e-33,
0.0020931768,
0.0086562,
0.0200993791,
-0.0293898024,
-0.0213533286,
-0.0293485913,
-0.0228944439,
0.0360874385,
-0.0221719667,
-0.0173313878,
0.0071420628,
-0.0033185876,
-0.0135617629,
-0.0296051856,
0.0334413201,
0.0152467545,
0.0122398688,
-0.0291940179,
-0.0322230123,
0.0257653091,
-0.0315855108,
-0.0286819842,
-0.0041691978,
-0.0723805502,
0.0022486381,
0.0033535238,
0.0188958906,
-0.0232518166,
-0.0650363639,
0.0374959745,
-0.0570098273,
0.0116439955,
0.0398471169,
-0.0094814077,
0.0174069796,
-0.0002057701,
-0.0345436744,
-0.0108623002,
-0.0074584549,
-0.0466431119,
-0.067314662,
-0.0702215433,
-0.0006296607,
0.0001545939,
-0.0508129783,
-0.0194478277,
0.0321224369,
-0.0329422839,
0.0514196306,
-0.0166995227,
-0.067646116,
0.0021513533,
-0.0211978909,
0.0176813155,
-0.024693815,
0.0109259505,
-0.0362613909,
-0.0011481657,
-0.0181429312,
0.0431313217,
-0.0237300768,
-0.0845745206,
0.0388256833,
0.0177782215,
0.0447290801,
0.0536199212,
0.0363797136,
0.0219992734,
-0.011319356,
-0.035043858,
0.0420980789,
0.0809510276,
0.0608653538,
0.0470639132,
-0.0021528974,
-0.0237449817,
-0.0302570853,
0.019916445,
0.0078288,
0.0108715603,
0.040794868,
-0.01227667,
-0.0558603555,
0.004172259,
0.0067459322,
-0.0167848319,
-0.026266126,
-0.0056301779,
0.0328067914,
-0.0308962241,
-0.0046702242,
-0.016903216,
0.0395686142,
-0.0432626642,
0.062913157,
-0.0085994788,
0.0209027119,
0.0573538952,
0.0197596811,
-0.0139167458,
-0.0065416754,
-0.0431287959,
0.006060685,
0.0535135381,
0.0573869385,
-0.0187760554,
-0.0191283319,
0.0248418618,
0.0220786892,
-0.0184565913,
-0.0194473099,
-0.0108701233,
0.0741536766,
-0.0441703871,
0.0368166156,
0.0160371438,
0.0231097564,
-0.0659274906,
-0.00446021,
0.0172211379,
-0.0366652161,
0.0428102016,
-0.0515423678,
-0.0050255684,
-0.0019603639,
-0.0107513582,
-0.0707248449,
-0.0039423169,
-0.0292673148,
-0.0513304472,
-0.0035234066,
0.0146177234,
3.196e-7,
-0.0358842537,
0.0689522848,
-0.0236198865,
-0.0038917502,
0.014172256,
0.013524116,
-0.0139026996,
0.0036190515,
0.0400625393,
-0.0015729866,
0.0308527201,
0.0572616644,
0.0442302376,
-0.0008188376,
-0.0351322852,
-0.0461727455,
0.0221004393,
-0.0142355384,
-0.0476551875,
0.0164312422,
0.0089998618,
0.0313079618,
0.0313770473,
0.0096750623,
0.0388412774,
-0.0492603183,
0.0044601196,
-0.0447030403,
0.0223425701,
-0.0060545136,
-0.0152373752,
0.0368641205,
0.0032937005,
0.0762747303,
0.0270976741,
-0.0028820774,
-0.0188773032,
0.0118490309,
0.0189534649,
0.0296113584,
0.000308025,
-0.0040668556,
0.0346270464,
-0.0472783744,
0.0838587731,
-0.0737183467,
-0.033989545,
-0.0046861828,
-0.019808935,
-0.0055562467,
0.0011818791,
0.0112948017,
0.0094109233,
0.0098620681,
-0.0105401287,
-0.0189787522,
0.0113777183,
-0.0179240219,
0.0252507776,
0.0289970674,
-0.0652109608,
-0.0852023661,
-0.037742164,
-0.0257367995,
0.0323244035,
-0.0664487332,
-0.0083877407,
2.909064747e-34,
0.0071093128,
0.0257342029,
0.0324384086,
-0.0588431247,
0.0825359821,
-0.0161906779,
-0.034675248,
-0.0200130381,
0.0205723513,
-0.0643945783,
0.0170580372
] |
http://arxiv.org/abs/2402.02541v1 | http://arxiv.org/pdf/2402.02541v1 | Knowledge Generation for Zero-shot Knowledge-based VQA | 2024-02-04T15:41:35 | 2024-02-04T15:41:35 | Rui Cao, Jing Jiang | cs.CL | cs.CL, cs.AI, cs.CV | Previous solutions to knowledge-based visual question answering~(K-VQA)
retrieve knowledge from external knowledge bases and use supervised learning to
train the K-VQA model. Recently pre-trained LLMs have been used as both a
knowledge source and a zero-shot QA model for K-VQA and demonstrated promising
results. However, these recent methods do not explicitly show the knowledge
needed to answer the questions and thus lack interpretability. Inspired by
recent work on knowledge generation from LLMs for text-based QA, in this work
we propose and test a similar knowledge-generation-based K-VQA method, which
first generates knowledge from an LLM and then incorporates the generated
knowledge for K-VQA in a zero-shot manner. We evaluate our method on two K-VQA
benchmarks and found that our method performs better than previous zero-shot
K-VQA methods and our generated knowledge is generally relevant and helpful. | Knowledge Generation for Zero-shot Knowledge-based VQA: Previous solutions to knowledge-based visual question answering~(K-VQA)
retrieve knowledge from external knowledge bases and use supervised learning to
train the K-VQA model. Recently pre-trained LLMs have been used as both a
knowledge source and a zero-shot QA model for K-VQA and demonstrated promising
results. However, these recent methods do not explicitly show the knowledge
needed to answer the questions and thus lack interpretability. Inspired by
recent work on knowledge generation from LLMs for text-based QA, in this work
we propose and test a similar knowledge-generation-based K-VQA method, which
first generates knowledge from an LLM and then incorporates the generated
knowledge for K-VQA in a zero-shot manner. We evaluate our method on two K-VQA
benchmarks and found that our method performs better than previous zero-shot
K-VQA methods and our generated knowledge is generally relevant and helpful. | [
0.0209809281,
0.0032313976,
-0.0101334434,
0.0634522885,
0.0064949919,
-0.0151217515,
0.0034058071,
0.0389070176,
0.0100361174,
-0.0629924387,
-0.0235225018,
0.005464775,
-0.0465852469,
0.0346566774,
0.0613256842,
-0.0096827177,
0.0324947834,
-0.0214874484,
-0.0202110484,
-0.0161401685,
0.0126929907,
-0.0149530852,
-0.0191230401,
0.0084598064,
-0.0556543656,
-0.0376326069,
-0.0397044197,
-0.0020372653,
0.0193296913,
-0.0355343185,
-0.0372857377,
0.0188359879,
0.0345319696,
-0.0239492152,
0.0000019834,
-0.0058553256,
0.0059915739,
-0.0010731264,
-0.0430721082,
-0.0263717156,
-0.0488411561,
-0.0212859921,
0.0295087565,
0.0232752189,
-0.0583430082,
-0.0049377889,
0.0382195264,
0.0179053228,
0.0546819419,
0.038476903,
-0.0177551135,
0.00741988,
0.0044935998,
0.0204305612,
0.0641372725,
-0.0539911091,
-0.0025462529,
0.0312583484,
0.018272765,
-0.0223476011,
-0.0481015556,
0.0499167852,
0.0210971311,
-0.0014324537,
0.017480677,
0.0534614474,
-0.0330465361,
-0.0039368002,
-0.0170605779,
0.1251903921,
0.0118249264,
0.0114466818,
-0.0046018912,
0.0039499975,
-0.0041912314,
0.0277271196,
-0.0602494814,
0.058347512,
0.0002283006,
-0.0134849157,
-0.0665890947,
0.034222573,
-0.0146014281,
-0.0298938099,
-0.0007736455,
-0.0089858435,
0.0047993441,
0.0010584417,
-0.0176351853,
0.0097371396,
-0.0485305376,
-0.0677976236,
-0.090127781,
0.0931898579,
0.0320162512,
0.0223144181,
-0.0008097755,
-0.0139123285,
0.0101214219,
-0.0297236741,
0.0548241176,
0.031954024,
-0.0407685637,
0.0405167118,
-0.0473473333,
0.0580052994,
-0.0075097042,
0.0088640125,
-0.042389933,
-0.0139952144,
-0.0553165935,
-0.0199404154,
-0.0247394182,
0.0481091291,
-0.0276372377,
-0.0429833494,
-0.0114921834,
0.0211489648,
-0.0106662307,
0.0023195937,
0.0021826946,
0.0479692109,
0.0170389544,
0.0148282181,
-0.0782269314,
0.0480646454,
-0.0532778911,
0.0184465051,
-0.0494084321,
-0.0319579504,
-0.0062583541,
0.0355975591,
0.0427112058,
0.0102431355,
-0.003312062,
0.0735046938,
-0.0396189354,
-0.0063199955,
-0.0441565774,
-0.0441628247,
-0.0123126069,
-0.0298747364,
0.0533749647,
-0.045794744,
0.009869176,
-0.0105799157,
-0.024217803,
0.0057003056,
0.0272482876,
0.0345566124,
-0.0304470398,
0.0732093602,
-0.0669998974,
-0.0115927281,
-0.0005908579,
0.0192876589,
0.015002179,
0.0820805952,
-0.0026391107,
0.0067464067,
0.0318440013,
0.0179072265,
0.0463248789,
0.0208499264,
0.0297902878,
0.013867477,
-0.0418759026,
0.0299235713,
0.0042652399,
0.0080931494,
0.0090662595,
0.0081614899,
-0.0612669252,
-0.0277948081,
0.0061138752,
0.0386878252,
0.0454141162,
-0.0104115605,
0.0268209949,
0.0015073395,
0.0239655338,
-0.0082129976,
-0.0000201,
0.0889541432,
-0.04849023,
0.0162707008,
-0.0190879963,
0.0006444668,
-0.0643185377,
-0.0415997505,
0.0004558501,
-0.0138508845,
-0.0165932439,
0.0218695328,
0.0404112451,
0.0518023483,
-0.0365142114,
0.008078997,
-0.0026712534,
-0.0213691816,
-0.0256538037,
-0.0213964805,
0.1069300175,
0.059182629,
0.0302628204,
0.0059942342,
-0.0514138676,
0.0127074756,
0.0044395174,
0.0233182646,
-0.0567775257,
-0.0108383866,
-0.0076925172,
-0.0499674529,
-0.0492723472,
-0.0570391715,
-0.0067034019,
-0.0097737284,
0.0168338437,
0.0377839915,
0.0036665678,
-0.0344422907,
0.0603288896,
0.0115871243,
0.004378581,
-0.0275054835,
0.0181449652,
0.0152240153,
-0.0282539725,
0.0228969678,
0.004978789,
-0.0013449679,
0.0185609255,
-0.0297613367,
-0.0131683275,
-0.0082937069,
-0.0118680233,
-0.0777490437,
0.0312312692,
-0.0517451726,
0.0186442062,
-0.011003335,
-0.0440061055,
-0.0320197493,
0.0213257112,
0.0214775987,
-0.0223722756,
-0.049592521,
-0.0574986599,
0.0530302823,
-0.008482025,
-0.0064849323,
0.0765695274,
-0.0097064916,
-0.0271657184,
0.0070465519,
0.0049216514,
0.0494115688,
-0.0273825936,
0.0150586711,
-0.0131427711,
0.0089099733,
0.015497636,
-0.0359605737,
0.0277986545,
-0.0582173616,
0.0151679041,
-0.0982867405,
0.0077354587,
-0.0083005745,
-0.019312432,
0.032139346,
-0.0211774278,
0.0026404052,
-0.0322378092,
-0.0710754767,
0.0587469302,
0.0099374084,
-0.0201707594,
0.0562293567,
0.0343777649,
0.026758058,
-0.0159862936,
0.0113520641,
0.0193779618,
0.0825016722,
0.0207310282,
0.0224537663,
-0.0024644202,
0.0221943166,
-0.0952876434,
0.0342864022,
0.0081553916,
-0.0206655692,
0.0296571627,
-0.0205089077,
-0.0046182168,
0.0277683344,
-0.0094857831,
-0.002997936,
0.0128060225,
-0.0151040833,
-0.0278813187,
0.0201724544,
0.0085662585,
-0.0484744161,
0.0304883588,
-0.0165300854,
-0.0373167247,
0.015918199,
-0.0261191502,
-0.024506893,
-0.0812587067,
0.0802725926,
0.054961849,
-0.0244807452,
-0.069826819,
0.0448283367,
0.0241409838,
0.0647002384,
-0.0269659646,
-0.0267626997,
0.0074658371,
-0.0107275313,
0.0158111732,
0.0431919284,
0.0141959591,
-0.0436480679,
0.059780661,
-0.0148980105,
0.0193396062,
-0.0167222265,
0.031273026,
-0.0433324054,
-0.035643518,
0.0004494764,
-0.0058934605,
0.0359688364,
-0.03841234,
0.0405536965,
-0.0249395017,
0.0114559149,
-0.0701386854,
0.002394571,
-0.0335228555,
0.03012117,
-0.0701384619,
-0.0052319956,
0.0335078761,
0.0093915043,
0.0323882997,
-0.0232554376,
-0.0141292932,
-0.0005697419,
-0.0147702526,
0.0645419806,
0.0731038302,
0.0273764059,
-0.0271408372,
-0.0081816493,
0.0355674066,
0.0677760392,
0.0119004184,
0.0132567063,
-0.0052157566,
0.0067254393,
-0.0109730428,
0.0156704132,
0.0017109218,
0.0350222662,
0.0114722773,
-0.0183149166,
0.0089924494,
0.0660153106,
0.0333337486,
0.0083485274,
-0.0067109936,
-0.0331803896,
-0.0128926514,
0.0376381576,
0.0605838858,
-0.1586944014,
-0.0421820991,
-0.048332572,
0.01163569,
0.004782625,
-0.0192600135,
-0.0229245014,
-0.096175842,
0.0603255741,
-0.0108927358,
-0.0397077873,
0.0045996178,
0.0237468947,
-0.0166155212,
0.0074060531,
-0.0328544267,
0.010985923,
-0.0156717729,
0.0341765732,
-0.0054878294,
0.0167393573,
-0.0204503834,
0.0387022831,
-0.03398351,
0.002937105,
0.0288061276,
0.0239108931,
-0.0127245737,
-0.0720793307,
-0.0577947982,
-0.0485986806,
0.0375288725,
-0.0080079148,
0.0524354242,
-0.0100592803,
-0.0480196215,
0.01559838,
0.0148914633,
0.0011964263,
0.0010225779,
0.0263117105,
0.0354780704,
-0.0167201068,
-0.0259965751,
-0.0058182729,
0.0148711093,
-0.0067801424,
0.0382868908,
0.0341596529,
0.0065905578,
-0.0438173376,
-0.0059545957,
-0.0081662647,
-0.0625407621,
-0.0734172389,
0.0048191831,
0.0279398635,
-0.011587671,
-0.1032064259,
-0.0177324284,
0.0826950148,
0.0887004212,
-0.0381668136,
-0.0111132292,
-0.0205067415,
0.0190656055,
-0.0404131822,
0.0129743442,
-0.0365473181,
0.0152871106,
-0.0323180817,
0.0022431437,
0.0209226776,
-0.059694428,
-0.0186808314,
-0.043697346,
0.0023648113,
-0.0152123645,
-0.054465007,
0.1012479886,
0.00467883,
0.0195217393,
0.0318808891,
-0.0140165724,
0.0768675134,
0.0411630683,
0.0092795817,
0.0328006446,
-0.0077766296,
0.0176733248,
0.0055731242,
0.0030641663,
-0.0536419377,
0.002380515,
-0.0098859202,
0.0546183102,
-0.0689430162,
0.0043334211,
0.0020133429,
0.0201082099,
-0.0235537048,
0.0024292639,
0.0041198423,
0.0082290592,
0.0186858177,
-0.0266506262,
0.0049952194,
-0.094155252,
0.0077180648,
-0.0059104348,
-0.0222969763,
0.0159922596,
-0.0350513384,
-0.0747421235,
0.0309877098,
-0.0054284935,
-0.0255063586,
0.0536296293,
0.0108076073,
-0.0035269365,
0.0012069342,
0.0076627294,
0.0059100953,
0.0013027984,
0.0599055439,
0.0216455571,
0.0731041282,
0.018372966,
-0.0028375387,
-0.0597530566,
0.0268347338,
-0.006324023,
-0.0694880635,
0.0146410074,
0.0697214007,
0.0723811686,
0.0356901847,
-0.0279321186,
0.0153074786,
-0.0122341216,
0.0306156538,
-0.007251041,
-0.0217246059,
-0.0395448022,
0.0790845901,
0.0277231429,
0.0111064892,
-0.0279391259,
0.0974219516,
0.0084853861,
0.0285278298,
-0.0558012128,
0.0128002642,
-0.0238528308,
0.0073870597,
0.048078578,
0.0019993046,
0.0364366807,
0.0459928177,
-0.0519321561,
-0.0664682165,
0.0444430746,
0.004455816,
-0.0485060476,
0.0100039504,
0.0496146418,
-0.0071398909,
0.0648768991,
-0.0116808843,
-0.0163938105,
-0.004672267,
-0.0350626968,
0.0067766886,
0.0333175585,
-6.344301703e-33,
-0.0436788276,
-0.0787472948,
0.0861474648,
-0.0244079437,
-0.0012621128,
0.0223128777,
-0.0100859078,
-0.0108630154,
-0.0104694646,
-0.0103116343,
0.0158921629,
-0.0484698415,
-0.0118513936,
0.0228260364,
-0.0259136427,
-0.0485249795,
-0.0149320671,
-0.0205086693,
0.0128349755,
-0.001175547,
0.0037627418,
0.0107886903,
0.0124810394,
-0.0054249666,
-0.0172644556,
-0.0191568974,
0.0019799548,
0.0209885705,
0.0168897677,
0.0606902763,
-0.0256951712,
0.0077384147,
0.0161987934,
-0.0051079178,
0.017487945,
-0.022403948,
-0.0501262397,
-0.0623662621,
0.0388858095,
-0.0162336305,
-0.0885578841,
-0.0011848005,
0.0116565507,
-0.007472123,
0.035156779,
-0.029223185,
0.0408208221,
-0.0036867254,
0.0023522552,
-0.0395382047,
-0.0177876242,
-0.0106090736,
-0.0125309611,
-0.0493169911,
0.0178024303,
0.0494689681,
0.0350245908,
-0.0089454968,
-0.0651108176,
0.0152468868,
0.0152430236,
-0.0229254197,
0.087046802,
0.0151948575,
0.0172591414,
0.0257116742,
-0.0010882807,
-0.0389400832,
-0.092019558,
0.0050783316,
-0.0146820955,
0.0563628115,
0.0482946411,
0.029248381,
0.0031758901,
0.0219413564,
-0.0596649572,
0.0166527573,
-0.0737885386,
0.0539759062,
-0.0350366421,
-0.0400707424,
0.0152562214,
-0.0124050425,
-0.042756021,
0.0187959429,
-0.0446124785,
-0.0087577756,
0.051439438,
0.0305265263,
-0.0076858336,
0.0154099492,
0.0368162803,
-0.0223952178,
0.03848828,
0.0256052669,
-0.0116818529,
0.050775487,
0.0011680669,
0.0033532104,
-0.0580097213,
0.0439596586,
0.07539244,
-0.0192526057,
-0.0130304061,
0.0490693264,
0.0270314589,
0.0061717289,
-0.023460269,
-0.0229234714,
-0.0251213498,
0.0241162814,
-0.0147134392,
-0.033220049,
-0.0322454646,
0.0314034671,
-0.0423308052,
-0.0063080583,
-0.0109993741,
-0.0019440292,
0.0380282849,
-0.0419698767,
-0.0081377234,
0.0301799141,
-0.0497933701,
0.0241096895,
-0.0141910538,
-0.0066098138,
-0.0236532874,
-0.0610953793,
-0.0034010075,
0.0600324646,
2.54e-7,
0.0362492688,
-0.0211951509,
-0.0151618142,
0.0379507504,
-0.0019150805,
0.0394382998,
-0.0170742944,
-0.0068965829,
0.0012875071,
-0.0355579033,
0.0120835258,
0.0940291882,
-0.0109381368,
0.0081599606,
-0.0402257629,
0.0422071107,
0.0180527773,
-0.0455451161,
-0.0503834635,
-0.0034572298,
0.078143239,
-0.000188415,
-0.0228039976,
-0.0074992976,
0.0436234027,
-0.0309474822,
0.08492852,
0.0439005792,
0.0064001605,
-0.0353152491,
0.0492647849,
-0.0081940601,
-0.0169069655,
0.0268683024,
0.0216006059,
-0.0036256013,
0.0039578546,
0.0140499612,
-0.0285904445,
-0.0419123657,
-0.0036206199,
0.0064187991,
0.0094996011,
0.0068979817,
0.1176488325,
-0.0259338245,
-0.0214858055,
0.0104997009,
-0.0533270575,
0.0016079143,
0.0114542395,
0.0079631833,
-0.0098397518,
0.0187847316,
0.0392219462,
-0.031462349,
-0.0167756677,
-0.0331053697,
0.0234487392,
-0.0288180374,
-0.0027800587,
-0.002209235,
-0.0172814988,
-0.0240982175,
0.0125830034,
-0.0019588091,
-0.047939416,
1.62976265e-34,
0.0107893888,
-0.0016312769,
0.0091973683,
0.0355031528,
-0.0024217383,
-0.0189011712,
-0.0179434158,
-0.0282430984,
0.0316201411,
-0.0528599285,
-0.0132430149
] |
http://arxiv.org/abs/2402.02522v1 | http://arxiv.org/pdf/2402.02522v1 | Absolute convergence and error thresholds in non-active adaptive sampling | 2024-02-04T15:10:34 | 2024-02-04T15:10:34 | Manuel Vilares Ferro, Victor M. Darriba Bilbao, Jesús Vilares Ferro | cs.CL | cs.CL, cs.AI, cs.LG | Non-active adaptive sampling is a way of building machine learning models
from a training data base which are supposed to dynamically and automatically
derive guaranteed sample size. In this context and regardless of the strategy
used in both scheduling and generating of weak predictors, a proposal for
calculating absolute convergence and error thresholds is described. We not only
make it possible to establish when the quality of the model no longer
increases, but also supplies a proximity condition to estimate in absolute
terms how close it is to achieving such a goal, thus supporting decision making
for fine-tuning learning parameters in model selection. The technique proves
its correctness and completeness with respect to our working hypotheses, in
addition to strengthening the robustness of the sampling scheme. Tests meet our
expectations and illustrate the proposal in the domain of natural language
processing, taking the generation of part-of-speech taggers as case study. | Absolute convergence and error thresholds in non-active adaptive sampling: Non-active adaptive sampling is a way of building machine learning models
from a training data base which are supposed to dynamically and automatically
derive guaranteed sample size. In this context and regardless of the strategy
used in both scheduling and generating of weak predictors, a proposal for
calculating absolute convergence and error thresholds is described. We not only
make it possible to establish when the quality of the model no longer
increases, but also supplies a proximity condition to estimate in absolute
terms how close it is to achieving such a goal, thus supporting decision making
for fine-tuning learning parameters in model selection. The technique proves
its correctness and completeness with respect to our working hypotheses, in
addition to strengthening the robustness of the sampling scheme. Tests meet our
expectations and illustrate the proposal in the domain of natural language
processing, taking the generation of part-of-speech taggers as case study. | [
-0.0086190999,
0.0357789658,
-0.0188675337,
-0.0334570482,
-0.0468722321,
0.0246782992,
-0.0027233625,
0.0159822386,
0.0051235845,
-0.0072674612,
0.0413894951,
-0.0585207976,
-0.0325195901,
0.0048986468,
0.0151201403,
-0.0165224057,
0.0095349532,
-0.0447372459,
-0.0036676258,
-0.0258690584,
0.0032783106,
-0.0081299869,
0.0042708456,
0.0449100696,
0.0240539238,
0.0294623859,
-0.0087763397,
-0.0262890495,
-0.0295564011,
-0.0181538053,
0.0019144438,
0.0284357797,
-0.0220461376,
0.0010936598,
0.0000018215,
-0.0553875417,
0.0038941607,
0.0043694372,
-0.0194645636,
-0.0261809044,
0.0485900529,
0.0183430687,
0.0151726613,
0.0385375097,
-0.0187526122,
-0.0140075739,
0.0212274604,
-0.0281016659,
0.0505213849,
0.0083337305,
0.0083356937,
0.0243604239,
0.0191571135,
-0.034584552,
-0.0235400256,
-0.0748574361,
0.0329090506,
-0.018701341,
0.0094999997,
-0.0350774862,
0.009909777,
0.0329678506,
-0.0042936779,
0.0015347077,
0.017466763,
0.0268800464,
-0.0198381171,
-0.0490873083,
0.0058032591,
-0.0063428334,
-0.0407166146,
-0.0215711407,
0.0048638182,
-0.0457823351,
0.0074717291,
0.0452775136,
-0.0450381674,
0.0029651872,
-0.0307510588,
0.047381144,
-0.0016058624,
-0.0410947315,
0.0218421314,
0.0150675317,
0.0020151699,
0.0526715182,
-0.0257114973,
-0.0474136136,
0.0494470932,
0.0471680351,
-0.0430714786,
0.0055372613,
0.0086770449,
0.0346788168,
-0.025917327,
0.0027391352,
-0.0512053333,
-0.0105063543,
0.0063491971,
-0.038550172,
0.0744010061,
-0.032997068,
0.0537262447,
0.0408634804,
0.0068219863,
-0.0095565971,
0.0190662425,
-0.0189135876,
-0.0230140667,
-0.0003322928,
-0.028687777,
0.0270160642,
-0.0446403809,
0.0355984382,
0.0001680596,
0.0340566151,
-0.0664999336,
0.026265312,
-0.0077032153,
-0.0175005775,
0.0214465763,
-0.0212146286,
-0.0837773234,
0.0606697723,
-0.0132742934,
-0.0232233014,
-0.0194116142,
0.013996751,
-0.0094990954,
-0.0446414538,
0.0064568599,
-0.0189312212,
-0.0141289365,
-0.0249225143,
-0.0055731642,
0.0507937036,
0.0011441326,
-0.0141743431,
-0.0812196061,
-0.0271482468,
-0.0185210947,
-0.0070274938,
0.0028963271,
-0.028144246,
-0.007770061,
-0.0003895336,
0.0370367728,
-0.0221241172,
-0.0289548729,
0.0039572623,
-0.0326055251,
0.0522335768,
-0.0162658766,
0.0310959052,
-0.0141454078,
-0.0480285101,
-0.0014691338,
0.1043854356,
0.0684967935,
0.0125167891,
0.0624305233,
0.0197555516,
-0.0192396566,
0.0250477456,
0.0132379625,
-0.0141361607,
-0.0368750729,
-0.0501580052,
-0.0182461236,
-0.0116062621,
-0.0316166803,
0.0062051709,
-0.0318242721,
-0.0088140955,
0.0376544334,
0.0076585752,
0.0185804274,
0.0408214219,
0.0472339205,
-0.0129720094,
-0.0081576426,
0.0359760672,
0.0102963988,
-0.0501033626,
-0.0108084716,
0.0100817177,
-0.0272376016,
0.0287884809,
-0.0308434702,
-0.0517266057,
0.0103071025,
-0.0014185907,
-0.0129825696,
-0.013707418,
0.0816665813,
0.0158287156,
-0.0209751576,
0.0101698609,
0.0376069322,
0.0021872274,
0.0427020378,
-0.0714014471,
0.0015405163,
-0.0304561965,
0.0307909604,
-0.0652655736,
-0.0001263005,
-0.0286316033,
-0.002692678,
-0.0389244929,
0.0455968194,
-0.0203422401,
0.0331073813,
-0.0116025498,
-0.0518786535,
0.0854189396,
0.0007064315,
-0.0303617343,
-0.0282657761,
0.0566647016,
-0.0230016448,
-0.0616610721,
-0.0875198692,
0.0359950177,
-0.0578482486,
-0.0365440249,
0.0302315392,
-0.0262176879,
0.0795523599,
0.0085282493,
-0.0506477952,
-0.0180378482,
0.0486641675,
-0.0213703457,
-0.0201514643,
0.010812277,
-0.0148647306,
-0.0097983982,
0.0329702087,
-0.0225634929,
0.026387414,
-0.013171332,
-0.0356577747,
-0.0258306414,
0.0100561343,
-0.0460079238,
-0.0165675506,
-0.0672743022,
0.0566840693,
-0.0543239191,
0.0464963205,
-0.0631301031,
0.0276710708,
-0.0125933886,
0.0076837461,
0.0127991717,
-0.0339885727,
-0.0217934977,
-0.0140608931,
-0.0674781576,
-0.0232607946,
-0.0189898554,
0.0044205775,
0.0440236926,
0.0097937183,
-0.008866176,
-0.0181737058,
-0.0162869487,
-0.0073475726,
-0.0398104973,
0.0128814513,
-0.0344098061,
-0.0061593666,
0.0266729053,
-0.0112703508,
0.0310093425,
0.0073006074,
0.009458025,
0.0340779684,
0.030187102,
-0.0005182982,
-0.0245296136,
-0.0105575258,
-0.0421070568,
0.001157842,
0.0818956792,
0.0646771267,
-0.0235564038,
-0.0439709164,
-0.0387014188,
0.033195734,
0.0018186586,
0.0368415006,
0.0205087028,
-0.0004111646,
-0.0069428338,
-0.0475825891,
0.078181535,
0.0221041795,
0.0574188121,
0.0075987647,
0.052652631,
-0.0327436179,
0.0199265592,
-0.0653112084,
-0.0196867939,
-0.0229617786,
-0.0215336755,
0.0519081429,
0.0994838104,
-0.0177192483,
-0.0079209879,
-0.0163946077,
0.0269397423,
-0.0120602613,
-0.067681253,
-0.0012448811,
0.0495584309,
-0.0569317602,
0.0164648406,
-0.0091919657,
-0.042934414,
0.0223538671,
-0.0058177253,
0.0168475676,
-0.031136591,
-0.014157895,
0.0014253573,
-0.0384564847,
0.0519449562,
0.005486398,
0.0453872979,
0.0148297576,
-0.009954975,
-0.0549328066,
0.0459436141,
-0.0102737695,
0.0018588236,
-0.0276461281,
0.0202562939,
0.0254716631,
-0.0136646712,
-0.0299509447,
0.012838034,
-0.0068621957,
0.0320953317,
-0.0181892775,
0.0156816114,
-0.010667867,
0.0526757948,
-0.0235927962,
0.0488063693,
-0.0172540788,
-0.0333617665,
-0.0006630772,
0.0322076976,
0.0460121296,
0.0233523827,
-0.0262807421,
-0.0534844436,
0.0223602094,
-0.0065958491,
0.0779513046,
0.0478394069,
-0.0423796847,
0.021914253,
0.0221445765,
-0.059303049,
-0.017708268,
0.0182620138,
-0.0196136776,
-0.0095950384,
-0.0088333338,
0.0886972174,
0.0073244805,
-0.0142638488,
0.0310213789,
0.0279075094,
0.0433634147,
0.0376047194,
0.0121198092,
0.0041707223,
-0.0176586341,
0.003545196,
0.1112925783,
-0.0431393757,
-0.0514961779,
0.0207509995,
-0.0736645162,
0.0573415495,
-0.0259755719,
-0.065275684,
0.0221039485,
-0.034959197,
0.0168939382,
-0.065146789,
0.0008698057,
0.0013468332,
-0.0242575556,
0.0194415003,
0.0691610649,
-0.0390046053,
-0.0164174903,
-0.0566899851,
-0.0813465491,
0.0486197881,
0.0347015001,
0.0475785099,
0.0000451998,
-0.0057761096,
0.0326595046,
-0.0544143617,
-0.0050279717,
-0.0237250049,
0.011484609,
0.0564506277,
-0.0186228231,
0.0032183391,
0.0084759835,
0.0303245634,
0.0387219265,
0.0236418601,
0.1018521413,
0.0040316591,
-0.0417319648,
-0.002468409,
0.00259862,
0.0117040686,
0.0105952434,
0.0258103739,
-0.0602548495,
-0.0250455495,
0.0258738622,
0.0359731279,
-0.0695937052,
-0.0349372402,
0.0539107062,
-0.0918254554,
-0.0201874189,
-0.0576307587,
0.0050868564,
0.0194958318,
0.1294172108,
0.0099461982,
0.0641477108,
0.0735563934,
0.0195272174,
-0.0146084838,
0.0012311218,
-0.0028018944,
-0.028572116,
0.0001191958,
-0.0420098417,
0.02545847,
-0.0303131863,
-0.0342504233,
-0.0001987561,
-0.0287460871,
-0.0554547533,
0.0675858557,
0.0352155268,
0.0313452668,
0.0364896208,
0.0017090713,
-0.1115327626,
0.0708649755,
0.0287167206,
0.072738491,
0.0154426564,
0.0271006152,
-0.0143510653,
-0.0355287232,
0.0400565118,
-0.0526399761,
0.0127823828,
0.0120981792,
0.0061875777,
-0.0401495323,
-0.0404574163,
0.008172038,
0.056365177,
0.0060638287,
-0.0397963189,
-0.0719534233,
-0.0142853418,
0.0112015,
0.109898001,
0.0280682743,
-0.0721507072,
0.0370596163,
-0.0013213633,
0.0031652811,
-0.0317966379,
-0.0098992111,
0.020351924,
0.0284877513,
-0.0465937108,
0.0111634629,
-0.0048886524,
0.0335853249,
-0.0018572055,
0.0004978623,
0.0041808682,
0.0170376711,
0.0254102107,
0.0220913701,
0.0106607927,
0.0387459733,
0.0047534998,
0.0940836444,
-0.0161277466,
-0.0065205535,
-0.0509786122,
-0.0245612394,
-0.0767307654,
0.0623935796,
0.0164662339,
-0.0119295381,
0.042992834,
-0.0087982174,
0.0160423033,
-0.0183325354,
0.0256712362,
0.0135869756,
-0.0222109873,
0.0101246275,
0.0136695104,
-0.016099656,
0.0365826823,
0.050746236,
0.0045732944,
0.0487369597,
-0.0241385251,
-0.004662245,
-0.0358533189,
-0.0254937783,
0.0060956175,
-0.0537068471,
-0.0110844839,
0.0261991266,
-0.024718063,
-0.0594406649,
0.0248095002,
-0.0161513612,
0.0244044978,
-0.0348665565,
0.0183422808,
0.0413217768,
0.0015497856,
0.0552877188,
-0.0321504287,
0.0037680594,
0.0129758539,
0.0208821427,
0.0096789487,
-5.879691602e-33,
-0.0142967347,
-0.006401631,
-0.0431926921,
0.0287898686,
-0.0029707884,
-0.0094386414,
0.0170577746,
-0.0217380412,
0.0304185487,
-0.0072007733,
-0.038621515,
-0.0463506021,
0.0259781498,
0.0356467105,
0.0101601463,
-0.0204089731,
0.0612634681,
-0.0411500148,
-0.0118820816,
0.0324882232,
0.0285205357,
0.0383163765,
0.0001559343,
-0.0407054611,
-0.0271544755,
0.0244731754,
-0.0337433107,
0.0177447796,
0.0258560423,
0.0529290251,
0.0374507345,
0.041372072,
0.0098551521,
-0.0704619437,
0.0092618428,
-0.03896375,
-0.0710215122,
-0.0270306412,
-0.0298776124,
0.0344447978,
-0.0186073147,
0.039125964,
0.0380218923,
0.0041655838,
-0.0436576828,
0.0372566059,
0.0342959315,
-0.0339165963,
-0.0674047992,
-0.0685015917,
0.0025800394,
0.024036739,
-0.0265431181,
0.0360328853,
0.007501374,
0.0234015416,
-0.0206652693,
0.0036814683,
0.0121471239,
-0.0275645442,
0.0627921894,
0.0202738997,
0.0246741455,
0.0539884381,
0.0172975324,
0.0453406759,
0.0393745638,
0.0030478295,
0.0894128531,
0.0137574868,
0.0107650887,
0.005433213,
0.000672069,
0.023911126,
0.036093723,
0.0178390648,
-0.0208119527,
0.0082831671,
0.1151414216,
0.0063898969,
0.052825734,
-0.0412812382,
-0.0369679928,
-0.0156817287,
-0.0298386235,
-0.0505463146,
-0.017616991,
0.003942803,
0.0207605083,
0.0431005843,
-0.094468154,
0.04005621,
0.0115929917,
-0.0314217955,
0.011138632,
-0.0648650825,
0.0203711446,
-0.0020617405,
0.0151551208,
-0.0090809995,
-0.0148073137,
-0.0804769918,
-0.0314197838,
-0.0045167105,
0.0320210159,
-0.0176156517,
0.0056996299,
0.020326674,
-0.0192932636,
0.0064336434,
-0.0118776271,
0.0662692413,
0.0358950533,
0.0453766733,
0.0606169887,
-0.0362476557,
-0.0149318967,
0.0353712216,
0.0326370783,
-0.0441798717,
-0.0202351362,
-0.0083786007,
0.0307078846,
0.0218409933,
-0.0495573543,
-0.0269204043,
0.0115740765,
-0.0209172219,
-0.0144154457,
-0.0622767843,
-0.0334572196,
0.0175385438,
2.642e-7,
0.0027211388,
0.0148452101,
-0.0297677722,
-0.012148344,
0.003041107,
0.0193648636,
-0.016216781,
0.0200888533,
-0.0040291771,
0.0588686615,
-0.0182771105,
-0.0205043629,
0.0290446058,
0.0513522439,
-0.0806071237,
-0.0306039844,
-0.0360033698,
0.0730149299,
-0.008768863,
0.0072337692,
-0.0189837702,
0.0498083606,
-0.0154792853,
-0.060355667,
0.0012505648,
-0.1254091561,
-0.020540081,
0.0182226151,
0.0183471832,
0.0348369144,
0.0632007346,
0.0065307855,
0.008141364,
0.0322056897,
-0.0298877098,
0.0163140167,
-0.0053730756,
0.032150887,
-0.0116395717,
0.0043143868,
0.0537943691,
-0.0165641885,
0.0438827835,
-0.0009763796,
0.0917207375,
-0.0465868376,
0.0106178457,
0.0032867838,
0.0131769627,
-0.047344245,
-0.0034955307,
-0.0544050746,
0.0257274676,
0.0157868173,
0.0166442785,
-0.0200104173,
-0.0406922735,
-0.0012774391,
0.0341181755,
-0.0260918457,
-0.0325622857,
-0.0779411122,
-0.0229649488,
0.0206894521,
0.0162576567,
-0.0460661352,
0.0324846096,
2.191778323e-34,
0.0626522228,
-0.0082088327,
0.0065952754,
0.047101073,
-0.0463085398,
-0.0138454065,
0.0196352005,
0.0032542702,
0.006363899,
-0.0503306538,
-0.0294744018
] |
http://arxiv.org/abs/2402.02516v1 | http://arxiv.org/pdf/2402.02516v1 | Adaptive scheduling for adaptive sampling in POS taggers construction | 2024-02-04T15:02:17 | 2024-02-04T15:02:17 | Manuel Vilares Ferro, Victor M. Darriba Bilbao, Jesús Vilares Ferro | cs.CL | cs.CL, cs.AI, cs.LG | We introduce an adaptive scheduling for adaptive sampling as a novel way of
machine learning in the construction of part-of-speech taggers. The goal is to
speed up the training on large data sets, without significant loss of
performance with regard to an optimal configuration. In contrast to previous
methods using a random, fixed or regularly rising spacing between the
instances, ours analyzes the shape of the learning curve geometrically in
conjunction with a functional model to increase or decrease it at any time. The
algorithm proves to be formally correct regarding our working hypotheses.
Namely, given a case, the following one is the nearest ensuring a net gain of
learning ability from the former, it being possible to modulate the level of
requirement for this condition. We also improve the robustness of sampling by
paying greater attention to those regions of the training data base subject to
a temporary inflation in performance, thus preventing the learning from
stopping prematurely.
The proposal has been evaluated on the basis of its reliability to identify
the convergence of models, corroborating our expectations. While a concrete
halting condition is used for testing, users can choose any condition
whatsoever to suit their own specific needs. | Adaptive scheduling for adaptive sampling in POS taggers construction: We introduce an adaptive scheduling for adaptive sampling as a novel way of
machine learning in the construction of part-of-speech taggers. The goal is to
speed up the training on large data sets, without significant loss of
performance with regard to an optimal configuration. In contrast to previous
methods using a random, fixed or regularly rising spacing between the
instances, ours analyzes the shape of the learning curve geometrically in
conjunction with a functional model to increase or decrease it at any time. The
algorithm proves to be formally correct regarding our working hypotheses.
Namely, given a case, the following one is the nearest ensuring a net gain of
learning ability from the former, it being possible to modulate the level of
requirement for this condition. We also improve the robustness of sampling by
paying greater attention to those regions of the training data base subject to
a temporary inflation in performance, thus preventing the learning from
stopping prematurely.
The proposal has been evaluated on the basis of its reliability to identify
the convergence of models, corroborating our expectations. While a concrete
halting condition is used for testing, users can choose any condition
whatsoever to suit their own specific needs. | [
0.0332427621,
0.0060552261,
-0.0177486539,
-0.0163009614,
-0.0338541977,
0.0178028215,
0.0039290646,
0.0033932051,
0.0147961434,
-0.0357475355,
0.0178902708,
-0.0232208427,
-0.0186059158,
-0.0025827677,
-0.0286392812,
-0.0735510662,
0.0382262543,
-0.0481249765,
0.0008002878,
-0.0106283128,
-0.0053348877,
-0.0104819285,
-0.0581992082,
0.0304128248,
-0.0057079601,
0.0229633935,
-0.0096700508,
-0.0491977818,
-0.0337554738,
-0.0476321168,
-0.0113461195,
-0.0047808262,
-0.0200904999,
0.0476119146,
0.0000019682,
-0.0691139102,
-0.0382680967,
0.022160992,
-0.0153558375,
-0.0489403084,
0.0563986972,
0.0135757457,
-0.0202801954,
0.048423294,
-0.0243503582,
0.0220900457,
0.0554295033,
0.0166601557,
0.0155359562,
0.0370713659,
-0.0120737795,
0.0000561041,
-0.0009939264,
0.0183050614,
0.0174464211,
-0.0623850636,
0.0022970929,
-0.0078663873,
0.0506541319,
-0.0229779836,
-0.0160951708,
0.0429855064,
-0.0379895642,
-0.0083494429,
0.0309913028,
0.0085367225,
-0.0137027809,
-0.0287117437,
-0.0258212704,
0.0070716008,
-0.0112864114,
-0.0095311999,
0.0358480923,
-0.0106083481,
-0.0069673476,
0.0005211023,
-0.0089010028,
-0.0203739721,
-0.0303223636,
0.0491410457,
0.0204098448,
0.0061408821,
0.0024347133,
-0.013422153,
-0.0230881777,
0.0868160948,
-0.0033705935,
-0.0663186386,
0.0300994013,
0.0360134281,
-0.0276461821,
-0.0069706584,
0.0337415785,
0.0198642127,
-0.0284499098,
0.0331905335,
-0.0544640832,
-0.0235077012,
-0.0139483884,
-0.0478989929,
0.0405541509,
0.0019733605,
0.0317242257,
0.0270012803,
-0.0193053689,
-0.0232785214,
0.0252628177,
0.0028760466,
-0.0330928788,
-0.0111200046,
-0.0897349045,
0.0064430539,
-0.074531734,
0.0493824817,
0.0048162513,
-0.0043749101,
-0.0685825348,
0.0042067659,
-0.0102257803,
-0.0357785448,
0.0079246499,
-0.0286703072,
-0.0549508668,
0.0365498289,
-0.0337420143,
0.0009312839,
-0.0297982749,
0.0024168375,
-0.0328713544,
-0.0366387628,
-0.0032798809,
-0.0003372503,
-0.0054838005,
0.004228937,
0.0161899142,
0.0226490404,
-0.0052541327,
-0.0091493502,
-0.0916007012,
-0.0219899993,
-0.0434683338,
0.0059156013,
-0.0151001252,
-0.0326159112,
-0.0018703206,
0.0117100496,
-0.0091571072,
-0.0051776785,
-0.0056077386,
0.0216295924,
-0.0295312759,
0.0796420351,
-0.0492243432,
0.0123592578,
0.0228239391,
-0.0438284017,
-0.0010823426,
0.0759372637,
0.0800824612,
-0.0055788662,
0.0656567067,
0.0164807234,
-0.0121204117,
0.0805395916,
-0.0252197403,
-0.0383331999,
0.0039922167,
-0.0863783658,
0.0461586043,
0.0127015617,
0.0335601456,
0.0467993952,
0.0044061146,
0.0147188529,
-0.0011265047,
0.0369683616,
0.0274321213,
0.0267796516,
0.0524296016,
0.0287455618,
-0.0139228841,
0.0097963409,
0.0310061406,
-0.0693782344,
-0.0378019214,
0.0108862305,
-0.0265271552,
0.0072687953,
-0.0102252197,
-0.0379392058,
-0.0053731548,
-0.0329922885,
-0.0036051935,
-0.0009141229,
0.110527873,
0.0359668359,
-0.0498646721,
0.0425442792,
0.0515066907,
-0.0017467092,
0.0164973233,
-0.0539226271,
-0.0484867245,
0.0355609208,
0.0269420687,
-0.0717358515,
0.0124489693,
-0.0192540344,
-0.0174111612,
0.0107695172,
0.0607127808,
-0.0099289417,
-0.0030401512,
-0.0504792258,
-0.0281854831,
0.0854479969,
-0.0297371335,
0.0052920571,
0.0012394243,
0.0349670388,
-0.0156239672,
-0.058438316,
-0.0314243808,
0.0545885414,
-0.0252354909,
-0.0441032611,
0.0180251021,
-0.0140847629,
0.0145073971,
0.0441083945,
0.0001966891,
0.0139834406,
0.0377016366,
-0.039048072,
0.0322445966,
0.0011777406,
-0.0182405077,
0.0204366501,
0.0578585938,
0.0369808674,
-0.0018139214,
-0.0242221057,
-0.0001823315,
-0.0496806763,
0.0185175817,
-0.0466784835,
0.0372060165,
-0.0380218327,
0.0499418043,
-0.0088849757,
-0.0001977621,
-0.0554914959,
0.0062146438,
0.034321744,
0.0261673145,
0.0028040248,
0.0167818107,
0.0153689012,
-0.0575304851,
-0.0795694888,
0.0031566813,
-0.0328511372,
-0.0262707043,
-0.0043743043,
-0.0165341608,
-0.0149552012,
-0.0454022475,
-0.0798997059,
-0.0168763008,
0.0044681728,
0.0325803533,
-0.0356510244,
-0.0087389601,
0.0063334247,
-0.0554832667,
-0.0042884313,
-0.013408035,
0.0228980705,
0.0004908532,
0.0021821542,
0.0090851877,
-0.0648682937,
0.0018648694,
0.0010719672,
-0.0192015469,
0.0161894504,
0.0616925582,
-0.0753434449,
-0.0701034218,
-0.0453512371,
0.026469605,
0.0460489132,
0.0528371185,
0.0003160478,
0.0074779801,
0.0201877393,
0.0172790475,
0.0423691832,
0.0234884787,
0.0070500937,
-0.0027937931,
0.0267391019,
-0.0369614884,
0.0011163136,
-0.0398617312,
-0.0212061591,
-0.0319387242,
-0.0106509654,
0.0196990874,
0.0695818886,
-0.0007693268,
0.0297271311,
-0.022894064,
-0.0409554131,
-0.0273701921,
-0.0617677793,
-0.0079764854,
0.0036961078,
-0.0698959604,
-0.0191935617,
0.0041218237,
-0.0265869554,
0.0262681711,
-0.0225727204,
0.0048271208,
-0.0875417292,
-0.0357095785,
0.0045532007,
0.0227921735,
0.0223026574,
0.0173499733,
0.074756451,
-0.0352065116,
-0.0140769407,
-0.0012869522,
0.0347559452,
-0.0348864011,
-0.0396476611,
-0.033497829,
0.0221219659,
0.0169234276,
-0.0283224043,
-0.0582130738,
0.0112985643,
-0.0345049128,
0.0246177353,
-0.0200424548,
0.0004761267,
0.0031339889,
0.0364377424,
-0.009538346,
0.0420964509,
0.0033118073,
-0.0023372825,
-0.0119440164,
0.0249937512,
0.0362406261,
0.0331021249,
-0.0381911844,
-0.0069591287,
-0.0068313209,
-0.0015839094,
0.1061934382,
0.0009150015,
-0.0343186297,
0.0267706458,
0.0211075731,
-0.0623180531,
-0.0399565808,
0.0307956096,
-0.0255326144,
-0.0205756892,
-0.0152467545,
0.0960327983,
0.0347169973,
-0.0192154422,
-0.0109383184,
0.0209307689,
0.0304560382,
0.0084255403,
0.0400632285,
-0.0344181657,
-0.0254189204,
0.0303268768,
0.0945167914,
-0.0254669096,
-0.0176792406,
0.0106240381,
-0.1281553805,
0.0176685918,
-0.0064993701,
-0.0355718397,
0.0198461358,
-0.0159584451,
-0.0464306213,
-0.0186602138,
-0.0436314717,
0.0122467391,
-0.031674426,
0.0214763023,
0.0957154483,
-0.0620774515,
0.0010711098,
-0.0265736077,
-0.0756716728,
0.0720404387,
0.0670290738,
0.0038637281,
0.0101088546,
-0.0486131795,
0.0130865127,
0.0094293449,
-0.0021945988,
-0.0227107424,
0.0093431026,
0.0003705163,
-0.0037022722,
0.0015916976,
0.0017243004,
0.0059942757,
-0.0040935422,
0.0194727089,
0.0931784883,
-0.0092309499,
0.0348151959,
-0.0028060186,
0.0507843122,
0.0393575765,
0.0452988334,
-0.0201891959,
-0.0814780593,
-0.0296134353,
0.0068187281,
0.0153527912,
-0.0812848285,
0.00991794,
0.0569435619,
-0.0829775333,
-0.0494005084,
-0.0944368094,
0.0220458135,
0.0205969438,
0.0980604663,
0.0249146223,
0.0576511547,
0.0393499024,
0.0604875013,
-0.0585623421,
-0.035798382,
-0.0182445161,
-0.014547375,
-0.0691565126,
-0.0528377183,
0.0375389718,
0.003504988,
-0.0238523092,
-0.0120248599,
-0.0090428665,
-0.0642973781,
0.0974168852,
0.0391758569,
0.0317704864,
0.0259035379,
0.0020109084,
-0.0682046339,
0.0364258289,
0.0602465048,
0.0191026609,
0.0017493651,
0.038463328,
0.0057860087,
-0.021263307,
0.0117312679,
-0.0302773174,
0.0204197224,
-0.0303802676,
0.0244243033,
-0.0468703881,
0.0065578124,
0.0046453769,
0.0584221445,
0.038830746,
-0.0562511347,
-0.0479303449,
-0.0150338355,
0.0251941234,
0.0278940201,
0.0231137555,
-0.0419869348,
0.0244689416,
-0.0141485985,
-0.0022651805,
-0.0018130026,
-0.006909091,
-0.0198456161,
0.0051897904,
0.0261614081,
0.0236540269,
-0.0022568007,
0.0047245361,
-0.0408334322,
0.0142793711,
0.0311810579,
0.0133309178,
0.0526498146,
0.0247264709,
-0.018360788,
0.0322774798,
0.0179034937,
0.0633265823,
-0.03376849,
-0.023625521,
-0.0174873173,
0.0095861387,
-0.0547268502,
0.0360662639,
0.0121127134,
-0.001755396,
0.0059099523,
-0.0163065828,
0.0159068201,
0.0005026648,
0.0296611432,
-0.0248525608,
-0.0085868658,
0.0051564598,
0.0086151948,
-0.0173743591,
0.0094576199,
0.0573689863,
-0.0069819372,
0.0378782451,
-0.0232036859,
-0.0238070916,
-0.037690036,
-0.0131472154,
-0.0189832579,
-0.0025386747,
-0.0120270206,
0.0156285074,
-0.0367524363,
-0.0264914092,
-0.0056928853,
0.0130105549,
0.0254506171,
-0.0243479777,
0.0772013813,
0.006674822,
-0.0289284643,
0.0278642513,
-0.0460605137,
-0.0036757432,
0.045004949,
-0.0174254905,
0.0145884529,
-5.591720465e-33,
-0.0345272981,
-0.0243252702,
-0.0096492423,
0.0293859169,
-0.0243914742,
-0.0368182287,
0.0398854241,
0.0222807247,
0.0210291147,
-0.0241782218,
-0.0331229493,
-0.0560980886,
0.057152953,
0.029623067,
0.0178252459,
0.0159289967,
0.082293123,
-0.0592626296,
0.0030086793,
-0.0048313877,
0.0738912448,
0.0533559658,
0.0258317776,
-0.0177407116,
0.0111288894,
0.027336454,
-0.01815008,
0.002524168,
0.0651854128,
0.0716242567,
0.0147725474,
0.0389290601,
0.0062394678,
-0.0218322687,
-0.0006882598,
0.049399741,
-0.1100968868,
-0.0166759379,
-0.0114182057,
0.0517043062,
0.0031338192,
0.0165776797,
0.023824973,
-0.0187568553,
-0.0545378625,
0.0551884286,
0.0270568412,
-0.045462925,
-0.0459160879,
-0.0082901912,
-0.0248867422,
0.0332159959,
0.0202800278,
0.0295498949,
0.067924127,
-0.0235531274,
-0.0219450761,
-0.0382566229,
-0.0140660675,
0.0189793836,
0.0572999455,
0.0348620601,
0.0288856216,
0.046725709,
0.0125077954,
0.0634235218,
0.0632904992,
-0.026896432,
0.0602027923,
0.010352077,
0.0519310087,
0.0029950093,
0.0250628218,
-0.0161827616,
0.0919379815,
-0.0174431354,
-0.0141099319,
0.0111983931,
0.0622632243,
-0.0171632376,
0.0182259697,
-0.0028926968,
-0.065065898,
-0.0083341887,
-0.0198804587,
-0.0249993149,
-0.0240657851,
-0.0364174917,
0.0270790309,
0.029856069,
-0.0633852854,
0.0382824987,
0.0270431694,
-0.0479865447,
0.0179085117,
-0.0931832865,
0.0132409418,
0.0203474294,
-0.0260769036,
-0.0292629916,
-0.0476831011,
-0.0981801972,
0.0002347497,
0.0008835644,
0.0478406847,
-0.0244434159,
0.023910705,
-0.0087022185,
-0.0843804404,
0.0351594761,
-0.001697268,
0.0706405565,
0.0262800269,
0.0212878268,
0.0602219105,
-0.0176160354,
-0.0043926784,
0.0151703954,
0.0221613366,
-0.0289021321,
-0.0187493786,
-0.0233166628,
0.0213106498,
0.0275260806,
-0.0306381397,
0.0262600537,
0.0083940309,
0.0148594854,
-0.0074594403,
-0.0459117144,
-0.0173082352,
0.0341228507,
2.744e-7,
0.0169347748,
0.0489040911,
0.0354644693,
-0.014815527,
-0.0004122794,
0.0089047616,
0.0139012104,
0.027632717,
-0.0152310058,
0.0406870432,
-0.0303415935,
-0.03032328,
0.0170855839,
0.0298852865,
-0.1031858847,
-0.0437977128,
0.0047918675,
0.0610951744,
-0.0269701052,
0.0398269296,
0.0322979391,
0.045091249,
0.0014573389,
-0.0366868563,
-0.0008646164,
-0.0843062624,
0.0096532479,
0.0326001272,
0.0169529952,
0.0252535772,
0.0467365868,
0.0567628704,
0.0089924727,
0.0264620651,
-0.0531984679,
0.0068156971,
0.0144215664,
-0.0004442049,
-0.0097173071,
-0.019589074,
0.0395212211,
-0.0172557421,
-0.0126561066,
0.0173101351,
0.0519107729,
-0.0423620455,
-0.0095281489,
-0.0073433886,
0.0276578385,
-0.0450427569,
-0.0343660414,
-0.0258999206,
-0.0270372462,
0.0213979762,
0.0381172225,
-0.0141649013,
-0.0230858475,
-0.0296197757,
0.0283485893,
0.0229823086,
-0.0194058008,
-0.0521945022,
-0.0229971223,
0.0151033895,
0.0175451823,
-0.0337036923,
0.0080897501,
2.342023491e-34,
0.0378053673,
-0.0319462977,
-0.0026650205,
0.0482862219,
-0.0193079449,
-0.0105206864,
0.0179785136,
0.0418615006,
0.0138979321,
-0.0466218516,
-0.028992096
] |
http://arxiv.org/abs/2402.02515v1 | http://arxiv.org/pdf/2402.02515v1 | Modeling of learning curves with applications to pos tagging | 2024-02-04T15:00:52 | 2024-02-04T15:00:52 | Manuel Vilares Ferro, Victor M. Darriba Bilbao, Francisco J. Ribadas Pena | cs.CL | cs.CL, cs.AI, cs.LG | An algorithm to estimate the evolution of learning curves on the whole of a
training data base, based on the results obtained from a portion and using a
functional strategy, is introduced. We approximate iteratively the sought value
at the desired time, independently of the learning technique used and once a
point in the process, called prediction level, has been passed. The proposal
proves to be formally correct with respect to our working hypotheses and
includes a reliable proximity condition. This allows the user to fix a
convergence threshold with respect to the accuracy finally achievable, which
extends the concept of stopping criterion and seems to be effective even in the
presence of distorting observations.
Our aim is to evaluate the training effort, supporting decision making in
order to reduce the need for both human and computational resources during the
learning process. The proposal is of interest in at least three operational
procedures. The first is the anticipation of accuracy gain, with the purpose of
measuring how much work is needed to achieve a certain degree of performance.
The second relates the comparison of efficiency between systems at training
time, with the objective of completing this task only for the one that best
suits our requirements. The prediction of accuracy is also a valuable item of
information for customizing systems, since we can estimate in advance the
impact of settings on both the performance and the development costs. Using the
generation of part-of-speech taggers as an example application, the
experimental results are consistent with our expectations. | Modeling of learning curves with applications to pos tagging: An algorithm to estimate the evolution of learning curves on the whole of a
training data base, based on the results obtained from a portion and using a
functional strategy, is introduced. We approximate iteratively the sought value
at the desired time, independently of the learning technique used and once a
point in the process, called prediction level, has been passed. The proposal
proves to be formally correct with respect to our working hypotheses and
includes a reliable proximity condition. This allows the user to fix a
convergence threshold with respect to the accuracy finally achievable, which
extends the concept of stopping criterion and seems to be effective even in the
presence of distorting observations.
Our aim is to evaluate the training effort, supporting decision making in
order to reduce the need for both human and computational resources during the
learning process. The proposal is of interest in at least three operational
procedures. The first is the anticipation of accuracy gain, with the purpose of
measuring how much work is needed to achieve a certain degree of performance.
The second relates the comparison of efficiency between systems at training
time, with the objective of completing this task only for the one that best
suits our requirements. The prediction of accuracy is also a valuable item of
information for customizing systems, since we can estimate in advance the
impact of settings on both the performance and the development costs. Using the
generation of part-of-speech taggers as an example application, the
experimental results are consistent with our expectations. | [
0.0049672043,
-0.0050042444,
-0.0086820703,
-0.0115270186,
-0.0310855713,
0.0039187949,
0.0204442646,
0.0171439946,
-0.0046188496,
0.011549836,
0.0416364931,
-0.0313105322,
0.0157019757,
0.011600594,
-0.0228580702,
-0.0669733882,
0.0301211495,
-0.0467395149,
0.0051392373,
-0.0068351552,
-0.0067941318,
-0.056263838,
-0.043554578,
0.061465472,
-0.0739495754,
0.0235968679,
-0.0046470556,
-0.0410951376,
-0.0239714105,
-0.0598364398,
0.0037627963,
0.0043975743,
-0.0046447576,
0.0759898201,
0.0000020821,
-0.0756252855,
-0.0518265665,
0.0221943539,
-0.0225062873,
-0.0397610515,
0.0482952148,
0.032945551,
-0.0367353112,
0.0265232176,
-0.0173083339,
0.0111227324,
0.0639582202,
-0.0187083241,
0.0378081687,
0.0336230695,
-0.0103488239,
-0.0372400582,
0.0047434988,
0.0084504187,
0.0265674423,
-0.052049946,
0.0278191064,
0.0226240978,
0.0180552509,
-0.0563777499,
-0.0335991271,
0.063582249,
-0.0211703312,
-0.0204165746,
0.0276519731,
0.0098389043,
0.0304366089,
-0.0130035859,
-0.0025424703,
-0.020354094,
-0.0114297196,
-0.0025367509,
0.0126425168,
-0.0161884483,
-0.0271531995,
-0.019771155,
0.0035154682,
-0.0344576463,
-0.0319390073,
0.0157030001,
-0.0070649409,
-0.0098389937,
-0.022887202,
-0.0210127756,
-0.0289969835,
0.0719552636,
-0.0107766362,
-0.0786349252,
0.0493131839,
0.0072500855,
-0.0038352727,
-0.0100655165,
0.0127518401,
0.0325999185,
-0.0044266637,
0.0358171761,
-0.0358838849,
-0.0273300689,
-0.0180587117,
-0.0605705641,
0.0423239656,
-0.0165302102,
0.0266725104,
0.0277252719,
-0.0394855626,
0.0119949458,
0.0067919111,
-0.0285120811,
-0.0537589081,
0.0094594331,
-0.1026389524,
0.0180002805,
-0.0443769284,
0.0380868092,
0.0323586687,
0.0080447216,
-0.0887192264,
-0.0185747705,
0.0067577818,
-0.0013935931,
-0.0132133756,
-0.0252934918,
-0.0251320135,
0.0271637123,
-0.0669851303,
0.0074101817,
-0.0060788249,
-0.0025795829,
-0.0360072851,
-0.0038632348,
0.0198718682,
0.0100172786,
0.0088932766,
0.0392442085,
0.0412209108,
0.0349288508,
0.0186466854,
-0.0092057968,
-0.0878189653,
-0.0376068391,
-0.0212307535,
0.015029774,
-0.0118365046,
-0.0329039283,
-0.0337489881,
0.0133243212,
-0.0238618981,
-0.0116085745,
-0.0209612176,
-0.0105533982,
-0.0029379597,
0.0654066503,
-0.0495467447,
0.0504274704,
0.040472541,
-0.0484242477,
-0.0495663136,
0.0589657798,
0.065736331,
-0.0077498965,
0.0524578728,
-0.0235372167,
-0.0138300154,
0.0756773949,
-0.0128363287,
-0.0485168621,
0.0045859935,
-0.0708930641,
0.0371960402,
0.0147102149,
-0.0047003734,
0.0250367075,
-0.0102362623,
0.038499549,
0.0368624814,
0.057725098,
0.073107928,
0.0133554535,
0.039302174,
0.0064444169,
0.0226340052,
0.009546781,
0.0365675874,
-0.0525845215,
-0.0407286398,
0.0228680111,
-0.0243014619,
-0.0243313257,
-0.0071940818,
-0.0631538182,
-0.0195538979,
-0.0304449536,
-0.0075974059,
-0.0421996377,
0.0835992545,
0.007926465,
-0.0728655979,
0.0293913856,
0.0334505774,
-0.0084809158,
-0.0071759233,
-0.0558134206,
-0.0105678337,
0.031075811,
0.0393456891,
-0.0699805692,
0.0135855665,
-0.0246546511,
-0.0165821407,
0.0005738822,
0.0415079743,
-0.0012702381,
-0.0260645803,
-0.0357009582,
-0.0392575525,
0.0806063414,
-0.0302350149,
-0.0039243875,
-0.0086700823,
-0.0042211181,
-0.0117878336,
-0.019652715,
-0.0368787013,
0.0545851514,
-0.0474053249,
-0.0290484708,
-0.0051361243,
0.0097434428,
0.0157469474,
0.0397366062,
0.0082544452,
0.0064334199,
0.0383254923,
-0.0137804346,
0.026421139,
0.0159902144,
0.0251950044,
0.0358003005,
0.0732536912,
0.0131378472,
-0.015214866,
-0.0532229096,
0.0060575693,
-0.0539900847,
-0.0030811885,
-0.0683450401,
0.0423948802,
-0.0272158869,
0.0614939928,
-0.0232223812,
-0.0525791161,
-0.0333030187,
0.0188003257,
0.0290020835,
0.0712639838,
-0.0190288778,
0.001418321,
0.0586934052,
-0.040547505,
-0.065067105,
-0.0257594623,
-0.0349714085,
-0.0305408835,
-0.0015635416,
-0.0094912378,
-0.0140090883,
-0.0216978136,
-0.0747751743,
-0.0497610234,
0.0246956516,
0.0659478307,
-0.0185704418,
-0.0377417654,
-0.0064751464,
-0.0230398644,
0.0045122476,
0.0089236628,
0.0007298753,
-0.00491388,
0.0319973305,
0.0223549865,
-0.0766083747,
-0.03461257,
0.0061108726,
-0.0183369797,
0.0239729937,
0.0654769912,
-0.0567914397,
-0.0626027733,
-0.0167382695,
-0.0179863553,
0.0533735082,
0.0404672734,
-0.015495616,
0.0403234027,
0.0217181779,
0.0472726934,
0.0161354765,
0.0181101188,
-0.0223898925,
0.0063835024,
0.0147656603,
-0.0417318605,
-0.0090015996,
-0.0531200059,
-0.0047232369,
-0.0290775597,
-0.0104789529,
0.0073551582,
0.069120042,
0.0046827849,
0.0039598346,
-0.0314014442,
-0.0263169371,
-0.0337124653,
-0.0287628286,
-0.0240240842,
0.010391132,
-0.0407685786,
-0.0133013893,
-0.0067562419,
-0.0549489856,
0.0443010032,
-0.0098742824,
0.0189135615,
-0.0545496009,
-0.0299190897,
0.0119705172,
0.0390160084,
0.0386464819,
0.0148014883,
0.040200673,
-0.0116420994,
-0.0238461792,
0.0312641747,
0.0486756265,
-0.0505725965,
-0.0084729781,
-0.0627782121,
0.0275028255,
0.0104491496,
0.0242842585,
-0.0199461244,
0.0201691017,
0.0051993779,
0.0259979963,
-0.0244717076,
-0.0041696103,
0.0555798598,
0.0334211402,
-0.0001584701,
0.066960074,
0.0355815627,
-0.0016817992,
-0.0012460295,
-0.0168802366,
-0.0126571264,
-0.0158334412,
-0.0351842232,
-0.0348796621,
-0.0013991235,
-0.0190222543,
0.0633789226,
0.028407814,
-0.064728111,
-0.0262337681,
-0.0117450878,
-0.0609232411,
-0.0490657054,
0.0416002497,
-0.0224757064,
-0.0368852951,
-0.0072790086,
0.0904374421,
0.031415727,
-0.0042361179,
-0.0174361803,
0.0261834692,
0.0053726989,
0.0075674704,
0.0135970926,
-0.0151430657,
0.0104237515,
0.0140050771,
0.0761182904,
-0.0272861794,
-0.0255140532,
0.0125668542,
-0.1153796911,
0.0273509081,
-0.0241272878,
0.0009574306,
0.0014901956,
-0.0205969792,
-0.0279723648,
-0.0101063708,
-0.087100938,
0.0288462304,
-0.0191143323,
0.0201740339,
0.0748873428,
-0.0898840278,
-0.0287070554,
-0.0373491384,
-0.0647951365,
0.031142801,
0.0803235024,
0.0129510583,
0.0031314187,
-0.0801188648,
0.0194759257,
0.0272518396,
0.0308721773,
-0.0214963146,
0.0609101206,
0.0295426603,
0.0094420062,
0.0108996369,
-0.0091648968,
0.0367445052,
0.0235831216,
-0.0017001982,
0.063043341,
-0.0093145762,
0.0206900779,
0.0135997124,
0.0222080294,
0.0298946984,
0.0342458375,
-0.0020579661,
-0.0861388072,
-0.0254339688,
-0.0033673758,
0.0631567016,
-0.0192597359,
0.0205659643,
0.0224751849,
-0.0939043462,
-0.0464620143,
-0.0977541953,
0.0557057112,
0.0069372226,
0.0932507664,
0.0041455044,
0.0415128209,
0.0027174659,
0.0335676633,
-0.0397069491,
-0.0179752205,
0.027662266,
0.0171524212,
-0.0935467035,
-0.0263552088,
0.0548365675,
0.0164820496,
-0.0332721844,
-0.0609102286,
-0.0260601025,
-0.0504845381,
0.0606425963,
0.0511118136,
0.0324885324,
-0.0010102841,
-0.0005613416,
-0.0722799748,
0.0414838791,
0.034088783,
0.0322361253,
-0.0145366602,
0.0487583913,
0.0268397592,
-0.0184709951,
-0.0050886432,
0.0108502116,
0.0293502193,
-0.0177569464,
0.0004317715,
-0.0332785994,
-0.0058786338,
-0.0176122542,
0.046925392,
0.0798280537,
-0.0500203185,
-0.0975059643,
-0.0290476121,
0.0222235676,
0.048343081,
0.0075500258,
-0.0307436381,
-0.0140504828,
0.0061597326,
0.0260350816,
0.0034973596,
0.0007113892,
-0.0210283697,
-0.0108853653,
0.011041428,
0.0448709168,
0.0224757772,
0.0378569663,
-0.0137683777,
-0.0015063668,
0.0350693762,
-0.0088757547,
0.015365269,
-0.0017753403,
0.0204054061,
0.0400770865,
0.0011149619,
0.0575801246,
0.0331557058,
-0.0014906648,
0.0072045419,
0.008478282,
-0.0694798306,
0.0182893071,
0.0080437493,
0.0258344077,
0.008167455,
-0.0172071066,
0.0154418908,
-0.0068812133,
0.0161280092,
-0.0346115865,
0.0007336395,
0.000780283,
-0.0050021848,
-0.0229083467,
0.0033849715,
0.0488556251,
-0.0105915014,
0.0406441055,
-0.0817760825,
-0.019500548,
-0.02407833,
0.0076888148,
0.0080878884,
-0.017739173,
-0.0252455622,
0.0214052629,
-0.0554640181,
-0.0278195646,
0.0034921363,
0.0313257426,
0.0438123308,
0.0008176153,
0.0700254664,
0.0038717007,
-0.0457027815,
0.038095817,
-0.0261485707,
-0.025246717,
0.0246048681,
0.0090783341,
0.0069676396,
-5.433267867e-33,
-0.0187646486,
-0.0363933668,
-0.0049926532,
0.0638339221,
-0.0212030802,
-0.0113265794,
0.0274708904,
0.0266262759,
0.0132324984,
-0.0386283696,
-0.0110011185,
-0.0587164126,
0.0201453548,
-0.0099917818,
-0.0095773824,
-0.0009194507,
0.0645572171,
-0.0517639369,
0.0074948524,
-0.0683055893,
0.0642631277,
0.0557475649,
0.0415330231,
-0.0431795046,
0.0196267236,
0.0231019258,
-0.0439476371,
0.0195936169,
0.0340909734,
0.0487424843,
0.0229799487,
0.0179034229,
-0.003387904,
-0.0488152131,
-0.0049976096,
-0.0102153355,
-0.037518885,
0.0002119359,
-0.0305905361,
0.0342460722,
0.0151490299,
0.031898167,
0.0442394763,
-0.0093511706,
-0.071723856,
0.0396204554,
0.0403115265,
-0.029154608,
-0.0026177503,
-0.0137020145,
-0.0372669436,
0.03443335,
0.0377260149,
-0.0008372931,
0.0614757985,
0.0128984535,
0.0008392804,
-0.0706663728,
-0.0700449795,
0.034222845,
0.0642776862,
0.0362266637,
0.017641509,
0.0281838514,
0.0426881202,
0.0373886339,
0.0571019575,
-0.0340310298,
0.0380199216,
-0.0061855316,
0.0514017604,
0.0135641685,
0.0476978868,
-0.0479715802,
0.0857994109,
-0.0256774407,
-0.0324351005,
0.0418479145,
0.0035056164,
0.0028221826,
-0.0333819613,
0.0285853166,
-0.0498196818,
0.0125788627,
-0.0247449055,
-0.009524839,
0.0112696989,
-0.0521794222,
0.0230370611,
-0.0003778652,
-0.0651511177,
0.021853976,
0.0062469849,
-0.0292249471,
0.0240113046,
-0.0942961425,
0.0070068967,
0.0704992041,
0.0002718394,
-0.0001716036,
-0.0732866526,
-0.097163558,
0.004413004,
0.0055699539,
0.0666410029,
-0.0331766792,
-0.0129459975,
0.0219606794,
-0.0872962698,
0.0291209128,
0.0284924731,
0.0207675714,
0.0005489878,
-0.0005216156,
0.0608542673,
-0.0398074426,
-0.002729313,
0.0266224742,
0.0179659817,
0.039130237,
-0.0273564234,
-0.0172466002,
0.0136775039,
0.0173964724,
-0.0009852273,
0.018906951,
0.022023825,
0.01525224,
-0.0221880563,
0.0111900121,
-0.0113163162,
0.0484196767,
2.764e-7,
0.0345859639,
0.0502647236,
0.0200529005,
0.0114840679,
0.0186211467,
0.0199930482,
-0.0049172607,
-0.0278865919,
0.0078949109,
-0.025871722,
-0.0097417627,
-0.053450387,
0.0264551286,
0.0207128767,
-0.0928464606,
-0.0896215588,
0.0266569108,
0.0520145595,
-0.0282442998,
0.0596727952,
0.0524682626,
0.0207225382,
0.0164212547,
-0.021981502,
0.0029491449,
-0.0435778983,
0.020636389,
0.0281603839,
0.0192507356,
0.0182425436,
0.0551398247,
0.0160181336,
0.0137772756,
0.028845977,
-0.0484335534,
0.0239808392,
0.0201064274,
-0.0193991717,
-0.0134773115,
0.0220949315,
0.0500264578,
0.0134644555,
-0.0002926908,
0.0011421054,
0.046627786,
-0.0029504464,
0.0018233324,
-0.0220384281,
0.0276805311,
-0.0094916373,
0.0266846698,
-0.0286297519,
-0.0176982787,
-0.0072442186,
0.0105276788,
-0.0147447167,
-0.0226215534,
-0.0086622154,
0.0061688381,
0.0110614737,
0.0000475284,
-0.0663234293,
-0.0225225668,
0.0343321413,
0.014799023,
-0.0393963419,
0.0077449451,
2.352249603e-34,
0.0092999386,
-0.0502123982,
0.0303203557,
-0.0037565196,
-0.0073239575,
0.0095268236,
0.0498271435,
0.0189912766,
0.0256550815,
-0.0163363758,
-0.011592675
] |
http://arxiv.org/abs/2402.02513v1 | http://arxiv.org/pdf/2402.02513v1 | Early stopping by correlating online indicators in neural networks | 2024-02-04T14:57:20 | 2024-02-04T14:57:20 | Manuel Vilares Ferro, Yerai Doval Mosquera, Francisco J. Ribadas Pena, Victor M. Darriba Bilbao | cs.LG | cs.LG, cs.AI, cs.CL, cs.NE | In order to minimize the generalization error in neural networks, a novel
technique to identify overfitting phenomena when training the learner is
formally introduced. This enables support of a reliable and trustworthy early
stopping condition, thus improving the predictive power of that type of
modeling. Our proposal exploits the correlation over time in a collection of
online indicators, namely characteristic functions for indicating if a set of
hypotheses are met, associated with a range of independent stopping conditions
built from a canary judgment to evaluate the presence of overfitting. That way,
we provide a formal basis for decision making in terms of interrupting the
learning process.
As opposed to previous approaches focused on a single criterion, we take
advantage of subsidiarities between independent assessments, thus seeking both
a wider operating range and greater diagnostic reliability. With a view to
illustrating the effectiveness of the halting condition described, we choose to
work in the sphere of natural language processing, an operational continuum
increasingly based on machine learning. As a case study, we focus on parser
generation, one of the most demanding and complex tasks in the domain. The
selection of cross-validation as a canary function enables an actual comparison
with the most representative early stopping conditions based on overfitting
identification, pointing to a promising start toward an optimal bias and
variance control. | Early stopping by correlating online indicators in neural networks: In order to minimize the generalization error in neural networks, a novel
technique to identify overfitting phenomena when training the learner is
formally introduced. This enables support of a reliable and trustworthy early
stopping condition, thus improving the predictive power of that type of
modeling. Our proposal exploits the correlation over time in a collection of
online indicators, namely characteristic functions for indicating if a set of
hypotheses are met, associated with a range of independent stopping conditions
built from a canary judgment to evaluate the presence of overfitting. That way,
we provide a formal basis for decision making in terms of interrupting the
learning process.
As opposed to previous approaches focused on a single criterion, we take
advantage of subsidiarities between independent assessments, thus seeking both
a wider operating range and greater diagnostic reliability. With a view to
illustrating the effectiveness of the halting condition described, we choose to
work in the sphere of natural language processing, an operational continuum
increasingly based on machine learning. As a case study, we focus on parser
generation, one of the most demanding and complex tasks in the domain. The
selection of cross-validation as a canary function enables an actual comparison
with the most representative early stopping conditions based on overfitting
identification, pointing to a promising start toward an optimal bias and
variance control. | [
-0.0305190571,
0.0516566224,
0.009638017,
-0.0289697647,
-0.010911081,
0.0554692224,
0.0100845965,
-0.0261329301,
-0.0122296857,
-0.0031441564,
0.0499762073,
0.005271622,
-0.0130744101,
0.0342814215,
0.0060706879,
-0.0114397621,
0.060823217,
0.000818108,
0.0637996271,
-0.0096495626,
-0.0079800179,
-0.0110535156,
-0.0029245645,
0.077270709,
-0.0403957888,
-0.0356775597,
0.0272673685,
-0.0156388152,
-0.0032865447,
-0.0386907645,
0.0283868443,
-0.0131805791,
-0.0050974539,
0.076603353,
0.0000021696,
-0.0712421238,
-0.048023086,
0.0166661702,
-0.0302050468,
-0.0377081707,
-0.024876222,
-0.0216690563,
0.0142675154,
0.0165854823,
-0.0184139535,
-0.0012935417,
0.0441405289,
0.0480963066,
0.0661719888,
0.0565720089,
-0.0053629815,
-0.0037165119,
0.0498540848,
0.0001122918,
-0.0164395105,
-0.0378059819,
-0.0081950333,
-0.014731125,
-0.0231069531,
-0.1256380975,
-0.0013287362,
0.027884759,
0.0118154902,
0.0397858322,
-0.0111126537,
0.0408342592,
-0.0475382432,
-0.0366321951,
-0.0226398651,
0.0132644102,
-0.0528308116,
-0.0192468148,
0.0301922001,
-0.0165321212,
-0.0160741676,
-0.0010336636,
-0.0210246406,
0.0137638794,
-0.0220533796,
0.0431491099,
0.0518267751,
0.028609436,
0.0151108308,
0.0364690796,
-0.0147758722,
0.1033257917,
-0.0112599581,
-0.0646887422,
0.0741335526,
0.0958354548,
-0.0475152656,
-0.01317163,
0.0136719756,
0.0198099017,
-0.0560989901,
-0.0042097624,
-0.0551654138,
0.0075075179,
-0.0126034971,
-0.051219482,
0.0270538051,
0.0328593105,
0.0011088463,
0.0091829496,
-0.0451884121,
0.0212695207,
-0.0284767672,
-0.0022578978,
0.0010756794,
0.0280085541,
-0.047983449,
0.0084431367,
-0.0636629537,
0.0233444553,
0.0456190631,
0.0047417739,
-0.0637125745,
0.0141239082,
-0.0381138846,
0.0628117099,
-0.0400713012,
0.0246960763,
-0.0102699734,
0.0364471301,
-0.0098262699,
-0.0099089267,
-0.031944409,
-0.0107614072,
0.0109599764,
-0.041274745,
-0.0191126484,
0.0375641361,
-0.0017056733,
-0.0336858258,
0.0068739681,
0.0681362599,
0.0012972847,
0.0127688674,
-0.0392179117,
-0.0754581615,
0.0032728601,
-0.0410190523,
0.0244528316,
-0.02378661,
-0.0098111369,
0.0028706507,
0.0218206588,
-0.0028371778,
-0.0155158024,
0.0017643373,
0.0251794253,
0.0144532472,
-0.0272662081,
-0.0210268758,
0.0363135487,
-0.0330681801,
-0.0387720317,
0.0688687339,
0.0249702595,
-0.0020686369,
0.0442137495,
-0.0034712011,
-0.007070743,
0.0097115263,
0.0226243474,
-0.0528856665,
-0.0674115345,
-0.0473466478,
0.0101588769,
0.0264361911,
0.0130928904,
0.0188937895,
-0.049788788,
0.0145751406,
0.0406945869,
0.0317174494,
0.0283804256,
0.0481246077,
0.1147244945,
0.018922234,
-0.02546007,
0.0483679585,
-0.0271637402,
-0.0135223996,
-0.0396758765,
0.0085551096,
-0.0508404076,
0.0083890473,
-0.0275084097,
-0.0649633855,
0.0059342915,
0.0045656855,
-0.0447161011,
-0.0168215316,
0.047719162,
0.007826034,
-0.0379993133,
0.0027059051,
-0.011878497,
-0.0034208633,
0.0038753687,
-0.0558320023,
0.0236198734,
0.0594500564,
0.0318498202,
-0.0274102539,
-0.0147862621,
-0.0291568507,
0.0145816691,
-0.005913185,
0.0166148618,
-0.0070400243,
0.0431289747,
-0.0628592968,
-0.0131726116,
0.0325524844,
0.0219634064,
-0.0343898982,
-0.0098829567,
0.0404389836,
0.0100365039,
-0.0054171523,
-0.0858333558,
0.0580113344,
-0.0020884471,
-0.0300333016,
0.0135291414,
0.0310353041,
0.0290524289,
-0.022007348,
-0.0150589989,
0.0009295925,
0.0199707542,
-0.0074058459,
-0.0202097986,
0.0482951701,
-0.0512221381,
-0.0345428586,
-0.0042784014,
-0.0601898059,
-0.00121656,
0.0112796091,
-0.0243493859,
-0.0581356362,
0.0220299643,
-0.062136177,
0.0402952842,
-0.0105592189,
0.0229977779,
-0.0176925343,
0.0320361815,
-0.0260784607,
0.0621117577,
-0.0068232762,
0.0439646319,
0.0112097021,
-0.0419909805,
-0.0081561068,
-0.0546499901,
-0.0941151455,
-0.0083872993,
-0.0378780067,
0.0041176272,
0.016919272,
0.0155722052,
-0.04588392,
-0.0126071395,
-0.0724374503,
-0.0071913521,
-0.0250979047,
0.0212618094,
-0.0198621899,
-0.0517630465,
-0.0036668647,
-0.0292858928,
0.0138943922,
0.0389458016,
0.0413662717,
0.0032595498,
0.0479702391,
0.0095724314,
-0.0527394451,
-0.0082218535,
-0.0334978513,
-0.0300735924,
0.074251309,
0.0274143275,
-0.060299214,
-0.038141042,
0.0197106078,
-0.0043935399,
0.0320411548,
0.0194959193,
0.0045612431,
0.0491096787,
-0.0221060645,
-0.0537087657,
0.0893000811,
0.0019081142,
0.0029340244,
0.0230520908,
-0.0009451169,
-0.0526700392,
0.0362218246,
-0.0164653528,
0.0229500774,
0.0069822837,
0.0016656863,
0.0194312707,
0.1090660915,
-0.017432563,
0.0137867043,
-0.0382672735,
0.018903004,
-0.0261600297,
-0.0311724208,
-0.0018548295,
-0.0171882473,
-0.0022025688,
0.0276092216,
-0.0135637904,
-0.0842989907,
0.0692016706,
0.0075116386,
-0.010168856,
-0.021106232,
-0.0139962779,
-0.0179660693,
0.0107913539,
-0.000659487,
-0.0111431684,
0.0485985577,
-0.0293947365,
-0.004104469,
-0.0029782897,
0.0069753942,
0.0058576791,
0.0051382403,
-0.0406272523,
0.0156480558,
0.0116439406,
-0.0181824658,
-0.0515991226,
-0.0191315114,
-0.0216922462,
-0.0209008399,
-0.0509186536,
0.0163078737,
-0.0172654372,
-0.0278604943,
-0.0655581132,
0.0659616888,
-0.0079178028,
-0.0253246669,
-0.0330304764,
0.0558138676,
0.0003663847,
0.0158685558,
0.038229648,
-0.0378671624,
0.0814776048,
-0.0001811371,
-0.010069387,
0.0263084695,
-0.0502224304,
0.0237391256,
0.0179098658,
-0.0606163368,
-0.0421413891,
0.0369964987,
-0.0117540024,
-0.0826192424,
-0.0660204217,
0.1030540764,
0.004605975,
-0.0440913625,
0.0071627088,
0.0264519919,
0.031394992,
0.0069303373,
0.0320813805,
-0.0195917077,
-0.0086547611,
-0.0006005702,
0.061364226,
-0.0617935061,
-0.0400608219,
-0.0243874993,
-0.0703732371,
0.0259571429,
-0.0114956321,
-0.0292751063,
0.0450693145,
0.0203742925,
-0.0475400686,
-0.0444295853,
0.0064063999,
-0.0069925808,
-0.0100491187,
0.0225848425,
0.0514682047,
-0.0248572789,
-0.0636230409,
0.0014565118,
-0.0530244447,
0.0337758251,
0.0745331571,
0.0117645217,
-0.0246024597,
-0.0325129554,
0.0180088915,
-0.0235212687,
0.0335170403,
-0.0054960931,
0.0584940538,
-0.0000381356,
-0.020975342,
-0.0336366519,
0.0071850824,
0.0102616018,
0.0029348729,
0.0213221833,
0.0952196568,
-0.0317531377,
-0.0745369866,
0.0009876022,
0.0359193571,
0.0332179479,
0.0351168513,
0.0056338911,
-0.0728765577,
-0.0846665278,
0.0027704926,
0.0393153355,
-0.1094342321,
0.0442384407,
0.0128747579,
0.0339489058,
-0.0541302301,
-0.0702360123,
-0.0156738386,
-0.0182331689,
0.0579260662,
-0.019479068,
-0.0216675811,
-0.001751657,
0.0260100979,
-0.0428120121,
0.0521851517,
0.0465308018,
-0.0082250843,
-0.0895772353,
0.0478756204,
0.0181081761,
-0.0162961446,
-0.0312729329,
0.0045200237,
-0.0488481633,
-0.0578446127,
0.02797902,
0.0624766238,
0.0224403068,
0.0240899976,
-0.0244355071,
-0.1044127643,
0.0564607196,
0.0281399824,
-0.0083365161,
-0.0067806896,
0.0727253109,
-0.0163976587,
-0.0012814827,
0.0250202026,
-0.0010024132,
0.0319212377,
-0.0035484619,
-0.0112932194,
-0.0321193635,
0.0526530482,
0.0391936116,
0.0007259977,
0.0314292014,
-0.04195356,
0.0026367302,
-0.0220628548,
-0.0155200167,
0.0688688233,
-0.0012367213,
-0.0509033836,
0.0114485407,
0.036699187,
0.002881628,
-0.0263259951,
0.0187409651,
0.0120989894,
-0.026238516,
-0.0445715412,
0.0294888504,
-0.0107150199,
0.0502931215,
-0.0030877457,
0.0006938961,
-0.0396197699,
0.0194266178,
-0.0090063289,
0.0288645495,
-0.0091928691,
0.0518369749,
0.0320496596,
0.0847609416,
-0.0032227214,
0.0009653546,
-0.0481036194,
-0.027969243,
-0.0634916946,
-0.0088668009,
0.0288881864,
-0.0123909861,
0.0589327961,
-0.0069766631,
0.015733324,
-0.0062809186,
0.0362666026,
-0.0121263741,
-0.0305954199,
-0.003756593,
0.0157111902,
-0.0257960018,
-0.0476571694,
0.0505043194,
-0.0038594818,
0.0372352824,
-0.0235119313,
-0.0083692241,
-0.0312183611,
-0.0163975209,
-0.0123643009,
-0.0034318392,
-0.0317004696,
-0.0166667234,
-0.0235663652,
-0.0253015794,
0.0301870164,
0.0298184343,
0.0673860237,
0.0096126739,
0.0267946124,
0.0216607843,
-0.0198248941,
0.0253159329,
-0.02210778,
-0.0151372673,
0.0474615395,
0.008193125,
-0.0547842793,
-5.975217783e-33,
-0.0143256634,
-0.0297163129,
0.0354429334,
-0.009118489,
-0.0336659253,
0.0237319134,
-0.0222649965,
0.0003277466,
0.0099571804,
0.0095584085,
-0.0023515117,
0.0174758211,
0.0240528863,
0.0011578733,
0.0330413543,
-0.0233148225,
0.0349418707,
-0.0015526877,
-0.0104134781,
-0.0086568044,
0.0331181772,
0.0369597971,
0.0545535535,
-0.0344945937,
-0.0210574027,
0.0132805416,
-0.0519898832,
-0.0226875842,
-0.034541592,
0.0145263756,
-0.0085263737,
0.0364758186,
0.0302059706,
0.0005422228,
-0.0014070323,
0.0320613831,
-0.0387462303,
-0.0475033447,
-0.030529201,
0.055843506,
-0.0092648715,
-0.0034125168,
0.0755462274,
-0.0123300096,
-0.0559967011,
0.0289958436,
0.0130680725,
0.0108493194,
-0.0114931483,
-0.069288291,
-0.0493595563,
0.0346257687,
-0.0011180029,
-0.0088579888,
0.0323221944,
0.0930887982,
-0.0113837617,
-0.0167668555,
-0.0348682217,
0.0300039705,
0.00657079,
-0.0209908579,
0.0233608801,
0.010560127,
-0.0158111714,
0.0423302911,
0.0470829867,
0.014720927,
0.0213874578,
-0.0087691853,
0.0552386306,
0.0221973751,
0.0207355898,
-0.0225521009,
0.0906391516,
-0.0294094272,
-0.0480924174,
0.0244244523,
0.029170027,
0.0547031164,
0.0316647775,
-0.012307236,
-0.0487574376,
-0.0364849977,
-0.0074885124,
-0.0844352767,
-0.0189397503,
-0.0020532261,
0.0412024558,
0.0333548635,
-0.0322779231,
0.0091065727,
0.0093245106,
-0.0385815576,
0.0743315741,
0.0099911587,
-0.0122444611,
0.0096632428,
0.0183449294,
-0.0123679312,
-0.0273584239,
-0.0838335752,
-0.0184847564,
-0.0053521022,
0.0806139857,
-0.0414983258,
0.0446071513,
0.0293426402,
-0.0197701436,
0.0345703363,
-0.0300405454,
0.0353503376,
0.0427610762,
0.055456575,
0.0789624676,
-0.0199117009,
-0.0157447476,
0.0285878852,
-0.0077113206,
0.0224325117,
-0.0448755249,
-0.0224287361,
0.0504769683,
0.0297157411,
-0.0149376551,
-0.0261233486,
0.0269330852,
0.0162650179,
-0.0173779633,
-0.0497543886,
-0.0134562561,
0.0444634855,
2.903e-7,
-0.0019937085,
0.0186663009,
0.0445074253,
-0.0069741057,
0.0297627877,
-0.0014586031,
0.0105412919,
-0.0016122265,
-0.0370328128,
0.029613914,
0.0468149632,
-0.0305479299,
0.0528219603,
0.0542654134,
-0.0719303116,
-0.0810055658,
-0.0270888899,
0.0489609726,
-0.0017184378,
-0.0226566661,
0.0281583089,
0.022729706,
-0.0218378399,
-0.0323527344,
0.0336217768,
-0.0519299731,
0.0265377369,
0.0098109888,
0.0500403382,
-0.0056867674,
0.0113754002,
0.012142906,
-0.0095455693,
0.0304167625,
-0.0098333117,
0.049675148,
-0.006950519,
0.0240751859,
-0.0000025715,
0.0265711527,
0.0136595732,
0.0056310426,
0.0193533059,
-0.0269611068,
0.0573111661,
-0.0770596117,
-0.0135333911,
-0.0262291934,
0.0005741903,
-0.0619302429,
0.0608578101,
-0.0459868833,
-0.0092335679,
0.0192448143,
0.0103228549,
-0.0110087357,
-0.0117172617,
-0.0399257876,
-0.0131226927,
0.0561933257,
-0.0302558467,
-0.0376276225,
-0.0460781343,
0.0134646418,
0.033527486,
-0.0478901677,
0.0042576385,
2.614144234e-34,
0.0236769151,
-0.023723999,
0.0422091894,
0.0246095695,
-0.0073011941,
0.0165992528,
-0.0406753831,
0.0123727238,
0.0355129912,
-0.0163214337,
-0.0100341961
] |
http://arxiv.org/abs/2402.02503v1 | http://arxiv.org/pdf/2402.02503v1 | GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering | 2024-02-04T14:28:23 | 2024-02-04T14:28:23 | Ziyu Ma, Shutao Li, Bin Sun, Jianfei Cai, Zuxiang Long, Fuyan Ma | cs.CV | cs.CV, cs.CL | Knowledge-based visual question answering (VQA) requires world knowledge
beyond the image for accurate answer. Recently, instead of extra knowledge
bases, a large language model (LLM) like GPT-3 is activated as an implicit
knowledge engine to jointly acquire and reason the necessary knowledge for
answering by converting images into textual information (e.g., captions and
answer candidates). However, such conversion may introduce irrelevant
information, which causes the LLM to misinterpret images and ignore visual
details crucial for accurate knowledge. We argue that multimodal large language
model (MLLM) is a better implicit knowledge engine than the LLM for its
superior capability of visual understanding. Despite this, how to activate the
capacity of MLLM as the implicit knowledge engine has not been explored yet.
Therefore, we propose GeReA, a generate-reason framework that prompts a MLLM
like InstructBLIP with question relevant vision and language information to
generate knowledge-relevant descriptions and reasons those descriptions for
knowledge-based VQA. Specifically, the question-relevant image regions and
question-specific manual prompts are encoded in the MLLM to generate the
knowledge relevant descriptions, referred to as question-aware prompt captions.
After that, the question-aware prompt captions, image-question pair, and
similar samples are sent into the multi-modal reasoning model to learn a joint
knowledge-image-question representation for answer prediction. GeReA unlocks
the use of MLLM as the implicit knowledge engine, surpassing all previous
state-of-the-art methods on OK-VQA and A-OKVQA datasets, with test accuracies
of 66.5% and 63.3% respectively. Our code will be released at
https://github.com/Upper9527/GeReA. | GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering: Knowledge-based visual question answering (VQA) requires world knowledge
beyond the image for accurate answer. Recently, instead of extra knowledge
bases, a large language model (LLM) like GPT-3 is activated as an implicit
knowledge engine to jointly acquire and reason the necessary knowledge for
answering by converting images into textual information (e.g., captions and
answer candidates). However, such conversion may introduce irrelevant
information, which causes the LLM to misinterpret images and ignore visual
details crucial for accurate knowledge. We argue that multimodal large language
model (MLLM) is a better implicit knowledge engine than the LLM for its
superior capability of visual understanding. Despite this, how to activate the
capacity of MLLM as the implicit knowledge engine has not been explored yet.
Therefore, we propose GeReA, a generate-reason framework that prompts a MLLM
like InstructBLIP with question relevant vision and language information to
generate knowledge-relevant descriptions and reasons those descriptions for
knowledge-based VQA. Specifically, the question-relevant image regions and
question-specific manual prompts are encoded in the MLLM to generate the
knowledge relevant descriptions, referred to as question-aware prompt captions.
After that, the question-aware prompt captions, image-question pair, and
similar samples are sent into the multi-modal reasoning model to learn a joint
knowledge-image-question representation for answer prediction. GeReA unlocks
the use of MLLM as the implicit knowledge engine, surpassing all previous
state-of-the-art methods on OK-VQA and A-OKVQA datasets, with test accuracies
of 66.5% and 63.3% respectively. Our code will be released at
https://github.com/Upper9527/GeReA. | [
0.0404124483,
0.0327752009,
-0.0119374795,
0.0591356307,
-0.0095933275,
0.0147695057,
0.0354381166,
0.0136729805,
0.0347188264,
-0.0703191459,
-0.0066930368,
0.0197656881,
-0.0253002327,
0.0612369403,
0.052283112,
-0.0628257096,
0.0361577123,
-0.0547577478,
0.0042559458,
-0.0046509756,
0.0504995137,
0.0249914657,
-0.0096944887,
0.0418376699,
-0.0287125576,
-0.0488753356,
-0.0342108048,
0.0057583889,
-0.002243679,
-0.007426335,
-0.014804055,
0.0349680707,
0.0646002367,
-0.0315890461,
0.0000019052,
-0.0030096339,
0.0295749772,
-0.0173795298,
-0.0001884382,
-0.0114817536,
-0.0396502092,
-0.0169674139,
0.0402845703,
-0.0053557139,
-0.0396846347,
-0.0019332517,
0.0467548221,
0.0475228354,
0.086726144,
0.0241045505,
-0.0177844353,
0.0013167498,
-0.0083773527,
-0.0040956475,
0.0470047109,
-0.0260773078,
-0.020450227,
0.0612331517,
-0.0124856653,
0.0233454984,
-0.0501720905,
0.0597461499,
-0.0000926379,
-0.033239726,
0.0015769253,
0.0540796444,
0.0122761726,
-0.0456206575,
-0.0336044729,
0.1064672396,
0.0374637246,
0.0135772564,
-0.0359652825,
0.008619803,
0.0314377025,
0.0441903435,
-0.0493700542,
0.0433265008,
-0.0126621481,
-0.0149672888,
-0.0497041792,
0.0108663421,
-0.0155434757,
0.0007974825,
0.0231266189,
0.0033637469,
0.0181572195,
0.0154413814,
-0.0246858038,
0.0089853508,
-0.0677681789,
-0.0606236346,
-0.055037424,
0.0662283674,
0.0096739214,
0.0174814835,
-0.0178778358,
0.0031055375,
0.0045091691,
-0.015968224,
0.0061900392,
0.0125875073,
-0.0534126759,
0.0650146082,
-0.0315967537,
0.0706348792,
-0.004470448,
0.0229673032,
-0.046267163,
-0.0377270505,
-0.0184770357,
-0.0187198129,
-0.0496746711,
0.0055322084,
-0.0656377301,
-0.0467497744,
-0.0030436085,
0.0504957773,
-0.0204846524,
0.0253923945,
-0.0014245054,
0.0506992042,
0.0020489851,
-0.0003971461,
-0.0737983733,
0.050381992,
-0.0462623015,
0.0170261934,
0.0034870273,
-0.0148944575,
0.0071997237,
0.0334779173,
0.0298258886,
-0.0246553961,
-0.0097781178,
0.057250306,
-0.0542401001,
-0.0053606443,
-0.0055690967,
-0.0568503104,
-0.0239914544,
-0.0172541253,
0.0300347209,
-0.0369851701,
0.03068983,
-0.0037882703,
0.0051443861,
-0.0150353853,
0.0062413691,
0.0343432128,
-0.0130534293,
0.093159087,
-0.0284281578,
0.0018559194,
0.0268122032,
0.0238971543,
-0.0274149571,
0.0487738363,
-0.0125741754,
0.0263426285,
0.0584717318,
-0.0165805481,
0.0203079842,
0.0128886644,
0.0239756033,
-0.0001494173,
-0.0252067707,
0.0509747565,
0.0105585549,
0.0134158507,
0.0365910046,
0.0223599877,
-0.0098829577,
-0.0110151963,
0.0797286183,
0.0190263148,
-0.0136371972,
0.0166531913,
0.0196889434,
-0.0071410453,
0.0091758557,
-0.0152762514,
-0.0509484075,
0.0973571464,
-0.0372953415,
0.0157825928,
-0.0049569895,
0.0267408509,
-0.0500053205,
-0.0032125001,
0.0085215457,
-0.009681209,
-0.0068766591,
0.0515669994,
0.0208691824,
-0.0032674386,
-0.047103405,
0.0451823846,
-0.0044986461,
-0.0174716506,
-0.0137427868,
-0.0298166648,
0.0848763809,
0.0403030217,
0.0121809086,
-0.0026623404,
-0.036944136,
-0.0011348327,
0.0216024164,
0.0339908972,
-0.0128324591,
-0.0163794439,
0.0121975923,
-0.0362863764,
-0.0431487225,
-0.021267863,
-0.0236980729,
-0.0061739176,
0.021131441,
0.0095309671,
0.0151629848,
-0.0077876891,
0.0257920511,
0.0181559138,
0.0353835002,
-0.010627253,
-0.0359938741,
0.0323155075,
-0.0117629319,
-0.0078832908,
0.0101618385,
-0.001547189,
0.0114104953,
-0.0348330997,
-0.0216404758,
0.0075053382,
-0.0687478408,
-0.0810260847,
0.0223762654,
-0.0366590545,
-0.0020862864,
-0.0008924602,
-0.0362863354,
-0.0205166545,
-0.0049290746,
-0.0216176324,
-0.0167047866,
-0.0784442499,
-0.0575830191,
0.0469260588,
0.0216753837,
-0.0175909474,
0.0511499792,
-0.0007919733,
-0.0278417207,
0.031519752,
0.0010707962,
0.0200233236,
-0.0054667466,
-0.006704336,
0.0154158017,
0.0064696684,
0.018775288,
0.0142959617,
0.0373938233,
-0.0568840317,
0.007678112,
-0.1150643006,
-0.027108077,
-0.0074781096,
-0.0504805669,
0.0066143172,
0.0177182052,
0.0098941661,
-0.0479354337,
-0.0677502155,
0.0436154418,
-0.0023520291,
0.0145429941,
0.0499899127,
0.0253974851,
0.0162886363,
0.0019043294,
0.0075467988,
0.0141762067,
0.0977948382,
0.0229270831,
0.0203179233,
-0.0521286391,
0.0025241971,
-0.0665681362,
0.0487845242,
-0.003734862,
-0.0208734889,
0.009171037,
-0.0270799305,
0.039631851,
0.0593906716,
0.0090848608,
-0.0106217451,
0.0065039182,
-0.0278721415,
-0.043941319,
0.0246963724,
0.0051631564,
-0.0772391558,
0.0190058388,
-0.0437454693,
-0.0374674276,
0.0108045936,
0.0046100495,
-0.0244903173,
-0.0734944865,
0.0531042777,
0.0414793044,
-0.0178768244,
-0.1009973958,
0.0093089454,
0.0240733493,
0.0412645563,
-0.0447100885,
-0.0617904998,
0.0621730983,
-0.0311390199,
-0.0181113686,
0.0208568927,
-0.0129808635,
-0.0368933454,
0.0888077915,
0.0010609594,
0.0297989473,
0.0044777445,
0.0126053439,
-0.0263158269,
-0.0306214225,
0.0135354744,
0.0143695781,
0.053442359,
-0.0684555471,
0.023683982,
-0.019151371,
-0.0378213972,
-0.0235578474,
-0.0442852937,
-0.0199822858,
0.00788981,
-0.0755908713,
-0.0159394611,
0.0265220739,
0.0348363332,
0.0586178079,
-0.0023130362,
-0.0232484303,
-0.0136919636,
-0.0147492625,
0.0582697578,
0.0801254585,
0.0672060102,
-0.0317717455,
-0.0238819476,
0.0542806536,
0.0379602984,
-0.0086679747,
-0.0164658204,
-0.0156131433,
0.0364026688,
0.0052785217,
0.0446078219,
0.0020142579,
0.036922425,
0.0874190256,
-0.0314829759,
0.0045763096,
0.1004104316,
0.0002452459,
-0.0335024595,
-0.0397944152,
0.0066572451,
0.0336660333,
0.0409224406,
0.0366081633,
-0.0801367313,
-0.0625990778,
-0.0378432274,
-0.0128205968,
-0.0081937127,
0.0062899827,
-0.0196654927,
-0.0758144557,
0.0346126072,
-0.0267549288,
-0.0342424735,
-0.0564345382,
0.0085209114,
-0.0253173076,
0.0302366558,
-0.0119238896,
-0.0079304203,
-0.0154474434,
0.0048954245,
0.0063793566,
0.0241870321,
0.0027283474,
0.0418807864,
-0.0557066239,
-0.0041510649,
0.0352519751,
0.0243051033,
-0.0054835314,
-0.0597757213,
-0.0181448516,
-0.0213087797,
0.0229628328,
-0.0066418769,
0.053464558,
-0.011406024,
-0.0513156913,
-0.0167473294,
0.0218393505,
-0.0123530701,
-0.0107710687,
0.0181852765,
0.0732603669,
-0.0237795189,
-0.0075164372,
-0.0354553461,
0.0247197784,
0.0034004385,
0.0620326586,
0.0290502328,
-0.0112246899,
-0.01513121,
-0.008430779,
-0.0149473716,
-0.0268014781,
-0.0045835483,
0.0175954252,
0.031036716,
0.0223342348,
-0.0728225559,
-0.019103609,
0.0647324249,
0.048021365,
-0.0054821731,
-0.0030668455,
-0.0246440079,
0.0369704105,
-0.0501770861,
0.0634561852,
-0.0345453694,
0.0165913999,
-0.0212458707,
-0.0013508517,
0.0365698114,
-0.0492171235,
-0.0214136858,
0.033748053,
-0.0353687555,
-0.0052628173,
-0.0526267737,
0.0776281804,
0.0082058851,
0.0373738371,
0.0387084037,
-0.0229384955,
0.0432969145,
0.0229356308,
0.0015478998,
0.0048384033,
-0.0147157861,
0.0092259645,
-0.0045099249,
0.0197457448,
-0.044863686,
-0.0235320944,
-0.0189904273,
0.0018540787,
0.0094661182,
-0.0410861671,
0.0167785436,
0.0364879631,
-0.0142648267,
-0.0161366183,
0.0034166956,
-0.0235596728,
0.0017084022,
0.0004257882,
-0.0063347653,
-0.0729592144,
0.0384270027,
-0.0087706773,
-0.0232499745,
0.0182136409,
-0.0314839333,
0.0085512428,
0.0146973794,
-0.0282951612,
-0.0062390999,
0.0423293263,
0.0040102196,
0.0027707629,
0.0173525177,
-0.017503025,
0.012507109,
-0.0237076059,
0.0354014374,
-0.009132795,
0.1226753592,
0.034860447,
-0.0486690663,
-0.0382657498,
0.0333496705,
-0.0200186726,
-0.058121752,
0.0000608067,
0.0615174584,
0.0446766466,
0.0076496745,
-0.0308438204,
-0.0040938309,
-0.0164234806,
0.0151994787,
0.0323461816,
0.0126618613,
-0.041678872,
0.0921506956,
0.0239364877,
0.0070086354,
-0.0054612849,
0.0547071844,
-0.00794484,
0.0307259914,
-0.0105570816,
-0.0177772585,
-0.0417754687,
0.0157292262,
0.0605020262,
0.0030535734,
-0.0059684343,
0.0634023249,
-0.0231519286,
-0.055760324,
-0.0037669032,
0.0209437832,
-0.0472644195,
0.0247878861,
0.0535423122,
0.0483825803,
0.0326480381,
-0.0540825203,
-0.0328540839,
0.0263676103,
-0.0536669753,
0.031644959,
0.0020750612,
-6.139163245e-33,
-0.0555477254,
-0.0397496074,
0.046513427,
-0.0287138633,
-0.01809402,
0.0262379553,
-0.0343641117,
-0.0085692564,
-0.0266828593,
-0.0011803275,
0.0176940877,
-0.0453618467,
-0.0073552816,
0.0392520092,
-0.0085487226,
-0.0368006378,
-0.005472424,
0.0026713358,
-0.0122463945,
0.0316762663,
0.0399798155,
0.0075719976,
0.028304046,
-0.0114496555,
0.0177829377,
-0.0030473529,
-0.0339045636,
0.0419416204,
-0.0166136622,
0.0548299327,
-0.0399849936,
0.0263258517,
0.0234661363,
-0.0211964324,
-0.0063823606,
0.030308431,
-0.0870752558,
-0.030285703,
0.0566802621,
0.0076143332,
-0.0883571878,
-0.0006777572,
-0.0085632689,
-0.0448796414,
0.04018718,
-0.0462800711,
0.065569438,
0.004527085,
0.0153769571,
0.0019637139,
0.00266913,
-0.0270442385,
-0.0204441473,
-0.042373918,
-0.0025615499,
0.0548870116,
0.0458406471,
0.0349219628,
-0.0217160583,
0.0160141885,
0.0011680713,
0.0116758393,
0.0820826292,
0.0348302238,
-0.0081666782,
0.0481894426,
0.0102297962,
-0.0400766879,
-0.0511851758,
0.0006135039,
-0.0050499723,
0.0617918074,
0.0416306667,
0.0423566364,
-0.0278863981,
0.0321123451,
-0.0447877236,
0.024750486,
-0.0573592857,
0.030249605,
0.0037657588,
-0.019335473,
0.0109442556,
0.0012005486,
-0.0604638942,
0.0147830928,
-0.0523782074,
-0.0022141493,
0.0542132556,
0.0264194347,
0.0144153554,
0.0422546081,
0.0628027171,
-0.032596834,
0.0169418603,
0.008106729,
-0.020897679,
0.024088718,
-0.0040910156,
0.0294916425,
0.0010494125,
0.0197720639,
0.0621245988,
-0.0435248613,
-0.0128457351,
0.0300138928,
0.0133255757,
0.0235235281,
-0.0307673626,
-0.0234924555,
-0.0656161755,
0.0025045858,
-0.0569229014,
0.0015156588,
-0.0196018275,
0.0402605422,
-0.0765626281,
-0.0106746163,
-0.0260361601,
-0.0182322711,
-0.0036494671,
-0.0488535687,
0.0070269094,
0.031501174,
-0.0412048548,
0.0057546282,
-0.0424900912,
-0.0654529557,
-0.0262197889,
-0.0505492315,
-0.0120180286,
0.0714156404,
2.461e-7,
0.0189409424,
0.0046437983,
-0.0016869684,
0.0442943648,
0.0212806221,
0.0711218119,
-0.0285977665,
0.0085281935,
0.0298267417,
-0.0466859527,
0.0056082052,
0.0593826063,
0.0042865961,
-0.0084647704,
-0.0327969044,
0.0118298586,
0.0007126099,
-0.0629977435,
-0.0618859082,
0.0130692255,
0.0893070474,
0.0548977926,
-0.0124991741,
-0.0331248976,
0.0424235687,
-0.0715482607,
0.0734313354,
-0.0195818059,
-0.0100270268,
-0.0169254802,
0.0542008467,
0.0053803413,
-0.0191017203,
0.0515521765,
0.0104642138,
-0.0447078981,
0.0300911516,
0.0371972471,
-0.0129164839,
-0.0390456542,
-0.0490560494,
0.0298998319,
-0.0079590473,
-0.0747689977,
0.0904649794,
-0.0219346806,
-0.0225981213,
-0.0102056693,
-0.0377403013,
-0.0374604128,
0.0227795374,
-0.0323210172,
-0.0469154157,
0.0143280709,
0.0278190449,
-0.021253949,
-0.0119601125,
-0.0728764683,
-0.0094820065,
-0.0079598222,
-0.0406215005,
-0.0304294862,
-0.0633938015,
-0.0203737821,
0.034304738,
-0.0535482727,
-0.0450144485,
1.802908491e-34,
0.0096098278,
0.0065058414,
-0.0015547668,
0.0135422191,
0.0287618656,
-0.0169530101,
-0.0131220622,
-0.0332308672,
0.0112700555,
-0.0845900029,
-0.0165935196
] |
http://arxiv.org/abs/2402.05123v1 | http://arxiv.org/pdf/2402.05123v1 | A Survey on Data Selection for LLM Instruction Tuning | 2024-02-04T13:32:01 | 2024-02-04T13:32:01 | Jiahao Wang, Bolin Zhang, Qianlong Du, Jiajun Zhang, Dianhui Chu | cs.CL | cs.CL | Instruction tuning is a vital step of training large language models (LLM),
so how to enhance the effect of instruction tuning has received increased
attention. Existing works indicate that the quality of the dataset is more
crucial than the quantity during instruction tuning of LLM. Therefore, recently
a lot of studies focus on exploring the methods of selecting high-quality
subset from instruction datasets, aiming to reduce training costs and enhance
the instruction-following capabilities of LLMs. This paper presents a
comprehensive survey on data selection for LLM instruction tuning. Firstly, we
introduce the wildly used instruction datasets. Then, we propose a new taxonomy
of the data selection methods and provide a detailed introduction of recent
advances,and the evaluation strategies and results of data selection methods
are also elaborated in detail. Finally, we emphasize the open challenges and
present new frontiers of this task. | A Survey on Data Selection for LLM Instruction Tuning: Instruction tuning is a vital step of training large language models (LLM),
so how to enhance the effect of instruction tuning has received increased
attention. Existing works indicate that the quality of the dataset is more
crucial than the quantity during instruction tuning of LLM. Therefore, recently
a lot of studies focus on exploring the methods of selecting high-quality
subset from instruction datasets, aiming to reduce training costs and enhance
the instruction-following capabilities of LLMs. This paper presents a
comprehensive survey on data selection for LLM instruction tuning. Firstly, we
introduce the wildly used instruction datasets. Then, we propose a new taxonomy
of the data selection methods and provide a detailed introduction of recent
advances,and the evaluation strategies and results of data selection methods
are also elaborated in detail. Finally, we emphasize the open challenges and
present new frontiers of this task. | [
0.0167049263,
0.0663679093,
-0.0204016361,
0.0480134897,
-0.0223503225,
0.0334127322,
0.0218572449,
0.0124981282,
-0.0070402501,
-0.0497374386,
-0.0708139911,
-0.0483516678,
0.0050930949,
-0.0598207265,
0.0319345593,
-0.0297760926,
0.0447729379,
-0.0575165711,
-0.0693406239,
0.012322587,
-0.0224112216,
0.0299006179,
-0.0038059996,
0.0132094538,
0.0327174291,
0.0082389275,
0.0105918963,
0.0267777592,
-0.0161773283,
-0.0463573225,
0.0246168822,
0.045855429,
0.0392126106,
-0.0276526958,
0.0000019046,
-0.0393095762,
-0.0135889361,
-0.0256671049,
-0.047501184,
0.0422099419,
0.0261168405,
-0.0021017042,
-0.0144675719,
0.0311185233,
-0.0048101186,
0.0028650044,
0.0375214145,
-0.0165431648,
0.0611195117,
0.0911613628,
0.0272783637,
-0.0304223821,
0.0034238941,
-0.0041833385,
-0.006725484,
-0.0168812182,
0.0279937051,
0.0365655571,
0.000785646,
0.0406280458,
-0.045611389,
-0.0188055616,
-0.0015934728,
-0.0000576505,
-0.0189155824,
-0.0005828358,
-0.0491103232,
-0.008979599,
-0.0197733287,
0.0650716573,
0.0071799918,
-0.0253839381,
-0.0085618971,
-0.0167237241,
-0.0225674696,
0.0006990358,
-0.0180875137,
0.0248072352,
0.0079383012,
0.0537717529,
0.059315674,
-0.0376150422,
0.015044353,
0.0033513904,
-0.0196046922,
0.0490413457,
0.01945851,
-0.0460380279,
0.0281679388,
-0.0079734419,
-0.0403605849,
-0.0258851256,
-0.0332508385,
0.0292969141,
-0.0322168991,
0.0438195728,
-0.0733183622,
0.002621474,
0.0025884835,
-0.0163292941,
0.0957603306,
0.0378691331,
-0.0049525392,
0.0247466024,
-0.0005199731,
0.0423391722,
0.038409099,
0.0519754738,
-0.0483268388,
0.0232851878,
-0.0503315851,
0.0429032519,
-0.0272185095,
0.0295225699,
-0.0192157365,
0.0430330634,
-0.0470538065,
0.0246716347,
0.0079805674,
-0.0096774604,
0.0405605324,
-0.0757967532,
-0.0529730059,
0.0335909016,
-0.0134814866,
0.0013734011,
-0.0117075481,
0.0218765661,
-0.0320261084,
-0.081483297,
0.0059091593,
0.0280162506,
-0.0146644097,
-0.0359378755,
0.0307311323,
0.0531308018,
-0.0016234881,
-0.0442261808,
-0.0872068703,
0.0106460014,
0.0140973497,
-0.062447384,
0.0245117769,
-0.0521726795,
0.0009506263,
0.0041962415,
0.0246249661,
-0.017817039,
0.0069711721,
0.0127921775,
0.0166073553,
0.0201071799,
-0.0185085405,
0.0239771642,
-0.0152668357,
-0.0254904237,
0.0181823894,
0.0844050571,
0.0415165834,
-0.0250896625,
0.0510959588,
-0.0527143776,
-0.0265940875,
0.0355228297,
0.0139725292,
0.0340519324,
-0.013061611,
0.0315666795,
0.0237328615,
-0.0098728491,
0.023806408,
0.0226050783,
-0.0106616775,
-0.0052873408,
0.0396320224,
0.0184505619,
0.0200782083,
-0.0144246947,
0.0506266728,
0.0258401949,
0.026726298,
0.0402795672,
-0.0277601089,
-0.0016055262,
-0.0043429453,
0.006447535,
-0.0096457666,
0.0290398393,
-0.0427743159,
-0.0534093343,
-0.0161257293,
0.0179502405,
-0.0112011014,
0.0197116435,
0.0410928503,
-0.0166298542,
-0.0564749204,
0.0466488041,
0.0611557215,
-0.0301329661,
-0.0049493364,
-0.0491937771,
0.0061878171,
0.0480498672,
0.0232347529,
-0.027473256,
-0.0414187349,
0.0007100325,
0.0829438567,
0.0043692812,
0.0524700657,
-0.0225977004,
0.0256480481,
-0.0861805528,
-0.0333229527,
-0.0196781494,
0.0113971252,
0.002149702,
0.0233264957,
-0.0304057021,
-0.0165010411,
-0.0182951558,
-0.0956619382,
0.0731153414,
-0.0425353833,
-0.0571131632,
-0.0131193968,
0.015746152,
0.0491145775,
-0.0122287767,
-0.0128004737,
-0.0202510785,
0.0455788299,
0.0048575224,
0.0640214607,
-0.0329988897,
-0.047167141,
-0.0332551189,
0.0092228027,
-0.0181029495,
-0.0163241215,
0.0108471746,
-0.0734982565,
-0.0440132618,
-0.0258975737,
-0.0490173735,
0.0151460422,
-0.0152926939,
0.0545045175,
-0.0279776864,
-0.0099170059,
-0.0428463295,
0.0303393118,
-0.0268539954,
0.017153278,
0.0413370915,
0.026992131,
0.0033793254,
-0.0489320047,
-0.0001308897,
0.0051345094,
-0.0043035569,
0.0443239361,
-0.0032678759,
0.0544768237,
-0.0377715975,
-0.0439839549,
-0.0437833369,
0.0198928975,
-0.0402649306,
0.0359277874,
0.0341962539,
-0.0322015695,
-0.0270687807,
0.006222832,
-0.0168219823,
-0.0265411511,
-0.0232566595,
0.0117995339,
0.0106727006,
0.0050343554,
-0.0426224135,
-0.0423237272,
-0.0341343433,
0.00200903,
0.0063287676,
-0.012923629,
0.0137147401,
-0.0762317032,
0.0715874285,
0.0161133744,
0.017917892,
-0.0163518097,
-0.0579535514,
0.0161679741,
0.0152468411,
-0.0043215002,
0.0735829175,
-0.0136579275,
0.0167339649,
0.0494646244,
-0.0051240157,
-0.0523099117,
0.0068802009,
-0.0231371243,
-0.0392710529,
-0.0666692182,
-0.0584775768,
0.0112518929,
0.0674275309,
-0.0023247162,
-0.0085673267,
-0.0336312801,
-0.022585528,
-0.0396674834,
0.0061845495,
0.0055227769,
0.0387198962,
-0.0054882625,
0.0117897196,
-0.019792784,
-0.0847318992,
0.0585021526,
-0.0322901234,
0.0155342314,
-0.0669853762,
-0.0074965809,
-0.0581786558,
-0.0043918053,
0.0047443793,
-0.0084943827,
-0.0213596541,
-0.0159179252,
-0.056053821,
-0.0025627888,
0.0386948101,
-0.0064653456,
0.0655593649,
-0.0626788884,
0.0059748259,
-0.0008208531,
-0.0018086545,
-0.0171241052,
0.0014860823,
0.0187405441,
0.0153176226,
0.0490653999,
-0.0077965907,
-0.0219871476,
-0.0136794643,
-0.0096987039,
0.0258891247,
-0.0140152052,
0.021930106,
-0.0314770304,
-0.0370713063,
0.0202949904,
0.0065575661,
-0.0208545737,
-0.0803911462,
0.0050659212,
-0.0071516745,
0.0471964851,
-0.0153619228,
-0.0001053414,
0.0377477631,
0.0639933497,
-0.0536967479,
-0.0229469258,
0.0510204174,
0.0907859281,
-0.0583236888,
-0.0028659953,
0.108139202,
0.0271182526,
-0.0255840644,
0.0107735852,
0.0205561873,
0.0207535718,
0.0323208794,
0.0227393303,
-0.0471801013,
-0.0273093581,
-0.0295835305,
0.0530694798,
-0.0160514712,
0.0046626953,
-0.0274580792,
-0.1155511886,
0.036266081,
0.0310283732,
-0.0435686894,
-0.0320546255,
-0.058206901,
0.0120962281,
-0.0296235159,
0.055598598,
0.0157753602,
-0.0692690834,
-0.0201503988,
0.0661356375,
-0.0197354704,
-0.0841654241,
0.0315281898,
-0.0245504696,
0.0633477271,
0.0305711422,
0.0218376536,
0.0000946401,
-0.026097279,
0.0190032106,
-0.0261921026,
-0.0000495255,
-0.0415566228,
0.0594580248,
-0.0079252301,
-0.0012483185,
0.0234055556,
0.0405530296,
-0.0046560033,
0.0289814025,
0.0172210243,
0.0643174648,
-0.0248351675,
0.0165655632,
0.0176331699,
-0.0177495759,
0.0161386523,
0.0492879935,
0.0475635417,
-0.060171172,
-0.0019991675,
-0.0241660345,
-0.0315964781,
-0.0767950714,
0.0043197614,
0.0542629138,
-0.0776758939,
0.0086715091,
-0.0309815742,
-0.0250160359,
0.0708395466,
0.0117957415,
0.0332523659,
0.0244513694,
0.0329497457,
-0.0099187633,
-0.0140205603,
0.0476921834,
-0.0464605503,
0.0090728011,
-0.1206844896,
-0.0884766653,
0.0158239454,
0.0022146483,
0.0029825699,
-0.0030880305,
-0.0691825375,
-0.0591828078,
-0.0007408923,
0.0611310303,
0.0093814451,
0.0519610867,
0.0294960607,
-0.0588881858,
0.0009988606,
-0.0078915311,
-0.0149379401,
0.017427288,
0.0525604673,
-0.0228694305,
-0.0253288765,
0.012227959,
-0.0275681131,
0.031679038,
0.0745356828,
0.0202717558,
-0.0428917594,
0.011334368,
0.019651752,
0.0301651172,
-0.0036416561,
-0.0155443698,
-0.0138263926,
-0.055104401,
0.0583570898,
0.024899438,
0.0225043427,
-0.0426148325,
-0.0034961498,
-0.006367173,
-0.0148944696,
-0.0114343828,
-0.0387506075,
-0.0313300677,
0.0304386448,
-0.0285401568,
0.0782654807,
-0.0032990437,
0.0437834002,
0.0391474776,
0.0236321595,
-0.0131928381,
0.0087648528,
0.0086607402,
0.044177331,
0.026205631,
0.0568686649,
-0.0035095282,
0.0489982255,
0.0080960914,
0.0088312915,
-0.0175230913,
-0.0137335416,
-0.0537942946,
-0.0107061062,
0.0590057895,
-0.0526675656,
0.046390675,
0.0014022435,
0.0136399167,
-0.0373517759,
0.0495153926,
0.0001532444,
-0.022623837,
0.0084469998,
0.0429559313,
-0.0141204363,
0.0043935939,
0.0754740089,
0.0490776747,
0.0465739667,
0.003273529,
0.0216260646,
-0.0441000573,
-0.0019906545,
0.0317075476,
-0.0165695678,
-0.0063046608,
0.0185031034,
-0.034471415,
-0.0582561344,
-0.0482795946,
0.0000688837,
-0.0041278391,
-0.0213927962,
0.0488703921,
0.0247454196,
-0.023219537,
-0.0001530865,
-0.0589236766,
-0.032873407,
-0.0061512329,
-0.0211248323,
-0.0325244814,
-5.666552803e-33,
-0.0490208343,
0.0094522275,
0.0241751727,
0.0078490954,
0.0134567264,
0.0039400123,
-0.0216215774,
-0.0327537917,
-0.0246864259,
-0.024712855,
-0.0000786825,
-0.0082468092,
0.006170189,
0.0294413641,
-0.0002888908,
-0.0035943456,
0.0396231934,
-0.0217070431,
-0.0231118295,
0.0233227387,
0.0291684028,
0.0064523458,
0.0790240988,
-0.0177639872,
-0.0057973536,
-0.0097509203,
0.0274717268,
0.0106335506,
-0.0174874403,
0.0910540298,
0.0358169898,
0.0484966449,
0.01305681,
-0.0462687686,
-0.0254372284,
-0.0434100628,
-0.0246326346,
-0.0540791117,
0.0244019814,
0.0346436761,
0.0125210658,
-0.0446019508,
0.0077959583,
-0.0189663731,
-0.0307890996,
-0.0329421908,
0.0228721723,
-0.028879568,
-0.0412927791,
0.0160845481,
-0.0435126387,
0.0338928364,
-0.0112275397,
-0.0276954733,
0.0267026536,
0.0170565601,
-0.0029507948,
-0.0145683764,
-0.0295734294,
0.0528463274,
0.0806715339,
0.1001383439,
0.0665300936,
0.03513401,
0.0220333952,
0.0265412349,
-0.0063840039,
0.0278894864,
0.0576146506,
0.0075653912,
0.0508972183,
0.0724100173,
0.0123799313,
-0.0275962334,
0.0247446261,
-0.011582382,
-0.0804780349,
-0.006446539,
0.033123374,
0.016546417,
0.0090550557,
-0.0085134394,
-0.0437862426,
-0.0225928016,
-0.0403279997,
-0.0053054607,
0.0042591859,
-0.0110442201,
0.0456659794,
0.0028180394,
-0.0387660712,
-0.0296823159,
0.0415975563,
-0.0265249666,
-0.0450217389,
-0.0094112996,
-0.0185070019,
0.0160181765,
0.0025608332,
0.0125232758,
-0.0899103954,
-0.0906382576,
0.0274658147,
-0.0528071932,
0.0425811969,
-0.004940975,
0.0261847749,
0.0255438276,
-0.0369745418,
-0.0113145942,
-0.0395788848,
0.0123534426,
-0.0073582716,
0.0316880383,
0.0265754592,
0.0115971593,
0.0021247081,
0.0694521815,
-0.003640177,
-0.0380528346,
-0.0275749471,
-0.0191856362,
-0.0086641433,
0.0547276996,
-0.0113466606,
0.0057529383,
-0.0328786075,
-0.011591482,
0.0115182176,
-0.0183020141,
-0.0021342649,
0.0152467759,
2.469e-7,
0.0142072262,
0.0543849654,
-0.014337115,
0.0479775555,
-0.0213772301,
0.002991752,
-0.0360993147,
-0.0202977601,
0.0247122925,
-0.0118894093,
0.0012557417,
0.031377241,
0.0846517906,
0.0213547219,
-0.0717021674,
0.0053377114,
0.0027598981,
0.0410759822,
-0.0441091806,
0.0019462584,
-0.0001672599,
0.1027155891,
-0.0103099709,
-0.0040102261,
-0.0278714411,
0.0149641167,
0.0013022944,
-0.0456697121,
-0.0038959333,
0.0750602856,
0.0255596135,
0.0338783227,
-0.0182856116,
0.0334924385,
-0.0526263639,
0.0109883873,
-0.0450079143,
0.0780579671,
0.0071468432,
-0.0071971575,
0.0353675,
-0.0038525211,
-0.0472256728,
-0.0187952723,
0.053305652,
-0.0432450995,
-0.0124268793,
0.0145533988,
0.0121656777,
-0.0226652138,
0.0016174132,
-0.0270999782,
0.0151382135,
0.0281120632,
0.0365592949,
-0.0511744283,
-0.0169374272,
-0.0231691673,
0.0174462255,
0.0251897182,
-0.0211679395,
-0.0815084428,
-0.0299602374,
0.0032317331,
-0.0024650684,
-0.0232016966,
0.0102822613,
2.442067703e-34,
0.0275853872,
-0.0321287252,
0.0405545942,
0.0363596752,
0.0121882185,
-0.0166343171,
-0.0217967089,
0.0333972834,
-0.0114156641,
-0.0814314857,
-0.0000054283
] |
http://arxiv.org/abs/2403.09674v1 | http://arxiv.org/pdf/2403.09674v1 | Navigating the Peril of Generated Alternative Facts: A ChatGPT-4 Fabricated Omega Variant Case as a Cautionary Tale in Medical Misinformation | 2024-02-04T13:21:19 | 2024-02-04T13:21:19 | Malik Sallam, Jan Egger, Rainer Roehrig, Behrus Puladi | cs.CL | cs.CL, cs.CY, cs.IR, cs.LG | In an era where artificial intelligence (AI) intertwines with medical
research, the delineation of truth becomes increasingly complex. This study
ostensibly examines a purported novel SARS-CoV-2 variant, dubbed the Omega
variant, showcasing 31 unique mutations in the S gene region. However, the real
undercurrent of this narrative is a demonstration of the ease with which AI,
specifically ChatGPT-4, can fabricate convincing yet entirely fictional
scientific data. The so-called Omega variant was identified in a fully
vaccinated, previously infected 35-year-old male presenting with severe
COVID-19 symptoms. Through a detailed, albeit artificial, genomic analysis and
contact tracing, this study mirrors the rigorous methodology of genuine case
reports, thereby setting the stage for a compelling but entirely constructed
narrative. The entire case study was generated by ChatGPT-4, a large language
model by OpenAI. The fabricated Omega variant features an ensemble of
mutations, including N501Y and E484K, known for enhancing ACE2 receptor
affinity, alongside L452R and P681H, ostensibly indicative of immune evasion.
This variant's contrived interaction dynamics - severe symptoms in a vaccinated
individual versus mild ones in unvaccinated contacts - were designed to mimic
real-world complexities, including suggestions of antibody-dependent
enhancement (ADE). While the Omega variant is a product of AI-generated
fiction, the implications of this exercise are real and profound. The ease with
which AI can generate believable but false scientific information, as
illustrated in this case, raises significant concerns about the potential for
misinformation in medicine. This study, therefore, serves as a cautionary tale,
emphasizing the necessity for critical evaluation of sources, especially in an
age where AI tools like ChatGPT are becoming increasingly sophisticated and
widespread in their use. | Navigating the Peril of Generated Alternative Facts: A ChatGPT-4 Fabricated Omega Variant Case as a Cautionary Tale in Medical Misinformation: In an era where artificial intelligence (AI) intertwines with medical
research, the delineation of truth becomes increasingly complex. This study
ostensibly examines a purported novel SARS-CoV-2 variant, dubbed the Omega
variant, showcasing 31 unique mutations in the S gene region. However, the real
undercurrent of this narrative is a demonstration of the ease with which AI,
specifically ChatGPT-4, can fabricate convincing yet entirely fictional
scientific data. The so-called Omega variant was identified in a fully
vaccinated, previously infected 35-year-old male presenting with severe
COVID-19 symptoms. Through a detailed, albeit artificial, genomic analysis and
contact tracing, this study mirrors the rigorous methodology of genuine case
reports, thereby setting the stage for a compelling but entirely constructed
narrative. The entire case study was generated by ChatGPT-4, a large language
model by OpenAI. The fabricated Omega variant features an ensemble of
mutations, including N501Y and E484K, known for enhancing ACE2 receptor
affinity, alongside L452R and P681H, ostensibly indicative of immune evasion.
This variant's contrived interaction dynamics - severe symptoms in a vaccinated
individual versus mild ones in unvaccinated contacts - were designed to mimic
real-world complexities, including suggestions of antibody-dependent
enhancement (ADE). While the Omega variant is a product of AI-generated
fiction, the implications of this exercise are real and profound. The ease with
which AI can generate believable but false scientific information, as
illustrated in this case, raises significant concerns about the potential for
misinformation in medicine. This study, therefore, serves as a cautionary tale,
emphasizing the necessity for critical evaluation of sources, especially in an
age where AI tools like ChatGPT are becoming increasingly sophisticated and
widespread in their use. | [
0.0064000245,
0.0369178317,
-0.0140448436,
-0.0097382441,
-0.0124336509,
0.0033181733,
0.0150653236,
-0.0171878058,
0.0544810407,
0.0240425766,
0.0351017006,
0.0373760872,
0.0416350774,
0.0875687003,
-0.0100160278,
0.008803416,
0.037952438,
0.0351275131,
0.0336169302,
-0.0054296916,
-0.0773465335,
0.0357803106,
0.0123042874,
0.0324297473,
0.0421074256,
-0.0429212153,
0.0468083285,
-0.0179095864,
0.0182861332,
-0.0540718772,
0.0075630182,
-0.0136082573,
-0.0095764166,
0.0600245707,
0.0000024483,
-0.0395861603,
0.0020335391,
0.0294262003,
-0.0368254855,
-0.0167830382,
0.0418946743,
0.0080775162,
0.0030700483,
0.0207951088,
-0.0398850776,
-0.0371646695,
0.0665324554,
0.0185605567,
0.0139292702,
0.084776938,
-0.0061830231,
-0.027846355,
0.0912638158,
0.0286806244,
0.035512995,
-0.0070076147,
0.046235323,
0.0117338337,
0.0535332486,
0.0176439732,
0.0578201376,
0.056531895,
-0.0104681225,
-0.0080780191,
-0.0166740101,
0.0751181021,
-0.007751781,
-0.0452741235,
0.0283526853,
0.0159901809,
0.0678401589,
-0.047471121,
0.0243348982,
0.0211043023,
-0.022798758,
-0.0031703641,
-0.0010365733,
-0.0115997,
-0.0491653085,
0.0087661846,
0.040658284,
-0.0054490496,
0.0521989726,
0.0129086338,
-0.0089913616,
0.011602887,
-0.0477668643,
-0.0312975496,
-0.0430987664,
-0.0371755473,
-0.0162213873,
-0.0871608332,
0.0223551635,
0.038809035,
-0.0498701967,
0.0148933558,
-0.004034989,
-0.0250045247,
0.0581753328,
-0.121180132,
0.0018852772,
-0.0019467202,
0.0043913075,
-0.0273692254,
0.026702961,
-0.0204025749,
-0.0483947136,
0.0725684389,
-0.0219428893,
0.0656042919,
-0.0382437408,
-0.0016344906,
-0.0265889764,
-0.0207686517,
0.0051807011,
0.0213661976,
-0.04874501,
-0.0348579288,
-0.0341045968,
0.0441873819,
-0.0382729843,
0.0485008769,
-0.0418925174,
0.0407292061,
0.0093592824,
0.0215637069,
0.0093815699,
0.0165919792,
0.0066912165,
-0.0613644458,
0.0019494944,
-0.0377522856,
0.0086899213,
0.0157891698,
0.0116915302,
0.040876586,
-0.0082650539,
0.0174118783,
0.0510231145,
-0.0164498333,
-0.0470239893,
-0.0657151416,
-0.0122029632,
-0.0368493497,
0.0129982643,
0.0440069661,
-0.0452100895,
-0.0299647544,
-0.0238532405,
-0.0145196347,
-0.1169697642,
0.0368607901,
-0.0798195004,
-0.0088484213,
0.0574240722,
0.0385192893,
-0.0127053615,
-0.0108833387,
0.0382334776,
0.0056594037,
0.0142523507,
0.0064123734,
0.0466524437,
-0.0405146852,
0.021591194,
0.0407974459,
0.0434888005,
-0.0019946175,
-0.0162515137,
0.0603663363,
-0.0290696807,
0.0034088048,
-0.0064644855,
-0.0217667408,
0.0361749306,
-0.0150234392,
-0.0279541258,
0.0027042152,
0.0277905185,
0.0373366252,
0.0338515602,
0.0238017198,
0.0010048926,
-0.0217478145,
-0.058247555,
0.0124580665,
-0.0364785455,
0.0126478616,
-0.0023923828,
-0.0197955482,
-0.0165014546,
-0.0081050256,
0.0022922039,
-0.0015955792,
-0.030091051,
-0.0429361872,
-0.0223572589,
-0.0551478229,
0.0680256709,
0.0137260668,
-0.0150320958,
-0.0980134308,
-0.02917891,
0.0627955422,
0.0322559513,
0.0051834229,
-0.0510454252,
-0.0396877155,
-0.001276052,
-0.0124796843,
0.0495242923,
-0.0049985629,
-0.032116551,
0.0115160979,
0.0440534987,
-0.0594594181,
0.0422558226,
0.0050143306,
-0.0418961756,
0.0185938831,
-0.0203394629,
0.0027993401,
-0.0404796228,
0.0036430731,
-0.1004016176,
-0.0191629715,
-0.0115844635,
-0.0038521804,
0.0621432886,
0.0012685717,
0.0183034576,
0.054036878,
0.0520459153,
-0.0248836521,
-0.0247989316,
0.0154909464,
0.0371492654,
0.038540829,
-0.0368588157,
-0.0195719898,
-0.0200643428,
-0.0048225923,
-0.0266907252,
0.0242494233,
-0.0124722775,
-0.0235900749,
0.0211377256,
0.0135248648,
-0.0022006752,
0.005819587,
0.039130345,
0.0191832073,
0.0046946411,
-0.0091165043,
-0.0294124372,
0.0188194793,
-0.0286653638,
0.0078979917,
-0.0415728055,
0.0576431863,
-0.0260751545,
0.0152045023,
0.0006656905,
-0.0336788967,
-0.0538209826,
0.0225675739,
-0.0016686105,
-0.0374635719,
-0.00969001,
0.0401809551,
0.0006959403,
0.0158734974,
-0.0357431434,
-0.0301353857,
0.0174293742,
0.0389358513,
-0.0382997543,
0.0572479963,
0.024078412,
0.0159710366,
0.0191014688,
-0.0149160679,
-0.0386548974,
0.0175670572,
-0.0789197758,
0.0262224972,
0.0528172702,
-0.0020390153,
-0.0477648564,
-0.038477283,
-0.0607285239,
0.0262378603,
-0.012376165,
-0.0114513543,
0.0176797956,
0.0054482133,
-0.0263605416,
0.0323032588,
0.0102694696,
0.0566770732,
0.0567277037,
0.046581991,
0.0390439741,
0.0039988249,
-0.0223297477,
-0.0163256247,
0.0124139804,
-0.0481139421,
-0.0200945362,
0.0312722176,
0.0418359414,
0.0114330463,
-0.0181834418,
0.0521922931,
-0.0075152847,
-0.0551607087,
0.0476933382,
-0.0885500088,
-0.0397787243,
0.020719219,
-0.033041399,
-0.016083112,
0.0430733487,
-0.0053476701,
0.0490029529,
-0.0926669613,
0.0093398336,
0.013287859,
-0.0199781489,
-0.0099894097,
0.0193103477,
0.0825041533,
0.0383521281,
0.0409149416,
-0.0251125395,
-0.0111342361,
-0.0528276376,
-0.0302725285,
-0.0074634627,
0.0111181065,
0.0487174615,
-0.0017819001,
-0.0328683183,
-0.0074942447,
-0.0461342745,
-0.025951203,
0.0208220761,
0.0004091211,
-0.0704413876,
-0.0470995419,
0.0205906034,
0.0514618456,
0.0142334029,
0.0013099099,
0.0158479158,
-0.0313324519,
0.0181630906,
-0.0367954038,
0.0490698889,
-0.0216348413,
0.0190163683,
0.0179618746,
-0.0369169079,
0.0035480685,
-0.0431429669,
-0.0043319492,
-0.032119479,
-0.002904752,
-0.021669412,
0.0519539379,
0.0081196371,
-0.0382946134,
0.0036864185,
0.0652535632,
0.0001728726,
0.0052707265,
0.036636319,
-0.018175438,
-0.0631681681,
-0.0363065898,
0.0495372377,
-0.0753815621,
0.0190531667,
-0.0021356093,
-0.0102467164,
-0.084757261,
-0.0487425514,
-0.02963477,
0.0355121046,
0.086751014,
0.0150138261,
-0.0344901383,
-0.0405418389,
-0.0632011592,
-0.0273742024,
-0.0123752644,
0.0225120783,
-0.0226140358,
0.0501886681,
0.0096402289,
-0.0200516973,
-0.0074481335,
-0.0187721886,
0.0250793882,
0.014197642,
0.0358601995,
0.0430826023,
0.06252148,
-0.0314396657,
-0.026669709,
-0.0295467135,
0.0682131276,
0.0165926646,
-0.0057790237,
0.0964393243,
-0.0114730606,
0.024623936,
-0.0704785958,
0.0229704808,
-0.0157033615,
0.0086970264,
0.0073407455,
0.0705211237,
-0.0314595737,
-0.0989115536,
-0.0361314267,
-0.0001333624,
0.0304565541,
0.040785145,
-0.0035610015,
0.0458932817,
0.030707702,
0.0219415464,
0.0993065163,
0.0079775508,
-0.0161749348,
0.0336864144,
-0.00612775,
-0.1120460257,
-0.0147172352,
-0.0445525981,
-0.0312241111,
-0.0219480544,
0.0108064869,
-0.0663224906,
0.0170047637,
-0.03403854,
-0.0212079044,
0.0257803053,
0.0377285443,
0.0357308723,
0.0279466305,
-0.026905667,
0.0241372604,
-0.0119248684,
0.0023306671,
0.0457201228,
0.0230423175,
-0.0507480241,
0.0419569947,
0.0209513903,
0.0296218693,
0.0220691301,
0.0109606655,
0.0261398163,
0.0457393453,
0.0372129567,
-0.025392795,
-0.017050989,
-0.0385176651,
-0.0207969528,
-0.0064234971,
-0.0112528047,
0.0334592015,
0.0277754273,
-0.0115765892,
-0.033138562,
-0.0463837497,
-0.0001615111,
0.0446608253,
0.0073331902,
0.0418789834,
-0.0444989204,
-0.0316485353,
-0.0107780369,
-0.0006855337,
-0.0389489271,
0.01692014,
0.0186548419,
-0.0541747622,
-0.0294712279,
0.0052681672,
-0.0416381024,
0.0223112293,
-0.0871640071,
-0.0063342671,
-0.0567400753,
0.053899508,
0.0311756097,
0.0783864483,
0.0449210368,
-0.0352715664,
0.0111683253,
0.0521997958,
-0.0176127609,
0.0139357978,
0.0142138517,
-0.0154899852,
-0.0068735299,
0.0047709453,
0.0010240072,
0.0192847997,
-0.0599698722,
-0.0495528318,
-0.038005624,
0.0289522428,
-0.0104513522,
0.0474394858,
-0.005717142,
-0.0092096869,
-0.0136115346,
-0.0414137766,
0.0139625007,
-0.0079223309,
0.0105641698,
-0.0357509814,
-0.0104013989,
-0.0605801009,
-0.0056173597,
0.036985334,
-0.0136078885,
0.1192104071,
-0.0217860844,
-0.0569223985,
-0.0148879923,
-0.052830752,
-0.0209561996,
0.0180881638,
-0.0482799485,
-0.0182599351,
-0.0706398636,
-0.0214525852,
0.0327619463,
-0.0019707608,
-0.0132174212,
0.0034563867,
-0.0490564443,
0.0392955579,
0.0210724585,
-0.0479319133,
0.0137380268,
0.0629794225,
-0.039337635,
-0.022663828,
0.0016388847,
-7.066619472e-33,
0.0372949913,
-0.0258378629,
0.0311396383,
0.0029275604,
0.0049258913,
0.073749885,
-0.0631405711,
-0.0303678773,
-0.0427382886,
0.0475826114,
-0.0166025143,
-0.0097883837,
0.007550831,
-0.0313590765,
-0.0048417612,
0.0002698936,
0.0670082942,
-0.0000712389,
-0.0120620504,
-0.0028532569,
0.0092364065,
0.0195958391,
0.0059857839,
-0.0904901028,
0.0394625887,
-0.02727158,
0.0127255004,
-0.0159910955,
-0.04668146,
0.0212174393,
0.0353701338,
-0.0057738917,
0.0107820462,
0.0332247093,
0.0047502429,
-0.0280723311,
-0.0001540125,
0.0335833281,
0.0014706518,
-0.0491199233,
-0.0180242937,
-0.0345592499,
0.0403566137,
0.0121585587,
-0.0146503178,
0.0095228357,
0.0202144813,
-0.0549949519,
-0.0044131801,
-0.0124480566,
-0.0518235117,
-0.00268144,
-0.0098159229,
0.0167638231,
-0.0584396347,
-0.0039801807,
-0.0379355513,
0.044627022,
0.0127186757,
0.0272903182,
0.071005553,
-0.0260664672,
0.0256359521,
-0.0334736146,
0.0083247516,
0.0074972264,
0.0644944683,
0.0024469455,
-0.0872588754,
-0.0030869872,
-0.0044360468,
0.0663294718,
0.0086380141,
0.0421799943,
-0.024248872,
-0.0585145429,
-0.0701815933,
0.0461351871,
-0.0334914662,
0.02300434,
-0.0274966303,
-0.0245651565,
-0.0565702356,
0.0003556915,
-0.0132825244,
-0.0414020158,
-0.0428513885,
-0.0094094006,
-0.0209098253,
-0.0133765619,
-0.0265745744,
0.0563851297,
-0.0171564072,
-0.0229553804,
0.0336009227,
-0.0478828177,
0.0389520936,
-0.0186259877,
-0.0153972218,
0.0152918873,
-0.0359806344,
0.0133361109,
0.0495910048,
0.0002137566,
0.0467920341,
0.0106955245,
0.0188831743,
-0.0047475975,
-0.0127389235,
0.0165705364,
-0.0067627821,
0.0172631256,
0.0147041623,
-0.0250250623,
0.0336101949,
0.0071754465,
0.0233876482,
-0.0091700805,
0.0266218893,
0.0727168024,
-0.0080382721,
-0.0235944185,
-0.0564597994,
0.0338568725,
-0.0378413126,
0.03251395,
-0.0040967735,
0.0339386202,
-0.0195485856,
-0.1061669961,
-0.004988926,
-0.011767582,
3.378e-7,
0.0418820567,
0.0907657444,
0.0038956008,
-0.0258020051,
0.0146591431,
0.0332933478,
-0.0297440831,
-0.0086058751,
0.0188974254,
-0.0325384103,
0.0290309247,
0.0293167271,
0.0068672597,
0.0143611953,
0.0169269964,
0.0096047483,
-0.0503832288,
0.0488230325,
-0.0278338399,
-0.0028698139,
0.0203580149,
0.0009026749,
-0.0038268592,
-0.0189850889,
0.0135551197,
0.0223119855,
-0.0005997295,
0.0376939774,
-0.010529764,
0.0201820042,
-0.0151677309,
-0.0103346277,
0.0254675075,
0.0450784154,
-0.0724986494,
-0.0729434267,
-0.0399162509,
0.0402822904,
-0.0276302602,
0.04808053,
0.0212477148,
-0.0455139503,
0.0312965885,
-0.025787292,
0.0215277113,
-0.0348748118,
-0.0552586168,
-0.0074553033,
-0.0502343103,
0.0291219726,
0.0277386494,
0.0138145462,
-0.0210415423,
0.0540795475,
-0.0304249153,
0.0553135686,
-0.0043351599,
-0.0174719393,
0.0049923779,
-0.0177253243,
-0.0371814892,
-0.039684467,
0.0439150967,
-0.0048875585,
0.0370617397,
-0.0398241878,
0.0260173269,
3.057147418e-34,
0.0321267173,
0.022631092,
-0.0258820802,
-0.0544018559,
0.0280262586,
-0.0053149653,
0.0162593089,
0.0147406952,
0.0250765383,
0.0022208057,
-0.0260476656
] |
http://arxiv.org/abs/2402.02479v2 | http://arxiv.org/pdf/2402.02479v2 | BRAIn: Bayesian Reward-conditioned Amortized Inference for natural language generation from feedback | 2024-02-04T13:16:29 | 2024-06-10T10:18:46 | Gaurav Pandey, Yatin Nandwani, Tahira Naseem, Mayank Mishra, Guangxuan Xu, Dinesh Raghu, Sachindra Joshi, Asim Munawar, Ramón Fernandez Astudillo | cs.LG | cs.LG, cs.AI, cs.CL, cs.HC | Distribution matching methods for language model alignment such as Generation
with Distributional Control (GDC) and Distributional Policy Gradient (DPG) have
not received the same level of attention in reinforcement learning from human
feedback (RLHF) as contrastive methods such as Sequence Likelihood Calibration
(SLiC), Direct Preference Optimization (DPO) and its variants. We identify high
variance of the gradient estimate as the primary reason for the lack of success
of these methods and propose a self-normalized baseline to reduce the variance.
We further generalize the target distribution in DPG, GDC and DPO by using
Bayes' rule to define the reward-conditioned posterior. The resulting approach,
referred to as BRAIn - Bayesian Reward-conditioned Amortized Inference acts as
a bridge between distribution matching methods and DPO and significantly
outperforms prior art in summarization and Antropic HH tasks. | BRAIn: Bayesian Reward-conditioned Amortized Inference for natural language generation from feedback: Distribution matching methods for language model alignment such as Generation
with Distributional Control (GDC) and Distributional Policy Gradient (DPG) have
not received the same level of attention in reinforcement learning from human
feedback (RLHF) as contrastive methods such as Sequence Likelihood Calibration
(SLiC), Direct Preference Optimization (DPO) and its variants. We identify high
variance of the gradient estimate as the primary reason for the lack of success
of these methods and propose a self-normalized baseline to reduce the variance.
We further generalize the target distribution in DPG, GDC and DPO by using
Bayes' rule to define the reward-conditioned posterior. The resulting approach,
referred to as BRAIn - Bayesian Reward-conditioned Amortized Inference acts as
a bridge between distribution matching methods and DPO and significantly
outperforms prior art in summarization and Antropic HH tasks. | [
0.0336299203,
0.1246237457,
0.0158300288,
-0.0304671209,
-0.0176890269,
0.0182576869,
-0.0131826028,
-0.0188291166,
-0.0426920354,
-0.0004452313,
-0.0397559889,
-0.0666607916,
0.0622040108,
0.0200397559,
0.0436114781,
-0.0623248108,
0.0087273316,
-0.0055883052,
0.0172513947,
-0.0076266332,
0.006566179,
0.0255168695,
-0.0394090675,
0.0626867712,
0.0285372604,
0.0074880887,
0.0034224209,
-0.0013139684,
0.0381621122,
-0.0290708151,
-0.0036046507,
0.0565287247,
0.0556010418,
0.0064241602,
0.0000019392,
-0.0227953102,
0.0023194123,
-0.0198583212,
0.0128543973,
-0.0362444595,
0.0663498193,
-0.016037561,
-0.007770909,
0.0231755711,
-0.0017094156,
0.0136214364,
0.0565877743,
0.0377902538,
0.0275682155,
0.0585472099,
0.0069232215,
-0.0339094922,
0.0234327279,
-0.0597133711,
0.0384751298,
0.0010899251,
0.0163387656,
0.0289422814,
0.0271706078,
-0.0305637531,
-0.0208697896,
0.0163995177,
0.026475165,
-0.0139802052,
-0.0072235549,
0.0695840567,
0.0419421531,
0.0089610163,
-0.0158579256,
0.0578561202,
0.0011288482,
-0.030543454,
0.0187675282,
-0.0169087593,
0.0015207187,
0.0050999238,
-0.0100525459,
0.0932430103,
-0.0304174889,
0.0511592962,
0.0370969065,
-0.0342528895,
0.0165249594,
0.0185879339,
-0.0059491815,
0.0685382038,
0.0336709097,
-0.0173929259,
-0.0013495642,
0.0470773689,
-0.0819748715,
-0.0714743659,
0.007054023,
0.0093848789,
-0.0683918968,
0.0016323521,
-0.1004563868,
-0.0395241231,
-0.0179538578,
-0.0024256974,
0.0407067239,
0.0623363443,
0.0166054573,
-0.0178641975,
0.0017913453,
0.0531730764,
-0.0305346027,
-0.0057453699,
-0.0378458574,
0.0467278771,
-0.0125194099,
-0.0004042831,
-0.0103837205,
0.0265978947,
-0.0458863042,
-0.0333854072,
-0.0808847398,
0.0132462531,
-0.0147266844,
-0.0031043245,
0.0058037126,
0.0377317742,
-0.0539121702,
0.0467218608,
-0.0368513539,
-0.0096176183,
-0.0512835048,
0.0136797838,
0.052873835,
0.0226635877,
-0.0393815003,
0.0492685176,
-0.0009754673,
-0.0179512091,
0.0068157935,
0.0539366268,
-0.0220945459,
0.0100115137,
-0.0361441821,
-0.035246145,
-0.0425287187,
-0.0430568233,
0.0123990029,
-0.0797690451,
0.0155837471,
-0.0261458438,
-0.0508863926,
-0.0237251483,
0.0043293764,
0.0138925808,
-0.0297678877,
0.0281600617,
0.0113121094,
-0.0116321659,
0.0386671647,
-0.0256032627,
0.0061037843,
0.0522406325,
0.0759282038,
0.0273237992,
-0.0453454144,
0.0583604835,
-0.0044580521,
0.0091724219,
0.0396939144,
0.017973857,
-0.0563106686,
-0.0393761843,
-0.0582964681,
0.068340838,
0.1202152967,
0.0370179452,
-0.0459141284,
0.0052288324,
0.046754431,
0.0242716689,
-0.0040960177,
-0.0534739867,
0.0616208501,
-0.0039458866,
-0.0362016819,
0.0419702567,
-0.0378070883,
0.0346694253,
-0.0076524983,
0.0431707725,
0.0581279211,
0.0012738605,
-0.0011825218,
-0.0397389121,
0.0243959762,
-0.0077676717,
-0.0091971438,
0.0073001957,
0.0201841425,
0.0223893411,
-0.0411365405,
0.0356801935,
0.0595522337,
-0.0195467714,
0.020483844,
-0.0299313832,
0.0441600643,
0.0301295873,
0.0515147448,
0.0359325148,
-0.0147723053,
-0.0102768457,
-0.0551367737,
0.0911069959,
-0.0076454855,
0.0088831261,
-0.028503729,
-0.0262397155,
-0.0405474566,
-0.0083546918,
0.0140786627,
0.0283881985,
0.0009929157,
0.0041141859,
0.0103121679,
-0.0081063509,
-0.0507757775,
0.0421699062,
-0.0006894557,
-0.010046443,
0.0273154769,
-0.0026871681,
0.0392843299,
-0.0098227179,
-0.0143509395,
0.0380147025,
-0.0227273479,
-0.013657446,
-0.0265677702,
0.0205314476,
-0.0573342182,
-0.0813411102,
0.0223370753,
-0.029564146,
0.0124110766,
-0.0035515772,
-0.008667673,
-0.0507035218,
0.0388931893,
-0.0245803315,
-0.0003859616,
-0.0122769829,
0.033567369,
0.0208901074,
0.0207492933,
-0.0464489162,
0.0572762936,
-0.0287928265,
-0.0222770479,
0.031566754,
0.0179329123,
-0.0287222676,
-0.0493159629,
-0.0046482785,
-0.0043117497,
0.0110568842,
-0.0176816303,
0.049925372,
0.0408868529,
-0.0384631529,
0.0019145148,
-0.0436826684,
0.0344986655,
-0.0099809701,
0.0425808504,
0.0605968647,
0.0078837955,
-0.0094581265,
0.0055789337,
-0.0489835329,
-0.0107377814,
0.0363604128,
0.0093076974,
0.0232424922,
-0.0137254102,
0.0126386778,
-0.0589013509,
-0.0004631227,
-0.0287424047,
0.0217677094,
0.0042160852,
0.0230989829,
-0.06236507,
0.0481561162,
-0.0504571609,
0.0243389215,
0.0011647567,
-0.0072577088,
0.0324207284,
0.009764798,
0.0810110569,
0.0540751033,
0.0353269838,
0.0039450401,
0.0614991933,
0.0076135965,
-0.0422101729,
0.0082993703,
-0.0614339449,
-0.0102449823,
-0.0097226296,
-0.028364202,
0.0219985377,
0.0893699974,
0.0404291861,
0.0376118906,
-0.0700824633,
0.0213043261,
-0.0017227979,
-0.0229352489,
-0.0164373219,
0.0024781812,
-0.0275288206,
0.0051059499,
-0.0415074043,
-0.0504077077,
0.102352865,
-0.0452138595,
0.0347432569,
-0.0530936606,
0.0017778932,
-0.0234406255,
-0.0148266852,
0.0289646853,
0.0288962461,
0.0057489593,
-0.0327983424,
-0.0592160858,
-0.0335182771,
0.0186115485,
-0.0073889447,
0.0378927588,
-0.0532379486,
0.0130379023,
-0.0087640211,
-0.0121844523,
-0.0622198991,
-0.0568338893,
0.0094749779,
0.0180732589,
0.0398910977,
-0.047129415,
-0.0365653858,
0.0277962796,
0.051649116,
0.0447619483,
-0.0341981538,
-0.0265717041,
-0.0238017291,
0.0506451316,
0.1127860025,
0.0167808365,
0.020532269,
-0.0290823486,
-0.0524638258,
0.0240760613,
0.0590222664,
0.0590441264,
-0.0056722742,
0.0298756007,
0.0211069938,
-0.0407052264,
0.0244952943,
0.0343242176,
0.0389260985,
-0.0073776781,
-0.0140311606,
0.1173605546,
-0.000611319,
-0.091142945,
0.0127949743,
-0.0015532572,
0.1024770141,
-0.0293724257,
0.0133505417,
-0.0950118005,
-0.0054283119,
0.005493497,
0.0493932106,
-0.0021550816,
-0.0480085611,
-0.0158226285,
-0.087644577,
-0.0029605215,
0.0268860031,
-0.0568806939,
-0.0172309224,
-0.0135012772,
0.0188110415,
-0.0049941214,
0.020409124,
0.014777719,
0.0062514157,
0.0490200631,
0.0382329486,
0.0260858685,
-0.037806619,
-0.013838226,
-0.0485561043,
0.0395693853,
0.1024668217,
0.0245486293,
0.0278373547,
-0.0152295204,
-0.0280242097,
-0.0258952156,
-0.0248788297,
-0.0069404608,
0.0160194896,
-0.0253442638,
0.0205404945,
-0.0156829469,
0.0714827701,
-0.0207296908,
-0.0153124267,
-0.0137988031,
0.0322480723,
-0.0328343995,
-0.0015198783,
-0.0362806544,
0.0518420748,
0.0509377718,
0.0590297319,
0.0423239209,
-0.0608712696,
-0.0324549265,
-0.0058733649,
-0.0282858666,
-0.0475645848,
-0.0129389428,
0.0132521959,
-0.0107056946,
-0.0357367881,
-0.1253807396,
-0.0215894356,
-0.024471771,
0.0394469984,
0.0050592567,
-0.0180890057,
0.0071975072,
-0.0585078225,
-0.0193643216,
0.0241811275,
-0.022432996,
-0.0006115991,
-0.0663730949,
-0.0319016464,
-0.0483393557,
-0.0463126749,
-0.0273167156,
0.0316917114,
-0.0787245184,
-0.0857531875,
-0.023730889,
0.0332323425,
0.013850892,
0.0497175604,
0.0104044909,
-0.01291057,
0.0593777224,
0.0213231072,
0.0029281261,
0.0488580987,
-0.0290619619,
0.0384639837,
0.0083200438,
0.029876411,
-0.0196657181,
0.0212877449,
0.0260659363,
0.0750214234,
0.0386677124,
-0.0252499711,
0.0238323156,
-0.0119281774,
0.0022139535,
-0.0943473279,
0.0059723766,
-0.0427957326,
-0.048162967,
-0.0215682779,
-0.0181405209,
-0.0351609774,
-0.0176030993,
0.0184144787,
0.0091028335,
0.0060150358,
-0.0228216387,
-0.0174014792,
0.0098586809,
-0.0149126826,
-0.0347318389,
0.0050554611,
0.0044907061,
0.0096671265,
0.0156854559,
0.020781571,
-0.0150358463,
-0.0432982296,
0.0628847554,
-0.01043001,
0.0730198175,
0.04079853,
-0.0276397113,
-0.003714957,
0.009685182,
-0.0622514524,
-0.0150031066,
-0.0162524823,
-0.0033269944,
0.0574782565,
-0.0067348839,
0.0702301189,
-0.01601943,
0.0190799702,
-0.0077440976,
0.0389596708,
-0.034772981,
-0.0124523221,
0.0307996944,
0.0048450334,
-0.0004175325,
-0.0271574166,
0.0789268687,
0.0237928014,
0.0112506095,
-0.0446936823,
0.0023494337,
-0.0193365254,
-0.0437659174,
-0.0293015223,
-0.0019186366,
-0.0291624367,
-0.0170662198,
-0.0251584016,
-0.053088814,
-0.0286956113,
0.0352747999,
0.0057752514,
0.0007854202,
0.057462465,
0.0320183374,
-0.0565088131,
-0.0367973931,
-0.0461165346,
0.0288798995,
-0.0462255888,
-0.0320644341,
-0.0857112333,
-6.252159109e-33,
-0.0515019745,
-0.0455409437,
0.0325779542,
-0.021548558,
-0.052014716,
-0.0177464653,
-0.0611795001,
-0.0021419616,
0.0396879092,
-0.0097644366,
0.0006330229,
0.0008712928,
0.0092395451,
0.0236926842,
0.0163353384,
0.0176822394,
0.0293623861,
-0.0092301741,
-0.0801603347,
0.0199849252,
0.0408607014,
0.0042552729,
0.0535077937,
-0.0068610483,
0.0281630848,
0.0115148658,
-0.0173688866,
-0.0050752698,
0.0001015502,
0.0290363543,
-0.0101654138,
0.0111357886,
0.0010725906,
-0.0013729364,
-0.021053059,
0.0182838161,
-0.0230342541,
-0.0147073762,
0.0407404564,
-0.0263251308,
-0.0643713772,
-0.0184976496,
0.0419970341,
-0.009442416,
-0.0123026054,
0.0141410828,
0.0114498446,
-0.0209347773,
-0.0022264521,
-0.0079723084,
-0.0619291104,
0.0074743368,
0.0088870302,
-0.0012988133,
-0.0156186353,
0.0223798342,
-0.0220072083,
0.0148465978,
-0.0251622852,
0.0356056169,
0.0342315361,
0.0423286222,
0.0749555379,
0.0515747368,
0.035263937,
0.0315184109,
0.0283556506,
-0.0277479254,
-0.0435362123,
0.0044571823,
-0.0021326167,
0.0028921762,
-0.0228784122,
0.0168890562,
0.0470152572,
0.0375257619,
-0.0209392123,
0.0526449606,
-0.0029931469,
0.0318458714,
0.0110772233,
0.0154648768,
-0.0262142662,
-0.0151139433,
-0.0406681001,
-0.026995331,
-0.0162246097,
-0.0521368161,
0.0069404584,
0.0163492449,
-0.0100698564,
-0.0550637655,
0.0154977692,
-0.0000360425,
0.0052109081,
-0.0053027347,
-0.0423287973,
-0.0083987257,
0.0121463081,
-0.0045959763,
-0.0893876404,
0.0116645377,
0.0210516844,
0.0084299454,
-0.0122747142,
0.0376799516,
0.0003080293,
0.0113017382,
-0.0075563551,
0.0257112961,
-0.0386976749,
0.0168838892,
-0.0118052065,
0.0227208287,
0.0456817448,
-0.0167250894,
0.0007701277,
-0.0056568556,
-0.004018974,
-0.0380140282,
-0.0025449081,
0.0471201204,
0.0038742803,
0.0562510192,
-0.0271636397,
0.0253720507,
-0.0013173093,
0.0325484127,
0.0124976933,
-0.0767586678,
-0.0383426324,
0.0159152243,
2.688e-7,
0.0077849333,
-0.0063079577,
0.0157872681,
0.0223569609,
0.0006404327,
0.0206512641,
0.0279643107,
0.0014226536,
0.0058994009,
-0.029523775,
0.0262046754,
0.0242061485,
0.0441477261,
-0.0075046849,
-0.0483519696,
-0.0243669488,
-0.1002390385,
0.0208780523,
-0.0134171927,
-0.062059328,
0.0135178678,
0.027610762,
0.0149600366,
-0.0315754265,
0.0322222337,
0.0247235652,
0.0620210804,
0.0194317531,
-0.0234065987,
0.0087332064,
-0.0116326464,
-0.0221722648,
-0.0103708729,
0.0075365603,
-0.0034818244,
-0.0014583567,
-0.029473247,
0.0337292142,
0.0244393516,
-0.0276250821,
-0.0238411129,
0.0101878829,
0.0189224891,
-0.0365052745,
0.0190764274,
-0.0450486615,
-0.0317297801,
-0.0404703803,
0.0120958053,
0.0008711586,
0.0035775253,
-0.0222074371,
-0.0218311604,
-0.0087469993,
0.0306777637,
0.0516434871,
0.0008589139,
-0.0417221412,
-0.030617753,
0.0525626615,
-0.0487875268,
-0.0033656333,
-0.0570374168,
-0.0045936201,
-0.0105712591,
-0.08135584,
-0.0246382561,
2.604542374e-34,
0.012282921,
-0.0221732315,
0.0112996129,
-0.0167548023,
0.0131991645,
0.0318781212,
0.0033574342,
-0.0025674915,
0.0393813662,
-0.0632486567,
-0.0614859276
] |
http://arxiv.org/abs/2402.02456v2 | http://arxiv.org/pdf/2402.02456v2 | tnGPS: Discovering Unknown Tensor Network Structure Search Algorithms via Large Language Models (LLMs) | 2024-02-04T12:06:13 | 2024-06-01T15:54:54 | Junhua Zeng, Chao Li, Zhun Sun, Qibin Zhao, Guoxu Zhou | cs.LG | cs.LG, cs.CL | Tensor networks are efficient for extremely high-dimensional representation,
but their model selection, known as tensor network structure search (TN-SS), is
a challenging problem. Although several works have targeted TN-SS, most
existing algorithms are manually crafted heuristics with poor performance,
suffering from the curse of dimensionality and local convergence. In this work,
we jump out of the box, studying how to harness large language models (LLMs) to
automatically discover new TN-SS algorithms, replacing the involvement of human
experts. By observing how human experts innovate in research, we model their
common workflow and propose an automatic algorithm discovery framework called
tnGPS. The proposed framework is an elaborate prompting pipeline that instruct
LLMs to generate new TN-SS algorithms through iterative refinement and
enhancement. The experimental results demonstrate that the algorithms
discovered by tnGPS exhibit superior performance in benchmarks compared to the
current state-of-the-art methods. | tnGPS: Discovering Unknown Tensor Network Structure Search Algorithms via Large Language Models (LLMs): Tensor networks are efficient for extremely high-dimensional representation,
but their model selection, known as tensor network structure search (TN-SS), is
a challenging problem. Although several works have targeted TN-SS, most
existing algorithms are manually crafted heuristics with poor performance,
suffering from the curse of dimensionality and local convergence. In this work,
we jump out of the box, studying how to harness large language models (LLMs) to
automatically discover new TN-SS algorithms, replacing the involvement of human
experts. By observing how human experts innovate in research, we model their
common workflow and propose an automatic algorithm discovery framework called
tnGPS. The proposed framework is an elaborate prompting pipeline that instruct
LLMs to generate new TN-SS algorithms through iterative refinement and
enhancement. The experimental results demonstrate that the algorithms
discovered by tnGPS exhibit superior performance in benchmarks compared to the
current state-of-the-art methods. | [
0.0171649698,
0.073723875,
-0.0328993052,
0.0306657944,
-0.0344535187,
0.0093802167,
0.0283508487,
0.0199404974,
0.0397639871,
0.0297073778,
-0.0103449617,
-0.0339267254,
-0.0074622338,
0.0234036949,
0.0159659144,
-0.0017142601,
-0.0248640738,
-0.0173838381,
0.0024882602,
-0.0269939918,
-0.0376034789,
0.0143499272,
0.0212023687,
-0.0063071195,
0.0108665088,
0.0168586951,
-0.0142371096,
-0.0112223495,
-0.0018849393,
-0.0195440259,
-0.0030537639,
0.0667062104,
0.0203903709,
0.0684814751,
0.0000019799,
0.0110123167,
-0.033012338,
-0.0075076488,
0.0258416571,
0.0106058093,
0.0176553763,
0.0109362816,
0.0233262125,
-0.0097162779,
-0.0405748636,
0.0282997731,
0.0522184074,
0.0061323312,
0.0829171315,
0.054161381,
-0.0083254799,
-0.0024226648,
-0.0066182301,
0.002612517,
-0.0045660329,
-0.0488485545,
0.015889924,
0.0127405971,
0.000117758,
-0.001058779,
0.015612917,
0.0445776395,
-0.0164340604,
-0.0222732536,
0.0355399959,
0.032420136,
-0.0119341305,
-0.0202739201,
0.0012993694,
0.1150525436,
-0.0137379253,
-0.0253745243,
0.054165218,
0.0094055263,
-0.0486185774,
0.0430701636,
-0.0548613816,
-0.0056284354,
-0.0209692884,
-0.0227996409,
0.0125463093,
-0.0368365385,
0.0368930437,
-0.005429545,
-0.0239105094,
0.0446846932,
0.0559334531,
0.0085503031,
-0.0125996452,
-0.0305663235,
-0.0021316339,
-0.0151186287,
0.0341182724,
-0.0059842705,
-0.0471065603,
-0.0408585146,
0.0375551991,
-0.0964401439,
-0.0315519534,
-0.0373143144,
0.0832776949,
0.0468084477,
0.0140577108,
0.0600536317,
-0.0795968398,
0.0240747109,
-0.0909801349,
-0.0324331261,
-0.0411728621,
0.0080082081,
-0.0452088416,
-0.0199834518,
-0.0294115655,
-0.0134885944,
-0.0269215014,
-0.0471853651,
-0.0180008002,
0.0471336953,
-0.0178551059,
-0.0243079402,
0.0155869825,
0.0478274673,
0.0501398668,
-0.0122128939,
-0.0572709329,
0.0091275731,
-0.0051420354,
-0.0123830233,
-0.0652683228,
0.0520150438,
-0.0058082906,
0.052367799,
0.0050073699,
-0.010925781,
0.0637988895,
0.0828057677,
0.0041497364,
-0.0394080654,
-0.0096321236,
0.0177741367,
0.0161629766,
-0.0580314435,
0.0635223389,
-0.0415969416,
0.0291129649,
0.0084295794,
-0.0054506697,
-0.0357979424,
-0.0144952824,
-0.0001983181,
-0.0448451266,
0.0184670705,
-0.0055071213,
0.039421387,
0.0481201187,
-0.0147699323,
0.0100987172,
-0.0030066224,
0.0024065396,
0.0263030026,
0.0435778983,
-0.0360035598,
0.0094565796,
0.034847606,
0.0068443599,
-0.004401518,
-0.0197812151,
0.0127625456,
-0.0023196952,
0.0732658952,
-0.0053704032,
0.0353148729,
-0.0644033551,
-0.0121859964,
0.0315747783,
0.0763686374,
0.0279151797,
0.0600599684,
0.0279639754,
-0.014741105,
-0.0256669577,
-0.029842671,
-0.0126634408,
0.0697978139,
-0.0193356331,
-0.0268952213,
0.0432572886,
0.0116888434,
-0.0336983874,
0.0207801983,
0.0107303197,
0.0069739814,
0.0120507032,
-0.013057977,
0.0631094128,
-0.0163713694,
-0.0330668651,
0.0250664279,
-0.0148144495,
-0.0171188004,
-0.0182698667,
-0.0311303865,
0.0327730291,
0.0639802516,
0.0669324398,
-0.0474848375,
0.0090837721,
0.0167828854,
0.000559653,
-0.0101864981,
0.0070135021,
0.0045348378,
-0.033474315,
-0.0396974385,
-0.0356620476,
-0.0121430866,
-0.0070887981,
-0.0248439759,
0.0109786214,
0.0013886699,
0.0051479037,
-0.0388388224,
-0.0390171818,
0.0332174264,
-0.0488754399,
0.0017161125,
0.104970023,
-0.005037732,
0.0097659938,
0.0536515377,
-0.0074150292,
-0.0051334207,
0.02609195,
-0.0273010619,
-0.0373674855,
-0.0278587136,
0.0150380516,
0.0354644768,
0.0315844864,
-0.0443545431,
-0.041006241,
-0.0137296645,
-0.0319444314,
-0.026455285,
0.1271285415,
-0.0043380004,
-0.0118912365,
-0.020558754,
-0.0102066938,
-0.0288705006,
-0.0314764678,
-0.0200821031,
0.057663288,
0.0172995236,
-0.014842255,
0.0499107465,
-0.0272456482,
0.0303317867,
-0.0126817795,
-0.0295354705,
0.0061317948,
0.0219942741,
0.0088261543,
0.0447941497,
0.0015366604,
0.0245846547,
-0.0395814441,
-0.0750803724,
0.0188276358,
-0.012166732,
0.0192916654,
0.022648029,
-0.0160456318,
-0.0047166869,
-0.0165944248,
0.0096873362,
0.0077514611,
0.0012481468,
-0.0327136777,
0.0458722524,
0.0106781647,
-0.0733479708,
-0.0493034124,
-0.0561534725,
-0.0145290624,
0.0616963319,
-0.0032087713,
-0.010482952,
-0.0330883488,
-0.0131607903,
-0.0130884228,
0.0285158064,
-0.0266331229,
-0.0007949013,
0.0242140274,
-0.0143878488,
0.0181895178,
0.0624713004,
-0.0019921211,
0.0013853989,
0.0548256896,
0.020436639,
0.0159404576,
0.0043165823,
-0.0604724474,
-0.013804269,
-0.0112307658,
-0.0100218765,
0.0275500212,
0.0281218309,
-0.00355713,
-0.0064644278,
-0.0380111039,
-0.0098341126,
0.0068208077,
0.0464326814,
-0.0603374094,
-0.0284988023,
0.0234542508,
0.0665557832,
0.0489897989,
-0.0199555419,
0.0644821823,
-0.0036746624,
-0.0115062492,
-0.0426887088,
-0.0390829891,
-0.0509957969,
0.0385658853,
-0.0257051867,
0.052147083,
-0.0142320301,
0.0344971977,
-0.01629976,
0.0316288248,
0.0566314831,
-0.0107682394,
-0.0052749384,
0.0027207544,
0.0034638848,
0.0274532009,
-0.0359292962,
-0.0130965319,
-0.0099469991,
-0.029289227,
0.0128112566,
0.0282858852,
-0.0809574425,
-0.0197352655,
0.0066620898,
-0.0015204288,
0.0396468453,
-0.0041922652,
-0.0079013398,
-0.0240287893,
0.0418863632,
0.0388014056,
-0.0623809546,
0.0649386942,
-0.0098369755,
0.0336553007,
0.023837965,
0.0084836297,
0.0124758668,
-0.0853353441,
-0.0024227414,
-0.0035978889,
-0.0240129922,
0.0079607414,
0.0147172101,
-0.0394836627,
-0.0079632094,
0.0193345137,
0.0587384067,
0.0135916919,
-0.0605540164,
-0.0147369448,
0.0250492226,
0.0118769389,
0.0269595813,
0.0527296476,
-0.0839456692,
0.0148045411,
-0.0088246642,
0.0115676178,
-0.0277915522,
-0.0546683073,
0.0195946489,
-0.0424546935,
0.058346577,
0.0017724897,
0.0265130121,
0.0018668125,
-0.0689139962,
-0.0185148604,
-0.0323252156,
-0.0153229535,
0.0195175111,
0.0142386621,
-0.0255198423,
-0.0101786302,
0.0028966726,
-0.0597876646,
-0.005260515,
-0.0033290482,
0.0129433842,
-0.0007442103,
0.0096209822,
-0.0239633992,
-0.0426837467,
0.0071280217,
0.0805156901,
0.0483027399,
-0.0183202978,
0.0277163628,
0.0105603309,
-0.0249307007,
-0.0745645091,
0.053888116,
-0.075504601,
0.0285233464,
-0.0132686561,
0.0834125876,
-0.0082809972,
-0.0445272624,
-0.0393242165,
-0.0240543317,
0.0105795329,
0.0607744865,
-0.0191498436,
-0.0656487271,
0.0002655732,
0.0246009082,
0.032853961,
-0.0033697311,
-0.0536477976,
-0.0475310944,
0.0292784534,
0.0413107909,
-0.0380005464,
0.1016325802,
0.0144019928,
0.0231047273,
0.0143010123,
0.0036595347,
0.0014606871,
-0.0331883468,
-0.0434721336,
0.0634368509,
0.0121761262,
0.0085662361,
-0.0425716229,
-0.012353207,
0.0108283386,
-0.0126604028,
-0.0467331409,
0.0297940224,
-0.105471395,
-0.050311178,
-0.0873243213,
0.0417528711,
0.0056174998,
0.0028497588,
-0.0049197893,
-0.0329916887,
0.0547899492,
-0.0084785037,
-0.0580938272,
0.0019537087,
-0.0129364431,
-0.0645280033,
-0.0516789518,
-0.0062408471,
-0.0013751095,
0.0011841968,
0.0620985217,
0.0129419761,
0.0059711845,
-0.0105279535,
0.0389899276,
0.0168719403,
0.0375793204,
-0.0385439321,
0.0340516269,
-0.055706203,
-0.0015848231,
0.0618839636,
-0.0098008197,
-0.0517147221,
0.0086105457,
0.0003257315,
-0.0160056688,
0.0276480485,
-0.0325233079,
0.0351149067,
0.0380576737,
-0.0123773068,
-0.0040209419,
0.0269210003,
0.0246620271,
0.0439646654,
-0.0303810053,
0.0163700711,
0.0162794143,
-0.0414646417,
0.0679201186,
-0.0345199108,
0.0448549576,
0.0293197799,
-0.0393961444,
-0.0480037145,
0.0281188674,
-0.0283707008,
-0.1328044236,
-0.0300581437,
-0.0004041405,
0.0834700093,
0.0400392115,
-0.0198482051,
-0.0456618443,
-0.0053374884,
0.0007094492,
0.0077752797,
-0.0301108342,
0.0238742605,
-0.0153677976,
-0.0085764164,
-0.0152055481,
-0.0084777437,
0.0323144644,
0.0076508983,
0.020285327,
-0.0158126894,
-0.0412873402,
0.0024216252,
-0.0307072625,
-0.0040588556,
0.0670867637,
-0.0459323153,
0.010557848,
-0.0238942578,
-0.0191980135,
-0.0314433277,
-0.0407217219,
-0.0384986848,
-0.0172922369,
0.048701372,
0.0623027384,
0.01514774,
0.0269243084,
0.0064118584,
-0.0400424562,
-0.0256444421,
-0.0630566403,
-0.0531881638,
-6.26506016e-33,
0.0411192179,
-0.0242074188,
0.0162140336,
-0.0344845355,
0.0263595842,
0.0030446963,
-0.0601304471,
0.0302958861,
0.0227817055,
0.0221827049,
-0.010536423,
-0.0134345209,
-0.0272458438,
-0.0353825204,
0.0715424195,
0.0488842987,
0.0184446946,
-0.0109331757,
-0.027120078,
0.0432889089,
0.026303824,
0.0024069184,
0.0518950745,
-0.095297955,
-0.0523984209,
0.0018021389,
0.0101199597,
0.0561909229,
0.0305706114,
0.0711069703,
-0.0297912639,
-0.0194636676,
-0.0104277907,
-0.0460601747,
-0.006173199,
-0.0299655255,
-0.0155121665,
0.0172856264,
0.0187607259,
-0.0526666753,
-0.0546792857,
-0.0004334834,
-0.0146460682,
0.0110055096,
-0.026713334,
-0.0014439909,
0.0209443495,
0.0103958156,
0.0123126134,
0.0201646425,
-0.0073795472,
0.0152247734,
-0.047903575,
0.0303772129,
0.0294008702,
0.0118205072,
-0.0383584462,
-0.0258651655,
-0.0414032042,
0.0610127077,
-0.02205332,
-0.0195262916,
0.0273406357,
0.0128964772,
0.0380805023,
0.0593948029,
0.0259571336,
0.0229408816,
0.0163663607,
-0.0235691555,
0.0047775004,
0.0559812896,
0.0268232264,
-0.0017685663,
-0.0201681554,
0.0020174368,
-0.0644943863,
-0.0372320004,
0.0025343504,
0.0047717183,
0.0901335776,
0.0106715662,
-0.0160325542,
-0.0355787128,
-0.0179477632,
-0.0175957289,
-0.0535780191,
-0.0195029471,
-0.0079861796,
0.0246234071,
-0.0189761817,
-0.0332670137,
0.0366533808,
-0.0592695139,
0.0679272115,
0.0470822267,
-0.0405343771,
0.0147002488,
0.042954836,
-0.0209858268,
-0.0306240674,
-0.0729832277,
-0.0335195549,
-0.0613766722,
0.0319565125,
0.0047154031,
-0.0352651849,
0.0341157839,
-0.0336610675,
-0.0105011258,
-0.0255041569,
-0.0124060679,
-0.0057756798,
0.0125124194,
-0.0209472459,
0.0346315429,
-0.0159474704,
0.0351180658,
0.0254163556,
0.0649970174,
-0.0297595132,
0.0364454761,
0.0282847807,
0.0067723743,
-0.0017215902,
-0.0328260288,
-0.0282085352,
-0.0052029998,
0.091941528,
-0.049445454,
-0.0289313626,
0.0056159459,
2.865e-7,
-0.015807841,
0.0660907254,
-0.0448554046,
0.0108125042,
0.0540433787,
0.0307396706,
0.0057926611,
-0.0177943371,
0.0113126589,
-0.013658735,
0.055560641,
0.0165058319,
0.0025123155,
-0.0137440199,
-0.0699492618,
-0.031305708,
-0.0745769665,
-0.0081940573,
-0.0631453991,
-0.0196350515,
-0.0326547436,
0.049337469,
0.042106729,
-0.0173333865,
0.0104974713,
-0.0512569249,
0.0450337268,
-0.0242638439,
0.0625586137,
0.0502055585,
-0.0185931828,
0.0171734542,
-0.0277184006,
0.0278112106,
0.0034840989,
0.0460631661,
-0.011408451,
0.0518736243,
0.0177207291,
0.0346938707,
0.0321762376,
-0.007290645,
-0.0036180376,
-0.0573713779,
0.0283797812,
-0.0161269549,
-0.0372714847,
-0.0688427016,
0.017624449,
-0.062828593,
0.027192682,
-0.0296252687,
0.0562886,
0.0457609668,
0.008948978,
-0.0077333553,
0.0540692881,
-0.0512619764,
0.0369863622,
0.0366074294,
-0.0476187095,
-0.0038932897,
-0.043545723,
-0.0203918014,
0.0502580144,
0.0254697297,
0.0579942092,
2.374150303e-34,
0.0214616265,
-0.0026175643,
0.0232507363,
0.0713542253,
0.0526814461,
-0.0223980304,
-0.0221128594,
-0.0025608474,
0.0105804633,
-0.0752527863,
-0.0282535069
] |
http://arxiv.org/abs/2402.02449v1 | http://arxiv.org/pdf/2402.02449v1 | Surfing the modeling of PoS taggers in low-resource scenarios | 2024-02-04T11:38:12 | 2024-02-04T11:38:12 | Manuel Vilares Ferro, Víctor M. Darriba Bilbao, Francisco J. Ribadas-Pena, Jorge Graña Gil | cs.CL | cs.CL, cs.LG, 68, 68T50 | The recent trend towards the application of deep structured techniques has
revealed the limits of huge models in natural language processing. This has
reawakened the interest in traditional machine learning algorithms, which have
proved still to be competitive in certain contexts, in particular low-resource
settings. In parallel, model selection has become an essential task to boost
performance at reasonable cost, even more so when we talk about processes
involving domains where the training and/or computational resources are scarce.
Against this backdrop, we evaluate the early estimation of learning curves as a
practical mechanism for selecting the most appropriate model in scenarios
characterized by the use of non-deep learners in resource-lean settings. On the
basis of a formal approximation model previously evaluated under conditions of
wide availability of training and validation resources, we study the
reliability of such an approach in a different and much more demanding
operationalenvironment. Using as case study the generation of PoS taggers for
Galician, a language belonging to the Western Ibero-Romance group, the
experimental results are consistent with our expectations. | Surfing the modeling of PoS taggers in low-resource scenarios: The recent trend towards the application of deep structured techniques has
revealed the limits of huge models in natural language processing. This has
reawakened the interest in traditional machine learning algorithms, which have
proved still to be competitive in certain contexts, in particular low-resource
settings. In parallel, model selection has become an essential task to boost
performance at reasonable cost, even more so when we talk about processes
involving domains where the training and/or computational resources are scarce.
Against this backdrop, we evaluate the early estimation of learning curves as a
practical mechanism for selecting the most appropriate model in scenarios
characterized by the use of non-deep learners in resource-lean settings. On the
basis of a formal approximation model previously evaluated under conditions of
wide availability of training and validation resources, we study the
reliability of such an approach in a different and much more demanding
operationalenvironment. Using as case study the generation of PoS taggers for
Galician, a language belonging to the Western Ibero-Romance group, the
experimental results are consistent with our expectations. | [
-0.0014314436,
0.0504672937,
0.0029314237,
0.0205113515,
-0.0316017307,
0.0044341492,
0.009388213,
0.0003239726,
-0.0096857548,
0.0006839357,
0.0261909533,
-0.0573288128,
0.0037873858,
0.0237755068,
-0.0279852301,
-0.05010923,
0.0290766843,
-0.0662505105,
0.0206970945,
-0.0229037106,
-0.0221850872,
-0.0107242418,
-0.0221304372,
0.0487475507,
-0.051837381,
0.0108918445,
-0.0022983216,
-0.0266254134,
-0.0438149124,
-0.0382063799,
0.0056944201,
0.0240933578,
-0.0188964736,
0.0720295832,
0.0000020342,
-0.0665119141,
-0.0674787685,
0.0139123034,
-0.0122127682,
-0.0151457898,
0.0139223626,
-0.030765349,
-0.0417272598,
0.0030246668,
-0.0393919051,
-0.0331860743,
0.0651399866,
0.0001281626,
0.0222912114,
0.0285965595,
-0.0192667767,
-0.0424898677,
-0.0056733144,
-0.007935537,
-0.0265486874,
-0.0424667411,
0.0240588207,
0.002259159,
0.0033835776,
-0.0074397717,
-0.0254594367,
0.0354302377,
-0.0081655188,
-0.0280923303,
-0.0042173536,
0.0198573228,
-0.0172764994,
-0.0142198252,
-0.0005418561,
0.0239706393,
-0.0487848334,
-0.0058521805,
0.000025149,
-0.025713779,
-0.0509801954,
-0.0476121157,
-0.0136698037,
-0.02217035,
-0.0037385204,
0.0385492742,
0.0482245833,
-0.0133481277,
0.0242267735,
0.0057970393,
-0.0209044311,
0.0746793747,
0.0001502603,
-0.0442332067,
0.0520244613,
0.0485813022,
-0.0527463108,
-0.0140732443,
-0.0022627898,
0.0177031104,
0.0021247806,
0.0159587469,
-0.0340826735,
-0.0101024052,
-0.0116721066,
-0.0506962873,
0.0436363295,
-0.0072172931,
0.0089671537,
0.0511672907,
-0.0543137193,
0.0050146598,
-0.0065027247,
-0.0385099836,
-0.000311339,
-0.026124334,
-0.0779147446,
0.0129957208,
-0.0768690631,
0.0309607964,
-0.006726597,
0.0009123774,
-0.0465081893,
-0.0314713269,
-0.0015106775,
-0.0192077663,
-0.0211268552,
-0.0167470556,
-0.0301300287,
0.0334459133,
-0.049772419,
0.0259138942,
-0.0111289546,
-0.0102608548,
-0.0325615294,
-0.0770906731,
0.0346942768,
0.0214467719,
0.0098941559,
-0.0219579358,
0.0254954305,
0.0328208804,
-0.0057034069,
-0.0129706925,
-0.0689204857,
-0.0313990936,
-0.0645167753,
-0.0260092635,
-0.0083404146,
-0.0675098896,
0.0011693622,
0.0094987676,
0.0083982628,
-0.0501358248,
0.0024916218,
-0.0141491946,
-0.0125098648,
0.0447380468,
-0.0290701445,
0.0057086521,
-0.0010626954,
-0.0161625985,
-0.031696666,
0.0405029319,
0.0529453866,
0.0094592068,
0.0654013008,
-0.0340414532,
-0.0142633338,
0.0771396309,
-0.0156672057,
-0.0114675835,
-0.0091214385,
-0.0477079563,
0.0219581351,
0.0160025246,
0.0520759188,
0.0698084608,
0.0165663064,
0.0367975719,
0.033138264,
0.0421506986,
0.0739505589,
0.0341533981,
0.0935943052,
-0.0151782688,
0.0046986402,
-0.0202985369,
0.0274598747,
-0.058267273,
0.0008729002,
0.0109666642,
-0.0149083091,
0.002601502,
-0.0172761641,
-0.0457432494,
-0.0187679417,
-0.022740107,
-0.0370329618,
-0.0426750854,
0.1016112119,
-0.0290011838,
-0.0597197637,
0.0370581746,
0.0190588832,
0.0047062743,
0.0061343587,
-0.0471804552,
-0.0068765339,
0.0286658909,
0.0480879582,
-0.0278507974,
0.0029139635,
-0.0135638621,
-0.0115627348,
0.0198050253,
0.0709603131,
-0.0098776808,
0.0041554598,
-0.0519318543,
-0.0245972443,
0.074889712,
-0.0036084694,
-0.002549048,
0.0006051593,
0.0199893154,
-0.0040477938,
-0.038023632,
-0.0815896019,
0.052492369,
-0.0363242626,
-0.0787403435,
0.0298602413,
0.0094941882,
0.0153817339,
0.0523689426,
0.0158733446,
-0.0007953253,
0.049711559,
-0.0505837649,
0.0387227237,
-0.0227854755,
-0.0294241719,
0.0308443476,
0.0455835611,
0.0043108319,
0.0159876756,
-0.0138131902,
-0.0005579911,
-0.057309553,
-0.021682173,
-0.0480293036,
0.0570579693,
-0.0215316657,
0.0437504128,
-0.0237798803,
0.0067229029,
-0.0349333026,
0.0542777143,
-0.0036374037,
0.0657885224,
0.0034896189,
-0.0082895365,
0.0350494497,
-0.045007322,
-0.059700042,
0.0073830388,
-0.0197224654,
-0.0098280963,
0.0002047369,
0.008329886,
-0.0012911529,
-0.0099485628,
-0.0742414072,
-0.0228301529,
-0.009867494,
0.0393565707,
-0.0277425461,
-0.0037000703,
-0.0007166703,
-0.03739921,
-0.0230615251,
-0.0030527811,
0.028490033,
0.0361456312,
0.0391490646,
0.039087221,
-0.0493828468,
-0.0372380018,
-0.0192144755,
-0.0032314756,
0.0306314174,
0.0266589336,
-0.0587055348,
-0.0461932868,
0.0031823975,
-0.0222905967,
0.0637694374,
0.0202632677,
-0.0254060086,
0.0239060689,
0.0071001612,
0.0512155518,
0.0107746255,
-0.0281808488,
0.0004547527,
0.0151466588,
0.0291121956,
-0.0433743037,
-0.0150414286,
-0.0348841324,
-0.0232187491,
-0.0398099609,
-0.0371584632,
0.0054784976,
0.1070319191,
-0.0267565232,
0.020284785,
-0.0212401077,
-0.0276146196,
-0.0449162051,
-0.0230240989,
-0.0029297499,
0.0334281027,
-0.0076218802,
0.0264607035,
-0.015662292,
-0.061097987,
0.0395562723,
-0.0244870912,
-0.0042841155,
-0.0460660048,
-0.0501575023,
-0.0233759936,
0.022261139,
0.0286693461,
0.0015327729,
0.0096430127,
-0.016564969,
-0.0205080081,
0.0160035621,
0.0331974626,
-0.0210141689,
0.0214344375,
-0.0346412212,
0.0463028215,
0.0011964047,
-0.0162716154,
-0.0285748541,
-0.0202013515,
-0.0348370858,
0.0531926192,
0.0240313206,
-0.0066311383,
0.0420580059,
0.0370619521,
-0.0097569134,
0.0692119375,
0.0398225784,
-0.0020716425,
-0.0214052945,
0.0376029983,
-0.0007194176,
-0.0380031019,
0.0249616206,
-0.0572800636,
0.0147154843,
-0.0162953287,
0.0509323701,
-0.0179783236,
-0.0596875437,
-0.0122013614,
0.0243123602,
-0.0310457908,
-0.070110999,
0.0252507105,
-0.0124671003,
-0.0528350845,
-0.0293836482,
0.0890937001,
0.0145364795,
-0.0327551849,
0.0206683595,
0.0187460184,
0.0789582357,
0.0089801177,
0.0371933207,
-0.0425118692,
-0.002669082,
0.0282124914,
0.0582792684,
-0.017777564,
-0.040573284,
0.015252105,
-0.1483514905,
0.0304071773,
-0.0131806787,
-0.0078020906,
0.0388537087,
-0.0438463837,
-0.0212988555,
-0.0339649543,
-0.0670056865,
0.0002621878,
-0.0187745001,
-0.0016857646,
0.0687767714,
-0.067493327,
-0.0368242078,
0.0112432214,
-0.0699284747,
0.001138702,
0.0691733807,
-0.0019553008,
-0.0111329174,
-0.0453977585,
0.0675857961,
0.0407706648,
0.0280543361,
-0.0332610011,
0.0197598655,
0.0316602066,
-0.0111618079,
0.018323563,
0.0030589218,
-0.0307972953,
0.0278208312,
-0.0199907217,
0.1045723483,
-0.0062421472,
-0.0057570022,
0.0180471744,
-0.0147145763,
0.0381459258,
0.0360048376,
0.0067130844,
-0.0642442554,
-0.068872869,
-0.0106269848,
0.0406055264,
-0.0349388793,
-0.0111417193,
0.0444749668,
-0.0527309179,
-0.0662126094,
-0.0617296174,
0.0095104221,
0.0485891066,
0.0834064484,
-0.0070387283,
0.0267170556,
-0.0047296043,
-0.0023727363,
-0.0321155936,
0.0305200722,
-0.0144720031,
-0.0054741246,
-0.0992255136,
-0.0102685746,
0.0307563115,
0.0335157327,
-0.0370897278,
-0.0587527715,
-0.0554871969,
-0.0556076951,
0.0519545637,
0.067002736,
0.0368882604,
0.004610755,
-0.0100905001,
-0.0637535304,
0.0322280452,
0.0475073084,
-0.0252407845,
-0.020715721,
0.0262112506,
-0.0018746923,
-0.0302723367,
0.0166993756,
-0.0036146923,
0.0320491418,
-0.0259825103,
0.0339392014,
-0.0111447554,
-0.0411561802,
0.0056763282,
0.0239569712,
0.0745203793,
-0.025429491,
-0.0765918419,
-0.0238033067,
-0.0035521893,
0.0537236184,
0.0081770662,
-0.0189002641,
-0.0003684865,
0.0058601932,
0.050992243,
-0.0131219588,
0.0074174637,
0.0049345023,
0.0432705991,
-0.0130385524,
0.0355236344,
0.0758802742,
0.0118734902,
0.0194951557,
0.0150607191,
-0.0029574186,
0.0377568565,
0.0016587619,
0.0370036177,
-0.001784025,
0.007737881,
-0.0157510992,
0.0745381713,
0.020902615,
-0.0332803838,
0.0035799779,
-0.0447981916,
-0.0551305301,
-0.0100896331,
0.06164461,
-0.0229448918,
0.0257857982,
-0.0286807772,
-0.0006842187,
-0.0264948122,
0.0079317549,
0.0079029212,
-0.0179298632,
0.0020516335,
0.0205855109,
-0.026312124,
0.0313994549,
0.0675225928,
0.0405017957,
0.043960616,
0.0006663233,
-0.0009235824,
-0.0373735987,
0.0132393427,
0.0348656252,
-0.0056341258,
-0.0437138416,
0.0009551595,
-0.0518641956,
-0.0512868091,
0.0197608192,
-0.0131245563,
-0.0049207583,
-0.0142167266,
0.0452601835,
0.0561825037,
-0.0114095472,
0.0333160609,
-0.0523926169,
-0.0077165668,
-0.0033235333,
0.0023138195,
-0.0199406091,
-5.838748766e-33,
0.0200849008,
-0.035427995,
-0.0005159896,
0.0022569411,
-0.0021302751,
0.021454962,
0.0040240856,
0.0199113619,
-0.0066826474,
0.0018964781,
-0.0206953194,
-0.0440251045,
0.0387197807,
0.0322928019,
0.0198346861,
0.0341777354,
0.0629243478,
-0.0366063379,
0.0286820717,
-0.0248739924,
0.0417082272,
0.0433410332,
0.0354154371,
-0.0584335849,
0.0434339643,
0.0117409099,
-0.0090858527,
0.0279386509,
0.0513415672,
0.0757118315,
-0.0219837222,
0.0215006862,
0.0127296345,
-0.0007407857,
0.0000749013,
0.0078631686,
-0.0612180643,
0.0031334101,
-0.0256161653,
0.0344582088,
-0.00543204,
-0.0007721568,
0.0337365642,
-0.001563856,
-0.0151179405,
0.047039438,
0.0259557646,
-0.0264363904,
-0.0354705825,
-0.0217778347,
-0.0360600911,
0.0486874618,
0.0510079414,
0.0238305181,
0.0568810292,
0.0487873517,
-0.0208621267,
-0.0182543974,
-0.0806222185,
0.013877538,
0.0391309336,
0.0582785234,
0.0593678355,
0.036232207,
0.0037913276,
0.0532748178,
0.0412186347,
-0.0226081815,
0.0261454973,
0.0064664735,
0.0742014498,
0.0417109057,
0.0545478351,
-0.0046587386,
0.0855658129,
0.0216712356,
-0.0650227591,
-0.0135532133,
0.0191273708,
0.0032354202,
0.0052156872,
0.0218030252,
-0.0936067253,
-0.0038599968,
-0.0096818609,
-0.0477300398,
-0.0062251273,
-0.0418524258,
0.0194579754,
0.0053596711,
-0.0618059002,
0.0283211488,
0.0480771698,
-0.0394592695,
-0.0075268643,
-0.0608817153,
0.0090738749,
0.0692094788,
-0.0209893119,
0.03613469,
-0.0754970908,
-0.1163210049,
-0.0279641841,
0.0088481642,
0.0482490994,
-0.0034054765,
0.0172029939,
-0.0069126664,
-0.0852678642,
0.0124373753,
-0.0128301131,
0.0360731557,
0.0085616531,
0.0762814656,
0.0565456599,
-0.0481489003,
0.0021195556,
0.0575420447,
0.0278910771,
0.0147242965,
0.0344492011,
-0.0126475561,
0.0097502116,
0.0140561378,
-0.0101719536,
0.0084634097,
-0.007951905,
0.0359136052,
0.0003100848,
-0.0134400045,
-0.0120223695,
0.0150831304,
2.727e-7,
0.0396812223,
0.0296332836,
0.0312988535,
-0.0096199345,
0.0047069113,
0.0238334388,
0.0236503892,
-0.012376301,
-0.0727980882,
-0.02304882,
-0.0048184106,
-0.0208704509,
0.0521243252,
0.045025982,
-0.0792536065,
-0.0338519663,
-0.0035354195,
0.0443658754,
-0.0381679051,
-0.0077102194,
0.0377182961,
0.0812159032,
-0.0183763728,
-0.0339515433,
-0.0156973749,
-0.0779381767,
0.0430727825,
0.038599655,
0.0088024586,
0.0253300648,
0.0629570708,
0.0393842049,
-0.0032913384,
0.0163193308,
-0.0279181339,
0.0139063848,
0.0200257357,
0.0359097198,
-0.0000237972,
0.0120017175,
0.057710398,
-0.0194555689,
-0.0206413269,
-0.0278875902,
0.055189155,
-0.0285689719,
-0.0023045321,
-0.0080633024,
0.0084980801,
-0.0080652088,
0.0025081129,
-0.0507351086,
-0.0412881412,
-0.0064548729,
0.0396996215,
-0.0483964495,
-0.043354325,
-0.0021177144,
0.0007908643,
0.0178066548,
-0.0055994247,
-0.0384776592,
-0.0485574752,
0.0360384956,
0.0226638652,
-0.0231761523,
0.0251618829,
2.784096151e-34,
0.0365274772,
-0.0290030614,
0.0050190883,
0.0272586904,
-0.0104576591,
-0.0108936569,
0.0329792872,
0.0027103922,
0.0131289475,
-0.0338887386,
-0.0177670997
] |
http://arxiv.org/abs/2402.02447v1 | http://arxiv.org/pdf/2402.02447v1 | Breaking MLPerf Training: A Case Study on Optimizing BERT | 2024-02-04T11:12:17 | 2024-02-04T11:12:17 | Yongdeok Kim, Jaehyung Ahn, Myeongwoo Kim, Changin Choi, Heejae Kim, Narankhuu Tuvshinjargal, Seungwon Lee, Yanzi Zhang, Yuan Pei, Xiongzhan Linghu, Jingkun Ma, Lin Chen, Yuehua Dai, Sungjoo Yoo | cs.LG | cs.LG, cs.CL | Speeding up the large-scale distributed training is challenging in that it
requires improving various components of training including load balancing,
communication, optimizers, etc. We present novel approaches for fast
large-scale training of BERT model which individually ameliorates each
component thereby leading to a new level of BERT training performance. Load
balancing is imperative in distributed BERT training since its training
datasets are characterized by samples with various lengths. Communication cost,
which is proportional to the scale of distributed training, needs to be hidden
by useful computation. In addition, the optimizers, e.g., ADAM, LAMB, etc.,
need to be carefully re-evaluated in the context of large-scale distributed
training. We propose two new ideas, (1) local presorting based on dataset
stratification for load balancing and (2) bucket-wise gradient clipping before
allreduce which allows us to benefit from the overlap of gradient computation
and synchronization as well as the fast training of gradient clipping before
allreduce. We also re-evaluate existing optimizers via hyperparameter
optimization and utilize ADAM, which also contributes to fast training via
larger batches than existing methods. Our proposed methods, all combined, give
the fastest MLPerf BERT training of 25.1 (22.3) seconds on 1,024 NVIDIA A100
GPUs, which is 1.33x (1.13x) and 1.57x faster than the other top two (one)
submissions to MLPerf v1.1 (v2.0). Our implementation and evaluation results
are available at MLPerf v1.1~v2.1. | Breaking MLPerf Training: A Case Study on Optimizing BERT: Speeding up the large-scale distributed training is challenging in that it
requires improving various components of training including load balancing,
communication, optimizers, etc. We present novel approaches for fast
large-scale training of BERT model which individually ameliorates each
component thereby leading to a new level of BERT training performance. Load
balancing is imperative in distributed BERT training since its training
datasets are characterized by samples with various lengths. Communication cost,
which is proportional to the scale of distributed training, needs to be hidden
by useful computation. In addition, the optimizers, e.g., ADAM, LAMB, etc.,
need to be carefully re-evaluated in the context of large-scale distributed
training. We propose two new ideas, (1) local presorting based on dataset
stratification for load balancing and (2) bucket-wise gradient clipping before
allreduce which allows us to benefit from the overlap of gradient computation
and synchronization as well as the fast training of gradient clipping before
allreduce. We also re-evaluate existing optimizers via hyperparameter
optimization and utilize ADAM, which also contributes to fast training via
larger batches than existing methods. Our proposed methods, all combined, give
the fastest MLPerf BERT training of 25.1 (22.3) seconds on 1,024 NVIDIA A100
GPUs, which is 1.33x (1.13x) and 1.57x faster than the other top two (one)
submissions to MLPerf v1.1 (v2.0). Our implementation and evaluation results
are available at MLPerf v1.1~v2.1. | [
0.0207624547,
0.0429314747,
0.0186895151,
0.0670942888,
0.0161809642,
0.0183620583,
0.0369018838,
0.0232378058,
-0.0156008285,
-0.0281169023,
-0.0428696088,
-0.0332282223,
-0.0204135124,
-0.0342520773,
0.0461235009,
-0.0680687726,
0.0046335552,
0.00694197,
-0.0044054212,
-0.0278543029,
0.0226891357,
-0.0004980575,
-0.0169160943,
0.01207803,
-0.0218784809,
-0.0164098181,
-0.0241904967,
0.0453694463,
0.0481919721,
-0.0098966416,
0.0182443298,
0.0389144681,
0.0561711118,
0.0458318591,
0.0000020383,
-0.0059639416,
-0.0294673033,
0.0205955058,
-0.0048776739,
0.0286301784,
-0.0463432968,
0.0081592826,
-0.0372297987,
0.0288227759,
0.005811404,
0.0196408518,
0.0386839472,
0.0158695169,
0.06713368,
0.0707606822,
-0.0036015534,
-0.0316351242,
0.0171949714,
-0.0224662032,
-0.0407214239,
-0.0291443933,
0.019128114,
-0.0203930829,
-0.0159321316,
-0.0731373131,
-0.0276233144,
0.011776885,
0.0304939914,
0.0396008119,
0.0147240618,
0.0362006277,
-0.1098976284,
-0.0320622176,
-0.0432021022,
0.0849708244,
-0.0546237677,
-0.0039739818,
0.0238781199,
0.010559056,
-0.0153936353,
-0.025594091,
-0.0087495446,
0.0197148491,
-0.0079316348,
0.0291909315,
-0.0415004939,
-0.0472445115,
0.0414829738,
-0.0184041001,
-0.0432404205,
0.0332783349,
0.022582503,
-0.0277975947,
-0.0259467959,
0.0085065011,
-0.0462152474,
-0.0226456486,
0.0512062982,
0.0184279364,
-0.0313925855,
-0.0060372422,
-0.0404258892,
-0.0742481425,
-0.0108101871,
-0.0353233628,
0.0631909072,
0.0159956161,
0.0058105295,
0.0358126312,
-0.0397454724,
0.0200357903,
-0.0046340758,
-0.0004190083,
0.0000174736,
0.0112884147,
-0.0252492484,
-0.0040523503,
-0.0671064705,
0.0324905366,
-0.0346151106,
-0.0291703418,
-0.025071077,
0.0008054796,
-0.0150202233,
-0.0220470559,
0.0457221977,
0.0188357979,
-0.0207234789,
0.0185615011,
0.0249149892,
0.0469179749,
-0.0760425553,
-0.0152697945,
-0.0253013335,
-0.0479147397,
0.0029339623,
-0.0032848991,
0.0036072563,
-0.005394815,
0.041143585,
0.0350783654,
-0.0035412114,
-0.0130508086,
-0.0601872578,
0.0276008379,
0.0197783727,
-0.026646629,
-0.0170747247,
-0.043462377,
0.0521410145,
-0.0028655117,
0.03017506,
0.0476474501,
0.0187380034,
-0.0082306899,
0.010952536,
0.0459454358,
-0.0303383004,
-0.0029509861,
0.0549379922,
-0.0036837701,
-0.0141371684,
0.0846542865,
0.0580819473,
0.0238259304,
0.0103151584,
-0.0148078119,
-0.0214478001,
0.0446809903,
0.0449451432,
0.0193368103,
-0.0372851714,
-0.021476645,
-0.0128188953,
0.0383766182,
0.0438224524,
0.0545292012,
-0.0656956509,
-0.0429951772,
0.0566406213,
0.0346969627,
0.0139614297,
0.0493982136,
0.008207378,
0.0076013939,
0.0285657551,
-0.0111549636,
0.0045135901,
-0.0178774968,
0.0214083269,
-0.0008120538,
0.0385539718,
0.0535204522,
-0.0448430888,
0.0313208327,
0.0049681021,
-0.0118173184,
-0.051835794,
0.0177630987,
0.0190861914,
-0.009001554,
-0.0454030186,
0.0958020091,
0.0358882397,
-0.0483375639,
-0.018563807,
-0.022147404,
-0.0368337594,
0.0609094948,
0.0409817696,
-0.0445494391,
-0.0685643703,
0.0062701148,
-0.0040623015,
0.0302087627,
0.0031556718,
-0.0040674554,
0.0378529765,
-0.0806745812,
-0.0010985509,
0.0078239739,
-0.0311029516,
0.0588589087,
0.0307163615,
0.0071115671,
-0.0358812027,
-0.0140331183,
-0.024239812,
0.0327301398,
-0.0148998667,
-0.0241628103,
-0.0064051915,
0.0379607715,
0.0432226434,
0.0007208144,
-0.0509624816,
-0.0289236456,
0.0591533259,
-0.0399905741,
-0.0010953769,
-0.0123443296,
-0.022308927,
-0.0690235496,
0.0385976844,
-0.0146507099,
-0.019828029,
0.0522935949,
-0.0471316278,
-0.0280416794,
-0.004316201,
-0.0297590699,
0.0716078281,
-0.0178589821,
0.0376887172,
0.0107260086,
-0.0233235918,
-0.0098378481,
0.0409763567,
-0.0101676267,
0.0240029357,
0.0005588031,
0.0088283829,
0.0307595544,
-0.0642336085,
0.0091426326,
0.0140776699,
-0.0111782346,
-0.0112812864,
-0.0227732994,
-0.0000883678,
-0.0442730971,
-0.0211632308,
-0.1120658815,
0.0034096586,
-0.029117709,
0.0447284691,
0.0172189903,
0.0157338269,
-0.0162052158,
-0.0479053073,
-0.0069183419,
-0.0131660067,
-0.0590345748,
0.0224868078,
0.016233569,
0.0122894449,
-0.0409441441,
0.0134649435,
-0.0245520379,
0.0013183369,
0.0594873913,
-0.0184348859,
-0.0409036875,
-0.0258505009,
-0.0443480648,
-0.0428456925,
0.0530021787,
0.0075819623,
-0.0185058992,
0.0167092588,
0.0100444453,
0.0437872261,
0.0672918931,
0.0375173502,
0.0481088087,
-0.0332143903,
0.0268988237,
-0.0734703764,
0.0576247163,
0.0028786515,
0.0138644641,
-0.0291732866,
0.0245843753,
-0.0151109677,
0.0565693378,
-0.0437336527,
0.0204462744,
-0.0443891175,
-0.0544170514,
-0.009261556,
-0.0742724016,
-0.0615740418,
-0.0094230277,
-0.0432744138,
-0.0587750673,
-0.0174512509,
-0.0541528016,
0.0457701422,
-0.0396590345,
-0.0048259282,
-0.0559765473,
-0.0365877002,
-0.0191123728,
0.0223151539,
0.0009612253,
0.0660755709,
-0.0949840248,
-0.0655149966,
-0.0105547374,
0.0029415572,
0.0129259927,
-0.0103518888,
-0.0345510021,
0.0016321756,
0.0070508826,
-0.0421119705,
-0.0906933993,
-0.0140851568,
-0.0221176278,
-0.0147254812,
0.0070300377,
0.0322178006,
-0.0828720257,
-0.0526403897,
0.0442265868,
0.0168270543,
0.0625809729,
-0.0227324776,
0.0027937512,
-0.0071162754,
0.0517867468,
0.0091143083,
-0.0283245277,
-0.0065863854,
-0.0529255308,
-0.0320192799,
0.0547163486,
0.1141335666,
-0.0085474756,
-0.051704593,
-0.0056440812,
0.0087928157,
-0.0659108907,
-0.015588427,
-0.0216271579,
0.0156346317,
-0.0348695926,
0.0202560108,
0.078493394,
0.0398147628,
-0.0576175377,
-0.00212568,
0.0069086989,
0.0607877187,
-0.0144732557,
0.0245759469,
-0.0804830417,
-0.0300088637,
0.0148161901,
0.0273731854,
0.0384354591,
-0.0155060003,
0.0012211099,
-0.0832665637,
0.0110873254,
0.0007989515,
-0.0385375135,
-0.0200813953,
-0.0503697135,
-0.0301510207,
0.0122927967,
0.0087081548,
-0.007952746,
-0.000799484,
0.0058217156,
0.0730745792,
-0.0047751064,
0.0002894234,
-0.0309955645,
-0.0598099492,
0.0106703052,
0.0493196771,
0.0372802652,
0.0011118918,
-0.041382622,
0.0200247038,
0.0418296754,
-0.0003102637,
-0.0147850504,
0.0481389277,
-0.0731999427,
-0.0293880478,
-0.0061887237,
-0.0215340387,
-0.0156566575,
-0.0011629951,
0.0032381085,
0.0565529466,
0.0144922202,
0.0129823284,
0.0152324215,
-0.0080182841,
0.0106462855,
0.0513921045,
0.0205772575,
-0.0350355804,
-0.0087068845,
0.0031207665,
0.0049420726,
-0.0249440148,
0.0503161922,
-0.0187042188,
-0.0339900851,
-0.0371563248,
-0.0547887199,
-0.0029808911,
0.0378121957,
0.0835181326,
0.0123173082,
-0.0126230642,
0.0108716218,
0.0370047539,
-0.057332132,
0.0402127504,
-0.0103802923,
0.0137649421,
-0.0795568749,
-0.0438183546,
0.0374426022,
0.0287142936,
-0.0258552339,
0.0273617879,
-0.0359229259,
-0.0607256629,
0.0056375321,
0.0324812755,
0.0149070602,
0.0343060307,
0.0051545794,
-0.0375573523,
0.0430734903,
0.0936290845,
-0.0073581324,
0.0226551238,
0.0171850473,
-0.0481385402,
0.0080090743,
0.0360324793,
-0.0474820249,
0.0043059094,
-0.0152356569,
0.0556288399,
-0.015949674,
-0.0022236318,
-0.0327948928,
0.0497382991,
0.0034925505,
-0.034750957,
-0.0334918685,
-0.032436844,
0.059040565,
0.0311076809,
0.0083462587,
-0.0524634644,
0.0410661399,
0.0324691497,
0.03065194,
-0.0191695858,
0.0478461199,
0.0772764757,
-0.0495276526,
-0.0310383737,
0.0253315829,
-0.0176006109,
-0.02358808,
0.024720557,
0.0107356356,
0.0150880329,
-0.0290021487,
0.0255882163,
0.0203292705,
-0.0137625113,
0.0132127889,
0.0418748669,
0.0004571496,
-0.0433712117,
0.0566022284,
-0.0449180491,
-0.0560283177,
-0.0452385619,
0.0566731654,
0.0858185589,
0.0277594868,
0.0593686216,
0.001100669,
-0.0264404416,
0.0118804937,
-0.032648474,
-0.0175443776,
-0.002430239,
0.0099773444,
0.02776682,
-0.0038653926,
0.0261143204,
0.060539525,
-0.0237146076,
0.057214275,
0.0033366987,
-0.0046471576,
-0.0466115586,
-0.0096494025,
0.0520145185,
-0.0062301476,
-0.0239850041,
-0.0315243378,
0.0348552763,
-0.0444225781,
0.0038793376,
-0.0029694608,
-0.004618058,
-0.001544514,
0.0687667578,
0.0215267986,
-0.0066834539,
-0.0135730663,
-0.0223098621,
0.0314827263,
-0.0026387051,
-0.0030618669,
0.0147030605,
-6.338918674e-33,
-0.0121706827,
0.0210805889,
0.0170180276,
-0.0122056305,
-0.0331268497,
0.0497117713,
0.002012027,
-0.0448026471,
0.0031316448,
-0.0010998753,
-0.0250959452,
0.0318042524,
0.035463281,
0.0157562923,
-0.0408939831,
-0.0237174183,
0.0383509099,
-0.0245778151,
-0.0252742618,
0.0065753246,
0.0098558366,
0.1037425175,
0.0470000133,
-0.0128182983,
-0.0541766249,
-0.0029804248,
0.0027109105,
0.0207498409,
0.017151149,
0.0205666665,
-0.0213014483,
-0.0000832728,
-0.0104605164,
-0.0400495715,
0.004690934,
0.0025747695,
-0.0662379637,
0.0067015961,
-0.0148471445,
0.0462391451,
-0.0609336756,
0.0320470557,
0.0211873949,
0.0311157852,
0.0277680121,
0.0282564946,
0.0105041452,
0.0146219563,
-0.0366153978,
-0.0670213923,
-0.0066795996,
0.0366632864,
-0.0009749173,
-0.0181416068,
0.0427488387,
-0.0034165245,
0.0145351961,
0.008758788,
-0.0520588905,
-0.0013301487,
0.0506706797,
0.0167760439,
0.0497734882,
0.0688456371,
0.0453011617,
0.076254338,
0.0707050264,
-0.0202850699,
0.0211881101,
-0.0182653461,
0.0079392763,
0.037191689,
0.009763266,
-0.0207856745,
0.035448622,
0.0200729314,
-0.0752838179,
-0.0106299846,
0.0683143139,
-0.0294565689,
0.0089031579,
0.0372941084,
-0.0276483987,
0.0128915785,
0.0139075136,
-0.0250777304,
-0.0064234897,
0.0043796152,
0.0155323427,
0.0273917541,
-0.0296656601,
0.0556462705,
0.0067499639,
-0.0325146653,
-0.0090500517,
-0.003674923,
0.0331827365,
0.0067807082,
0.0179477539,
0.0310756993,
-0.070943661,
-0.0299209636,
-0.0195512399,
0.0078816423,
0.0160214156,
-0.0029129521,
0.0622437447,
0.0232208613,
-0.0515058488,
-0.0031230438,
-0.0217905156,
0.0113652954,
-0.0103312451,
0.0827649534,
0.047326941,
-0.0387865826,
0.0274206381,
0.0263214577,
0.0121623594,
0.0306434054,
0.0065449029,
0.024310546,
-0.0175029505,
0.011648722,
0.0048406725,
0.0037530474,
-0.0495298132,
0.0276269354,
0.0172738396,
-0.0856125653,
-0.0215032361,
-0.022996854,
2.813e-7,
0.0176159479,
0.0208995827,
0.0260498282,
-0.0397428796,
0.0027476803,
-0.0084938984,
-0.0191815533,
0.0405641347,
0.0167875234,
0.0562759414,
0.0185502786,
0.0365819708,
0.0268301461,
-0.0097848754,
-0.109719418,
-0.0329034477,
-0.0908097774,
0.02360004,
-0.0288316738,
-0.0268080905,
-0.0535588078,
0.0897986144,
0.0099771591,
-0.0344113298,
0.0747121572,
-0.0279718172,
0.0219140369,
-0.0246767607,
0.0031318564,
0.0189208146,
0.0252336338,
0.0107476544,
-0.0138659654,
0.0111102182,
-0.0117875021,
0.0040600589,
0.0111887204,
0.055629123,
-0.0434339941,
-0.0160538945,
0.0401256047,
-0.0049388283,
-0.0072342372,
-0.0165468398,
0.0140490923,
0.0180736482,
-0.020375086,
0.017447494,
-0.0167295001,
-0.0364794694,
0.0414978713,
-0.0523177683,
-0.008713536,
0.0171173159,
0.0185510591,
-0.0623182505,
-0.0249829423,
-0.0780398101,
-0.011381167,
-0.0176517777,
-0.0328758173,
-0.0707550272,
-0.0618938915,
-0.0155498767,
0.0165032297,
-0.0749477297,
-0.0069430987,
2.393415094e-34,
0.0336718075,
-0.0145882284,
0.0137463389,
0.0573507026,
0.0245616082,
-0.0342615508,
0.0020261374,
-0.0320125595,
-0.0363866054,
-0.04555998,
-0.0412584879
] |
http://arxiv.org/abs/2402.02446v3 | http://arxiv.org/pdf/2402.02446v3 | LQER: Low-Rank Quantization Error Reconstruction for LLMs | 2024-02-04T10:59:52 | 2024-05-30T09:49:47 | Cheng Zhang, Jianyi Cheng, George A. Constantinides, Yiren Zhao | cs.LG | cs.LG, cs.CL | Post-training quantization of Large Language Models (LLMs) is challenging. In
this work, we introduce Low-rank Quantization Error Reduction (LQER), which
combines quantization and low-rank approximation to recover the model
capability. LQER leverages an activation-induced scale matrix to drive the
singular value distribution of quantization error towards a desirable
distribution, which enables nearly-lossless W4A8 quantization on various LLMs
and downstream tasks without the need for knowledge distillation, grid search,
or gradient-base iterative optimization. Unlike existing methods, the
computation pattern of LQER eliminates the need for specialized Scatter and
Gather processes to collect high-precision weights from irregular memory
locations. Our W4A8 LLMs achieve near-lossless performance on six popular
downstream tasks, while using 1.36$\times$ fewer hardware resources than the
leading state-of-the-art method. We open-source our framework at
https://github.com/ChengZhang-98/lqer | LQER: Low-Rank Quantization Error Reconstruction for LLMs: Post-training quantization of Large Language Models (LLMs) is challenging. In
this work, we introduce Low-rank Quantization Error Reduction (LQER), which
combines quantization and low-rank approximation to recover the model
capability. LQER leverages an activation-induced scale matrix to drive the
singular value distribution of quantization error towards a desirable
distribution, which enables nearly-lossless W4A8 quantization on various LLMs
and downstream tasks without the need for knowledge distillation, grid search,
or gradient-base iterative optimization. Unlike existing methods, the
computation pattern of LQER eliminates the need for specialized Scatter and
Gather processes to collect high-precision weights from irregular memory
locations. Our W4A8 LLMs achieve near-lossless performance on six popular
downstream tasks, while using 1.36$\times$ fewer hardware resources than the
leading state-of-the-art method. We open-source our framework at
https://github.com/ChengZhang-98/lqer | [
-0.0595781915,
0.0968836397,
0.016821228,
0.0778804347,
0.0408877097,
0.0191990826,
-0.008773312,
0.0240040757,
0.0273287315,
-0.0152169606,
-0.0404733457,
-0.0608613789,
-0.00136508,
-0.0236627907,
0.0461593643,
0.012904278,
0.0411866978,
-0.0301520359,
-0.0061530774,
-0.0416308418,
-0.0149212005,
-0.004980234,
0.0072414353,
0.0109919906,
0.0097575476,
-0.0284885559,
-0.0064258627,
0.022937173,
0.0041951691,
-0.0247648675,
0.0239039045,
0.0371059105,
0.0611236021,
0.0300137252,
0.0000022137,
-0.010374831,
-0.0239466578,
-0.0236267187,
-0.0227389149,
0.0364159457,
-0.0175277963,
-0.0040209456,
-0.0411974154,
0.0114788311,
0.009735005,
0.0048466451,
0.0499850549,
0.0086884107,
0.0564649105,
0.0325936601,
0.0065628742,
-0.0177757889,
0.0316092819,
-0.0272021554,
-0.0174617451,
-0.0349505134,
0.0101972586,
0.0240211226,
-0.0668769851,
0.008515263,
-0.0242927894,
0.0395030938,
-0.0109958742,
0.0126163168,
-0.0358852334,
0.000973198,
-0.033772707,
0.0127827786,
-0.0112170186,
0.0575641692,
-0.0666790828,
-0.0313240103,
-0.0110458601,
0.023562612,
-0.0103349611,
-0.1143814847,
-0.0408802181,
0.059052743,
0.0446965769,
0.0499692,
-0.0374502838,
-0.0373563245,
0.053161554,
0.0150042726,
-0.0378616564,
0.0446569249,
0.0212279111,
-0.0073357001,
-0.0276336428,
0.0013546456,
-0.0737156644,
-0.0143260742,
-0.0020350199,
-0.0078266552,
-0.0238034036,
-0.0057664909,
-0.0458286256,
-0.0331536047,
-0.0236641373,
0.0090111503,
0.0461416468,
0.0474335402,
-0.0078110569,
0.0258682817,
-0.008287007,
0.07052432,
-0.0129191522,
0.0374197438,
-0.0359813347,
0.0240469333,
0.0129735591,
0.0053501544,
-0.0225563925,
0.0436719805,
-0.0526994281,
-0.0004838224,
-0.0333428569,
-0.0282024667,
-0.0188655499,
0.0019103757,
0.0089406557,
0.0320780799,
-0.034456443,
-0.0149496831,
0.0260358695,
0.0375471339,
0.0290991422,
0.0183247663,
-0.0005571035,
-0.0218909867,
-0.0339841768,
0.016012229,
0.014310156,
-0.0250568986,
0.0407919921,
-0.0103254328,
-0.0254299548,
-0.0512220711,
-0.0365403742,
0.0218813028,
-0.019089777,
-0.0610414259,
-0.0170006007,
-0.0647711083,
0.0174160302,
-0.0220756624,
0.023299858,
-0.0565904342,
0.0256319009,
-0.0069558257,
-0.0224877354,
0.0670848489,
-0.0366669521,
-0.0049458239,
0.0174299236,
0.0278433524,
0.0573666543,
0.0460989773,
0.0206149444,
0.0439190716,
-0.001568587,
0.0125346277,
-0.0229895059,
0.0491908714,
0.0441886485,
-0.0041680057,
-0.0566779003,
-0.0333126485,
-0.0392389484,
0.0396390669,
0.1110338867,
0.1024941877,
-0.0731041059,
-0.0054516261,
0.0411550738,
0.0902722701,
0.0347642042,
0.0219379142,
-0.0179304406,
0.0015860402,
0.0096534221,
-0.0339320041,
-0.0276980717,
0.0063149477,
0.0159523748,
-0.0140215196,
0.0621130578,
0.0459471121,
-0.0425138995,
-0.0254265927,
-0.0073911995,
-0.02177581,
-0.0109372288,
0.0050749346,
0.0417577736,
0.0086838966,
-0.0049145743,
0.1074007824,
0.0274595637,
-0.0769145116,
0.00362908,
-0.0761056319,
0.0165893026,
0.0411881544,
0.0304159205,
-0.0210645869,
0.0237668045,
0.0001481744,
-0.0104660373,
-0.0196598973,
-0.037287131,
0.0030860368,
0.0319865309,
-0.0496935137,
-0.0321122818,
-0.0108013246,
0.0155619206,
-0.0137187168,
0.024102211,
-0.0076074773,
-0.0054790163,
-0.0612337925,
-0.0172168277,
0.0122293066,
0.0202739984,
-0.0596477315,
0.0400035642,
-0.0212676749,
0.017923722,
0.0300865546,
-0.023763746,
0.0004683095,
0.0421183743,
-0.0424083248,
0.0252699461,
-0.075718157,
0.003480433,
-0.0456063785,
0.0696769282,
0.0240203887,
0.0443494841,
0.1138799489,
-0.0415312201,
-0.0647945926,
0.0300871115,
0.0048171906,
0.007673684,
-0.0675664619,
0.0070460015,
-0.0081575029,
0.0657071173,
-0.0196545403,
0.0322050788,
-0.039818041,
-0.0002379534,
0.0296893995,
0.0170242656,
-0.0389425233,
-0.0436521098,
-0.0740630254,
0.0038177355,
-0.0037339409,
0.0893116891,
0.0011513062,
-0.0029780674,
0.0070909266,
0.0097174207,
-0.112907365,
0.0184155498,
-0.0520430543,
0.0166666638,
0.0331291705,
0.0379024446,
0.0217976645,
-0.0315765142,
-0.0088679446,
0.0064152526,
0.0108558731,
-0.0154836383,
-0.0118397102,
-0.0130366823,
0.0102932295,
-0.0460580178,
-0.0397292562,
0.0230083335,
0.0272467248,
-0.0796161592,
-0.0035484554,
-0.1014245376,
0.0402731821,
-0.0243925378,
0.0468267351,
-0.0217536762,
-0.0311343465,
0.0033131223,
-0.0344681889,
0.0031381447,
0.0270831138,
-0.0032986363,
0.0322535001,
0.0770877302,
0.0163697135,
-0.0542971417,
0.0142172435,
0.0002776187,
-0.0511608645,
-0.0156719144,
-0.0082169855,
-0.0070111589,
0.0601000674,
-0.0021064258,
-0.003552845,
-0.0081312079,
-0.0014439675,
0.0227280203,
-0.0038294883,
-0.0767484754,
0.0248818509,
-0.0388560034,
0.0468645617,
-0.000043942,
-0.036169935,
-0.0105367592,
-0.0490792356,
-0.0062964112,
0.007113412,
-0.0144870048,
0.0039507775,
0.043794211,
0.0181684252,
0.0277044624,
-0.0485176928,
-0.0225345269,
-0.0396618545,
-0.0203405228,
-0.0464613251,
-0.0504938662,
0.0408325605,
-0.0335622504,
0.0295544676,
-0.0380987152,
-0.0291715804,
0.0139540937,
0.0127825532,
0.0028337524,
0.0160819963,
0.0471300371,
-0.0416540876,
0.0533472449,
0.0322243832,
0.0262304917,
0.0595789701,
-0.0049958564,
-0.0031212252,
-0.0282616299,
0.0881049186,
0.0293553025,
0.0169927143,
0.0524153784,
-0.0056840372,
0.012101342,
0.0472767055,
0.068965584,
0.0330351815,
-0.0542705469,
0.0039124107,
0.0562658943,
0.0509541146,
-0.0146641731,
0.0090821972,
0.0452008247,
-0.0478872061,
-0.0184381176,
0.0647009611,
0.0071393391,
-0.0671267658,
0.0077851755,
0.0000108702,
0.027328508,
0.018082587,
0.0254851431,
-0.088535428,
0.0012314387,
0.023882227,
0.0367149077,
0.0702470466,
-0.0113922972,
-0.001915706,
-0.0597102605,
0.0430946313,
0.0442731604,
0.0063592996,
-0.0316985026,
-0.007981386,
-0.0099703902,
-0.0209932681,
0.0874815881,
0.0597264022,
-0.0057251356,
-0.0096167466,
0.0231199861,
-0.0125622302,
-0.0635421872,
0.0221284591,
0.0047563557,
0.0772667006,
0.0422446355,
0.0228024516,
-0.0078961458,
-0.0041912352,
0.007092142,
-0.0110136243,
-0.025489701,
-0.0001080302,
0.0165132899,
0.0223673172,
-0.0213796198,
0.0268541947,
0.0312281679,
-0.0151855741,
-0.0251666903,
-0.0259065907,
0.0505833998,
-0.0243921541,
-0.0135974148,
0.0408264212,
-0.0566010773,
-0.0119988769,
0.0284709111,
0.0206890851,
0.0103985155,
-0.0475869477,
-0.016275635,
0.0393204466,
-0.0348209254,
-0.0683154464,
-0.0187482592,
-0.0074852179,
-0.007351988,
-0.0453798734,
0.0032899028,
0.0169810317,
0.0059960252,
-0.0596883632,
-0.0209467188,
-0.0140826302,
-0.0195275247,
-0.077910468,
0.003540989,
-0.0532482676,
0.0233857576,
-0.0196176209,
0.0090967482,
-0.0093010087,
-0.0254255366,
-0.013006351,
-0.0252776165,
-0.1174919009,
-0.0491340347,
-0.0006121302,
0.0266159941,
0.0082717426,
0.0448076837,
0.0188958719,
-0.0534853488,
0.0691338927,
0.0394554771,
-0.0056217024,
0.0517888591,
0.0191624314,
-0.002867982,
-0.018395381,
0.0460055023,
-0.0342295766,
0.0168310944,
0.0161613412,
0.0852682516,
-0.0472925007,
-0.0345883183,
-0.006792801,
0.0011923001,
0.0650757626,
0.0032784054,
-0.0116276843,
-0.0094993822,
0.0080085946,
0.0019373497,
-0.0010149324,
-0.0515245087,
-0.0040006926,
0.0148049686,
-0.029765157,
-0.0172428675,
-0.0071269395,
0.043387536,
-0.0078536002,
0.0082657756,
-0.0224019438,
-0.041694086,
0.0000725832,
0.0516702123,
-0.0045493413,
-0.022332184,
0.0160930101,
0.0190306921,
0.0385667048,
-0.0165860858,
-0.0177144371,
-0.0214317888,
-0.0037114099,
-0.0417063721,
0.0240193475,
-0.0267890189,
-0.0845376179,
-0.0551988855,
-0.0107596563,
0.0702321157,
0.0339657106,
0.0786682144,
0.0336405374,
-0.0197425559,
-0.0238347221,
-0.009038914,
0.0144247022,
0.0078178886,
0.0393982083,
0.0187018998,
-0.0169536974,
0.0175505262,
0.0741761699,
0.0182342473,
0.0238203499,
-0.0142548745,
-0.0111305853,
-0.0309989396,
-0.0628625825,
-0.0314276405,
0.0075938995,
-0.0281456374,
-0.0521104075,
0.0296043754,
-0.0584213436,
0.0075125317,
0.0317475498,
-0.0390484668,
-0.0257273577,
-0.0042534932,
0.0402205326,
0.0085617993,
0.0067723007,
-0.0311668869,
-0.0543273389,
0.0075502596,
-0.0230698623,
0.0025735621,
-6.775388542e-33,
-0.0311766863,
0.0153202778,
0.0131009221,
-0.0292278063,
-0.0180475339,
0.0185236372,
-0.0218935143,
0.0127836987,
0.011094314,
-0.005382875,
-0.0242617875,
0.045056425,
0.0162537899,
0.0293918103,
-0.0021679832,
0.0508350842,
0.0710690469,
-0.038471695,
-0.0169134978,
0.0382761769,
0.0277563687,
0.047841385,
0.0497034751,
0.0304520968,
-0.0285912212,
0.0054752217,
-0.0071608666,
0.0111378096,
0.0360552147,
-0.0074059707,
-0.0612166151,
0.0310597196,
0.0105144028,
0.0099550001,
-0.0390596949,
-0.0069417786,
-0.0724657327,
0.0201831069,
-0.0048273667,
0.0011921668,
-0.0289491639,
-0.000609471,
0.0263116863,
0.0083870972,
-0.0215569697,
-0.0542037785,
-0.0242507011,
-0.0260922033,
-0.0055851592,
-0.0788899288,
0.0100071561,
0.0090611232,
0.0079588145,
0.0116152493,
0.0653169677,
-0.0282835551,
-0.0201845802,
0.0004392349,
-0.1101968735,
-0.0171283707,
-0.0179820806,
0.0819344446,
0.0579914078,
0.034383513,
-0.0088995006,
0.0466281585,
0.0444142669,
-0.0452062674,
-0.0138860364,
0.0310360678,
0.0538672172,
0.0352819562,
0.0210316218,
0.0345575027,
0.0593441017,
-0.0014149097,
-0.0562933832,
0.0025723609,
-0.0238088127,
0.0342142098,
0.0221787021,
0.0636221543,
0.00823073,
-0.0487712882,
-0.0188118815,
0.0064371279,
-0.0618578531,
-0.0311964545,
-0.0056977365,
0.0160437226,
-0.0401517078,
-0.0083685154,
0.0304191634,
-0.0585302822,
-0.007602497,
-0.0006484509,
0.0166621301,
0.0117876753,
0.0003050569,
0.0297297779,
-0.0243568998,
-0.0196782723,
0.0168034676,
0.0048129386,
0.0009870018,
0.0007477607,
0.0241014697,
0.066814661,
-0.0614812858,
-0.0188094489,
-0.022708945,
-0.0079105208,
-0.0086135231,
0.0358978137,
0.0043146391,
0.0178585853,
0.0267854836,
0.0436635949,
-0.0014590377,
0.0392157063,
-0.0053752512,
-0.0051765051,
0.009015155,
0.0312280394,
-0.0521246195,
0.0090203807,
-0.0616103373,
0.0644963756,
0.0481491834,
-0.0477961749,
-0.0322425663,
0.0011435859,
2.878e-7,
0.0456314497,
0.0136269936,
0.0027958055,
-0.0041078907,
-0.0080135036,
-0.0130633526,
-0.0103930822,
0.0113944476,
-0.0145964921,
-0.0109684113,
0.0362273157,
0.0205573142,
0.0527683459,
0.012482008,
-0.0348302275,
0.0045446544,
-0.0657219961,
-0.0169903953,
-0.022834029,
-0.0640024766,
0.0224097427,
0.0895361528,
-0.0138906538,
-0.0392113589,
0.0162167232,
0.0258313026,
0.064621225,
-0.0011321624,
0.0204905905,
0.018910246,
0.0416210406,
0.0179016441,
-0.0049284389,
-0.0429038331,
0.002114668,
0.0016857516,
-0.0416382216,
0.061545223,
-0.0148842055,
-0.0219858754,
0.0059051798,
0.0310099479,
-0.0060747797,
-0.0495179743,
0.0380734056,
-0.0184684377,
0.0186144263,
-0.0863877088,
-0.0239958726,
-0.0147938197,
-0.008611178,
-0.03448838,
0.0197526775,
0.0087701492,
0.0063138092,
0.0119416118,
0.0057417559,
-0.0981559604,
-0.0139931496,
0.0207747072,
-0.0455718376,
-0.0466593765,
-0.0158217642,
-0.0223968811,
-0.0134591432,
-0.0645831451,
0.022782309,
2.629682571e-34,
0.0320613123,
0.0050274795,
0.0262226127,
0.0387542397,
0.0078936433,
-0.0410379246,
0.0006889176,
0.0068347757,
-0.037156608,
-0.0651816949,
-0.0313481912
] |
http://arxiv.org/abs/2402.02420v2 | http://arxiv.org/pdf/2402.02420v2 | Factuality of Large Language Models in the Year 2024 | 2024-02-04T09:36:31 | 2024-02-09T06:36:41 | Yuxia Wang, Minghan Wang, Muhammad Arslan Manzoor, Fei Liu, Georgi Georgiev, Rocktim Jyoti Das, Preslav Nakov | cs.CL | cs.CL, cs.AI | Large language models (LLMs), especially when instruction-tuned for chat,
have become part of our daily lives, freeing people from the process of
searching, extracting, and integrating information from multiple sources by
offering a straightforward answer to a variety of questions in a single place.
Unfortunately, in many cases, LLM responses are factually incorrect, which
limits their applicability in real-world scenarios. As a result, research on
evaluating and improving the factuality of LLMs has attracted a lot of research
attention recently. In this survey, we critically analyze existing work with
the aim to identify the major challenges and their associated causes, pointing
out to potential solutions for improving the factuality of LLMs, and analyzing
the obstacles to automated factuality evaluation for open-ended text
generation. We further offer an outlook on where future research should go. | Factuality of Large Language Models in the Year 2024: Large language models (LLMs), especially when instruction-tuned for chat,
have become part of our daily lives, freeing people from the process of
searching, extracting, and integrating information from multiple sources by
offering a straightforward answer to a variety of questions in a single place.
Unfortunately, in many cases, LLM responses are factually incorrect, which
limits their applicability in real-world scenarios. As a result, research on
evaluating and improving the factuality of LLMs has attracted a lot of research
attention recently. In this survey, we critically analyze existing work with
the aim to identify the major challenges and their associated causes, pointing
out to potential solutions for improving the factuality of LLMs, and analyzing
the obstacles to automated factuality evaluation for open-ended text
generation. We further offer an outlook on where future research should go. | [
0.0544419251,
0.0329420343,
-0.0183393769,
0.0577035174,
-0.0514047518,
-0.0343793444,
0.0187323987,
-0.0179762449,
0.0066312361,
-0.0652444512,
0.0140933907,
-0.0706802607,
0.0274676867,
0.0149982926,
0.0225680955,
-0.0157490708,
0.0293658935,
-0.012654596,
-0.0074969283,
-0.0188231841,
0.0031395894,
0.0398500636,
-0.0198332071,
0.0489179902,
-0.0268538129,
-0.0182988402,
-0.0146742193,
0.004094332,
-0.0078377835,
0.0122264614,
-0.0065034744,
0.0248681884,
-0.0038647098,
0.0566568933,
0.0000022258,
-0.042383261,
0.0139929932,
-0.0210055541,
-0.0482071564,
0.0444596261,
0.0426444076,
0.0187109467,
-0.0148904128,
0.0196940675,
-0.0278273337,
0.0150709907,
0.039550852,
0.0034957556,
0.084330745,
0.0747757256,
0.0084602023,
0.0037670725,
0.0189065989,
0.0153399818,
0.0103983441,
0.0249372441,
0.0190110113,
0.0401018895,
0.0229467675,
0.0136170657,
0.0225076079,
0.0463274941,
0.0177166201,
-0.0449607559,
0.0068629836,
0.046875488,
-0.0353481434,
-0.0294718072,
0.0276139062,
0.0536436774,
0.0325848833,
-0.0117740026,
-0.0017312121,
-0.0201232042,
-0.0100727454,
-0.0043554883,
-0.0298343021,
0.0765556991,
-0.0305083301,
0.0084266318,
0.0060354001,
-0.0178374574,
0.0161971319,
0.0250977427,
-0.0112008583,
0.0563791804,
0.0106554078,
-0.0261391923,
-0.0594652481,
0.0170548819,
-0.0057152873,
-0.085984461,
-0.0136784548,
0.0257811062,
-0.0402893238,
0.0349902287,
-0.0523112267,
0.022224851,
0.039473068,
-0.0659946427,
0.0274442956,
0.0280625187,
-0.0146766137,
-0.0001525557,
0.0230815671,
0.0303539578,
0.0054133558,
-0.0055047604,
-0.0221304726,
0.0344267301,
-0.0049844547,
-0.0026908466,
-0.015021489,
0.0252850167,
-0.0272464547,
0.002896664,
-0.0312593691,
0.0206706598,
0.0051887366,
0.0345654115,
-0.0260866657,
0.0115393568,
-0.0481797121,
-0.0037437922,
-0.0411956236,
0.018707145,
-0.0200877916,
0.0147834411,
0.0135991424,
-0.0926635936,
0.0173440743,
0.0493684076,
-0.0233975463,
-0.0001963426,
-0.0289903823,
0.0235765819,
-0.0076061208,
0.0052822307,
-0.037034791,
-0.0684094205,
-0.0388468802,
-0.0782903582,
-0.0307867024,
-0.0763823465,
0.0208327007,
0.012866945,
-0.0181974526,
-0.011520816,
0.0098878639,
0.0523795336,
-0.0192822255,
0.0360269994,
-0.0248974301,
0.0116144409,
0.0438729078,
0.004536089,
-0.0004405278,
0.0488444045,
0.069581382,
0.0347772725,
0.063973546,
0.0243999436,
0.0063297399,
0.0252664965,
0.0061488342,
0.0109383604,
0.0024115148,
0.0026530894,
-0.0185201727,
0.0585738197,
0.004301989,
0.0707399696,
-0.0290693101,
0.0036664943,
0.0953035355,
0.0077668941,
-0.0115085319,
0.0196828898,
0.0704773143,
0.0409684516,
0.0470769517,
0.0725490674,
-0.01736545,
0.0172633138,
-0.0344942585,
0.0475025289,
0.0484390557,
0.0503481738,
-0.0016385362,
-0.0729684159,
-0.0264437068,
0.0088240393,
-0.0237770136,
-0.0071240091,
0.0047271736,
-0.0108140483,
-0.0206692591,
0.0556908175,
0.0496575683,
-0.0728824586,
0.021304911,
-0.0483783856,
0.0307835843,
0.0625986457,
0.0711794347,
0.0066494541,
-0.0477293618,
-0.0022116716,
0.0259378143,
0.0726105571,
0.0534396879,
-0.0017984991,
-0.0166342407,
-0.0378674418,
0.0288356598,
-0.0168829449,
0.0145258373,
-0.0182435922,
0.0255818702,
-0.0177898146,
0.0198657848,
-0.0081833797,
-0.0389559194,
0.0186192691,
-0.0453025438,
-0.0024566618,
-0.0109860376,
0.0026444094,
0.0018393113,
-0.0031809881,
-0.0298517365,
0.0243027322,
-0.0172834229,
0.001917886,
-0.0164160486,
0.0074620852,
-0.0506623276,
-0.0222166907,
0.0526352748,
-0.0084462194,
-0.010699328,
0.0069056889,
-0.013359922,
-0.0154929683,
-0.0118249943,
-0.0653731376,
0.0304220449,
-0.0283466298,
0.025178954,
0.0353592522,
0.0034641644,
-0.0247164443,
0.0536828451,
0.0004204115,
0.0128367404,
0.0244490486,
0.0017834256,
0.0082448395,
-0.0448897816,
0.0047122152,
0.0422411151,
0.0093052667,
0.0313697867,
0.0470714793,
0.0194231309,
-0.0526288524,
-0.0168618523,
-0.08637546,
0.0077097053,
-0.0204757284,
0.0086910622,
0.0290312879,
-0.0049925628,
-0.0149453888,
0.0162573364,
-0.0588033982,
-0.0062163789,
0.021010533,
-0.001761036,
0.077019304,
-0.0102173006,
-0.0008841085,
-0.0453321263,
0.013352152,
-0.0160014089,
0.0163273662,
0.0519846715,
-0.0047786972,
-0.0587030314,
0.0471637547,
-0.0910673961,
-0.0040301238,
0.01452538,
-0.048049327,
0.0137021756,
0.0193214435,
0.0555006042,
0.0256018285,
0.0438178331,
0.0264015365,
0.0574104637,
0.0003838276,
-0.0394510999,
-0.0009879313,
-0.020044202,
-0.0340157561,
0.0101044104,
-0.0874802843,
0.020213306,
0.0279545952,
0.0438255742,
-0.0085081384,
-0.0471974909,
0.0208999421,
-0.0474157333,
-0.0269788634,
-0.0390582494,
-0.0096446341,
-0.0270858947,
0.0252924487,
-0.005411569,
-0.0503204465,
0.0850376412,
-0.0652898178,
0.0247340184,
-0.044280991,
-0.010650388,
-0.0255512092,
0.0171070993,
-0.0089191785,
0.0340148956,
0.0325429961,
0.0120139029,
-0.0374468341,
-0.0042856699,
0.049401246,
-0.0409885868,
0.0622390695,
-0.0834268928,
0.0137987891,
0.0153526915,
-0.0345507972,
-0.0407473296,
-0.0942972526,
0.009514207,
-0.0260436721,
0.0162211061,
-0.018172482,
-0.0365302414,
-0.0082158167,
0.0653615296,
-0.0014036372,
-0.0192345995,
-0.0239536874,
-0.0261249803,
0.0318202116,
0.0738794282,
0.0114626959,
0.0412768051,
-0.030695254,
0.0261099469,
0.0738011375,
-0.0132498555,
0.0240932759,
-0.0020097997,
-0.0400736295,
0.0179086141,
-0.0200194605,
-0.0412481204,
0.0353082232,
0.0765111968,
-0.0443698019,
-0.0438972041,
0.1120834649,
0.0435351357,
0.015027862,
0.0199194402,
-0.0029308079,
0.0339242667,
-0.020838039,
0.0198824033,
-0.0905458406,
0.0075912252,
-0.0046711997,
0.0257394612,
-0.0455391817,
-0.0346060768,
-0.0363696404,
-0.1255169362,
0.0438214391,
0.0241380204,
-0.0092419237,
-0.0438124016,
-0.0073189232,
-0.0060270568,
0.0289921984,
0.0385284387,
0.0106230127,
-0.0229964778,
0.0225588884,
0.0436433256,
-0.0606556535,
-0.0219197944,
0.0091643119,
-0.0170206111,
0.0521921813,
0.0159071814,
0.0321992822,
0.0088619478,
-0.0124742193,
0.0024563114,
0.0235701911,
-0.0103633516,
0.0131381406,
0.0136154471,
-0.0050572623,
-0.0028861454,
-0.0538363829,
0.010978722,
0.0254091304,
-0.0162055958,
0.0036320621,
0.0444210023,
-0.0416811109,
-0.0043286895,
-0.0026346594,
0.0551672056,
0.0165468324,
0.080982402,
0.0339290127,
-0.0132828308,
0.0033770276,
-0.0595612936,
-0.0164487679,
-0.0027449082,
-0.0459007248,
0.0124152415,
-0.023621764,
-0.0141391838,
-0.0919742882,
-0.0476569124,
0.0011458521,
0.0005077007,
-0.035127867,
-0.0429588519,
0.0406695195,
-0.030565029,
0.0278869849,
-0.0053268466,
-0.0559352785,
0.0308253281,
-0.0428798571,
-0.0065516038,
0.016051624,
-0.0498761721,
0.0339036919,
0.0111446679,
-0.0822362006,
-0.0653553903,
-0.0066743181,
0.0231783856,
0.0208990909,
0.0404558815,
-0.0196215212,
-0.0433257036,
0.0174745526,
0.0404040627,
-0.0523040742,
0.0400642492,
0.0035288136,
0.0281510856,
0.0116039021,
0.0433871411,
-0.0061758077,
0.001200766,
-0.0374150425,
0.050516136,
-0.0147747882,
-0.0832108632,
0.0466803499,
0.0687271506,
-0.0100735351,
-0.0195117481,
-0.0183765721,
-0.0240063947,
0.0155276936,
-0.0669198409,
-0.0480410792,
-0.0252376106,
-0.0022990506,
-0.019034585,
-0.002611432,
-0.0144978054,
-0.036682345,
-0.0538352504,
0.0482387319,
-0.0138989361,
-0.0078997361,
-0.0363022611,
0.0087559139,
0.038863834,
-0.0215278603,
-0.0602619722,
0.0291130003,
-0.0210346468,
0.0535775796,
0.0141581874,
0.0514369756,
0.015453428,
-0.0179937053,
0.0037982701,
-0.0023801997,
-0.0400736555,
-0.0281722248,
-0.035899289,
0.0668432787,
0.0602994524,
0.0204454418,
0.0728429705,
-0.0587065816,
-0.0071281241,
-0.0107991183,
0.0323701575,
0.0328067914,
-0.0597901121,
0.0039577321,
-0.0010676968,
-0.0432609469,
-0.0006955035,
0.0833756849,
0.0008786677,
0.0350928009,
0.0051625352,
-0.0325947329,
-0.0359558351,
-0.0458981059,
-0.0237939078,
-0.0062886388,
-0.0053771394,
-0.0105782421,
-0.0891478732,
-0.0674149767,
-0.0186075531,
0.0486084595,
0.0169333834,
-0.0403735787,
0.0095027313,
0.0752300099,
0.0225643609,
-0.0023152847,
-0.0709344894,
0.0591830984,
-0.0301738884,
0.0386993848,
-0.0161130838,
-6.969177602e-33,
-0.0396398902,
-0.0572298132,
0.0281131081,
-0.0140508097,
-0.0290593281,
0.0102182562,
-0.033602342,
0.0129765747,
-0.0141664101,
0.02365385,
0.0098641561,
0.0036193784,
0.020816233,
0.0244826656,
-0.0177832432,
-0.0183714405,
0.0479777493,
0.0045527099,
-0.0667585284,
0.0193567704,
0.050839629,
-0.0414307527,
0.029825313,
-0.0413803272,
0.0735452175,
-0.0398191549,
-0.022828687,
0.0239612367,
-0.0438607931,
0.0521786027,
-0.0177212264,
0.0782349035,
0.005555606,
-0.0244647544,
-0.0155064594,
-0.0115412334,
0.004536625,
-0.0343092829,
0.0527782887,
-0.0126990136,
-0.016670838,
-0.0218012407,
0.0036685979,
-0.0156572741,
-0.0201045442,
0.0057790941,
0.0217873231,
-0.0185402799,
0.0057353294,
0.0065037496,
-0.0654720515,
0.0384510681,
0.0176172461,
-0.0188169908,
-0.0529911406,
0.0354140513,
-0.0020488184,
-0.0396097042,
-0.0479505733,
0.041392941,
0.0357623138,
0.0347793214,
0.0773571283,
-0.0261456091,
-0.0088459942,
0.0440853201,
-0.0271577109,
0.037331257,
-0.0584960654,
0.0051808408,
-0.0130891996,
0.0309028216,
0.0176760275,
0.0878784359,
-0.0228418019,
-0.0096818451,
-0.0551394485,
0.0175082013,
0.0420055538,
0.0110642044,
0.004893315,
-0.0367448032,
-0.0664763153,
-0.0251866765,
-0.023260951,
0.0548297912,
-0.01305083,
-0.037595313,
0.0039593973,
-0.0260186177,
-0.0225112345,
-0.0504550561,
-0.0114729023,
-0.0193856917,
-0.0119752306,
-0.0027091119,
-0.0633092374,
0.0091662798,
-0.021227099,
0.0347537249,
-0.042914141,
0.016782403,
0.0044508781,
0.0011518773,
-0.0150833456,
0.050799083,
0.0055305534,
0.0220480692,
-0.0151409162,
0.0298106074,
-0.0452097282,
0.0012150973,
0.0100340359,
-0.0528090671,
0.0259875469,
-0.0148084937,
-0.0499819219,
0.0323696882,
-0.0015431393,
0.0143630318,
-0.0231668595,
-0.0370136239,
-0.0004676255,
0.0251461621,
-0.0444893762,
0.0355162434,
-0.0576058291,
0.0057406989,
0.0352031998,
-0.0412333161,
-0.0188280307,
-0.0354713351,
2.887e-7,
0.0074175606,
0.0402669758,
0.0388154127,
0.0493991189,
-0.0201158673,
0.0332724154,
-0.0095788566,
0.0025601017,
0.0207131933,
-0.0280516837,
-0.0081691295,
0.0257818662,
0.0598474331,
0.0445483811,
-0.0315706432,
-0.005339676,
-0.0836517587,
-0.0178477876,
-0.0639488026,
-0.0263620522,
0.0608022586,
0.0453588255,
0.0180924665,
0.0061479411,
-0.0072404956,
-0.0277344193,
0.0463382415,
-0.0069396216,
-0.0983739495,
0.0399280153,
0.0832853317,
-0.0257916152,
0.0184881557,
-0.0094623771,
-0.0012845329,
-0.0219206419,
-0.0199125931,
0.050557375,
0.0462709032,
0.0027150838,
-0.0114025986,
-0.045569025,
0.0016994964,
-0.0576319806,
0.0529987775,
-0.0309383795,
-0.0202774853,
-0.042293407,
-0.0027247632,
0.0028578704,
0.0082574654,
-0.023243973,
-0.0564373285,
0.0291830897,
0.0007212573,
0.0079322895,
0.0068563875,
0.0084902262,
-0.0553104989,
0.0004260348,
-0.0318931118,
-0.0986423939,
-0.0018837117,
0.0046687219,
0.009887632,
-0.0442938432,
0.0227873474,
2.672232711e-34,
0.0256832298,
0.0157181192,
0.0227865223,
-0.0147608174,
0.058662869,
0.0299627166,
-0.0007310247,
-0.0054739979,
0.0234365407,
-0.0574079044,
-0.0229154527
] |
http://arxiv.org/abs/2402.02416v4 | http://arxiv.org/pdf/2402.02416v4 | Aligner: Efficient Alignment by Learning to Correct | 2024-02-04T09:24:51 | 2024-06-24T18:55:16 | Jiaming Ji, Boyuan Chen, Hantao Lou, Donghai Hong, Borong Zhang, Xuehai Pan, Juntao Dai, Tianyi Qiu, Yaodong Yang | cs.CL | cs.CL, cs.AI, cs.LG | With the rapid development of large language models (LLMs) and ever-evolving
practical requirements, finding an efficient and effective alignment method has
never been more critical. However, the tension between the complexity of
current alignment methods and the need for rapid iteration in deployment
scenarios necessitates the development of a model-agnostic alignment approach
that can operate under these constraints. In this paper, we introduce Aligner,
a novel and simple alignment paradigm that learns the correctional residuals
between preferred and dispreferred answers using a small model. Designed as a
model-agnostic, plug-and-play module, Aligner can be directly applied to
various open-source and API-based models with only one-off training, making it
suitable for rapid iteration. Notably, Aligner can be applied to any powerful,
large-scale upstream models. Moreover, it can even iteratively bootstrap the
upstream models using corrected responses as synthetic human preference data,
breaking through the model's performance ceiling. Our experiments demonstrate
performance improvements by deploying the same Aligner model across 11
different LLMs, evaluated on the 3H dimensions (helpfulness, harmlessness, and
honesty). Specifically, Aligner-7B has achieved an average improvement of 68.9%
in helpfulness and 23.8% in harmlessness across the tested LLMs while also
effectively reducing hallucination. In the Alpaca-Eval leaderboard, stacking
Aligner-2B on GPT-4 Turbo improved its LC Win Rate from 55.0% to 58.3%,
surpassing GPT-4 Omni's 57.5% Win Rate (community report). | Aligner: Efficient Alignment by Learning to Correct: With the rapid development of large language models (LLMs) and ever-evolving
practical requirements, finding an efficient and effective alignment method has
never been more critical. However, the tension between the complexity of
current alignment methods and the need for rapid iteration in deployment
scenarios necessitates the development of a model-agnostic alignment approach
that can operate under these constraints. In this paper, we introduce Aligner,
a novel and simple alignment paradigm that learns the correctional residuals
between preferred and dispreferred answers using a small model. Designed as a
model-agnostic, plug-and-play module, Aligner can be directly applied to
various open-source and API-based models with only one-off training, making it
suitable for rapid iteration. Notably, Aligner can be applied to any powerful,
large-scale upstream models. Moreover, it can even iteratively bootstrap the
upstream models using corrected responses as synthetic human preference data,
breaking through the model's performance ceiling. Our experiments demonstrate
performance improvements by deploying the same Aligner model across 11
different LLMs, evaluated on the 3H dimensions (helpfulness, harmlessness, and
honesty). Specifically, Aligner-7B has achieved an average improvement of 68.9%
in helpfulness and 23.8% in harmlessness across the tested LLMs while also
effectively reducing hallucination. In the Alpaca-Eval leaderboard, stacking
Aligner-2B on GPT-4 Turbo improved its LC Win Rate from 55.0% to 58.3%,
surpassing GPT-4 Omni's 57.5% Win Rate (community report). | [
0.040761821,
0.0884810612,
-0.0183749795,
0.030770354,
-0.0577051677,
0.0085796649,
-0.0338701159,
-0.0298491418,
0.0165555794,
-0.0237210747,
-0.0185673982,
-0.0274652224,
0.0108083151,
0.0025992552,
0.0085440818,
0.0015081904,
0.0775027126,
-0.0187319145,
0.0310570151,
-0.0116711073,
0.0242232028,
0.0169546641,
-0.014089413,
0.0391534641,
0.0167328529,
-0.0157215763,
-0.0364892147,
-0.0010536524,
0.0272744372,
-0.0256403927,
0.0212397277,
0.0717081055,
0.0055899904,
0.0725997314,
0.0000022366,
-0.0622643754,
-0.0399055667,
-0.011106668,
-0.0241722036,
0.0256004613,
-0.0023104455,
0.0195387769,
-0.0169486403,
0.0383571386,
-0.0431771092,
0.0159753785,
0.0905769914,
0.0123255942,
0.0911047086,
0.0488850735,
0.0128413783,
-0.0176337548,
0.0160841644,
-0.0088552125,
0.0045523108,
-0.0606082305,
0.0092886444,
-0.0048198509,
-0.0314254723,
-0.0429931767,
-0.0083115688,
0.0190941468,
0.0027730372,
0.0200063866,
0.0137306862,
0.016041249,
-0.0398951732,
-0.0292265154,
-0.0393757112,
0.0710246041,
-0.0093536861,
-0.0471280701,
-0.0248522982,
-0.0227473862,
0.0209062826,
-0.0096066641,
-0.0208045673,
-0.0096435659,
0.0022263082,
0.0240658037,
0.0144690089,
-0.0206640568,
0.0101823062,
0.002673957,
-0.0223597307,
0.0845969915,
0.0277779475,
-0.0495785922,
0.0114265988,
0.0075542009,
-0.0244286675,
-0.0625765026,
-0.006779158,
0.0049333493,
-0.0786800906,
0.0244269576,
-0.0434314422,
-0.0063550975,
0.0564677529,
-0.0486232378,
0.0303385779,
0.0440520644,
-0.0111460416,
0.0392137468,
-0.0154700503,
0.0909093022,
-0.0044518914,
0.0353530496,
-0.0217511095,
0.0020779329,
-0.0008949331,
-0.0227587353,
-0.0005799186,
0.0223545879,
-0.0124153243,
-0.0009918545,
-0.0705476627,
0.0208262652,
-0.0325721093,
0.0576656833,
0.022359319,
0.0057503399,
-0.0251539182,
-0.00489002,
-0.0327501222,
0.049481824,
-0.0290815104,
0.0179428104,
0.021670552,
-0.06129447,
0.0062173889,
-0.0152980071,
-0.0295993797,
-0.0282585137,
0.0258339494,
0.0713569745,
-0.0056628422,
-0.0093098562,
-0.0077485056,
-0.0014396643,
-0.0251498651,
-0.0731725916,
-0.0377019532,
-0.0888551027,
0.0465897806,
0.0074566873,
-0.0129380524,
-0.0650704652,
0.0147897666,
0.018397985,
-0.039048668,
0.0711122006,
0.0049466006,
0.0259523392,
-0.0142703475,
-0.0011165567,
-0.0029340088,
0.0690816566,
0.0196612347,
0.0440784432,
0.0462682508,
-0.0513630398,
0.0214081127,
0.0458999127,
0.0047623892,
-0.0097066117,
0.0024794526,
0.0188032612,
0.0321642645,
0.0334543586,
0.0110694375,
0.0402181223,
0.022064928,
-0.0417362452,
0.0989793316,
0.0110114804,
0.0033814621,
-0.007471133,
-0.0285985619,
0.0187782962,
-0.0040329583,
0.0016471468,
-0.0921264067,
0.0491806753,
-0.0513036661,
0.0246776119,
0.012843986,
0.0801249743,
-0.0076531246,
-0.0304447785,
-0.0194075592,
0.0015084527,
-0.0089456607,
0.0432517715,
0.0272468105,
-0.036820896,
-0.0762013122,
0.0476622134,
0.0408640504,
-0.0543210469,
-0.0345891081,
-0.0575623177,
0.0487447716,
0.0557800084,
0.051777117,
-0.0038164342,
-0.0451099277,
-0.024788605,
-0.0299364217,
0.0470180325,
0.0411396883,
-0.0056323246,
0.000128281,
-0.0527375303,
-0.0377406925,
0.0209322833,
0.0015897072,
0.0353102647,
0.0149451587,
-0.0123503879,
-0.0066868654,
-0.0117766932,
-0.0142151304,
0.0500314161,
0.0131996684,
-0.0367981419,
-0.0321540125,
0.0115363263,
0.0447990149,
0.0134535776,
-0.0100328075,
-0.0006353367,
0.0500431545,
0.0280981325,
0.0096152062,
0.0021965476,
-0.0367947817,
0.0209418833,
0.0889266133,
0.0064759431,
-0.0234290697,
0.0750985369,
-0.0355089456,
-0.0086288676,
-0.021832034,
-0.0299276225,
0.0428538248,
-0.0158644188,
0.0641710609,
-0.0321646109,
-0.0151280612,
-0.0467465334,
0.0371939354,
-0.016693512,
0.0142623791,
-0.0243998803,
-0.0328702442,
0.0176500268,
-0.0124611547,
-0.030845385,
0.0031618774,
-0.018637158,
0.0120595358,
0.0319841206,
0.006913343,
-0.0507871658,
-0.033961039,
-0.0768630505,
0.0231422,
-0.0037588538,
0.0157376416,
0.0339642316,
-0.0015760887,
0.005713779,
0.0011365101,
-0.0466196537,
-0.0123020345,
0.0366425067,
0.0277274176,
0.0047681779,
-0.0131553859,
-0.0639571473,
-0.0641712621,
-0.0086966269,
-0.036913868,
-0.0002188702,
-0.0347285718,
-0.0178692788,
-0.0469149016,
-0.0056682732,
-0.0545012057,
0.0335472301,
-0.0070441258,
-0.0701769739,
0.012777755,
0.0156627279,
0.0019942455,
0.0676473901,
0.035395585,
-0.0226185266,
0.0660520121,
-0.003483718,
-0.0881024152,
0.0015908952,
-0.0156609584,
-0.0709896684,
-0.0228309613,
-0.0504414774,
-0.0225201603,
0.0944735631,
0.0088821268,
0.0140196718,
-0.0647348762,
0.0016441108,
0.0070502455,
-0.0404960997,
-0.025383722,
-0.0253537782,
-0.0278424732,
0.0135606918,
-0.0296848789,
-0.0830460936,
0.1030872837,
-0.0331761315,
-0.0027941968,
-0.0420556627,
0.0144340331,
-0.003901623,
0.0366214179,
-0.0061697909,
0.0221070088,
0.0205172468,
-0.0207540542,
-0.0520834289,
-0.0572830997,
0.0290056728,
-0.0705796629,
0.0613524765,
-0.0533948839,
0.0013519035,
-0.0073615033,
-0.0570363402,
-0.0583717115,
-0.0560029671,
-0.0148227597,
-0.0114167295,
-0.0066686822,
-0.0220960621,
-0.0353352353,
-0.0239274111,
0.0288859978,
-0.0118317436,
-0.016281791,
0.0115027437,
-0.0192603506,
0.0000635853,
0.1004078239,
0.0129502406,
0.0497237369,
-0.0264711678,
0.0085877869,
0.0626984835,
0.0532656871,
0.0003658205,
0.015572303,
-0.008749783,
0.0481362417,
-0.0316957831,
0.014818646,
0.0467734896,
0.044522617,
-0.0188699216,
0.0269926786,
0.1013914347,
0.0385706015,
-0.0230887327,
0.0017865804,
-0.0215448514,
0.066750899,
0.03068473,
0.0292203613,
-0.0896750018,
-0.0295687243,
0.0071941717,
0.0312676765,
-0.0067604925,
-0.0138264084,
0.0036370526,
-0.1309449226,
-0.0019417147,
0.0294359457,
-0.0443366729,
-0.0348914154,
-0.0319124088,
0.0056378334,
-0.0400221087,
0.0355554149,
0.0421357267,
0.0206107758,
-0.0199235138,
-0.0080858376,
-0.0158703197,
0.0169354714,
0.0215746816,
-0.0663890094,
0.0963054374,
0.0835381523,
0.0165486895,
-0.0116246417,
-0.0661573708,
0.0190359987,
0.0142683517,
-0.0249165446,
0.0180844162,
0.0374916196,
-0.0184537694,
0.0160313062,
-0.030824719,
0.0182315055,
-0.0082976324,
-0.0021081765,
0.0406894423,
0.0850466788,
0.0189529452,
0.0058958246,
-0.0067849993,
0.0082499813,
0.0293989442,
0.0791300759,
-0.0040771067,
-0.0366306491,
-0.0454619937,
-0.0254755896,
-0.0110978186,
-0.0274264254,
-0.0407557748,
0.0033610822,
0.0160373468,
-0.0012383972,
-0.0544563271,
-0.039698083,
0.0097985147,
0.0210415684,
0.0232299622,
0.0016067611,
-0.0289574843,
-0.0069620972,
-0.0390728004,
0.0144610535,
-0.0202028584,
0.0270703342,
-0.0540631078,
0.0041576251,
0.0029578679,
-0.0165692922,
0.0049981526,
0.0314484015,
-0.034444809,
-0.0778518841,
0.0444382206,
0.0551849976,
0.0313920937,
0.0404303931,
-0.0050721308,
-0.0250040684,
0.0120313279,
0.033444237,
-0.0451496765,
0.0254110768,
0.0164784323,
-0.0186168831,
0.0048783114,
0.018866716,
-0.0271955952,
-0.0208478719,
0.0358449332,
0.0311466511,
0.0054826909,
-0.0378925838,
0.0653287992,
0.0246396139,
0.0356849581,
-0.0235222559,
-0.0085204942,
-0.0052448008,
0.0319109224,
-0.039345704,
-0.0306955781,
0.0006158624,
-0.0174129736,
0.0112497006,
-0.0091713341,
0.0163099654,
-0.0030705337,
-0.0281243511,
0.0242216922,
-0.056332279,
-0.0084606037,
0.0043284446,
0.0071910336,
0.0329823717,
-0.0057104798,
-0.0367061086,
-0.0062322798,
0.0193696897,
0.0764644518,
0.0099476678,
0.0692213029,
0.001981237,
-0.0261030588,
-0.0139421495,
0.036112994,
-0.0260618217,
-0.0413335152,
-0.0327473953,
0.0243727118,
0.0582762584,
0.0032345371,
0.0580530502,
0.0002236571,
-0.0067532673,
-0.011953379,
0.0225529112,
-0.0198052321,
-0.0220564976,
0.0138141783,
0.0309826769,
-0.0277023148,
-0.0053619877,
0.0380310826,
-0.0174577944,
0.0369852073,
-0.0516006798,
-0.0128697371,
-0.0259071998,
-0.0330372266,
-0.0050371857,
0.0321229249,
-0.0609780625,
-0.0357979499,
-0.0788968503,
-0.049356956,
-0.0357954167,
0.055692777,
0.0000623772,
-0.0042261425,
0.0095418608,
0.067874752,
-0.0153461425,
-0.0421893671,
-0.0368128605,
0.0011823083,
-0.0282604061,
-0.0115013188,
-0.0328610912,
-6.626884668e-33,
-0.0322826132,
0.0063778246,
0.0163756944,
-0.0220226869,
-0.0324025676,
-0.0354095474,
0.0022307234,
0.0426093414,
-0.0507471114,
-0.0229090024,
-0.0212099981,
-0.000394243,
0.0205622762,
-0.0085386531,
0.0351270847,
0.0092262952,
0.0188151784,
0.0074520637,
-0.0542048402,
-0.0005266449,
0.029842129,
-0.000088602,
0.0576960668,
-0.0881817415,
0.0231565516,
0.0200576521,
-0.025205303,
0.0096512102,
-0.0717011467,
0.0598079152,
0.0146667231,
0.009894073,
-0.0010615173,
-0.0232451223,
-0.0031181283,
0.0188855398,
-0.0158006232,
0.002749193,
0.0536484048,
0.0110970056,
-0.0023914545,
-0.0139499968,
0.0352601558,
-0.0301126465,
-0.0412455052,
-0.0180261899,
0.0258395094,
-0.0316616334,
-0.0083961766,
0.045235794,
-0.077110894,
-0.0002775858,
-0.0184371471,
-0.0053768791,
-0.0243742038,
0.0706107691,
0.010951262,
0.0333206318,
-0.0783145875,
0.0361546911,
0.017337719,
0.0366523899,
0.1000379995,
0.0356941782,
0.00296203,
0.0665539876,
0.023695251,
-0.0267594475,
0.0064025563,
-0.0200342536,
0.00835477,
0.0201816279,
0.0928248689,
0.0533871576,
0.0335026942,
-0.0279978197,
-0.0403026864,
0.0409917459,
0.0450342894,
0.0017918599,
0.0358552895,
0.0320245624,
-0.0822679102,
0.0035709357,
-0.0212709829,
0.0275468286,
-0.0381908268,
-0.0530614667,
0.0473781526,
-0.0063421275,
-0.0153672202,
-0.0155275427,
0.0427011251,
-0.0358684063,
0.0018168609,
0.0267907903,
-0.0146449441,
0.0193000585,
-0.0274345614,
0.0066551194,
-0.0472964831,
-0.0052232319,
-0.0139673343,
0.0098670451,
0.0315107815,
0.0134395678,
0.0224475376,
0.0077379453,
-0.0780327246,
0.0315866582,
-0.0288309045,
0.0143071171,
0.0059708478,
-0.0406834222,
0.0391666479,
-0.0247243848,
-0.003037591,
0.0213719103,
0.0152070792,
-0.0103391679,
-0.0602530502,
-0.0154266311,
-0.0082776509,
0.0043539256,
-0.0066270786,
-0.0045838021,
-0.024117751,
0.0203456432,
-0.0165452007,
-0.0020054178,
-0.0146932676,
-0.0032941576,
2.917e-7,
0.0008487554,
0.0149896936,
0.0178677849,
-0.0021742992,
-0.0299387537,
0.0600392036,
-0.0203446932,
0.037881624,
0.0492676683,
-0.0283150543,
0.0109177697,
-0.0029731465,
0.0526333675,
0.0291556846,
-0.0380261205,
-0.0091495607,
-0.05324094,
0.0086735487,
-0.051767882,
0.0255363565,
0.055336684,
0.0972237736,
0.0337791666,
-0.0068002855,
-0.0037535636,
0.0141916526,
0.0199838206,
-0.0235327277,
-0.0303724073,
0.0271988828,
-0.0075146952,
-0.0343196467,
-0.0274163783,
0.0234875549,
-0.0103500346,
0.0246448815,
-0.0379259773,
0.0562055781,
-0.0303250235,
-0.0452833138,
-0.0165761914,
-0.0109008094,
0.0054900697,
-0.0860297009,
0.0558135286,
-0.0223903917,
-0.0340083204,
-0.0418070331,
-0.0489504002,
-0.018167669,
0.014446876,
-0.0123903872,
-0.0183992479,
0.00675264,
0.0133636985,
0.0116322031,
-0.0286625717,
-0.0486116558,
-0.0494262315,
0.0293304864,
-0.0118810143,
-0.0662240759,
-0.0107084094,
0.0265954863,
-0.0008555203,
-0.0854427889,
0.0172862057,
2.883658227e-34,
0.004190383,
-0.0174177829,
0.0597725399,
0.0246443935,
0.0388540626,
-0.0010827376,
0.0417942554,
0.0298241396,
0.0264078844,
-0.0554397441,
-0.0069336277
] |
http://arxiv.org/abs/2402.02408v1 | http://arxiv.org/pdf/2402.02408v1 | GLaPE: Gold Label-agnostic Prompt Evaluation and Optimization for Large Language Model | 2024-02-04T08:57:54 | 2024-02-04T08:57:54 | Xuanchang Zhang, Zhuosheng Zhang, Hai Zhao | cs.CL | cs.CL, cs.LG | Despite the rapid progress of large language models (LLMs), their task
performance remains sensitive to prompt design. Recent studies have explored
leveraging the LLM itself as an optimizer to identify optimal prompts that
maximize task accuracy. However, when evaluating prompts, such approaches
heavily rely on elusive manually annotated gold labels to calculate task
accuracy for each candidate prompt, which hinders the widespread implementation
and generality. To overcome the limitation, this work proposes a gold
label-agnostic prompt evaluation (GLaPE) to alleviate dependence on gold
labels. Motivated by the observed correlation between self-consistency and the
accuracy of the answer, we adopt self-consistency as the initial evaluation
score. Subsequently, we refine the scores of prompts producing identical
answers to be mutually consistent. Experimental results show that GLaPE
provides reliable evaluations uniform with accuracy, even in the absence of
gold labels. Moreover, on six popular reasoning tasks, our GLaPE-based prompt
optimization yields effective prompts comparable to accuracy-based ones. The
code is publicly available at https://github.com/thunderous77/GLaPE. | GLaPE: Gold Label-agnostic Prompt Evaluation and Optimization for Large Language Model: Despite the rapid progress of large language models (LLMs), their task
performance remains sensitive to prompt design. Recent studies have explored
leveraging the LLM itself as an optimizer to identify optimal prompts that
maximize task accuracy. However, when evaluating prompts, such approaches
heavily rely on elusive manually annotated gold labels to calculate task
accuracy for each candidate prompt, which hinders the widespread implementation
and generality. To overcome the limitation, this work proposes a gold
label-agnostic prompt evaluation (GLaPE) to alleviate dependence on gold
labels. Motivated by the observed correlation between self-consistency and the
accuracy of the answer, we adopt self-consistency as the initial evaluation
score. Subsequently, we refine the scores of prompts producing identical
answers to be mutually consistent. Experimental results show that GLaPE
provides reliable evaluations uniform with accuracy, even in the absence of
gold labels. Moreover, on six popular reasoning tasks, our GLaPE-based prompt
optimization yields effective prompts comparable to accuracy-based ones. The
code is publicly available at https://github.com/thunderous77/GLaPE. | [
0.0376276672,
0.0215607248,
0.0024512641,
0.0302156024,
-0.0518814139,
-0.0064234347,
0.0529933162,
-0.014469767,
0.0362073556,
-0.0316472352,
-0.0384964496,
-0.0125966575,
-0.0093239145,
0.0108497348,
0.00786574,
-0.0378141031,
0.070809707,
-0.0831536651,
0.0380881913,
-0.0362009555,
-0.0276381895,
-0.0171749555,
-0.0404506139,
0.0400034264,
-0.0118734501,
-0.0534963384,
-0.0006282916,
-0.0153300278,
0.0064404127,
-0.0306624658,
-0.0033024603,
0.0967149064,
-0.0417212248,
0.0501220226,
0.000002386,
-0.0750030652,
0.0082300166,
-0.021460155,
-0.0689266697,
0.010793183,
0.0058743684,
0.0305458643,
-0.0046371571,
0.0389930196,
-0.0253177732,
-0.0092672873,
0.039989844,
0.0183685087,
0.0692440644,
0.029737331,
0.0149186496,
0.0066913171,
0.0068717469,
-0.0312939659,
0.0236221123,
-0.0734137967,
-0.0020905894,
0.008606757,
0.0336128809,
0.0399109721,
-0.0099039786,
0.0510165542,
0.0379216373,
-0.0392522067,
0.0120448815,
0.0233799703,
-0.0163841378,
-0.0294854175,
-0.0370759182,
0.0612406358,
0.0094843097,
-0.0121595636,
-0.0350866988,
-0.0242123343,
0.0287756752,
-0.0122598894,
-0.0383351259,
0.0379330553,
0.0147132576,
-0.0200509299,
0.000901564,
0.0354829878,
0.0135950632,
0.0010732473,
0.0043035615,
0.1074758321,
0.0360718556,
-0.0301614888,
0.0210875031,
0.0059715929,
-0.047985401,
-0.0826191008,
0.0047829277,
0.0505857207,
-0.0863662362,
0.0056959884,
-0.0425518081,
0.0030337009,
0.0380823836,
0.0084949266,
0.0579500832,
0.0495699793,
-0.041859258,
0.0636937693,
-0.0161670595,
0.1126416177,
0.0018892923,
0.0087441774,
-0.0442072414,
-0.0056057093,
-0.0236340985,
0.0093605863,
-0.0459351204,
0.0098500326,
0.0184060875,
-0.0178501569,
-0.0421449207,
0.0416195653,
-0.0244926643,
0.0384811535,
0.0262246616,
-0.0014189257,
0.0028797602,
-0.0195573699,
-0.0264774524,
0.0215485711,
-0.0606215559,
-0.0037942186,
0.033144284,
-0.057917349,
-0.0033384124,
0.0049861516,
-0.0121747907,
-0.0369169377,
0.0044868565,
0.060196612,
-0.0067032846,
-0.0110026896,
0.0042281332,
-0.0210589245,
-0.0405135676,
-0.0579034723,
-0.0352337919,
-0.0756282285,
0.0414686948,
0.015394249,
-0.0047243731,
-0.0288693011,
-0.0004594862,
0.0212718453,
0.0124013228,
0.0598708577,
-0.0042637512,
0.0304527227,
0.0170862265,
0.0123168323,
-0.0466475263,
0.0379107967,
0.0026518169,
0.0435101017,
0.0571452379,
-0.0394689143,
0.010590666,
0.0526464917,
-0.0004448279,
-0.0260696877,
0.0135530531,
0.0085407188,
0.0678881928,
0.0305241756,
-0.0063019688,
0.0438585617,
0.0084075676,
-0.0090085566,
0.0827460587,
0.0460196398,
-0.0344997495,
-0.0531589054,
0.0202104487,
-0.011802204,
0.0060656373,
0.0058037094,
-0.0809921026,
0.0247482844,
-0.0539901145,
0.0150112873,
0.0453766957,
0.0136534143,
-0.0229335278,
-0.0402014516,
-0.0207855441,
0.004257021,
-0.0423508137,
0.0466678888,
0.0117332777,
-0.0015251609,
-0.036450237,
0.054210525,
0.0071998541,
-0.0396800041,
-0.0150535777,
-0.0525994264,
0.0666185021,
0.0394274145,
0.0372686908,
-0.0087707574,
-0.0291984566,
-0.0119237667,
0.0423772968,
0.0497391894,
0.053082604,
-0.0160616655,
-0.0035809285,
-0.0364727788,
-0.0355990715,
0.0000412878,
-0.0043600621,
-0.0046305624,
0.0296390671,
-0.0083822813,
0.0153272375,
-0.0267515909,
-0.0017732369,
0.0303044971,
0.0078563262,
-0.0465979949,
0.008596614,
0.0120489402,
0.0100338114,
0.0214613806,
0.0121667748,
0.0369062908,
0.0248175692,
-0.0188323203,
-0.0015071988,
-0.0425880775,
-0.0691435114,
-0.0120419227,
0.0495388806,
0.0090280166,
-0.043753285,
0.0337469392,
-0.0280151423,
-0.0542458147,
-0.0342983864,
-0.0293665007,
-0.0093313195,
-0.0219063628,
0.0404893272,
-0.0149458861,
-0.0027198989,
-0.0649278909,
0.0567635074,
0.0142451739,
0.0299303513,
0.0187135451,
0.062139377,
0.0259691551,
-0.0752504915,
-0.0520429425,
0.010484552,
-0.0144970808,
-0.0021945222,
0.0445505902,
0.0156625342,
-0.0551579818,
-0.035010919,
-0.1025018245,
0.0104920659,
-0.0151958335,
0.0017794075,
0.0260952525,
0.0093214037,
-0.0128512308,
-0.0291493535,
-0.0278936978,
-0.023307316,
0.0297566149,
0.0288144723,
0.011111863,
-0.0034966494,
-0.013914817,
-0.0261387844,
-0.0231816042,
-0.0052698143,
0.0169073697,
0.0075915279,
0.0163508505,
-0.0528280623,
-0.0038329712,
-0.0326306969,
0.0187432524,
0.0132401809,
0.0115807829,
0.0284805764,
-0.0075155669,
0.0498947129,
0.0650391281,
0.0252430327,
-0.0297614038,
0.0548720062,
-0.0186825823,
-0.0715343654,
0.0080477344,
-0.0160689484,
-0.0483883731,
0.0172968265,
-0.0345100574,
0.0178981777,
0.0821476877,
-0.0013070328,
-0.0040947343,
-0.054575596,
0.0071674054,
-0.0385480598,
-0.0247948598,
-0.0664730892,
-0.0005888153,
-0.0257643983,
0.0551066212,
-0.0230037272,
-0.0593731776,
0.0788417682,
-0.0676555857,
-0.0020267197,
-0.0199468881,
0.0141421631,
-0.0166396871,
0.0832637995,
-0.0123985112,
0.0072032334,
0.0475401804,
-0.0143522089,
-0.0227114838,
0.0089881141,
0.0211994499,
-0.0386409983,
0.0452359617,
-0.0428896546,
0.0182611346,
-0.0177618656,
-0.0591606088,
-0.0506061688,
-0.043364197,
-0.0327589475,
0.0261770282,
-0.0375940315,
0.0039155916,
-0.0142795285,
0.0095669534,
0.0211956725,
0.0339040197,
-0.0517686047,
0.0093170246,
-0.0346224979,
0.0299963597,
0.0616884828,
0.0146903228,
0.0155668827,
-0.0074305176,
0.0350268371,
0.0256305095,
0.03825799,
0.0337079614,
0.0040782569,
0.005044078,
0.0605776608,
-0.0410667248,
-0.0221284796,
0.0587495007,
0.0441818126,
-0.0527148433,
-0.0153002525,
0.1405798048,
0.0451644659,
-0.0500025339,
-0.0099978363,
-0.0275089368,
0.0347359814,
0.0318348557,
0.0282607488,
-0.0562043972,
-0.0490700901,
0.02689372,
0.0190322734,
-0.000828305,
-0.0271415152,
-0.0019769031,
-0.1307936758,
0.0258468818,
-0.0169943646,
-0.024030691,
0.0038588759,
-0.0381203964,
0.0053195837,
0.002582005,
0.0396467894,
0.0433349498,
-0.006435093,
0.0017154496,
0.0293949731,
-0.017539259,
-0.0445633046,
0.0702172071,
-0.0238529257,
0.0727184862,
0.0761879086,
-0.0096134432,
-0.0051725982,
-0.0813784301,
0.0453324206,
0.005957928,
-0.0153021067,
-0.0085453819,
-0.0047019101,
0.0079029612,
-0.013782803,
-0.0084246779,
0.0175111815,
-0.0148428427,
-0.0316662267,
0.0487528183,
0.0915381312,
-0.0077524353,
0.0044540823,
-0.0033885001,
-0.0164959393,
0.0258701816,
0.0952026695,
0.0299150348,
-0.0549273901,
-0.0337555595,
-0.0299317464,
0.0041014221,
-0.038401749,
-0.0375099182,
0.0364901796,
0.0147412186,
-0.0038661822,
-0.0242413543,
-0.0370099805,
0.0367608853,
0.0361702293,
0.0057044956,
0.0036541165,
-0.0583617464,
0.0024344097,
-0.0239971541,
0.0512805581,
-0.0749531016,
0.0428035185,
-0.110066101,
0.0276094768,
0.0117405178,
-0.0002215937,
-0.0181908011,
0.0600732565,
-0.0455223024,
-0.0565347001,
0.0084968433,
0.0595608689,
0.0249890126,
-0.002201952,
-0.034239132,
-0.0671095178,
0.0115486467,
-0.0097645205,
-0.0299962927,
0.0133315492,
-0.004987929,
-0.0074624489,
-0.0107632969,
0.0233154632,
-0.0402057469,
0.0271948185,
0.0258799084,
0.0625174344,
0.0260768197,
-0.0674015507,
0.0328883976,
0.0587130301,
0.0499439202,
-0.0286696181,
0.0242110379,
-0.0158268306,
0.0258912947,
-0.0022749384,
-0.0397849865,
-0.0217818543,
0.0217390805,
0.0057143588,
-0.0392998606,
0.0332650207,
-0.0003208037,
-0.0136852395,
0.0285550356,
-0.0514569469,
-0.0040560751,
-0.0047209938,
0.0337269008,
0.0277173184,
-0.0187885445,
-0.033757098,
0.0152820228,
0.0211138465,
0.0355437174,
0.0003514748,
0.0676711202,
0.0305207595,
-0.0103175556,
-0.0148143955,
0.0370444246,
-0.0564384237,
-0.0817814246,
-0.0004069387,
0.0227900054,
0.0604037382,
0.0094494699,
0.0108390367,
-0.0069001117,
-0.0206827633,
-0.0420029238,
0.0080153896,
0.0645343736,
-0.0467961803,
0.0455372669,
0.0318450108,
0.005195166,
-0.012120761,
0.041868899,
-0.0409827866,
0.0141177662,
-0.0068597053,
0.008702104,
-0.0364348963,
-0.0249369107,
-0.0221618563,
0.0059759994,
-0.0453274027,
-0.0229307171,
-0.0652248189,
-0.0535530411,
0.0059951409,
0.0408288389,
-0.0070938827,
0.0172163006,
0.04329836,
0.0627821237,
-0.0041659395,
-0.0308365524,
-0.0288739875,
0.0116892215,
-0.0137040345,
-0.0169486664,
-0.0224803928,
-6.574112318e-33,
-0.0443718284,
-0.0337641798,
0.0224682149,
-0.0375133939,
0.0014063079,
0.0198076144,
-0.0500301607,
0.030759789,
-0.016706394,
-0.010257368,
-0.0256167874,
-0.034355633,
0.0245035663,
0.0080312956,
0.0230895393,
-0.0145503683,
0.0336550921,
0.0230387282,
-0.0625290573,
0.004488186,
0.003225768,
-0.0140653672,
0.0143340426,
-0.0365789644,
0.0606102459,
-0.0134290215,
-0.0143422084,
0.0291777458,
-0.0339657851,
0.0683148131,
-0.0222073533,
0.0521537811,
0.0133041013,
-0.0706111789,
-0.0204193406,
0.0522561185,
-0.0533710979,
-0.055618953,
0.0202229209,
-0.0070196651,
-0.0678590164,
-0.0172481518,
-0.0187713392,
-0.0137370499,
-0.0506311841,
0.0022745198,
0.0150919501,
-0.0064091426,
-0.0258644577,
0.0243386682,
-0.0422028825,
0.003234532,
-0.0030581981,
-0.0282132663,
0.0084189745,
0.0523586497,
0.003145131,
0.0159730073,
-0.0262920242,
0.0223688502,
0.0175830107,
0.0370508954,
0.0710312501,
0.0469610803,
-0.0147092054,
0.0355057865,
0.0319854319,
0.0067441561,
0.0030421002,
-0.0187128317,
0.0109814387,
0.0137890205,
0.0743743479,
0.019270394,
-0.0091059953,
-0.0018210099,
-0.033589866,
0.0303441361,
0.0493699871,
-0.0051859068,
0.0244545657,
-0.0016959171,
-0.0239513032,
-0.0233013052,
-0.0825075731,
0.014273149,
-0.0506400317,
-0.0134991389,
0.0568367094,
-0.0228066854,
-0.030154448,
-0.0089149838,
0.0442420542,
-0.0102434233,
-0.0305266958,
0.0100923777,
0.0106110433,
0.0381470434,
-0.0051024179,
0.0344175249,
-0.0309971068,
0.0005724467,
0.0011032091,
-0.0181984305,
0.0279729068,
0.0108086616,
0.0000543226,
0.0098762745,
-0.0640374348,
0.0108129848,
-0.0449799448,
0.0254580751,
0.0033526896,
-0.0326776989,
0.0383956507,
-0.0169409122,
-0.0076280576,
0.0736587271,
-0.0029705444,
-0.0261326898,
-0.0200584456,
0.0107505964,
0.0034897926,
-0.0032407022,
-0.0041714893,
0.015663771,
-0.0235432778,
0.0365834497,
-0.0122672655,
0.0069482541,
-0.0007694747,
0.004979732,
2.876e-7,
0.0341537632,
0.0134331184,
0.0200523771,
0.033324033,
0.0149047095,
0.0423654355,
-0.0127461851,
0.0021275142,
0.0218941066,
-0.03075126,
-0.030467581,
0.0243987329,
0.0347245112,
0.0307173785,
-0.0880110338,
0.0159106012,
-0.0648056865,
0.0192940719,
-0.0642392486,
0.0353543051,
0.0435291193,
0.0538799576,
0.0188663993,
-0.0200959798,
-0.0067302892,
-0.0113921287,
0.0196079183,
-0.0321165659,
0.004758589,
0.0352358595,
0.0471796654,
-0.0380202308,
0.007471167,
0.0385642946,
0.0097973002,
-0.0128460135,
0.0122996019,
0.067404978,
0.0121420259,
-0.0119167604,
-0.0399031118,
-0.0212973785,
0.0183419846,
-0.0950194895,
0.0888074711,
-0.0027972481,
-0.0209218208,
-0.0576633252,
-0.001498223,
-0.0270501897,
0.0368934944,
-0.0171895847,
-0.0177860707,
0.0147995288,
0.0201200359,
0.0138450935,
-0.0384532213,
-0.0597278811,
-0.0513196737,
-0.0058017531,
-0.0394973643,
-0.036981605,
-0.0517184064,
0.0032677972,
-0.0029127372,
-0.107314676,
-0.004277708,
2.68660267e-34,
0.0113304891,
-0.0099065593,
0.0470837541,
0.0171113294,
0.0193045624,
0.0094373273,
-0.0046028919,
0.0182967503,
0.0261471253,
-0.0718908459,
-0.0370796509
] |
http://arxiv.org/abs/2402.02392v2 | http://arxiv.org/pdf/2402.02392v2 | DeLLMa: A Framework for Decision Making Under Uncertainty with Large Language Models | 2024-02-04T08:11:45 | 2024-06-09T05:04:13 | Ollie Liu, Deqing Fu, Dani Yogatama, Willie Neiswanger | cs.AI | cs.AI, cs.CL, cs.LG | The potential of large language models (LLMs) as decision support tools is
increasingly being explored in fields such as business, engineering, and
medicine, which often face challenging tasks of decision-making under
uncertainty. In this paper, we show that directly prompting LLMs on these types
of decision-making problems can yield poor results, especially as the problem
complexity increases. To aid in these tasks, we propose DeLLMa (Decision-making
Large Language Model assistant), a framework designed to enhance
decision-making accuracy in uncertain environments. DeLLMa involves a
multi-step scaffolding procedure, drawing upon principles from decision theory
and utility theory, to provide a rational and human-auditable decision-making
process. We validate our framework on multiple realistic decision-making
environments, demonstrating that DeLLMa can consistently enhance the
decision-making performance of leading language models, and achieve up to a 40%
increase in accuracy over competing methods. | DeLLMa: A Framework for Decision Making Under Uncertainty with Large Language Models: The potential of large language models (LLMs) as decision support tools is
increasingly being explored in fields such as business, engineering, and
medicine, which often face challenging tasks of decision-making under
uncertainty. In this paper, we show that directly prompting LLMs on these types
of decision-making problems can yield poor results, especially as the problem
complexity increases. To aid in these tasks, we propose DeLLMa (Decision-making
Large Language Model assistant), a framework designed to enhance
decision-making accuracy in uncertain environments. DeLLMa involves a
multi-step scaffolding procedure, drawing upon principles from decision theory
and utility theory, to provide a rational and human-auditable decision-making
process. We validate our framework on multiple realistic decision-making
environments, demonstrating that DeLLMa can consistently enhance the
decision-making performance of leading language models, and achieve up to a 40%
increase in accuracy over competing methods. | [
0.0084541691,
0.1022896171,
-0.0143566588,
-0.008660065,
-0.0332786329,
-0.0068437201,
0.0066043367,
0.0174013712,
-0.0189806428,
-0.0223228987,
-0.0012406755,
-0.0435437411,
0.0076865507,
0.0609114766,
-0.0082596857,
-0.0199028868,
0.0127520328,
-0.0712320134,
0.0667022243,
-0.017616488,
-0.0094323624,
-0.0267798342,
-0.00216165,
0.0317698494,
-0.0013479125,
0.0007016912,
0.007420064,
-0.0254855938,
0.0233253464,
-0.0517306775,
-0.0033052519,
0.0591171645,
0.0314041227,
0.0225683339,
0.0000022894,
-0.0649357513,
-0.0284547433,
-0.0143250162,
-0.0543135367,
0.0442046598,
0.052229479,
0.00827769,
-0.0140926652,
0.0146161942,
-0.0310470443,
0.0118501391,
0.0304782018,
0.0135631161,
0.0204044245,
0.0188974235,
-0.0110619562,
-0.0018439798,
-0.012490646,
-0.0601843894,
-0.0575427227,
-0.0778850168,
0.0309545156,
0.0150169116,
0.0431834348,
0.0001033875,
-0.0014067192,
0.0378074199,
-0.0005194936,
-0.0323427357,
-0.0352961086,
0.0295186006,
0.016236065,
-0.0162023343,
-0.0198257491,
0.0411247537,
0.0110255275,
0.0212956239,
-0.0127180181,
-0.0283755679,
0.0083442461,
0.0108787054,
-0.0238667484,
0.0122235902,
-0.0035550778,
0.0330430083,
-0.0179392453,
0.0267855767,
0.0222047642,
0.0372237861,
-0.0370574072,
0.108130537,
-0.0345458984,
-0.0007531898,
0.0387541056,
0.0397672243,
-0.0120528881,
-0.0631343126,
0.0129874703,
0.0483577698,
-0.0936187431,
0.0187974963,
0.0190117452,
0.0272661392,
0.0319974348,
-0.0614277646,
0.0161025226,
0.0222966634,
0.026864389,
0.0353396349,
-0.0072020791,
0.0315971524,
-0.0219758265,
-0.0307519604,
-0.0325810835,
0.0413076282,
0.0149104865,
0.0057758298,
-0.0201533642,
-0.0061489008,
-0.0628850311,
0.0007296737,
-0.0923772752,
-0.0067689298,
-0.0722925141,
0.0569294542,
0.0045939977,
0.030790206,
0.0158101711,
-0.0448061377,
-0.0784225389,
-0.0100276573,
-0.0111337341,
0.0670728683,
-0.0073334235,
-0.0012454867,
0.0451094545,
0.0182990786,
-0.0349250324,
0.0022579189,
0.0038908075,
0.0919438303,
-0.0187296998,
0.0051831109,
-0.0360555053,
-0.0181527231,
-0.0159367379,
-0.0796977058,
-0.065741159,
-0.0747838467,
-0.0085602701,
0.0102632651,
-0.0004824363,
-0.0559875704,
0.0216713585,
-0.013680188,
-0.008140116,
0.0725636035,
-0.0157004595,
0.0157655906,
0.0276020747,
0.0079474309,
0.0060015912,
0.0602725744,
0.0256146025,
0.0412830859,
0.0140129831,
-0.0017270034,
0.0173415001,
0.0178332515,
-0.0302075613,
0.0207115114,
-0.0815082267,
0.0151764723,
-0.0310560521,
0.0131474026,
-0.0298473947,
0.0348345712,
-0.0118804462,
0.0175384097,
0.0468802676,
0.0108674066,
0.0390270278,
-0.015596726,
-0.010515443,
0.0126283569,
-0.00025841,
-0.0547562577,
-0.0134608103,
0.0200120397,
-0.0143742403,
-0.0032957715,
-0.0054930407,
0.0208430905,
0.0010842701,
-0.0528585464,
-0.0251121335,
-0.0315561369,
0.0337621532,
0.0120677175,
0.0479200408,
-0.073031351,
-0.0170533936,
0.0050692922,
0.0386512913,
0.0174541418,
-0.0567537993,
-0.0794229582,
0.0495390557,
0.0921667442,
0.0251660459,
-0.0532575175,
-0.0107144378,
-0.0221001487,
-0.0169809423,
0.0148077952,
0.0263017751,
-0.0112617575,
-0.0319253951,
-0.0487374663,
-0.009965024,
0.0439687967,
0.0415078662,
0.0398178324,
0.0362644903,
0.0551321916,
-0.0119529953,
0.0120577831,
-0.0380248018,
-0.0069076316,
0.0022980119,
-0.0278596282,
0.0181585606,
-0.0074956277,
0.040438585,
-0.0103771798,
-0.0039448692,
0.0445969366,
-0.007614099,
0.0035707923,
-0.0526539423,
-0.0094054518,
0.0173341911,
0.0609646887,
0.0354006812,
0.0298956539,
-0.005219081,
-0.0461828597,
-0.021343058,
-0.016721271,
0.0153967058,
-0.0289098639,
0.0012600718,
-0.0054377615,
0.0379847288,
-0.0481929444,
0.0060424553,
-0.0595185719,
0.0610151663,
-0.003887332,
0.0031944967,
0.021993408,
0.0139448689,
-0.0075021228,
-0.0369774923,
0.0124568725,
0.0292171258,
0.0312059112,
0.0525590628,
0.0201550126,
0.0365641266,
0.004992988,
-0.026737906,
-0.0470462479,
-0.0306530409,
0.0258734897,
0.0182352383,
-0.0430618227,
-0.0433976538,
-0.0176176112,
-0.0075775781,
-0.0198425036,
0.0225971509,
0.0064382022,
0.0340770222,
0.0239745863,
0.0182728432,
-0.0217554271,
-0.0550005548,
-0.0309741665,
-0.0000171529,
0.0318251811,
0.0364814475,
-0.0216305163,
-0.0818724856,
0.0433176234,
-0.0269828886,
0.0068965959,
-0.0016668313,
-0.0080766985,
0.0047678337,
-0.0068372292,
0.042312324,
0.0119189089,
-0.0141245313,
0.0384374522,
0.0950309411,
-0.007679801,
-0.0647084489,
-0.0043555121,
-0.0170462374,
0.0024286399,
0.0040630898,
-0.0340171121,
-0.0088225501,
0.0550080352,
-0.0234959405,
0.0037703118,
-0.0298766065,
0.0413176455,
-0.0533694252,
0.0001010224,
-0.0482041724,
-0.023532426,
0.0049787117,
0.1428657919,
0.0457100458,
-0.0455569103,
0.0935471877,
-0.0021514464,
0.0278097577,
-0.031109523,
-0.0476334058,
-0.0532628559,
0.0017867594,
-0.0319060832,
0.006409334,
-0.0376903825,
0.0047454899,
-0.0092886966,
-0.0410387516,
0.0356542058,
-0.0457808003,
0.0786209553,
-0.0801931545,
0.0235370807,
-0.007308797,
-0.030457288,
-0.0141370641,
-0.0044098468,
-0.0836513042,
-0.0213578194,
0.0666265935,
-0.0448315926,
-0.0554314777,
0.0231195092,
0.0336801112,
0.0410729907,
-0.0286985375,
-0.0073277238,
0.0060111089,
-0.0268973261,
0.0110196639,
0.013491719,
0.0997704118,
-0.0187045448,
0.0324407555,
0.0245434139,
0.0039030495,
0.0049025943,
-0.0176318623,
0.0026042305,
0.0219254419,
-0.0049403687,
-0.0010259028,
0.0555569939,
0.0012838759,
-0.0481086634,
0.0049987077,
0.0802902207,
0.0172515381,
-0.0475616083,
0.0129391253,
-0.0899487808,
0.0177980289,
0.0037749754,
0.0460679419,
-0.0720409378,
-0.0192953199,
-0.0087820031,
0.0162666962,
-0.0218613911,
-0.0207749251,
0.0070466455,
-0.0877871215,
0.0799842626,
0.0064767837,
-0.0078228442,
-0.0124958577,
-0.0432121083,
0.0085827392,
-0.0388867483,
0.0264747329,
-0.0036830858,
-0.0037275532,
-0.0218351688,
0.0030571376,
-0.0458232984,
-0.0517279133,
0.0206298474,
-0.019822048,
0.0152105205,
0.0293056797,
0.0099396771,
-0.0148579245,
-0.0093613556,
0.0253036004,
0.0264212694,
-0.0154911056,
-0.0029285473,
0.0117333671,
0.0593809709,
-0.0163085982,
-0.0315443613,
0.0005208307,
-0.0075699254,
-0.0241837408,
0.0152781103,
0.0799065828,
-0.0229319092,
-0.0238215756,
0.0069396528,
0.0141659463,
0.0236764718,
0.0306938943,
-0.0220222883,
-0.0603350624,
-0.0459488146,
-0.0742584318,
-0.0158163533,
-0.009883373,
-0.0306786429,
0.0378003605,
-0.0099655623,
-0.0405692086,
-0.0250404608,
-0.0088367527,
0.0393792577,
0.0174246151,
0.0198795758,
-0.0005874954,
0.0586656481,
0.0106740408,
-0.0153236808,
0.029842766,
-0.0161868427,
0.0728406534,
-0.0667986721,
0.0434090383,
0.0372942612,
0.0018270583,
-0.0302316938,
0.0264133364,
0.000399222,
-0.0666311458,
-0.0011565782,
0.0524548851,
0.0442102738,
-0.0511586815,
-0.0200447459,
-0.0438918807,
0.0186599772,
0.031868346,
-0.018674802,
-0.0197801776,
0.0383199826,
0.0418173857,
-0.0144552719,
0.031196231,
-0.0035729154,
0.0021822348,
-0.0116810687,
0.0187435485,
-0.0596425012,
-0.064143002,
0.0036688212,
0.0209830813,
-0.0143906772,
-0.018362198,
-0.0025021168,
-0.0299881995,
0.0132986056,
-0.0068641957,
-0.0182928871,
-0.0067538028,
-0.0231816992,
0.0226906482,
-0.0434933007,
0.0238584504,
-0.0151172206,
-0.0090810331,
-0.0444203503,
0.0098234732,
0.0597503558,
0.0103624202,
0.0243533868,
0.1008018926,
-0.0184998177,
0.0100129498,
0.0005840965,
-0.0163942929,
0.0342716575,
0.0277097858,
0.0346737802,
0.0041156583,
-0.0312033594,
-0.0327254646,
0.0059216148,
0.0085503347,
-0.1110631078,
0.0045221928,
0.0225920379,
0.0176322293,
0.0141126653,
0.0311192535,
-0.0652510971,
-0.0064627593,
-0.0214420389,
-0.0121265911,
0.0050467285,
0.0128143849,
0.0318818539,
0.0237788502,
-0.0331275538,
-0.0027905423,
0.0729066581,
0.0104085458,
0.0187526774,
-0.0707292259,
-0.06465552,
-0.0427536108,
-0.0666555315,
-0.0147776529,
0.0071122479,
-0.0633546039,
-0.0162218865,
-0.1099135354,
-0.0340567455,
-0.0145560512,
0.0261267181,
-0.0376510993,
-0.0489717908,
-0.006558754,
0.0350692421,
0.0262884572,
0.0447212383,
-0.038081076,
0.0132512888,
-0.019469982,
0.017713163,
-0.0312803239,
-6.430155402e-33,
0.0234291945,
-0.0489548221,
0.0218141638,
-0.0550786257,
0.0283065625,
-0.018068295,
-0.0185372829,
0.0688620359,
0.0120740775,
0.0358915925,
0.0003714037,
-0.0059884293,
0.0225981567,
-0.0167302545,
0.0582814217,
-0.0100497929,
0.0056238305,
0.034951888,
-0.0226984266,
-0.0266943574,
0.0178492945,
-0.0132190688,
0.0245493334,
-0.0616956651,
0.0309148058,
0.0469350256,
-0.0372928195,
0.0245162379,
-0.0412304029,
0.054612577,
-0.014215908,
0.0340597294,
0.0007603471,
-0.0637680069,
-0.0092375819,
0.0317987613,
0.0125377402,
-0.0090493206,
0.0039773583,
0.0091265962,
0.0104769506,
-0.0221720617,
-0.0147794187,
0.0340824611,
-0.0579428375,
0.0648079962,
-0.0154447127,
-0.0598784499,
-0.0127618629,
0.0064830845,
-0.0456428379,
0.0348703228,
0.001075191,
0.00525459,
0.0185434017,
0.0763841346,
-0.0492084511,
0.0076270103,
-0.0179967005,
0.0424418375,
0.0513570607,
-0.0007162151,
0.0590888932,
-0.0210688114,
-0.0388433971,
0.0380725786,
-0.0250336248,
-0.034079276,
-0.041789405,
-0.0333384052,
-0.0331890248,
0.0789306164,
0.044846613,
0.042232804,
-0.0300464835,
-0.0135176266,
-0.0804019868,
0.0491377078,
0.0108651342,
0.0546134003,
0.0463281274,
0.0246772449,
-0.0394255333,
-0.0603914,
-0.0179303121,
-0.011333935,
-0.0393989384,
-0.0027192335,
0.0040771714,
0.0132574197,
-0.0298011657,
0.0346913449,
0.0166710652,
-0.017541647,
-0.0690907836,
-0.0057906015,
0.0018946939,
0.0369751826,
-0.0029603301,
0.0274049845,
-0.0390655026,
-0.0157557782,
-0.0331405178,
0.0052656345,
0.0062400792,
0.0116544059,
-0.0022042026,
-0.0252370667,
-0.0513980798,
0.0372924469,
-0.0142944381,
0.0256518442,
0.0150035471,
-0.0000789516,
0.0273307823,
-0.0178423915,
0.0037109011,
0.0534757487,
-0.0050008064,
-0.0203300733,
0.0149281071,
-0.0079919677,
-0.0174735319,
0.0347665772,
-0.0156673267,
-0.014604114,
-0.0464977138,
0.0364876352,
0.0469457395,
-0.0163623299,
-0.0006628147,
0.0009101448,
2.922e-7,
0.0373010822,
0.0846724659,
-0.0156256482,
0.0140192304,
0.0157623291,
0.0700082406,
-0.0363269113,
-0.042529691,
-0.0156194633,
0.0035684982,
0.032727465,
0.0055059004,
0.0373471938,
0.0628530607,
-0.0068893936,
-0.0458329208,
-0.0587212741,
0.0170337614,
-0.0451802909,
-0.0315820165,
0.0369054824,
0.0661741644,
0.0136049781,
-0.0233200323,
-0.0043452936,
-0.0145469811,
0.0357207656,
-0.0841594189,
-0.0091235274,
0.0311300009,
0.0371153615,
0.0069623226,
0.0030748562,
0.04038091,
0.0088852644,
0.0046537393,
-0.0191227812,
0.0936776325,
0.0217864234,
0.0286621712,
0.0094582411,
-0.0161709823,
0.0227348171,
-0.0379584096,
0.0369027965,
-0.0627141967,
-0.0643566102,
-0.0347825736,
0.0319634192,
-0.026636079,
0.0635420829,
-0.0290638972,
-0.0109118139,
0.0035406037,
-0.0207837615,
0.0224411879,
0.0102677131,
-0.026389759,
-0.0264581013,
0.0769108385,
-0.0423440337,
-0.0209756382,
-0.0235050414,
0.0428474955,
0.0420772023,
-0.074682869,
0.0454028398,
2.567628407e-34,
0.0599215701,
0.0140434233,
-0.0200566929,
0.0330062881,
0.0497864597,
0.039339859,
-0.0198517051,
0.026401639,
0.0333613269,
-0.0387144424,
-0.0099268649
] |
http://arxiv.org/abs/2402.02389v2 | http://arxiv.org/pdf/2402.02389v2 | KICGPT: Large Language Model with Knowledge in Context for Knowledge Graph Completion | 2024-02-04T08:01:07 | 2024-02-23T09:01:44 | Yanbin Wei, Qiushi Huang, James T. Kwok, Yu Zhang | cs.CL | cs.CL, cs.AI | Knowledge Graph Completion (KGC) is crucial for addressing knowledge graph
incompleteness and supporting downstream applications. Many models have been
proposed for KGC. They can be categorized into two main classes: triple-based
and text-based approaches. Triple-based methods struggle with long-tail
entities due to limited structural information and imbalanced entity
distributions. Text-based methods alleviate this issue but require costly
training for language models and specific finetuning for knowledge graphs,
which limits their efficiency. To alleviate these limitations, in this paper,
we propose KICGPT, a framework that integrates a large language model (LLM) and
a triple-based KGC retriever. It alleviates the long-tail problem without
incurring additional training overhead. KICGPT uses an in-context learning
strategy called Knowledge Prompt, which encodes structural knowledge into
demonstrations to guide the LLM. Empirical results on benchmark datasets
demonstrate the effectiveness of KICGPT with smaller training overhead and no
finetuning. | KICGPT: Large Language Model with Knowledge in Context for Knowledge Graph Completion: Knowledge Graph Completion (KGC) is crucial for addressing knowledge graph
incompleteness and supporting downstream applications. Many models have been
proposed for KGC. They can be categorized into two main classes: triple-based
and text-based approaches. Triple-based methods struggle with long-tail
entities due to limited structural information and imbalanced entity
distributions. Text-based methods alleviate this issue but require costly
training for language models and specific finetuning for knowledge graphs,
which limits their efficiency. To alleviate these limitations, in this paper,
we propose KICGPT, a framework that integrates a large language model (LLM) and
a triple-based KGC retriever. It alleviates the long-tail problem without
incurring additional training overhead. KICGPT uses an in-context learning
strategy called Knowledge Prompt, which encodes structural knowledge into
demonstrations to guide the LLM. Empirical results on benchmark datasets
demonstrate the effectiveness of KICGPT with smaller training overhead and no
finetuning. | [
0.0626362562,
0.0268165525,
0.0033002917,
0.0594129115,
-0.0053165308,
0.0017045711,
-0.0691918954,
0.0216013491,
0.0913595706,
-0.0309422426,
0.0204442665,
0.0138022164,
-0.0056081507,
0.0133014591,
0.0252397861,
-0.0146453278,
0.0154194441,
0.0007095071,
0.012330153,
-0.029599905,
-0.0198275149,
0.0479581356,
0.0155593557,
-0.0057831225,
-0.0079033282,
0.0059599346,
0.0111488756,
0.0070906216,
-0.021600768,
-0.0213255137,
0.0125322379,
-0.007225865,
0.0004506512,
-0.0259141885,
0.0000020708,
-0.0461370163,
-0.013505389,
0.0256842077,
-0.0345052369,
-0.0165341645,
0.0153408265,
-0.0185494311,
0.0321684591,
0.0019074873,
-0.0498049222,
0.0201266613,
0.0036292276,
0.0489454642,
-0.0023651053,
0.0458698869,
0.0049702423,
-0.0295766611,
0.03087819,
-0.0459738076,
0.0074725454,
-0.0293987636,
0.0612334572,
0.0198541544,
0.0454903096,
0.0199197866,
-0.0129897483,
0.0389968269,
0.0103828581,
0.0233003516,
0.0103007704,
0.0330887735,
-0.0283780117,
0.0256223232,
0.0264465753,
0.0765260682,
0.0086816335,
0.0244967267,
0.0136330528,
-0.0077663925,
0.0130926073,
0.0190004483,
-0.0303275883,
-0.0238153562,
-0.023335766,
-0.001874413,
-0.0291714836,
-0.0414109044,
0.0422669165,
0.0005714752,
-0.0306336656,
0.0728180408,
-0.0042781141,
-0.021169208,
0.0159346685,
0.0295282304,
-0.025081763,
-0.0315643735,
-0.0121178059,
0.0250117853,
-0.0282401331,
0.0559003279,
0.0039796489,
0.020915417,
0.0299559925,
-0.0213210788,
-0.0512172468,
0.0131572187,
-0.0052336087,
0.0677893311,
-0.0613288134,
-0.0020471956,
-0.0020026653,
0.0028753839,
-0.0353269614,
-0.0519406497,
-0.0482350774,
-0.0008976258,
-0.0315063037,
0.0566851236,
-0.0206112675,
-0.0265041608,
-0.025227068,
0.0427327193,
-0.0262780096,
0.0653654262,
0.0409533083,
-0.0010382842,
0.0019130063,
-0.0324962586,
-0.0280768033,
0.0401114523,
-0.011724473,
0.0186291989,
-0.027308356,
0.0075822095,
-0.0067531108,
-0.0274761021,
0.0178098213,
-0.0548375472,
0.0304897204,
0.0292447545,
0.0189181007,
-0.0575700514,
-0.0615171976,
-0.045765508,
-0.0091642151,
-0.0755217746,
-0.0365050063,
-0.0863402411,
0.0685768872,
0.0401882119,
-0.0024161323,
-0.0445134155,
0.0358094387,
-0.0152246179,
-0.0708262399,
0.035929691,
-0.0363620557,
0.0031760815,
-0.0049499045,
0.0069000782,
-0.023781918,
0.0553169884,
-0.0083821565,
0.0004647512,
0.0444193669,
0.013006933,
-0.0528567694,
0.0267305728,
0.0643613636,
0.0181208048,
0.0132466415,
0.0351012833,
-0.0004888236,
0.015286413,
0.0109164016,
0.0378607735,
-0.0173434131,
0.019432243,
0.0276549868,
0.0025592337,
0.0062426045,
-0.0140958717,
0.0380949266,
0.0011190496,
-0.0126950573,
0.0061442889,
-0.0382283442,
-0.0130249877,
-0.0651839599,
0.0519792102,
0.031434603,
0.0284995101,
-0.0103076939,
0.0025289522,
0.0310041886,
-0.0236256197,
0.0416700542,
0.0531003028,
0.0999715924,
0.0230610203,
-0.0260879565,
0.0226425938,
0.0622086227,
-0.0081622424,
0.0367734879,
0.0219035726,
0.0105756847,
0.0604940839,
0.049075488,
-0.0286311172,
-0.0055252528,
0.0122813191,
0.0184200685,
-0.0113362642,
0.0013014537,
-0.039955169,
-0.0123624234,
-0.0236031357,
-0.0599742197,
0.0244298521,
-0.0060363589,
-0.004561034,
0.0556233376,
0.0582941882,
0.0338708796,
-0.0999208465,
0.0072359438,
0.0145361777,
-0.0558717474,
0.0212784726,
-0.0116481949,
0.0144988913,
0.012017848,
0.0429923944,
0.0111437794,
-0.0125844358,
0.043056,
-0.0167856365,
-0.0446751565,
-0.055281803,
0.0102013927,
-0.0713917464,
0.0590707213,
0.0171517376,
-0.0082320785,
-0.0195199084,
-0.0338094607,
-0.0457827523,
-0.0204936136,
0.0137036415,
0.0953219011,
-0.0212975666,
0.0063651442,
0.0032632882,
-0.0234759282,
-0.0113346372,
0.0897680372,
0.0009320946,
0.0347475931,
-0.0039648493,
-0.0149901323,
-0.0194308907,
-0.0154535826,
-0.0509940349,
0.0168562103,
0.0224457365,
0.0128843663,
0.032834366,
-0.018453043,
-0.0620055124,
-0.012494728,
-0.1210621744,
0.0374088101,
0.0077697937,
-0.0163247511,
-0.0452752821,
0.0316340141,
0.0112117287,
-0.0088997884,
-0.000547752,
0.0103729498,
-0.0288047791,
0.0274753887,
-0.0101931114,
0.0421056524,
-0.0367722996,
0.0160989277,
-0.0253787264,
0.0025263301,
0.0587991178,
-0.0092070838,
-0.022184927,
-0.0191719364,
0.041942697,
-0.0457863137,
0.0299494173,
-0.0358158723,
0.03312001,
-0.0254299808,
-0.025480086,
0.0225656684,
0.0143178506,
-0.0106705911,
0.0041228123,
0.0033977218,
0.006262836,
-0.0697172806,
0.0041709952,
-0.0539081357,
-0.0212159175,
-0.0555040874,
0.0101060551,
-0.0218421463,
-0.0006531674,
0.0116275875,
-0.035524223,
-0.06895715,
-0.0265448056,
-0.0003131284,
0.0202665962,
-0.047572162,
0.0167429093,
0.0219742823,
-0.0144194635,
-0.0229132716,
0.0161776487,
0.0787868351,
-0.0286378488,
0.0132203,
0.0350621529,
0.0027646977,
-0.0351068415,
0.0671749711,
0.0079216678,
0.0108226556,
-0.0030107796,
-0.0194142964,
-0.0371134169,
0.0025937133,
0.0182753261,
-0.0279659647,
0.0345311351,
0.0065965853,
0.0421822555,
0.0278125573,
-0.0152605744,
-0.0141855087,
0.0048538884,
-0.0320592485,
0.0332726873,
0.009259833,
0.0017491431,
0.0106793959,
-0.0095670512,
0.013689721,
0.0194091871,
0.0144281993,
-0.0194279179,
-0.0257161018,
0.0129546234,
0.0558868237,
-0.0352102146,
0.0058675222,
-0.027635213,
0.0025452599,
0.0533833168,
0.0228396486,
-0.0287947562,
0.0252635013,
-0.0131949773,
0.0078560375,
-0.0550659075,
-0.0290396642,
0.0178634636,
0.0033704978,
-0.0416246541,
0.0383258946,
0.1001052782,
-0.0132916039,
-0.0372946151,
-0.00353884,
0.0270798523,
0.0123983975,
0.0370155424,
0.0215948187,
-0.1067145169,
0.0002816962,
-0.0016559118,
0.0383557007,
0.042487476,
-0.0713993087,
-0.0033128997,
-0.1255491078,
0.0227377638,
0.023021793,
-0.0224069022,
-0.0661257058,
-0.0495814197,
0.0129987095,
0.0052438974,
-0.0276630204,
0.0200112667,
-0.0251224488,
-0.0036223053,
-0.0072726146,
0.0070560393,
-0.0658909157,
0.0428542048,
-0.0415673368,
0.0515194871,
-0.0204220954,
0.020219665,
0.0001076774,
0.018948555,
-0.0112384381,
-0.0074668843,
-0.0311782435,
-0.0039139357,
0.0290806852,
-0.0377988331,
0.0017696769,
-0.0063384403,
0.0179584417,
-0.0368466713,
0.0020047883,
0.0209508277,
0.0995798334,
0.0219553858,
-0.02109305,
0.01167004,
-0.0478627533,
0.0236180276,
0.0568199046,
-0.0079381643,
-0.013855434,
0.0172432587,
0.0040736576,
-0.0132046239,
0.0193297081,
0.0028737311,
0.0521693639,
-0.038752459,
0.0469740555,
-0.0274299681,
-0.0572819449,
0.0806413293,
0.0062179351,
-0.0049291193,
0.0933593884,
0.0157375578,
-0.0185228866,
-0.0541274734,
0.0254742131,
-0.0544312336,
0.0271232743,
-0.0378203131,
-0.0034278356,
0.0021398901,
-0.054559011,
-0.0332233869,
-0.0008321895,
-0.0575485416,
-0.0758680925,
-0.042094253,
0.0320902318,
0.0053869132,
0.0942682996,
-0.0284093972,
-0.0428713039,
0.0502854884,
0.0562467389,
-0.0455872864,
0.0108283767,
-0.010096373,
-0.0135136358,
0.0005970405,
0.0485842377,
-0.0891632438,
-0.0295258071,
0.0086251246,
0.0809928402,
-0.0376746766,
-0.0434748493,
0.0197630953,
0.0574729629,
0.0042448863,
0.0036155984,
0.0097307544,
-0.063882567,
0.0216155667,
-0.0153762288,
-0.0318115279,
-0.0540519729,
-0.0177929979,
0.0195497572,
-0.02026641,
0.0094840582,
-0.0529156402,
0.0209176615,
0.0518947206,
-0.0409720577,
-0.0151409553,
-0.0016844149,
-0.0536094531,
0.1198024303,
-0.036895521,
-0.0104442826,
0.0401992388,
0.0016125628,
0.0206672102,
-0.0431213975,
0.023665471,
0.0356386229,
-0.0223533995,
-0.0113188056,
-0.022553008,
0.0021295673,
-0.0247063711,
-0.0388885252,
0.0584451221,
0.0624007769,
0.0072387136,
0.0249893293,
-0.0366455875,
-0.0198054332,
-0.0282160826,
0.0120723862,
-0.0176239498,
-0.0502406098,
0.0289590955,
0.0276486948,
-0.0359712206,
0.0385078527,
0.072133325,
0.0320524387,
0.0651500076,
-0.0008410867,
-0.0351822525,
-0.0317023396,
-0.0312205572,
0.0304737855,
0.0232337229,
-0.0322645195,
0.0328864157,
-0.0266040806,
-0.0945065245,
0.0094155818,
-0.0434616022,
-0.0963237733,
0.0053188684,
0.0056056897,
0.0473969653,
0.0214439519,
-0.0328957923,
-0.0647423863,
-0.018291764,
-0.0227809306,
-0.020887062,
0.0253906753,
-6.261505759e-33,
-0.0346161202,
-0.0434896201,
0.0355928615,
0.0191532932,
-0.0272824969,
0.0148715023,
-0.0000049897,
-0.0163154434,
-0.0398127548,
0.0219965093,
-0.0131544247,
0.0092575559,
-0.0022816856,
0.0243558567,
-0.0034841408,
0.0229261015,
0.0195662584,
0.0249287095,
0.0290186536,
0.0199416354,
0.0387606919,
0.055007752,
0.0544157661,
-0.0838888288,
0.0377906598,
-0.021604564,
0.0209379103,
0.0101098474,
0.0261140652,
0.0383069143,
0.0016789832,
0.0669904724,
0.0178733822,
-0.0199222472,
0.0066441474,
-0.0383484289,
-0.1201926321,
-0.0212541055,
0.0131226499,
0.0240063835,
-0.0034734274,
0.0107260309,
0.0233556014,
-0.0332165323,
0.0711077303,
-0.0360802673,
0.0310964473,
-0.0256755725,
-0.0794890821,
-0.0551116243,
-0.0142224887,
0.03415538,
0.0280988049,
0.0187688209,
0.0068613011,
-0.0260091666,
0.0361989029,
-0.0147566721,
-0.0581136681,
-0.0166118797,
0.0127430875,
-0.0001235973,
-0.000533581,
0.0053986297,
0.0142209511,
0.0190228354,
-0.0222880449,
0.0346029587,
-0.034950994,
-0.0051806844,
-0.0029455412,
0.0541337915,
-0.0464636795,
0.0246291179,
-0.0448520146,
-0.0578069463,
-0.0477293842,
-0.0104468623,
0.013736017,
0.0082650222,
0.0144896973,
-0.0123090465,
-0.0447416008,
-0.0437512547,
-0.0344012938,
0.0001260299,
-0.0197684187,
-0.0381487571,
0.015944086,
0.0132607035,
-0.0020727252,
0.0054130768,
-0.0052781468,
-0.0698102117,
-0.0537553355,
0.0776387677,
0.0129786571,
-0.0027826347,
-0.0236803014,
0.0409861431,
-0.0479026213,
-0.0467741489,
0.0061011114,
-0.0628745481,
-0.0235745013,
0.0139005827,
0.0303990208,
0.0130919898,
-0.0277573317,
-0.0072693871,
-0.0268512145,
0.0242002234,
-0.0054758345,
-0.0151506271,
0.0594480336,
0.0597277768,
0.0143910907,
0.0348279998,
0.0638539046,
-0.011325391,
-0.0134229595,
0.0141953696,
0.0124960029,
-0.0343003646,
-0.0014217307,
0.002715175,
-0.0301158205,
-0.0068001854,
0.0270649493,
-0.0086683985,
0.0010908779,
0.0099465735,
2.856e-7,
0.035002593,
0.0042476985,
-0.0179083943,
0.053698577,
-0.0340138935,
-0.0045919181,
0.0107679898,
0.0116474098,
0.0103532784,
-0.0195392892,
0.046164535,
0.0044202986,
0.0037635958,
0.0349075757,
-0.127242893,
0.0000035459,
-0.0301376693,
-0.039550513,
-0.0574506037,
-0.0171679631,
0.0380063206,
0.0444923565,
0.0570161603,
0.0046787709,
-0.0096489266,
-0.0797030702,
0.0789746642,
0.0057238904,
-0.0053991005,
0.0304884966,
0.0146469027,
0.0460099801,
-0.0067025851,
0.0708872527,
0.0016958712,
0.0637765974,
0.0147169791,
0.046417661,
-0.0005192708,
0.0339857265,
0.0130034853,
-0.0324500762,
0.0008507345,
-0.0439499728,
0.0932808444,
-0.062643595,
0.0174278449,
0.0182447247,
-0.0477729514,
-0.0562329069,
-0.0272517316,
0.0079962732,
0.0314516574,
0.0084657995,
0.0249404833,
-0.0121812318,
0.0051226034,
-0.0774107352,
0.0168599449,
0.0362361073,
-0.0212335605,
-0.055752527,
0.0184027497,
-0.0256460421,
-0.0123441601,
-0.0648536906,
0.0268150195,
2.758957103e-34,
0.0336859711,
-0.0002579074,
0.0039972481,
0.006097341,
0.0464079678,
-0.0317255333,
-0.0114256414,
0.0192301404,
0.0073531144,
-0.0879706144,
-0.0178620219
] |
http://arxiv.org/abs/2402.02388v1 | http://arxiv.org/pdf/2402.02388v1 | Solution-oriented Agent-based Models Generation with Verifier-assisted Iterative In-context Learning | 2024-02-04T07:59:06 | 2024-02-04T07:59:06 | Tong Niu, Weihao Zhang, Rong Zhao | cs.CL | cs.CL, cs.AI, cs.LG, cs.SE | Agent-based models (ABMs) stand as an essential paradigm for proposing and
validating hypothetical solutions or policies aimed at addressing challenges
posed by complex systems and achieving various objectives. This process demands
labor-intensive endeavors and multidisciplinary expertise. Large language
models (LLMs) encapsulating cross-domain knowledge and programming proficiency
could potentially alleviate the difficulty of this process. However, LLMs excel
in handling sequential information, making it challenging for analyzing the
intricate interactions and nonlinear dynamics inherent in ABMs. Additionally,
due to the lack of self-evaluation capability of LLMs, relying solely on LLMs
is insufficient to effectively accomplish this process. In this paper, we
present SAGE, a general solution-oriented ABM generation framework designed for
automatic modeling and generating solutions for targeted problems. Unlike
approaches reliant on expert handcrafting or resource-intensive neural network
training, SAGE establishes a verifier-assisted iterative in-context learning
process employing large language models (LLMs) to leverages their inherent
cross-domain knowledge for tackling intricate demands from diverse domain
scenarios. In SAGE, we introduce an semi-structured conceptual representation
expliciting the intricate structures of ABMs and an objective representation to
guide LLMs in modeling scenarios and proposing hypothetical solutions through
in-context learning. To ensure the model executability and solution
feasibility, SAGE devises a two-level verifier with chain-of-thought prompting
tailored to the complex interactions and non-linear dynamics of ABMs, driving
the iterative generation optimization. Moreover, we construct an evaluation
dataset of solution-oriented ABMs from open sources.It contains practical
models across various domains. | Solution-oriented Agent-based Models Generation with Verifier-assisted Iterative In-context Learning: Agent-based models (ABMs) stand as an essential paradigm for proposing and
validating hypothetical solutions or policies aimed at addressing challenges
posed by complex systems and achieving various objectives. This process demands
labor-intensive endeavors and multidisciplinary expertise. Large language
models (LLMs) encapsulating cross-domain knowledge and programming proficiency
could potentially alleviate the difficulty of this process. However, LLMs excel
in handling sequential information, making it challenging for analyzing the
intricate interactions and nonlinear dynamics inherent in ABMs. Additionally,
due to the lack of self-evaluation capability of LLMs, relying solely on LLMs
is insufficient to effectively accomplish this process. In this paper, we
present SAGE, a general solution-oriented ABM generation framework designed for
automatic modeling and generating solutions for targeted problems. Unlike
approaches reliant on expert handcrafting or resource-intensive neural network
training, SAGE establishes a verifier-assisted iterative in-context learning
process employing large language models (LLMs) to leverages their inherent
cross-domain knowledge for tackling intricate demands from diverse domain
scenarios. In SAGE, we introduce an semi-structured conceptual representation
expliciting the intricate structures of ABMs and an objective representation to
guide LLMs in modeling scenarios and proposing hypothetical solutions through
in-context learning. To ensure the model executability and solution
feasibility, SAGE devises a two-level verifier with chain-of-thought prompting
tailored to the complex interactions and non-linear dynamics of ABMs, driving
the iterative generation optimization. Moreover, we construct an evaluation
dataset of solution-oriented ABMs from open sources.It contains practical
models across various domains. | [
0.0129529564,
0.0317039937,
-0.0233544819,
0.0397100151,
-0.0438442528,
-0.0356973968,
0.0028163709,
-0.0320643149,
0.0251572952,
-0.0379894674,
0.0211295988,
0.0122963013,
-0.0280785654,
0.074417457,
0.0263506901,
-0.0256325342,
0.0247980841,
-0.0442275628,
-0.0119976439,
-0.0040959967,
0.0019525385,
-0.0218936224,
0.0047402009,
0.0480284579,
-0.0011910989,
-0.0181206893,
-0.0487394854,
-0.0163196437,
0.013832733,
-0.0348588675,
0.0180991367,
0.0262409169,
0.0555803254,
0.0070660436,
0.0000027943,
-0.0599839836,
-0.058490824,
0.0168789811,
-0.0350466594,
0.0080003124,
0.0515852682,
0.0154677974,
-0.0206423514,
0.0002890072,
-0.0350017175,
0.0137515645,
0.0130921518,
0.0022582884,
0.0774592906,
0.0246096812,
0.0001852925,
-0.0042579095,
0.0055594016,
0.0174254682,
-0.0562946014,
-0.0569090396,
0.0478852093,
0.0205419995,
0.0269131325,
-0.0075414493,
-0.0164292865,
0.0026099104,
0.0281602219,
-0.0034453103,
0.0059184367,
0.0099249808,
0.0536775254,
-0.0137264105,
-0.031695921,
0.0428266004,
0.0157554206,
-0.0132540697,
-0.0032917876,
-0.0058754017,
0.0108112395,
0.0454187095,
-0.01485979,
0.0323602483,
-0.0089314152,
-0.0097830482,
0.0052142562,
0.0021828439,
0.0187802333,
0.0157634374,
-0.0657159984,
0.086287111,
-0.0468682833,
-0.0143375443,
0.0004478827,
-0.0393400937,
0.0117815817,
-0.0576154292,
-0.0080290651,
0.0678246245,
-0.0933178663,
0.0062893559,
0.0142763918,
-0.0062682615,
-0.0062498152,
-0.0551457182,
0.0878606886,
0.0042890706,
0.0392503999,
0.0103616901,
0.004009251,
0.0365508646,
-0.0506220683,
-0.0312988162,
-0.0152224684,
0.028455276,
-0.0802981332,
-0.0027236026,
-0.037862964,
0.0302936938,
-0.0148066869,
0.0067625386,
-0.021068925,
0.0469493344,
-0.0721039027,
0.0381378867,
-0.0586602129,
0.0221363194,
0.0270876363,
-0.0223836545,
-0.0824904144,
-0.0024622341,
-0.0162309371,
0.0228973478,
0.0192338098,
0.0061478396,
0.0312448852,
0.0039876988,
-0.040847145,
0.0083310939,
0.0174434651,
0.1035987139,
0.0117012272,
0.0390935913,
-0.0490104146,
-0.0467352383,
0.026296299,
-0.0806111321,
-0.0049777701,
-0.1059236676,
-0.0255789869,
-0.0155975902,
-0.0342057273,
-0.0488620847,
0.0044409833,
-0.0194010846,
0.0111936638,
0.0756878778,
-0.0216696169,
0.0332124829,
0.023009697,
0.0133612342,
0.0021696223,
0.0586650893,
0.0240098834,
0.0142149534,
0.0026722567,
0.0056792432,
0.0288039912,
0.0393250808,
-0.0070865578,
0.0014205949,
0.0090274252,
0.0187467523,
0.0183772165,
0.0003566419,
-0.0552958846,
-0.0087749055,
-0.0157529619,
0.0307392478,
0.0428257883,
0.0377360024,
-0.0042760288,
0.0306169223,
0.0651219934,
0.041520752,
-0.0169938244,
-0.0377476811,
-0.0005630018,
0.0312177166,
-0.0249812473,
0.0033764604,
-0.0671753362,
0.0400427878,
-0.0284564197,
-0.0689476579,
0.0228091311,
0.0153473299,
0.0289686397,
0.0351047739,
0.0624728911,
0.0108138761,
-0.0467935503,
-0.0022145212,
-0.0085396646,
-0.0029596847,
-0.0122642936,
-0.0827497616,
0.0461237915,
0.0885848552,
0.0612240732,
-0.0396256857,
-0.0281098094,
-0.0342419818,
0.0375326462,
0.036969129,
0.01981044,
-0.0331182368,
-0.0562228598,
-0.019057136,
-0.0502814427,
-0.0223933309,
0.0021995853,
0.0417269766,
0.0171604492,
-0.0187782608,
-0.0180696994,
0.0462294593,
-0.0544285513,
0.0385017842,
-0.0859819278,
0.0004460939,
0.0134026669,
0.0172898099,
0.0708045289,
0.0042880611,
-0.0151902791,
0.0439357199,
0.0337068588,
-0.0284421332,
-0.0507340506,
0.0091804191,
-0.0024365932,
0.0519290455,
0.0284029152,
-0.0408282727,
0.0145886512,
-0.0402905121,
-0.0174077768,
-0.0023802109,
0.0291817058,
0.0004694277,
0.0035804987,
-0.0266738255,
0.0331400819,
-0.0468737558,
-0.006092201,
-0.023788264,
0.0311974902,
0.0220135208,
-0.0562547632,
0.017916752,
0.0499809012,
0.0182040501,
0.009085048,
-0.0106454678,
0.0129458541,
-0.0145918326,
0.0501366407,
0.0478701144,
0.0226131566,
-0.0012537818,
-0.0110523012,
-0.0390439816,
0.0126098674,
0.0257426947,
-0.016697688,
0.0024706379,
-0.0412906297,
0.0026034496,
0.0400755405,
-0.003185265,
0.04911948,
0.0242833514,
-0.0027968918,
0.0207854621,
0.0263666529,
-0.0752668753,
-0.0144660808,
0.0243126266,
-0.0342132524,
0.0436763726,
0.0106016016,
0.0011846125,
-0.0697998852,
0.0614210963,
-0.0145461876,
0.009859575,
-0.011432345,
-0.0334660746,
0.0354951248,
-0.0176886879,
0.0041115484,
0.0133289266,
-0.0115603711,
0.0588136092,
0.0525870286,
-0.0536677912,
-0.0743095055,
0.0230594948,
-0.0415519811,
-0.012215659,
-0.0292940214,
-0.0003772749,
0.0159684792,
0.0473322757,
-0.0214976743,
-0.0003027559,
-0.0324807316,
0.0558231473,
0.0050911605,
-0.0118449302,
-0.0473748334,
-0.0336656943,
0.0113600921,
0.0915506929,
-0.0078602107,
-0.0136057297,
0.094216384,
0.007820582,
0.0276538935,
-0.0145398807,
-0.029226847,
-0.0374882594,
0.0238556303,
-0.0357789546,
-0.0064654155,
0.0263517238,
0.0619450063,
-0.0510823764,
-0.0209173039,
0.0254001319,
-0.0017648065,
0.019386258,
-0.0470527858,
0.0024487893,
0.0319827832,
-0.0285745226,
-0.0486300662,
0.001990363,
-0.0070334147,
0.0533553548,
-0.0429360569,
-0.004442363,
-0.0274091344,
0.0248498842,
0.0284796581,
-0.0039176675,
-0.0067038755,
-0.0514577441,
-0.0374736227,
-0.04537002,
-0.0101172766,
-0.0165073816,
0.0311566107,
-0.0231391247,
0.053170681,
0.0043711574,
0.0337143801,
-0.0385711081,
-0.0466467738,
0.0094965538,
-0.0158439726,
-0.0116201984,
0.0630727559,
0.0451184586,
-0.0382640995,
-0.0774789527,
-0.0230716337,
0.0604869314,
0.00397695,
-0.0063198376,
0.0245943144,
-0.0551107153,
-0.0052454849,
0.0292856898,
0.073143065,
-0.0636081472,
0.0075875898,
-0.0131260725,
-0.0034196915,
-0.0398984142,
-0.0471794903,
-0.0007932184,
-0.0930813551,
0.0725563392,
-0.000430725,
-0.0313792787,
0.0017956061,
-0.0438799076,
-0.0240857992,
-0.024208134,
-0.0170036517,
0.014682305,
0.0156567805,
-0.0385237485,
0.0101264007,
0.0046472889,
-0.0494508296,
0.0449232943,
-0.038562648,
0.0287143905,
0.0736868009,
0.0074335164,
0.0089493152,
0.0131482547,
-0.0015813387,
0.0225290712,
0.0346788205,
0.0141819231,
-0.0285101235,
0.0022992906,
0.0047688931,
-0.0124331471,
0.0291455258,
-0.0716938823,
-0.064892374,
0.0698243603,
0.0991920009,
-0.0083736479,
-0.0202572327,
-0.0253608692,
0.0097182021,
0.0252487175,
0.0290099513,
0.0089196041,
-0.0399426334,
-0.0571358278,
-0.0110703344,
-0.0590121225,
-0.0035358451,
-0.0336915515,
0.0149701294,
-0.0154167488,
-0.0099209938,
-0.0113099599,
-0.0078334594,
0.0186006483,
0.0368126705,
0.0200814866,
-0.0382409543,
0.0834015608,
0.0080841538,
-0.0164445229,
-0.0251103919,
-0.0186984707,
0.0530852005,
-0.0052957959,
-0.0181034841,
0.0296577699,
-0.0382698923,
-0.0149576273,
0.0305667929,
-0.0014100565,
-0.0756150857,
-0.0080701904,
0.0849171579,
0.0220910273,
0.0635994077,
-0.0018221732,
0.0015571871,
0.0970322937,
0.0211854205,
-0.009435066,
-0.0176788084,
0.0090935547,
0.0147691537,
0.0079514338,
-0.0016785355,
-0.0247483011,
0.0240807533,
0.0384037569,
-0.0026627241,
-0.0421151407,
-0.0113844881,
0.0413459651,
-0.0068450654,
-0.0420512818,
-0.0206950717,
-0.0854771361,
-0.0363468975,
-0.0089489333,
-0.0091224015,
-0.0511536114,
-0.0400892347,
-0.0027455362,
-0.0223579388,
-0.0270518549,
0.080477424,
0.0135899549,
0.0129299806,
0.017322069,
0.0445138551,
0.0474283397,
0.0018827766,
0.0412283763,
0.0364660323,
0.0049811578,
0.0190593153,
-0.0178556517,
-0.0148198139,
0.0211449526,
0.0086853961,
0.068606846,
0.0180802122,
0.0244689994,
-0.0439918414,
0.0149874482,
-0.0199346412,
-0.0706524104,
-0.0312440209,
0.0487620197,
0.0284177549,
0.0529823564,
-0.0396773219,
-0.0749356002,
0.0062040617,
-0.0065245824,
0.0361866914,
-0.0080264201,
-0.0273009073,
0.0014720467,
0.0460762382,
-0.0133535378,
-0.0288733039,
0.066390872,
0.0188775081,
0.0305718612,
-0.0277626365,
-0.0447392166,
-0.0500115566,
-0.0370367505,
0.0060937456,
0.0075446372,
-0.0163830724,
-0.0058897543,
-0.0664073527,
-0.0478721671,
-0.0275627635,
0.0560184643,
-0.0031614271,
0.0050136223,
0.003717128,
0.0639316142,
0.0067367144,
-0.0030101296,
-0.0096445912,
-0.0280682072,
-0.0450438708,
-0.019970974,
-0.0302925967,
-6.42989459e-33,
-0.0259577073,
-0.049329143,
0.0797162354,
0.0091063185,
0.0166818965,
-0.0210943576,
-0.0429367162,
-0.0095534306,
0.0090485085,
0.0391503647,
0.0152800241,
-0.0325044096,
0.0088363215,
-0.0278243273,
0.0217442364,
0.0263603758,
0.0319354609,
-0.0271571185,
-0.0122237587,
0.0103598665,
0.0082314592,
-0.0085477727,
0.0039438359,
-0.0774577633,
0.0254078936,
-0.0178874098,
-0.0251137465,
0.0453037098,
-0.0368587561,
0.0499675572,
-0.0201164931,
0.0491293222,
0.0173609313,
-0.0464907102,
0.0033505233,
0.0005014583,
-0.0314294696,
-0.0006839096,
-0.0256723873,
-0.0327205509,
-0.0098717669,
-0.0364903584,
0.0445228256,
0.0104722921,
-0.0514148623,
0.0138002466,
0.0499099232,
-0.0404076166,
0.0444289036,
0.0098839523,
-0.1120972335,
0.0051783943,
-0.0509322956,
-0.0341568962,
-0.0053874962,
0.044556886,
-0.0262995511,
-0.0402430221,
-0.0169468913,
0.0356386118,
0.0246482175,
-0.0551880114,
0.1021206304,
0.0148001257,
0.029884072,
0.0406374149,
0.0275941417,
0.0483246446,
-0.0290767048,
-0.0321736895,
0.0032588716,
0.073742941,
0.0184153672,
0.0739869848,
-0.0398895144,
0.0013166536,
-0.0734206736,
0.0367050618,
0.0222514011,
0.0448989794,
0.0650372729,
0.0042905286,
-0.002059191,
-0.039962206,
-0.0141337533,
-0.0232923105,
-0.0119966641,
-0.0321002342,
0.0345439911,
-0.0076194261,
-0.0475771278,
-0.0068549453,
0.0211881381,
-0.0811688453,
-0.0122394292,
0.0048974282,
0.0023650376,
0.0065606008,
0.0036968107,
0.0072771893,
-0.0556321219,
-0.0442796312,
-0.0397269651,
0.0145352371,
0.0147529598,
0.033673875,
-0.0054004886,
0.0191920884,
-0.0012103441,
0.0160900429,
-0.0529754125,
0.0277577043,
0.0485090762,
0.0100745969,
0.005027032,
0.0008456979,
0.0053063808,
0.0353085697,
-0.0065269321,
-0.0572706237,
0.0299281422,
0.0452680252,
-0.0240857042,
0.018355161,
-0.0182934701,
-0.0316292383,
-0.057834588,
0.0077887583,
0.0331639349,
-0.0070158248,
-0.0133046433,
0.0154089425,
3.372e-7,
-0.012081597,
0.0415951014,
0.007661595,
0.0120015452,
0.0303910132,
0.0764157176,
-0.0225824676,
-0.0142555088,
-0.0119743301,
-0.0323203839,
0.0307504237,
0.0521965809,
0.0349968933,
0.0064480947,
-0.0279674269,
-0.0291136056,
-0.0131763676,
-0.0007910923,
-0.0793771446,
-0.02147731,
0.0477317497,
-0.0076616602,
0.0212906469,
-0.0088798488,
0.0201651827,
-0.0244651996,
0.0383655429,
-0.0431619212,
0.0129746972,
0.0589012131,
-0.0940979347,
0.0046132142,
0.0158990584,
0.0383304767,
0.0197127163,
0.0221189614,
-0.0501999445,
0.0752803534,
0.0128851142,
0.0034194498,
0.0264903139,
-0.0163871534,
0.0120490845,
-0.0529834852,
0.0198250748,
-0.1067077145,
-0.0422190689,
-0.0554423183,
-0.0267957002,
-0.0439859331,
-0.0151060447,
-0.0197348129,
0.0129610682,
-0.0061140233,
-0.038756106,
-0.0108079854,
0.0031735317,
-0.0164778549,
0.0185499899,
-0.0142369084,
-0.0351835489,
-0.032795541,
-0.0053413594,
0.0251922254,
0.0514625162,
-0.039617002,
0.0604615919,
3.418527442e-34,
-0.0008956692,
0.0095283687,
0.015298713,
0.0489113219,
0.0373784937,
0.0276889428,
-0.0591281913,
0.0356035195,
0.0369326733,
-0.0432160832,
-0.0346172675
] |
http://arxiv.org/abs/2402.02380v3 | http://arxiv.org/pdf/2402.02380v3 | Evaluating Large Language Models in Analysing Classroom Dialogue | 2024-02-04T07:39:06 | 2024-02-23T02:19:09 | Yun Long, Haifeng Luo, Yu Zhang | cs.CL | cs.CL, cs.AI, cs.HC | This study explores the application of Large Language Models (LLMs),
specifically GPT-4, in the analysis of classroom dialogue, a crucial research
task for both teaching diagnosis and quality improvement. Recognizing the
knowledge-intensive and labor-intensive nature of traditional qualitative
methods in educational research, this study investigates the potential of LLM
to streamline and enhance the analysis process. The study involves datasets
from a middle school, encompassing classroom dialogues across mathematics and
Chinese classes. These dialogues were manually coded by educational experts and
then analyzed using a customised GPT-4 model. This study focuses on comparing
manual annotations with the outputs of GPT-4 to evaluate its efficacy in
analyzing educational dialogues. Time efficiency, inter-coder agreement, and
inter-coder reliability between human coders and GPT-4 are evaluated. Results
indicate substantial time savings with GPT-4, and a high degree of consistency
in coding between the model and human coders, with some discrepancies in
specific codes. These findings highlight the strong potential of LLM in
teaching evaluation and facilitation. | Evaluating Large Language Models in Analysing Classroom Dialogue: This study explores the application of Large Language Models (LLMs),
specifically GPT-4, in the analysis of classroom dialogue, a crucial research
task for both teaching diagnosis and quality improvement. Recognizing the
knowledge-intensive and labor-intensive nature of traditional qualitative
methods in educational research, this study investigates the potential of LLM
to streamline and enhance the analysis process. The study involves datasets
from a middle school, encompassing classroom dialogues across mathematics and
Chinese classes. These dialogues were manually coded by educational experts and
then analyzed using a customised GPT-4 model. This study focuses on comparing
manual annotations with the outputs of GPT-4 to evaluate its efficacy in
analyzing educational dialogues. Time efficiency, inter-coder agreement, and
inter-coder reliability between human coders and GPT-4 are evaluated. Results
indicate substantial time savings with GPT-4, and a high degree of consistency
in coding between the model and human coders, with some discrepancies in
specific codes. These findings highlight the strong potential of LLM in
teaching evaluation and facilitation. | [
0.0474956073,
0.0403052606,
-0.0247793384,
0.0485086963,
-0.011790893,
0.0124454889,
0.0187562071,
-0.0353051759,
-0.0175975244,
-0.0711457059,
-0.017568646,
-0.0266715176,
0.0091554197,
-0.008419075,
0.0105866436,
-0.0358410366,
0.007075178,
-0.0135304052,
0.0282385871,
0.0334562734,
0.0264925752,
0.003972502,
-0.020274451,
0.0421655588,
-0.0051060859,
-0.0398317538,
-0.0213729125,
-0.0325420573,
-0.0196265057,
-0.0275106747,
0.0208498631,
0.08249183,
0.0111140516,
0.0471413322,
0.0000023572,
-0.0284354854,
0.0029170527,
0.0070802807,
-0.0220508259,
0.0915340334,
0.0108237546,
0.0309213325,
-0.0083320606,
0.0066003655,
-0.0580632463,
-0.0076108589,
0.0430179499,
0.0068166195,
0.0594758056,
0.0040091695,
0.0024443881,
-0.0556980111,
0.0271516796,
0.0284228176,
-0.0327346921,
-0.0360798575,
0.0373122916,
0.1085343882,
0.0204745438,
0.0282635596,
-0.0594800673,
0.0290709753,
0.0162351485,
-0.0692620724,
-0.0160760023,
-0.0249759536,
-0.0512880348,
-0.0058876174,
0.0143816257,
0.0577673949,
0.0502498969,
0.019891154,
-0.0084860278,
-0.0179076642,
0.0121303517,
0.0262366142,
-0.0886202604,
0.029713152,
0.0065908707,
0.0237697661,
-0.0134771373,
0.0012727865,
-0.0297157578,
0.0046291281,
-0.0274826176,
0.0053300471,
-0.0044254595,
-0.0316815451,
-0.0334216766,
0.028540777,
-0.0459459014,
-0.0252716113,
-0.0252312254,
0.0360533111,
-0.0020418048,
0.0517683886,
0.0195797496,
0.0078514498,
0.0140909655,
-0.0384473801,
0.0012753227,
0.0533327907,
0.0105138021,
-0.0094898436,
0.0390510932,
0.0696521774,
0.0152886519,
-0.0213211887,
-0.0267415047,
0.0338539258,
-0.0894661769,
-0.0186459217,
-0.027981801,
0.028401209,
0.0158841275,
-0.0207400173,
-0.0388872139,
0.0068870927,
-0.0158981401,
0.0305021703,
-0.0306700654,
-0.06573838,
0.0124114202,
0.0110818846,
-0.013125658,
0.0762120858,
0.0131993582,
0.0383995846,
-0.0344131887,
-0.0864455104,
0.0411609747,
0.0401869565,
-0.0331865549,
-0.0167278834,
-0.0427298546,
0.015260024,
-0.0277240165,
0.007675604,
0.0149510866,
-0.0361979902,
-0.0202555768,
-0.0452167206,
-0.028448889,
-0.0626424924,
-0.0410801172,
0.0128312074,
0.0074253515,
-0.0320668556,
0.0250134468,
-0.0107714441,
0.0203622598,
0.0104444874,
-0.0244886335,
0.0468294844,
0.063434273,
0.0114213265,
0.0407807119,
0.0508018248,
0.0624947399,
-0.0792231932,
0.0147656417,
-0.0826283917,
0.0283050966,
0.0510047786,
-0.0068609826,
0.030579377,
-0.0253453255,
0.0035634385,
0.0355117768,
0.0325237699,
-0.044819124,
0.0601194836,
-0.0287599955,
0.0469085276,
0.0229818933,
0.0742552876,
-0.0329059586,
0.0234566461,
0.0172962565,
0.0340018459,
0.0571437143,
0.0049395352,
0.0131615391,
-0.0214342102,
0.0268915575,
-0.0196803827,
0.0566655621,
-0.0060336529,
-0.0029756422,
-0.0751515105,
-0.0420144014,
0.015799718,
0.021932736,
0.0194724351,
0.0574067906,
-0.033080522,
-0.0073877531,
0.0267155189,
-0.0083142612,
-0.0441977307,
-0.0038983971,
-0.0651105419,
0.0426946357,
0.0659734085,
0.0314419568,
-0.0029526157,
-0.0325846896,
-0.0262783393,
0.0267420057,
0.0392276905,
0.0249257851,
-0.0080446582,
-0.0442318805,
-0.0490842946,
-0.0157505199,
-0.0212660208,
0.01531203,
-0.0189799424,
-0.0122206677,
0.0035706898,
0.0189602841,
0.0094001722,
-0.0733914301,
-0.0016497097,
-0.0775752366,
-0.0255357474,
-0.0339836814,
-0.0296700448,
0.0400876813,
0.0399809219,
-0.0127117075,
0.0262971167,
-0.0088142315,
0.0202363562,
0.024289703,
0.0112628005,
-0.0316666216,
0.011585813,
0.0448309258,
-0.0339244828,
-0.0078078164,
-0.0114421481,
-0.0198282544,
-0.0505374297,
0.0205878671,
-0.0299349595,
0.031467583,
-0.0096134273,
0.0200018305,
-0.0557247289,
-0.019550547,
-0.0315357,
-0.005671191,
0.0400865376,
0.0244009066,
0.0201398283,
0.0288079642,
0.0693485588,
-0.0282571223,
-0.0183045622,
0.0428284816,
-0.0070213284,
0.0367877334,
0.0136066303,
-0.009234976,
-0.0342594795,
-0.0416778773,
-0.0873379931,
-0.0352742039,
0.0064398726,
0.0110353483,
-0.0058694296,
-0.0437651165,
-0.0137922671,
0.0320586376,
-0.0339230336,
-0.0495945252,
-0.0051236665,
0.0083889365,
-0.0086573195,
0.009301262,
-0.0185093563,
-0.0290233791,
-0.0145073673,
-0.0203691851,
0.0356848799,
0.0923762843,
-0.0041240496,
-0.0383536443,
0.0362699553,
-0.0237011872,
-0.030196039,
0.0134609751,
-0.0569676831,
0.047582563,
0.0054787211,
0.0338826887,
0.0329092368,
-0.0105260778,
0.0504921824,
0.0267803483,
-0.0899443626,
-0.052026473,
-0.0285255462,
0.0258140489,
-0.0403299481,
-0.0538508222,
-0.051613111,
-0.0405686498,
0.0671844557,
0.0384460613,
-0.0252460595,
-0.0447726585,
0.0342494398,
-0.0695409551,
0.0237090588,
-0.0277255941,
-0.0682587773,
-0.0292635821,
0.0461792909,
0.0334597938,
-0.0064033968,
0.0214878339,
-0.0206317566,
0.0104332538,
-0.0695726573,
0.0077865445,
-0.0223598853,
0.0199664161,
-0.0301145613,
-0.0297542904,
-0.0346949771,
-0.0215383209,
0.0082083419,
-0.0059314067,
0.0214078166,
-0.0611634739,
0.0558674149,
-0.0978575349,
0.0078299893,
-0.0189311523,
0.0137806106,
0.0001129345,
0.0129909236,
-0.0293212608,
0.0229281466,
0.0236645192,
-0.0189935789,
0.0432472043,
0.0682846829,
-0.0075879013,
0.0251886658,
-0.0445572659,
0.0155723076,
-0.0021442904,
-0.0177608151,
-0.0202663746,
-0.0081785871,
0.004557054,
-0.0296429005,
0.1067395806,
0.0289497469,
0.0047681015,
0.0524642766,
-0.0272498783,
-0.0320697911,
0.0155250402,
0.0460084677,
-0.0640855432,
0.025774749,
0.0447421931,
-0.0959260464,
-0.0073165991,
0.1077722609,
0.0066353218,
-0.0292269289,
-0.0269573648,
-0.0398876555,
-0.0245331768,
-0.024622703,
-0.0232039932,
-0.0768062323,
0.0072704549,
-0.0241755173,
0.0073200031,
-0.0270114075,
-0.0187686272,
0.0027204112,
-0.1041768044,
0.0512474291,
0.021115426,
-0.0250181817,
0.0044108541,
-0.011687194,
0.0212486424,
-0.0008224042,
0.0452502593,
0.0114943646,
0.0049815862,
0.026037734,
0.0613961741,
-0.0607424825,
-0.0475881621,
0.0230068248,
0.0383366831,
0.0495129228,
0.0152428653,
0.0368256047,
-0.037104331,
-0.0180597827,
0.0458995216,
0.0037958538,
-0.0639356077,
0.00417095,
-0.0039416659,
0.0301105306,
-0.0027564182,
-0.0108257812,
0.0399030522,
-0.012804741,
-0.0110459803,
0.0575069413,
0.0783636495,
-0.0177708361,
0.0341505371,
-0.0007227298,
0.050081376,
-0.0102858916,
0.0691163763,
0.0611432195,
-0.0452217422,
0.0324325189,
-0.0167445894,
-0.0138115948,
0.005418682,
-0.0472534522,
0.0062011611,
-0.0074690236,
0.0237969607,
-0.0095431078,
-0.0440535545,
0.0606425628,
-0.0085132662,
0.0187317058,
0.0370628573,
0.0373231173,
0.0007815865,
0.0044341236,
0.0108904913,
0.0065352111,
0.012099361,
-0.0679674,
-0.0030765776,
0.0330667906,
0.0142620252,
0.0047493149,
0.0247931797,
-0.0559562109,
-0.064472869,
-0.0046223779,
0.0036117458,
0.0150750652,
-0.01291773,
0.0522805564,
-0.0228417926,
-0.0645940602,
-0.0293395594,
0.0092630759,
-0.0051316712,
-0.0008994711,
0.0672178194,
0.0052254102,
-0.0201633181,
0.0422067232,
-0.0017757586,
-0.0002124921,
0.0096044429,
-0.0840255246,
-0.0876695141,
0.0046871006,
0.0258552767,
0.0069378717,
-0.0056198784,
-0.0099999774,
-0.0316638276,
-0.0122705661,
-0.0644863546,
-0.0225279257,
-0.0607846193,
0.006398832,
0.0327298716,
-0.0303529967,
-0.0108595807,
-0.049199827,
-0.0517692082,
0.037275143,
0.0067031994,
0.0600066669,
-0.0288909581,
-0.0170541108,
0.0133446511,
-0.0039333357,
-0.0135618001,
-0.0042786039,
0.0031142777,
0.0360106938,
0.0268860832,
0.0906832814,
0.0132840993,
0.0165486336,
0.0889469013,
-0.0143698612,
0.0292253084,
-0.0590042323,
-0.0179537255,
0.0070658857,
-0.020750057,
0.0152436076,
0.0345171019,
-0.030917,
0.0085224602,
-0.0367652588,
0.0507571809,
-0.0047090119,
-0.0266340002,
0.0544820614,
0.0224353634,
-0.0017027976,
-0.0281324759,
0.032658428,
0.0046171937,
-0.0032685536,
-0.0012482258,
-0.0132658612,
-0.0163696017,
-0.0120164575,
0.0619522482,
-0.0218564831,
0.0157950893,
0.0136186266,
-0.0724859685,
-0.029259881,
-0.0535306744,
0.0048000109,
0.0016405551,
-0.0234055556,
0.0124858879,
0.0556261875,
-0.0315746441,
0.0051250588,
-0.0425886102,
0.0043005822,
-0.0003797046,
-0.0260541644,
-0.0072198366,
-6.435282762e-33,
-0.0493149124,
-0.0750984848,
0.0342583954,
0.0153995221,
-0.0278649535,
-0.0243309326,
-0.0124165351,
0.025292268,
-0.0175483786,
0.0181783084,
0.0327774882,
0.0271110218,
-0.0096264156,
0.0387839228,
0.0570713058,
0.016294498,
0.0191616099,
0.0148996376,
0.0062608598,
0.0209849291,
0.0127290571,
-0.0527521446,
0.0533312969,
-0.0475607626,
0.0582700893,
-0.0206305943,
0.0031510598,
0.0008471445,
-0.0449884422,
0.0570868887,
-0.0199917015,
0.035326384,
0.0203765407,
-0.0437315665,
-0.008692435,
0.0298182126,
0.0073121432,
-0.0400871262,
0.0418966897,
0.0346273817,
0.021560071,
-0.0467075408,
0.0170244239,
-0.0250257961,
-0.040529713,
-0.0198075213,
0.0389276817,
-0.035157308,
-0.0189269725,
0.0028478203,
-0.0610618256,
0.0616849959,
-0.0539517403,
-0.0202877615,
0.0356140658,
0.0402504057,
-0.0307739954,
-0.0439137779,
-0.0289840493,
0.0494899154,
0.0931562856,
0.0090146419,
0.0102318721,
0.0072870259,
0.0153768035,
0.0206957832,
0.0158791263,
-0.0086570466,
0.0048116408,
-0.0066844043,
0.0433813632,
-0.0054386156,
0.0618758984,
0.0353002697,
-0.0298970584,
-0.0004443607,
-0.0489480197,
-0.0267539322,
-0.0046003871,
0.0004928128,
0.010622072,
-0.03681143,
-0.0415509492,
-0.0661356747,
-0.0440940931,
0.0165289938,
0.0219683927,
-0.0333458893,
0.0264021419,
-0.0741448328,
-0.0113874003,
0.0147149349,
0.0658732057,
-0.014987926,
-0.024355771,
0.0102577358,
-0.0281353146,
0.0551428944,
-0.0371303931,
0.0248298962,
-0.0613320693,
-0.0084668566,
0.0095864767,
-0.0072751613,
0.0210565589,
0.0140065877,
-0.0170368291,
-0.0037278912,
-0.0131396614,
-0.0047849701,
-0.0227258019,
-0.0200692434,
0.0006273486,
-0.0392300524,
-0.0218725316,
0.021462312,
0.0070564179,
0.0792616531,
0.0268471651,
-0.0329611897,
-0.0276757311,
-0.0296030734,
-0.0271307584,
-0.0352983214,
-0.0212126635,
0.0039021943,
-0.0018113761,
0.021775512,
0.0500772968,
-0.0106799984,
0.007065028,
-0.0131293209,
3.013e-7,
0.0663925037,
0.0735979304,
-0.0028331198,
-0.0426978618,
0.0136161325,
0.0026135733,
-0.0178303551,
-0.0458408408,
0.0596784577,
0.0264078807,
0.0380583405,
0.027248418,
0.0613135509,
0.0380228646,
-0.0338353142,
-0.0131205888,
-0.0519548841,
-0.0149285635,
-0.0502362661,
0.0312967002,
0.0704527497,
0.0125288935,
-0.0133920517,
0.0277777091,
-0.0656390712,
-0.0105084088,
0.0252445005,
-0.0261373464,
-0.0373468623,
0.0812506527,
0.0196660906,
-0.0216010865,
0.0218086056,
0.0176633019,
0.0208488237,
0.0254495572,
0.0189136155,
0.0622311831,
0.0178149883,
0.0358500592,
0.0158780888,
-0.0188023001,
0.0042008935,
-0.0553813577,
0.0201518852,
-0.064008005,
-0.0233843457,
0.0370675959,
-0.0210632551,
-0.0641048029,
0.0440168679,
-0.0306648184,
0.0200949442,
0.0379520915,
-0.0187157635,
-0.0139033813,
0.0506019443,
0.0299313869,
0.0041253092,
-0.0178150274,
-0.0577662252,
-0.0558077432,
0.0033367611,
-0.0530600213,
-0.0270496663,
-0.0035210403,
0.0169251841,
2.21192198e-34,
0.0528359488,
-0.0361364484,
0.0415382199,
0.0418439731,
0.0370919779,
0.0181198772,
-0.0073650805,
0.0498866402,
0.0399333686,
-0.054676801,
-0.0046753548
] |
http://arxiv.org/abs/2402.02379v1 | http://arxiv.org/pdf/2402.02379v1 | Rethinking the Evaluation of Pre-trained Text-and-Layout Models from an Entity-Centric Perspective | 2024-02-04T07:33:45 | 2024-02-04T07:33:45 | Chong Zhang, Yixi Zhao, Chenshu Yuan, Yi Tu, Ya Guo, Qi Zhang | cs.CL | cs.CL | Recently developed pre-trained text-and-layout models (PTLMs) have shown
remarkable success in multiple information extraction tasks on visually-rich
documents. However, the prevailing evaluation pipeline may not be sufficiently
robust for assessing the information extraction ability of PTLMs, due to
inadequate annotations within the benchmarks. Therefore, we claim the necessary
standards for an ideal benchmark to evaluate the information extraction ability
of PTLMs. We then introduce EC-FUNSD, an entity-centric benckmark designed for
the evaluation of semantic entity recognition and entity linking on
visually-rich documents. This dataset contains diverse formats of document
layouts and annotations of semantic-driven entities and their relations.
Moreover, this dataset disentangles the falsely coupled annotation of segment
and entity that arises from the block-level annotation of FUNSD. Experiment
results demonstrate that state-of-the-art PTLMs exhibit overfitting tendencies
on the prevailing benchmarks, as their performance sharply decrease when the
dataset bias is removed. | Rethinking the Evaluation of Pre-trained Text-and-Layout Models from an Entity-Centric Perspective: Recently developed pre-trained text-and-layout models (PTLMs) have shown
remarkable success in multiple information extraction tasks on visually-rich
documents. However, the prevailing evaluation pipeline may not be sufficiently
robust for assessing the information extraction ability of PTLMs, due to
inadequate annotations within the benchmarks. Therefore, we claim the necessary
standards for an ideal benchmark to evaluate the information extraction ability
of PTLMs. We then introduce EC-FUNSD, an entity-centric benckmark designed for
the evaluation of semantic entity recognition and entity linking on
visually-rich documents. This dataset contains diverse formats of document
layouts and annotations of semantic-driven entities and their relations.
Moreover, this dataset disentangles the falsely coupled annotation of segment
and entity that arises from the block-level annotation of FUNSD. Experiment
results demonstrate that state-of-the-art PTLMs exhibit overfitting tendencies
on the prevailing benchmarks, as their performance sharply decrease when the
dataset bias is removed. | [
0.0263786223,
0.0026225855,
0.0285769124,
0.0298492592,
0.0240691826,
0.047219038,
-0.0018540524,
0.0156373139,
0.0093012797,
-0.0566762425,
-0.0521399789,
0.0223844219,
-0.0034547762,
-0.0131788263,
0.0090862028,
-0.0440021493,
-0.0011630871,
-0.0022826397,
0.0018845415,
0.068702966,
0.0285730436,
-0.0017003161,
0.0076230089,
-0.0333354585,
-0.0470868833,
-0.0043749064,
-0.0020883165,
0.0592382364,
0.0468413271,
-0.0542754158,
-0.0190235432,
0.0236518066,
0.0086069116,
0.0468125492,
0.0000018098,
-0.0474565886,
-0.0361000188,
0.0431364216,
-0.041043222,
0.0137375547,
0.0057745087,
-0.0476423651,
-0.0159244034,
-0.0642251149,
-0.0017092959,
-0.0320289172,
0.0335755013,
0.0863449648,
0.0239292886,
0.0770366937,
-0.0349141583,
-0.0391698964,
0.0592552908,
-0.0452655815,
-0.0095167859,
-0.0273790117,
0.0213530865,
-0.0114795985,
-0.0068535996,
-0.0061069247,
-0.0008295476,
0.016046945,
0.0300187692,
0.0094407247,
0.0121662486,
0.0271361582,
-0.0175526813,
0.0112796212,
-0.0100331334,
0.1063778698,
-0.002465226,
-0.0036424669,
0.0312066637,
-0.0336914398,
-0.034640681,
-0.0097356401,
-0.0258415937,
-0.0133811999,
-0.0369911864,
-0.0088504888,
-0.00663496,
0.0409212001,
0.0242332071,
-0.0163324326,
-0.0154704666,
0.0000669624,
0.0036422035,
-0.0199852698,
0.0432850793,
0.0539554954,
-0.0078364946,
-0.0501884334,
0.0181811713,
-0.0008219968,
0.0426297784,
0.024182573,
-0.0691135079,
-0.0522678234,
-0.0134149538,
-0.0196665693,
-0.0330117978,
-0.0280071869,
-0.027065495,
0.0700076222,
-0.0760294795,
0.0383144878,
-0.0353325978,
0.0703840479,
-0.0000451148,
0.0229095742,
0.0213340521,
0.0202534311,
-0.0348390266,
0.0271058809,
-0.0273093637,
-0.0553472377,
0.0126715861,
-0.0022638212,
0.0173601191,
0.060157463,
-0.0092972154,
-0.0555968732,
-0.0231818892,
-0.0362154618,
-0.0016639768,
-0.0261084363,
-0.0027420961,
-0.0189436786,
-0.0186965689,
-0.0168965869,
0.0098806256,
0.0071553607,
0.0007286021,
-0.0688944608,
-0.0030287823,
0.0668461919,
0.0068540685,
-0.0792561024,
-0.0606442317,
-0.0060835206,
0.0263720769,
-0.0541333109,
0.0416776463,
-0.0223377291,
0.0304382611,
0.0140164346,
-0.0146910874,
-0.0226447508,
0.0100313053,
-0.0100910468,
-0.0518135056,
0.0620243065,
-0.0382637456,
-0.0041186004,
-0.0030921812,
0.0273216404,
0.0514458716,
0.0820043385,
0.0438907444,
0.0650526285,
0.0560082346,
0.017919274,
-0.0339467116,
0.0675581992,
0.0238747057,
-0.0418204181,
0.0311965961,
0.0472607724,
0.0230337791,
-0.009716426,
0.0280356817,
0.0267302804,
-0.0513367839,
-0.0341434106,
0.0463681333,
0.0100001013,
0.0428409316,
0.0031498095,
0.0429581068,
-0.0219685379,
0.0404674523,
0.0081409244,
-0.0562511608,
-0.0277932789,
-0.002221708,
0.0423818976,
0.0327859223,
0.0212740172,
-0.0149323987,
0.0047719572,
-0.0182561874,
-0.015305927,
0.0132022556,
-0.0106119569,
0.0848493129,
0.007933815,
-0.0694469661,
0.0386918075,
0.0080527542,
-0.0253409296,
-0.0307054333,
0.0203228462,
0.0189916939,
0.0347013772,
0.0465976559,
-0.0511448495,
-0.0371507816,
-0.0023693882,
0.0220407303,
0.0363930352,
0.0526011288,
-0.0273692235,
0.0431739688,
-0.0194367915,
-0.0237744041,
-0.0122970482,
-0.0125156455,
-0.0079149716,
0.0537842102,
0.0130607961,
0.0266071819,
-0.0505224206,
-0.0050982772,
0.0665517151,
-0.0152071277,
-0.0038917419,
-0.0427014865,
0.0211441722,
-0.0239942726,
0.0380274057,
-0.0434833653,
-0.014672651,
0.0207121782,
-0.0087266769,
0.0610675737,
-0.0126631064,
-0.0261826739,
-0.0398107208,
-0.0168144777,
-0.0042454833,
-0.0464537293,
-0.0045738639,
-0.050347425,
-0.0241226088,
-0.037967898,
0.0353256315,
0.04167578,
0.0019210748,
0.0054023918,
0.0355378762,
0.0275789332,
-0.0312912278,
0.0476835854,
0.0105097275,
0.016889492,
0.0117627857,
0.0000186495,
0.0118492208,
-0.0312137026,
-0.0296608191,
-0.0130256768,
-0.0164174233,
-0.0231600031,
0.0027689398,
0.0315835699,
-0.0693882778,
-0.0029899997,
-0.0835329741,
0.0205562748,
-0.0329769142,
0.0491919853,
0.016652327,
-0.0150037343,
-0.0406938754,
-0.0597303547,
-0.0082960213,
0.0008365291,
-0.0303216632,
-0.0025253124,
0.0295703541,
0.0343252681,
0.014558184,
-0.0410872474,
-0.0287529752,
0.0029566318,
0.0393171161,
0.0536741987,
-0.0499875173,
-0.0041372352,
0.0527425706,
0.0146589018,
-0.0098799095,
-0.0108672082,
0.0310572665,
-0.0188630652,
-0.0163490921,
-0.0018657013,
0.0962703228,
-0.0354911238,
0.0583295338,
-0.0495754443,
-0.0135708703,
-0.0310588423,
0.0306714792,
-0.0157194678,
0.0864976645,
-0.0064717936,
-0.0125339469,
0.0309698489,
0.0431272723,
-0.061956469,
0.0421747901,
-0.0160078947,
-0.0124153253,
0.0110827768,
0.0059174881,
-0.0102840392,
0.0013008084,
0.0437253751,
0.0339227356,
-0.0410625488,
-0.0470050462,
0.0104560843,
-0.0686105415,
-0.0101623284,
0.0225716438,
0.0327988155,
-0.0308325049,
0.0525466464,
0.0607064888,
-0.0071411328,
-0.0128305387,
-0.0300853383,
-0.0297156107,
0.0397518948,
0.0017248511,
-0.0120335296,
0.0245226081,
-0.03012757,
0.0733369216,
-0.0567138381,
-0.0252958611,
-0.0066952161,
-0.0256584976,
-0.006554347,
-0.0087952046,
0.0027047733,
-0.000733402,
-0.0066219931,
0.0651225001,
-0.0022873706,
0.0957390293,
-0.0055709849,
-0.0281457361,
-0.0060866098,
0.0836787969,
0.0239577666,
-0.0307966694,
0.0436570272,
-0.0340444632,
0.0137890773,
0.036801178,
-0.0050915424,
-0.0078057488,
-0.0194020327,
0.0034440614,
-0.0065797302,
-0.0359460972,
-0.0022055842,
-0.0145342136,
0.0553834662,
-0.0622942559,
-0.0306177605,
0.0979414582,
0.0263204928,
-0.0082182018,
-0.0621951073,
0.0428471416,
0.021134492,
-0.0143631287,
0.0069071148,
-0.0877541602,
-0.0185096934,
-0.0090910457,
-0.0197466109,
0.0452553704,
-0.0339515135,
0.0074724704,
-0.0829875097,
0.0077064494,
-0.0224779546,
-0.0624304041,
-0.0014844772,
-0.0588295646,
0.0134393396,
-0.0468699187,
-0.0244647842,
-0.0273044836,
-0.0219264887,
-0.0156838633,
-0.0027106486,
-0.0381555893,
-0.0300680883,
0.0115996636,
-0.0350833721,
0.0048091263,
0.053230498,
0.0116367713,
0.0008413131,
0.0162781235,
0.0291783251,
-0.0253585093,
-0.0295068268,
0.0051745824,
0.0854651183,
-0.0620898791,
0.0141226072,
-0.0038617863,
0.0003227343,
0.0009013241,
-0.023012992,
-0.0191348214,
-0.0027176484,
0.0254278835,
-0.0006685148,
-0.0181718078,
-0.0259266533,
-0.0116158035,
0.0380504318,
0.0285678953,
-0.0167539902,
-0.0811828822,
-0.0202075746,
0.030566873,
-0.0631208047,
0.0068786456,
0.0153957577,
-0.0794757158,
-0.0365292393,
-0.0906785578,
-0.0481475033,
0.0288274437,
0.0724227652,
0.0007654931,
-0.0412809886,
-0.0439051464,
-0.0140350033,
-0.0470321402,
0.0506061912,
0.0042772428,
0.0202406105,
-0.0615559258,
0.0036913187,
0.0144811273,
-0.0475129709,
-0.0315516852,
-0.0010047972,
-0.0768125504,
-0.0116574066,
0.0001539861,
0.0470928475,
0.0143367881,
0.0579338707,
-0.0224612001,
-0.0168711618,
0.0253437441,
0.0413589478,
-0.0669958293,
-0.0313979536,
0.0157639887,
0.012569394,
-0.0317928754,
-0.0062354458,
-0.0326315649,
0.0487025678,
0.0731359795,
0.0216788482,
-0.0117217908,
-0.0248766039,
-0.0263877288,
0.025764592,
0.0684519932,
0.0122249331,
-0.0069321892,
-0.0410811268,
0.0347384848,
-0.0033551208,
0.0184054021,
-0.0229876786,
-0.007688242,
0.005829385,
0.0484142452,
-0.0223042425,
0.0626471713,
0.0610402338,
0.0324265398,
-0.0598532222,
-0.0084887985,
0.0146091171,
0.0134730106,
-0.004915284,
0.0020645291,
0.008631289,
0.0047483863,
0.0292020161,
0.0622797944,
-0.0315500945,
0.0484017953,
0.0383126326,
-0.0026556007,
0.0229283422,
0.051992204,
-0.0360900015,
0.0214971919,
-0.070197776,
0.0177537166,
0.0302256402,
-0.0302779544,
0.0595134459,
-0.0156584419,
-0.022382969,
0.0134835122,
0.0015392906,
0.0194377508,
0.0072759092,
0.0198872499,
0.0213300642,
0.0015379625,
0.0123394495,
0.0319621712,
0.0312217847,
0.0198557973,
0.0110796159,
-0.0160511639,
-0.0485805124,
-0.0108560072,
0.0439556763,
0.0075685414,
-0.0076939017,
-0.0317244641,
-0.0325450897,
-0.0303971581,
0.0311578978,
0.0148748821,
-0.0192719661,
0.0461858474,
0.032380484,
0.0339387618,
0.0434837751,
-0.0027792356,
-0.0139038181,
-0.0264874212,
-0.0096313478,
-0.0108669503,
0.0194757096,
-5.751275089e-33,
-0.0680331364,
-0.0072478489,
0.0151543757,
-0.0021849072,
-0.0497544892,
0.0282525029,
-0.0288320798,
-0.0406653099,
-0.0277345851,
-0.0019501942,
-0.0113055026,
-0.0178241543,
0.0195167698,
0.0289568491,
-0.0653345808,
-0.0211794749,
0.0375448391,
0.0047231335,
-0.0505636968,
0.0104616014,
0.0128865195,
0.0251238272,
0.0690222532,
-0.0738733411,
0.0016072367,
0.0049183285,
0.0042306753,
-0.0253700651,
0.0463707633,
0.04634941,
-0.0032033406,
-0.0433895737,
0.0231118035,
-0.0483499616,
-0.0122577967,
0.0320438296,
-0.1448501199,
0.0064025819,
-0.0053788158,
0.0480832607,
-0.0281449817,
0.0076427758,
0.02409802,
-0.0098240124,
0.0381080061,
-0.0357157215,
0.0394635834,
-0.0100136911,
-0.0119311893,
-0.046198234,
-0.0709821805,
0.0690729544,
-0.0046721781,
0.0315667056,
0.0772024021,
-0.0346977226,
0.0149456719,
0.0261797067,
-0.090349786,
-0.0255654082,
0.029791845,
-0.0194206145,
0.0358504578,
0.0185835399,
0.0055286931,
0.0470910519,
0.0157182291,
0.0062560719,
-0.0214849282,
0.0432671532,
0.021371793,
0.0556701384,
0.0047854432,
0.0238384828,
0.0264734142,
0.0299801994,
-0.1052911654,
0.0610423349,
-0.0330635533,
-0.0050871801,
-0.0169726312,
-0.003342981,
-0.0439095646,
0.010070351,
-0.0263547692,
-0.0137944482,
-0.0058356742,
-0.0152516933,
0.0480861515,
0.0310702063,
-0.0325657465,
0.0786070526,
0.0459814854,
0.0016186107,
-0.0129043953,
0.0042013922,
-0.005208415,
0.0643444136,
-0.0241495371,
-0.0463693663,
-0.0833531693,
-0.0437585339,
0.0558010601,
-0.0058507062,
0.0111993365,
-0.030517349,
0.0833811164,
-0.0229937565,
-0.0519619435,
0.0106530897,
0.0066378065,
0.0005238057,
-0.0051084892,
0.0084954556,
0.0318133347,
-0.0277550519,
-0.027643932,
0.0462072901,
0.053982392,
-0.0068652714,
-0.0244578794,
0.0117056919,
0.0130647924,
0.0128056286,
0.0219334662,
-0.0149577446,
0.0075002331,
-0.0337522514,
0.0020643903,
-0.0646119192,
0.0134270107,
0.0183154587,
2.398e-7,
0.0566789247,
0.0065505346,
0.06243065,
-0.0248792432,
0.0059167384,
-0.0081324,
0.0103176348,
0.0260335375,
-0.0444732755,
-0.0099653769,
0.0131590385,
-0.0133168111,
0.0550786257,
0.0079983491,
-0.0486421064,
0.0155010708,
-0.0371059105,
-0.0280220918,
-0.0688976794,
-0.035548456,
0.0208867937,
0.0273111835,
-0.0287663992,
-0.0071840738,
-0.0135527272,
-0.0254727695,
0.0556558482,
-0.032625284,
0.0505288541,
0.0406450555,
0.0092013646,
0.0266459882,
-0.0552373566,
0.0541319251,
-0.0133639462,
0.0442829095,
0.0825974271,
0.0563878119,
0.0000828097,
0.0230047796,
-0.0175490957,
-0.0427602902,
0.0150394691,
-0.0350329801,
0.0747057348,
-0.0340270624,
-0.0011530609,
0.0186401065,
-0.0551600829,
-0.0246694498,
0.0030943954,
0.017208606,
0.0156512242,
0.0545462742,
0.0380218588,
-0.0349335372,
-0.0146213863,
-0.0840512142,
0.0115043093,
0.0151852183,
0.0141261015,
-0.0049191085,
-0.0388952605,
0.006062049,
-0.0105246911,
-0.0608541481,
-0.0412443578,
2.339129525e-34,
-0.0035544771,
0.0034387256,
-0.0157864299,
0.0309237074,
0.0461611301,
-0.0359235518,
-0.0265981965,
-0.0344538018,
-0.0370480791,
-0.0730795413,
0.0464209281
] |
http://arxiv.org/abs/2402.02370v2 | http://arxiv.org/pdf/2402.02370v2 | AutoTimes: Autoregressive Time Series Forecasters via Large Language Models | 2024-02-04T06:59:21 | 2024-05-23T01:58:28 | Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long | cs.LG | cs.LG, cs.CL | Foundation models of time series have not been fully developed due to the
limited availability of time series corpora and the underexploration of
scalable pre-training. Based on the similar sequential formulation of time
series and natural language, increasing research demonstrates the feasibility
of leveraging large language models (LLM) for time series. Nevertheless, the
inherent autoregressive property and decoder-only architecture of LLMs have not
been fully considered, resulting in insufficient utilization of LLM abilities.
To further exploit the general-purpose token transition and multi-step
generation ability of large language models, we propose AutoTimes to repurpose
LLMs as autoregressive time series forecasters, which independently projects
time series segments into the embedding space and autoregressively generates
future predictions with arbitrary lengths. Compatible with any decoder-only
LLMs, the consequent forecaster exhibits the flexibility of the lookback length
and scalability of the LLM size. Further, we formulate time series as prompts,
extending the context for prediction beyond the lookback window, termed
in-context forecasting. By adopting textual timestamps as position embeddings,
AutoTimes integrates multimodality for multivariate scenarios. Empirically,
AutoTimes achieves state-of-the-art with 0.1% trainable parameters and over 5
times training/inference speedup compared to advanced LLM-based forecasters. | AutoTimes: Autoregressive Time Series Forecasters via Large Language Models: Foundation models of time series have not been fully developed due to the
limited availability of time series corpora and the underexploration of
scalable pre-training. Based on the similar sequential formulation of time
series and natural language, increasing research demonstrates the feasibility
of leveraging large language models (LLM) for time series. Nevertheless, the
inherent autoregressive property and decoder-only architecture of LLMs have not
been fully considered, resulting in insufficient utilization of LLM abilities.
To further exploit the general-purpose token transition and multi-step
generation ability of large language models, we propose AutoTimes to repurpose
LLMs as autoregressive time series forecasters, which independently projects
time series segments into the embedding space and autoregressively generates
future predictions with arbitrary lengths. Compatible with any decoder-only
LLMs, the consequent forecaster exhibits the flexibility of the lookback length
and scalability of the LLM size. Further, we formulate time series as prompts,
extending the context for prediction beyond the lookback window, termed
in-context forecasting. By adopting textual timestamps as position embeddings,
AutoTimes integrates multimodality for multivariate scenarios. Empirically,
AutoTimes achieves state-of-the-art with 0.1% trainable parameters and over 5
times training/inference speedup compared to advanced LLM-based forecasters. | [
-0.0136733642,
0.1179824844,
0.0073750359,
0.005735355,
-0.0125545226,
0.0221264847,
0.0532637127,
-0.0207772721,
-0.0179108419,
-0.0132405777,
0.0211162362,
-0.0328562111,
-0.0111898081,
0.0568954237,
0.0049506063,
-0.068554461,
0.0461743958,
-0.0351875313,
-0.0026337302,
0.0029392303,
0.0040499265,
-0.0197567735,
0.0185189564,
0.0479209013,
-0.041116301,
-0.0050080419,
-0.0547113977,
-0.0367295071,
0.0306023266,
-0.0608280599,
-0.0138669787,
-0.0041215266,
0.0683807805,
0.0577878356,
0.00000199,
0.0046147835,
-0.0166771486,
-0.0196426362,
-0.0189239401,
-0.0057392251,
0.016778037,
-0.0666048005,
0.035322167,
0.017057877,
-0.0722493082,
-0.0262785982,
0.0387083963,
0.0244593136,
0.0790471286,
0.08081761,
-0.0158045683,
-0.0359571129,
0.0367071591,
-0.0478885472,
-0.0118572451,
-0.0584339947,
-0.0021089581,
-0.047853671,
-0.0130314808,
0.0138373757,
-0.0307061579,
0.0472193286,
0.0141812777,
0.0085145449,
0.0029643038,
0.0442197323,
0.029706534,
0.0005531373,
-0.0629432127,
0.0179360621,
0.0231569987,
-0.0308629852,
-0.0015829306,
0.0135985073,
-0.04466369,
0.0301789083,
-0.072774604,
-0.0048056217,
0.0442847908,
0.0511659719,
0.0413746275,
0.0353214629,
0.0227127727,
0.0171489511,
-0.0342295021,
0.0844302028,
-0.0028770105,
-0.0299485438,
-0.0483357795,
0.0155353518,
0.0321299806,
-0.0373449698,
0.0041831383,
0.0413010791,
-0.0202389732,
0.0000748651,
-0.02354284,
-0.0189228449,
-0.0451271534,
-0.0097440761,
0.0440059751,
0.0514054,
0.0002631801,
0.0154742096,
-0.005771508,
0.0284774303,
-0.0108712371,
-0.0946707204,
-0.0393001251,
-0.0260846391,
-0.0361407958,
-0.0434046574,
-0.0505594239,
-0.0340292752,
-0.0166571699,
-0.0111784544,
-0.0466848463,
0.0220024548,
-0.0123979729,
0.0066209254,
0.0315254405,
0.0117879473,
-0.0469640456,
-0.0301571935,
-0.0203535929,
0.0263041742,
-0.0270856582,
0.0138000539,
0.0011130497,
-0.0329160728,
-0.0065333773,
0.0108924871,
0.0049655167,
0.0077480483,
-0.0162096396,
0.0654214844,
0.0422338136,
0.0174217187,
-0.0595169961,
-0.0342644267,
-0.0275869593,
-0.0273141153,
0.0089886691,
-0.0059696613,
0.0149753196,
0.0132175712,
0.0316414721,
0.015265095,
-0.0013951325,
0.0026196353,
-0.0135105159,
0.0735744685,
0.0129607199,
-0.0018177012,
-0.0094139026,
-0.0384592712,
0.0440735742,
0.0739451945,
0.0487017967,
0.0141676208,
0.0411603115,
0.0004455182,
-0.0260866098,
0.0352731086,
0.0003139299,
-0.0157285817,
-0.0601377599,
-0.0004118929,
-0.0232213959,
0.0156443603,
0.0314678848,
0.0024229109,
0.0048488751,
-0.0198203549,
0.041240003,
0.0537636802,
0.0288618077,
0.0209519658,
0.0830133632,
-0.0362369977,
-0.0147590814,
-0.0100395912,
0.0059122378,
0.0432631373,
-0.0118520735,
0.0260959491,
0.0050619394,
0.0317657515,
-0.0288866609,
-0.0336842947,
-0.0043132082,
-0.0246274862,
0.0085168397,
-0.0405312963,
0.0414467975,
-0.0345815718,
-0.0454438664,
0.0467888266,
0.0141007062,
-0.0126692355,
0.0018386554,
-0.0066342358,
0.0130531378,
0.0945698991,
-0.0006130135,
-0.0021870397,
0.0195916723,
-0.0234382767,
0.0424050502,
0.0367775746,
0.0104189534,
-0.0032686733,
-0.0159982704,
-0.0548406877,
-0.0396273024,
0.0219322611,
0.0166058466,
0.0347465612,
-0.0436153896,
0.0638488457,
0.0257063601,
0.0057962299,
0.0206313916,
0.0010544127,
-0.0398934036,
-0.0557267815,
-0.0019315641,
-0.0038342434,
0.0411142819,
0.0034258186,
0.0109154694,
0.044471968,
0.0282298625,
-0.0598660037,
-0.0248470809,
0.0569855534,
-0.1143068373,
-0.052618362,
0.0050694617,
0.0094070518,
0.0278779566,
-0.0114658559,
-0.0468025096,
-0.0686341375,
0.0661878139,
-0.0271309689,
0.0988110155,
-0.0546820685,
-0.0148809077,
0.0467209406,
0.0018688109,
0.0138473706,
0.0227183253,
-0.0148598012,
-0.0195763037,
0.0024952374,
0.0309295412,
-0.0202195141,
-0.0049931151,
-0.0101613309,
0.0081377402,
0.0257929061,
-0.0043923347,
0.020501975,
-0.0301893223,
-0.0086892359,
-0.0038863067,
-0.0787004158,
-0.0443720035,
-0.032270845,
0.016174484,
-0.006281672,
-0.0203517526,
0.0237564556,
-0.0138172889,
-0.0108443368,
0.0089468146,
-0.0112628667,
0.0026275078,
0.0288240984,
-0.0064942203,
-0.0075442935,
-0.0095416317,
-0.0171078872,
-0.0279278755,
0.0792779997,
0.0039946022,
-0.0232483223,
-0.0312759578,
-0.0335403681,
-0.0076726666,
-0.0035731699,
-0.0216164365,
-0.0154053578,
-0.0105888061,
-0.0360902622,
0.0469293222,
0.006226914,
0.0212065112,
0.0407433026,
-0.0071242205,
-0.0710942298,
-0.0358946361,
0.0272574201,
-0.0242724139,
0.0298576839,
-0.0398057625,
0.007713831,
-0.0036064282,
0.0837537497,
-0.0066981656,
0.0113494899,
-0.0601611882,
0.010574257,
-0.0472048074,
-0.0113510517,
-0.0314807184,
-0.0164336488,
-0.0577122606,
0.0161683131,
0.0021478713,
-0.0805834234,
0.031384863,
-0.0385196283,
-0.026654046,
0.0266086254,
-0.0245420653,
-0.0660598502,
0.0234344658,
0.0166106932,
0.0005414741,
-0.0085935201,
-0.0136621464,
0.0013060495,
0.015240076,
-0.0094163315,
-0.0139679536,
0.0139069259,
0.0111209042,
0.0085775657,
0.0268058125,
-0.001308347,
-0.0299329981,
-0.034708105,
-0.0672293007,
0.0169593859,
-0.0260568634,
-0.0208758768,
-0.0349417366,
0.0561851673,
-0.0425896831,
0.0183941014,
-0.0115804337,
-0.0113604125,
0.0254285503,
0.0329067521,
0.0482071303,
-0.008233686,
-0.0156878456,
-0.0248528104,
0.0462623462,
-0.0160772409,
0.0657746643,
0.0041340301,
-0.0630133376,
0.0353508927,
0.0182032175,
-0.0263395496,
-0.0211203545,
0.0160601642,
-0.0131805334,
-0.0640422255,
-0.0102770021,
0.122082375,
0.0051253419,
-0.023662338,
0.0090059871,
0.0097789597,
0.0083788363,
-0.0038581162,
-0.0216590837,
-0.0727954954,
-0.0393196307,
0.0269588567,
0.0005618128,
-0.0006289707,
-0.0426787846,
-0.0222915821,
-0.0785792544,
0.0501677841,
0.0356917754,
-0.0280122869,
0.0039385534,
-0.0762163997,
-0.0243837591,
-0.0208273567,
0.0238760188,
0.0012118736,
-0.0246553142,
0.0315257125,
0.0492995158,
-0.0999529213,
-0.0367559157,
-0.0155487657,
0.0142664351,
0.0130136209,
0.0694180429,
0.0357256904,
-0.0083010113,
-0.0240763687,
-0.0426116809,
-0.0211949982,
0.042372264,
-0.0250200368,
-0.00309638,
-0.0291441847,
-0.0145466244,
-0.007930195,
0.0666746795,
-0.0150878029,
0.0092942184,
0.0054073832,
0.0991006717,
-0.0133896545,
-0.0305656586,
-0.0100735044,
-0.012475593,
0.0287926812,
0.0656110793,
0.0118587939,
-0.0368415527,
-0.0481946394,
-0.0099341059,
0.0174176302,
-0.0469085425,
-0.0283121765,
-0.0275704544,
0.0097967573,
-0.0277963653,
-0.0495434999,
-0.0491704829,
0.0090774419,
0.0562584065,
-0.008535306,
0.0383167192,
0.017911952,
-0.0177274756,
-0.0512827151,
-0.0108781606,
-0.0165961832,
0.0408495367,
-0.0437315665,
0.0198539458,
0.060382165,
0.0224452093,
-0.0223680362,
0.024922993,
-0.0972731486,
-0.0591658987,
0.0244896952,
0.0017356202,
0.0171117187,
0.029096894,
0.0510205738,
0.0194993,
0.1227043346,
0.0533869565,
-0.0074688122,
0.0089856088,
0.0548440553,
0.0607224517,
0.0237784274,
0.0370330811,
-0.0087977946,
0.0181384496,
-0.0494866893,
0.0365337133,
-0.0158180781,
-0.0283460878,
-0.0105356025,
-0.0192439388,
0.0496200621,
-0.0049750498,
0.021435041,
-0.0001585922,
-0.0120098144,
-0.0244857576,
0.012695346,
-0.0548591018,
0.0451243185,
0.0397742204,
0.0017891023,
-0.0275709964,
0.0282089282,
0.0405269153,
-0.0279529784,
-0.0172592551,
0.0532868616,
0.0036977334,
0.0270752143,
0.017832635,
0.01848769,
0.0206541326,
-0.0210014749,
0.0055251978,
-0.0136471977,
-0.01160016,
0.025136061,
-0.0046098335,
0.0533385538,
-0.0235772934,
-0.0101435324,
-0.0814415514,
-0.0146708665,
-0.0153277302,
0.01022927,
0.0286711436,
-0.0346230641,
0.0390794799,
-0.0608657002,
0.0177721083,
0.0190281235,
0.0426782034,
0.0014790029,
-0.01765633,
0.0417403914,
-0.008944015,
-0.0726950392,
0.0129954945,
0.0537839346,
0.0254439674,
0.0594449714,
0.0259210654,
-0.0285149794,
-0.0167609602,
-0.0142825842,
0.0178568102,
-0.0260200333,
0.0056216586,
-0.010583262,
-0.0438751169,
-0.0570034496,
0.0264014006,
-0.040794678,
-0.0276933592,
-0.0239366982,
0.0016542096,
0.0265911389,
0.0061059152,
0.0429087691,
-0.0166972019,
0.0197438449,
0.0050644721,
-0.049177859,
-0.0531525575,
-6.395052202e-33,
-0.0125842262,
-0.0153314741,
0.0252191853,
-0.0589216761,
-0.027788965,
0.0547493622,
-0.0081054512,
-0.0302214827,
0.0293621998,
0.0103820022,
-0.0156460106,
-0.0088566644,
0.0304329563,
0.0236582328,
0.0731462911,
0.0119972834,
0.0703655481,
-0.0399489999,
0.0289381836,
0.0108581129,
0.0697093755,
0.0491866097,
0.0582899675,
0.0117990766,
0.0145332934,
-0.0039761993,
-0.01737535,
0.0081122005,
0.0798249617,
0.0392634273,
-0.029561298,
0.0427809842,
0.0291849207,
-0.0341101177,
-0.0611114316,
0.0615952313,
-0.053171128,
-0.0637293458,
-0.0022815992,
-0.0014274641,
-0.0218923148,
-0.0009831763,
-0.0208698642,
-0.0161544289,
-0.0283722524,
0.0281784274,
0.0019021666,
-0.0231711008,
-0.0153968297,
-0.0456585959,
-0.0166023225,
0.0269483328,
-0.0231068842,
0.0077865636,
0.0616492927,
0.0664664283,
0.0236902796,
0.0141935321,
-0.0441406481,
0.0318318345,
-0.0120378742,
0.0129977055,
0.0361810103,
0.036704205,
0.0165310968,
0.0476719402,
0.0065797428,
-0.0640184358,
-0.0301520433,
0.0512362085,
0.0139953773,
0.0246006548,
0.0101414742,
-0.0016508099,
0.0509667732,
-0.0426216647,
0.003663898,
0.017026769,
0.0447191149,
0.0574119203,
0.048671186,
0.0286320169,
-0.0219664406,
-0.0310659818,
-0.0162012056,
-0.0244928673,
-0.0042785988,
-0.0241915397,
0.0404691696,
0.0381478854,
0.0359079428,
0.0299070124,
0.0615572967,
-0.0280030742,
-0.0370555706,
0.0533455461,
0.0113171376,
-0.0235114954,
0.0155767556,
0.0377418622,
-0.0618701391,
-0.0314946175,
-0.0032208352,
0.0165756103,
-0.0004270097,
-0.0510965511,
-0.0171354581,
0.069502309,
-0.0722501948,
-0.0235806312,
-0.0252705514,
-0.0067328983,
0.0167777948,
-0.0159585513,
0.0717675835,
-0.0408097692,
-0.0211937707,
0.0448669121,
-0.0199722089,
0.0014139112,
0.0328542888,
-0.0191157646,
0.0151018314,
0.0228623953,
-0.0192162506,
-0.0217282344,
-0.0159709956,
-0.0018145093,
0.0608948804,
-0.0471256338,
-0.0056674331,
0.0305739362,
2.735e-7,
0.0090680085,
0.055012349,
0.0326252207,
0.00772329,
0.0374560878,
0.0204065442,
-0.0365679935,
0.0271904841,
-0.0172818396,
0.0284643658,
0.0429672338,
-0.0080785472,
0.0727495551,
0.0213692989,
-0.045692578,
-0.0234934147,
-0.1180021092,
0.0104609923,
-0.0625815019,
-0.0322274677,
0.032926444,
0.1082506627,
0.0263313334,
0.0123673184,
0.0500723012,
-0.0587339066,
0.0473931469,
-0.0146804769,
-0.0159847941,
-0.0103400648,
0.010246696,
0.0356844068,
-0.0174636543,
-0.0388011485,
-0.0236060172,
0.0282258466,
-0.0286941752,
0.014512361,
0.0152981766,
-0.0230819639,
0.0092475172,
-0.0169262104,
-0.0358139649,
-0.0222679581,
0.0570492037,
-0.0438819453,
-0.0283875261,
-0.0624793284,
-0.0575931482,
0.044693131,
0.0145959025,
0.0123730283,
-0.0051144543,
0.0457795933,
-0.00210085,
-0.0124157732,
-0.0204956457,
-0.0719317943,
-0.02342779,
-0.0020784349,
-0.0486687571,
-0.004859637,
-0.0706624165,
0.0028036784,
0.0285761561,
-0.047096543,
0.0036884351,
2.487942977e-34,
0.0686982796,
-0.0167294014,
-0.0164257009,
0.0426685847,
0.0125447614,
-0.0235788561,
-0.0220863149,
0.0502206199,
0.0298875645,
-0.0837768242,
-0.0529491082
] |
http://arxiv.org/abs/2402.02369v1 | http://arxiv.org/pdf/2402.02369v1 | M$^3$Face: A Unified Multi-Modal Multilingual Framework for Human Face Generation and Editing | 2024-02-04T06:56:23 | 2024-02-04T06:56:23 | Mohammadreza Mofayezi, Reza Alipour, Mohammad Ali Kakavand, Ehsaneddin Asgari | cs.CV | cs.CV, cs.CL, cs.MM | Human face generation and editing represent an essential task in the era of
computer vision and the digital world. Recent studies have shown remarkable
progress in multi-modal face generation and editing, for instance, using face
segmentation to guide image generation. However, it may be challenging for some
users to create these conditioning modalities manually. Thus, we introduce
M3Face, a unified multi-modal multilingual framework for controllable face
generation and editing. This framework enables users to utilize only text input
to generate controlling modalities automatically, for instance, semantic
segmentation or facial landmarks, and subsequently generate face images. We
conduct extensive qualitative and quantitative experiments to showcase our
frameworks face generation and editing capabilities. Additionally, we propose
the M3CelebA Dataset, a large-scale multi-modal and multilingual face dataset
containing high-quality images, semantic segmentations, facial landmarks, and
different captions for each image in multiple languages. The code and the
dataset will be released upon publication. | M$^3$Face: A Unified Multi-Modal Multilingual Framework for Human Face Generation and Editing: Human face generation and editing represent an essential task in the era of
computer vision and the digital world. Recent studies have shown remarkable
progress in multi-modal face generation and editing, for instance, using face
segmentation to guide image generation. However, it may be challenging for some
users to create these conditioning modalities manually. Thus, we introduce
M3Face, a unified multi-modal multilingual framework for controllable face
generation and editing. This framework enables users to utilize only text input
to generate controlling modalities automatically, for instance, semantic
segmentation or facial landmarks, and subsequently generate face images. We
conduct extensive qualitative and quantitative experiments to showcase our
frameworks face generation and editing capabilities. Additionally, we propose
the M3CelebA Dataset, a large-scale multi-modal and multilingual face dataset
containing high-quality images, semantic segmentations, facial landmarks, and
different captions for each image in multiple languages. The code and the
dataset will be released upon publication. | [
-0.0095208809,
0.0494456515,
0.0075333761,
0.0155293969,
-0.0201050006,
0.0063423696,
0.0749106109,
0.0044796648,
-0.0393671431,
0.0051451577,
-0.019655738,
-0.0499266386,
-0.0007870213,
0.0816280842,
0.028643271,
-0.1152236089,
-0.0308418553,
-0.0237575211,
-0.0320805013,
-0.0104816146,
0.0411626548,
0.0182886664,
-0.0104323504,
0.0206548478,
-0.0863941535,
0.0075860559,
-0.0340838656,
0.0668534338,
-0.0027719825,
-0.0166664459,
-0.0257657655,
0.080646731,
0.0014109106,
0.0076141465,
0.0000021785,
-0.0249890462,
-0.0150481146,
-0.0085222712,
0.0213106889,
0.0034279402,
0.095914945,
-0.0429756492,
-0.045547273,
-0.0692447424,
-0.0311694853,
0.0106495004,
0.0906619728,
0.0665661246,
0.0876667127,
-0.0119081046,
-0.0269476809,
-0.0327400975,
-0.045533441,
-0.0193722583,
0.0505107418,
0.0013124658,
0.0100035295,
0.0042648399,
-0.0376107283,
-0.0487393439,
-0.0208847709,
0.0398568474,
0.0118300356,
-0.0342562683,
0.0287671629,
0.0241638962,
-0.0213298853,
-0.0854547992,
-0.0274040587,
0.0601692908,
-0.0086891549,
0.0278926641,
0.0007910296,
0.0000492213,
-0.0001477959,
0.0193524342,
-0.0130011085,
0.0076496429,
-0.0237942319,
-0.0007800062,
-0.0718217045,
-0.0272669997,
0.022415949,
-0.0253398046,
-0.0039881817,
-0.0330190323,
0.0443223268,
0.0131008914,
-0.056211926,
0.01044467,
-0.0182860065,
-0.0720146894,
0.0448567197,
-0.0109939901,
-0.0098089119,
-0.0190788396,
-0.0778405219,
-0.0377031937,
0.0083967159,
0.015763972,
0.0039648991,
0.0022389845,
-0.0363828801,
0.0473661795,
-0.0170577969,
0.0072165956,
-0.0153390458,
0.0016537242,
-0.0209011119,
-0.040939115,
-0.0022515445,
-0.0029725395,
-0.0006730143,
-0.0497913174,
-0.0344157182,
0.0412456952,
-0.0911179781,
0.0268255565,
-0.0310903527,
0.0104214391,
-0.0049652741,
0.0405584425,
-0.0590136684,
0.0118694538,
-0.0555096865,
-0.0138763189,
-0.0218874849,
0.0364340171,
-0.0127792181,
-0.0716028959,
0.0111849848,
-0.0060550137,
0.0458123945,
0.0208014846,
0.0161501989,
0.0435949266,
-0.0512767062,
0.0302759688,
-0.0110042756,
-0.0193779077,
-0.0153529244,
-0.034007933,
-0.0018005829,
-0.068641983,
0.0132357655,
-0.0448685624,
-0.0137919663,
-0.0066375798,
-0.0043802233,
0.012585043,
0.0663614646,
0.0621830113,
0.0068276483,
0.0064857793,
0.0402976908,
0.0131361531,
-0.0492429584,
0.0214273408,
0.0177888777,
0.0364705473,
0.0391506068,
-0.0066102985,
0.0286214631,
0.0386027955,
0.0203361064,
0.0297414735,
-0.0291149393,
-0.0190343838,
0.0472796001,
0.0108318022,
0.0040316172,
0.0210264232,
-0.0496213473,
-0.0068706949,
0.0630055293,
0.0261802562,
0.0414894708,
0.0396683887,
0.0205729473,
-0.0124912299,
0.0144105991,
-0.0377181284,
-0.0155760003,
0.0732050687,
0.0011659686,
0.0050252783,
-0.0428559035,
0.0551864468,
-0.0647762492,
-0.0100853704,
0.0581516959,
-0.0186792705,
0.0505753867,
-0.0263714865,
0.0527021997,
-0.0038939114,
0.0197336394,
-0.0096344398,
-0.0491637886,
-0.0974365547,
0.0262957364,
-0.0496008806,
-0.0081341052,
0.0433128625,
-0.0362061448,
0.00527112,
-0.0240623672,
-0.0000872154,
0.038243223,
0.0469917729,
0.0099651413,
-0.0075314986,
0.0839090124,
-0.0533518419,
-0.0195981432,
-0.0310937334,
0.0059684548,
0.0287723057,
-0.0170719158,
0.000029409,
-0.0074631078,
0.0220557749,
0.0875362381,
0.0385914966,
-0.0067586196,
0.0278112013,
0.009700642,
0.0460047871,
0.0024773018,
-0.0088069513,
0.0005230134,
-0.0197246093,
0.0174155831,
-0.0101881297,
0.0163814258,
0.041981671,
-0.0099538192,
0.0274624806,
0.040410284,
-0.056476064,
-0.046276506,
-0.033872012,
0.0030310748,
-0.0438685603,
0.0415887199,
0.0223534219,
0.0020297801,
-0.057183601,
-0.0169716403,
-0.0222507603,
-0.0084344335,
-0.0057949945,
0.0560837947,
-0.0092678424,
-0.0270170141,
0.0585538,
0.0096796844,
-0.0504527465,
-0.0204236209,
-0.0367616341,
-0.003452766,
-0.0074777473,
0.0119885905,
-0.0068950821,
0.0515121631,
-0.048297748,
-0.0148615185,
-0.0780698955,
-0.0255206209,
0.0029403318,
-0.0275750235,
0.025449479,
0.0376067981,
-0.0003357582,
-0.0546953678,
-0.0569302402,
0.0175276324,
0.0081056776,
-0.0284969117,
0.0168893449,
0.0118088052,
-0.0065501011,
-0.0369386002,
0.0131090684,
0.0372679047,
0.0162821077,
0.0424007587,
-0.0197640192,
-0.0142332101,
-0.0134338457,
-0.0434659943,
0.0468664803,
-0.0210882649,
-0.0173688792,
0.0057120444,
-0.0132149179,
0.065758571,
0.0570609234,
-0.0012247935,
-0.0091004325,
0.0165669415,
-0.0251260027,
0.004159831,
-0.0053358404,
0.0128689203,
0.0220117737,
-0.0187599901,
-0.0605534166,
-0.0166179743,
0.0069003659,
0.027813362,
0.0298561268,
-0.0419817232,
0.0567261539,
0.0343810394,
-0.0491918437,
-0.0126687586,
0.0467863157,
-0.0382719226,
0.0393251702,
-0.0643158853,
-0.0618892536,
0.0517181084,
-0.0180477407,
-0.0442055091,
0.0845287293,
-0.0035360672,
0.0150585659,
0.0595157035,
-0.0106461514,
0.0360588208,
-0.0006669509,
-0.0130919656,
-0.0395145305,
0.0252349284,
0.0424551666,
-0.0295587629,
0.0190650765,
-0.0606748797,
-0.035742607,
-0.0354006402,
-0.0122205075,
0.0093071992,
-0.0009927199,
-0.0278930366,
-0.0389915928,
0.0043916204,
-0.0249761511,
0.0111699719,
-0.0137514332,
0.0659866929,
0.0246407017,
-0.0279045328,
-0.0098736649,
0.0136393774,
-0.0309163854,
0.0202119313,
0.058591146,
-0.0573325902,
-0.0353301279,
0.0170750692,
0.0592990182,
0.0363795683,
-0.0403932668,
-0.0355034061,
0.0615831465,
0.0144058783,
0.0725798309,
-0.0271108672,
0.0107505526,
0.0571569428,
0.0018518383,
0.0454543903,
0.0185158141,
-0.010314866,
-0.0296714548,
-0.0407357961,
-0.0123436097,
0.0196168907,
0.0663198158,
-0.007943837,
-0.0725159347,
-0.028346708,
0.0198169611,
0.0055819373,
0.0051131034,
-0.0064704344,
-0.0022684191,
0.020844711,
0.0726321638,
0.0087539973,
-0.0664786473,
-0.0072260513,
0.027112199,
0.0573137216,
-0.0102650477,
-0.0667459816,
0.0162102673,
-0.0162897632,
-0.0019581039,
0.0026477193,
0.1029666141,
-0.0337488651,
0.0018335575,
-0.0477726385,
-0.0066187587,
0.0423624441,
0.0260977801,
0.0487432852,
-0.0681500584,
0.0051443805,
0.0264902748,
-0.0014536537,
-0.0303444974,
0.0401351117,
0.0406356975,
-0.0244844146,
-0.0367607176,
0.0761831179,
0.044064939,
0.0062720366,
0.0168987699,
0.0020019356,
0.0208422467,
0.0511948541,
-0.0223529153,
-0.0274624694,
-0.0133893136,
-0.0081623141,
0.004426443,
-0.035773728,
-0.0354446992,
-0.0339926444,
0.0283676051,
-0.0792354941,
-0.0471219122,
0.0293943994,
-0.0370098762,
-0.0904118344,
-0.0382954814,
0.022008162,
0.0705197379,
0.0529402532,
0.0022123633,
-0.0446100347,
0.0130522382,
0.0288316421,
-0.038706243,
-0.0077043283,
-0.0028926902,
0.0232470427,
-0.0162310675,
-0.0191485733,
0.0024578115,
-0.0742250383,
0.0205062944,
-0.0173895583,
-0.0161218271,
-0.0489911735,
0.0037959502,
0.078386806,
0.0126455799,
0.0201658346,
0.0197838116,
-0.0780184418,
0.0144277392,
0.0099246465,
0.0193864256,
0.0079592457,
0.0261543281,
0.0216784887,
0.0007051065,
0.0116453441,
-0.0397442728,
-0.0154387336,
0.0095412694,
-0.0144844223,
-0.0260855276,
-0.0238318797,
-0.0124521656,
0.0120057296,
0.0022690198,
-0.0317567252,
0.0273320824,
-0.0582472458,
0.0068225334,
-0.0393692404,
-0.0246632788,
-0.0669897497,
0.0084133269,
-0.0089152493,
0.0189800467,
0.0322198495,
-0.0104711847,
-0.0111155026,
0.0467388071,
-0.026857568,
0.0476457216,
0.0151444133,
0.0322021469,
0.0325037017,
0.0594663657,
0.0360571146,
0.0511345901,
-0.0033233857,
-0.0069464282,
-0.0142017612,
0.0281796306,
-0.0428227559,
-0.0446461886,
-0.0357972719,
0.0498063229,
-0.027765194,
0.0153453434,
-0.028798908,
-0.0027011775,
0.0452828817,
0.0357027575,
0.075021565,
-0.0083942777,
-0.0142246764,
-0.0154995089,
0.0750277564,
0.0052387952,
-0.0665155128,
0.0430704467,
-0.029315507,
-0.0214492604,
-0.0360525697,
0.0386326164,
-0.0533281267,
0.0064015822,
-0.018572554,
-0.033551231,
-0.0415467806,
0.0240355488,
0.0613917112,
-0.0289177243,
-0.0628302991,
-0.0331966467,
-0.0286728721,
-0.0047491114,
0.0070910798,
0.0304831713,
0.0002243788,
0.0261787977,
0.042691227,
0.0580727607,
-0.0231676586,
0.0037233061,
-0.0291412938,
0.0404146761,
-0.0256242324,
-0.0089392625,
0.0190566499,
-6.105806389e-33,
-0.0211192779,
0.0103689479,
0.0109358123,
-0.044372905,
-0.0274142828,
-0.0059905401,
-0.019121984,
0.0597218275,
0.0057494338,
-0.01345341,
0.0081345402,
-0.0400102213,
0.0016981367,
0.0373605005,
0.0032133379,
0.0098511679,
0.038700927,
-0.0119920513,
-0.0056262286,
0.0260297395,
0.0188367944,
-0.0068300534,
0.0468817428,
-0.0732221827,
-0.039071653,
0.0054747784,
-0.0503880344,
-0.0079638241,
-0.0208032411,
0.0473070778,
0.0072489679,
0.016131876,
0.0098258602,
-0.001402554,
0.0121891424,
0.0017291543,
-0.0311388541,
-0.0145141678,
-0.0711090118,
0.0267570075,
-0.0517323278,
-0.0357662812,
0.0266771987,
-0.0244889706,
-0.0063305837,
0.0056613516,
0.0459713899,
-0.0217302684,
-0.0041855006,
-0.006041477,
-0.0134378467,
0.0142747834,
0.0062053553,
-0.0296228509,
0.0324350223,
0.0583381243,
0.0224392693,
-0.0208565108,
-0.0244415794,
0.0374472737,
-0.0200831387,
-0.054748293,
0.0638125613,
0.0541789606,
0.0558315888,
0.0350522846,
0.0542150103,
-0.0549927019,
-0.0120561412,
0.0168390274,
0.0381441861,
0.0827333555,
-0.0133503927,
0.1067424193,
-0.0166441984,
-0.0351827629,
-0.0138700148,
0.036161609,
-0.0617613122,
0.0167396478,
-0.0202248972,
0.040303167,
0.0502074733,
0.0275736246,
-0.0220716838,
-0.0158496071,
-0.0544112884,
-0.0172817241,
0.0083269961,
0.026936356,
-0.0555467978,
-0.0123822317,
0.0179755408,
0.0353534333,
0.0331095941,
-0.0271898843,
-0.057251744,
0.064395979,
0.0058328123,
-0.0309267715,
-0.0027333978,
-0.0160487164,
0.05722069,
0.0050809984,
0.0437201038,
0.0159182288,
-0.0351149328,
0.019764673,
-0.0171734504,
0.0083791288,
-0.0336599499,
0.0327662006,
-0.0057140985,
0.0625734702,
0.0136421444,
-0.0278912168,
-0.0552105494,
-0.0311419852,
0.0103032142,
0.0167720262,
-0.0597651936,
-0.0670452267,
0.0174337495,
0.042598933,
-0.0505307168,
0.0450499244,
-0.0699715465,
-0.0340783373,
0.0810413882,
-0.0374636315,
-0.0330061056,
0.0077612605,
2.78e-7,
-0.0254655704,
0.0403461456,
-0.0067923204,
-0.0077957353,
0.0270491615,
0.0909932852,
-0.0356021486,
-0.0107319821,
-0.0025149419,
-0.0104419636,
-0.0030281125,
-0.0011420327,
0.0324485414,
0.0664775893,
-0.0074844547,
0.0346954055,
0.0435985513,
0.0283245314,
0.0110134464,
0.042634856,
0.0518742427,
0.0338129774,
0.0016844349,
-0.0141318701,
0.0750465244,
-0.0095248958,
0.0535963327,
-0.0391747616,
0.0081782183,
-0.0482385345,
-0.0103637706,
0.0225556772,
0.0079924418,
-0.0072150622,
-0.0266395845,
-0.0496836081,
0.0307323914,
0.0439296924,
-0.0316348262,
-0.0000871984,
0.0216280036,
0.0482629985,
-0.0223899856,
-0.0225211401,
0.025119625,
0.0312929042,
0.003064703,
-0.0469225347,
-0.0279790759,
0.0525349975,
0.025451038,
-0.0518706366,
-0.050346788,
0.0293680392,
0.0315340571,
-0.0101789832,
-0.0202299077,
0.0016266572,
-0.0235479958,
-0.0536407717,
-0.0357522927,
-0.0328756757,
-0.0841820985,
0.0174081381,
0.0101589514,
0.0253489744,
-0.0135231856,
1.990240276e-34,
-0.0350031182,
0.0031317831,
0.0278443787,
-0.0473013856,
0.0209460873,
0.0153870666,
-0.0241279267,
0.0062734084,
0.0015194075,
-0.0092040217,
-0.0511894785
] |
http://arxiv.org/abs/2402.02364v1 | http://arxiv.org/pdf/2402.02364v1 | The Developmental Landscape of In-Context Learning | 2024-02-04T06:23:05 | 2024-02-04T06:23:05 | Jesse Hoogland, George Wang, Matthew Farrugia-Roberts, Liam Carroll, Susan Wei, Daniel Murfet | cs.LG | cs.LG, cs.AI, cs.CL | We show that in-context learning emerges in transformers in discrete
developmental stages, when they are trained on either language modeling or
linear regression tasks. We introduce two methods for detecting the milestones
that separate these stages, by probing the geometry of the population loss in
both parameter space and function space. We study the stages revealed by these
new methods using a range of behavioral and structural metrics to establish
their validity. | The Developmental Landscape of In-Context Learning: We show that in-context learning emerges in transformers in discrete
developmental stages, when they are trained on either language modeling or
linear regression tasks. We introduce two methods for detecting the milestones
that separate these stages, by probing the geometry of the population loss in
both parameter space and function space. We study the stages revealed by these
new methods using a range of behavioral and structural metrics to establish
their validity. | [
-0.0371610522,
0.0462926663,
-0.0357318036,
0.0079236133,
0.0175549779,
0.0316186734,
0.012051411,
-0.0156382378,
-0.0889662877,
0.0234917253,
-0.0445457622,
-0.0375754386,
0.0008764869,
-0.0157126747,
0.0428931043,
-0.0416659415,
0.0406986885,
-0.0082022576,
-0.0446804278,
-0.0111554284,
-0.0397174843,
-0.0156361163,
-0.0212452449,
0.046641767,
-0.0138294632,
-0.0166978166,
0.0029501126,
-0.0380159058,
0.0051057879,
-0.0335603356,
0.0337461382,
-0.0085465303,
0.0485506542,
0.0576859228,
0.0000016385,
-0.030300837,
0.0048123705,
-0.0048137447,
-0.051100824,
-0.0476150848,
0.0370183028,
-0.0035003524,
-0.0383392796,
0.0207317956,
-0.0122740315,
-0.0098114852,
0.074877359,
0.0361533761,
0.0193941537,
0.0189705491,
-0.0140306028,
-0.0182217807,
0.0129343756,
-0.0124808447,
0.0099619795,
0.007400902,
0.0028875303,
-0.0414309092,
-0.0320989713,
-0.0348750241,
-0.0452016294,
0.0121316575,
0.0306977537,
0.0164794754,
-0.0015123663,
0.0421448462,
0.0393297337,
-0.0358852595,
-0.0198780969,
-0.0070286701,
-0.0088720955,
-0.0081207324,
-0.0121997315,
0.0256518703,
0.014976983,
-0.0170530006,
-0.0294365007,
0.0247663781,
0.0127210431,
0.0050378195,
-0.0429920815,
-0.0020612576,
0.0264256801,
-0.0396956429,
-0.0042661228,
0.0525796413,
0.0202851165,
-0.0283614118,
-0.0053959261,
0.0056440975,
0.00653265,
-0.011681539,
0.01299066,
0.0104555534,
-0.0147768045,
0.0498083979,
-0.0121307187,
-0.0132621052,
-0.0080433739,
-0.0080400305,
0.0572257265,
0.0147586623,
-0.0380060077,
-0.0035460272,
-0.0198463313,
0.0694994107,
-0.0219497196,
0.0235784966,
-0.003377412,
0.0223107543,
-0.0402217694,
-0.0555817336,
-0.0172208324,
-0.0183824711,
0.018629387,
0.0036991567,
-0.0236849282,
0.0672578588,
0.009470487,
-0.0131034367,
-0.0136263343,
0.0017011085,
-0.0918918028,
0.0245700348,
-0.0434031561,
0.0569667779,
0.0033373323,
0.0140487384,
0.0036981204,
0.0047184052,
-0.0256036296,
-0.0104665151,
0.0205484591,
-0.0035267291,
-0.0105072958,
0.0882895961,
-0.0135701932,
-0.0374739841,
-0.0453098081,
0.0424929,
-0.0113763018,
-0.0503495671,
0.0320205949,
-0.0273473356,
-0.0324814171,
0.0394674949,
-0.0175975412,
-0.0158296097,
-0.0190644227,
-0.0024509267,
0.0737620145,
0.0486254394,
-0.0048887446,
-0.0059741321,
0.0290560629,
-0.0139374612,
0.0099328477,
0.0666603297,
0.002363669,
-0.0492441319,
-0.0369624607,
-0.0037735405,
0.0016446074,
0.0418986678,
0.0198854487,
0.0190809108,
-0.0288506355,
-0.0445565581,
0.0205579158,
0.0081066778,
0.0589331873,
-0.0080014523,
-0.045177985,
0.0397507809,
0.0815467611,
0.0830422416,
0.052577842,
0.0320692509,
0.0111455964,
-0.0255568828,
0.0060121072,
0.0114533454,
-0.0710600764,
0.0245977789,
0.0177007373,
0.0043723602,
-0.0708552375,
0.0231612567,
-0.0240150224,
-0.0197002646,
0.0111530144,
-0.0010710252,
-0.0347229503,
-0.0372046232,
0.0773802847,
0.0484479405,
-0.0412707254,
0.0624119341,
0.0101257646,
-0.0340449587,
0.0296998322,
-0.0497492738,
-0.0061272425,
-0.0309366807,
-0.0246520713,
-0.046348989,
-0.0188212637,
-0.0139319375,
0.0057140025,
0.0428172573,
-0.033831615,
-0.0194334257,
0.0250241496,
0.0085658301,
-0.0243336335,
0.0103769349,
0.0174266659,
-0.0292987917,
-0.0163058639,
-0.052784849,
-0.0274660587,
-0.0730230585,
0.0078103826,
0.0656501129,
-0.0687415674,
-0.0033846935,
-0.0372368209,
0.016206637,
0.0229871739,
-0.0418846868,
0.0052853273,
-0.0069426391,
0.0279311892,
0.0216029081,
-0.0062165731,
0.0596280247,
-0.0659573451,
-0.017139975,
0.0317639969,
-0.0393025987,
-0.0170253105,
0.0462016277,
0.0181208905,
-0.0190641098,
0.0181858558,
-0.0938595086,
0.0975769609,
-0.0202733297,
0.0512218177,
-0.0639923736,
-0.0174903348,
-0.0575852618,
0.0735090449,
-0.0001418612,
-0.0071756826,
0.0096136406,
-0.0075485059,
-0.0025381611,
-0.0090739569,
-0.0101246973,
-0.0242632423,
-0.045156572,
0.029282229,
-0.0081523992,
-0.0284476597,
-0.0440793335,
-0.0094803916,
-0.02241401,
-0.0097609647,
0.000737951,
0.0435832329,
0.0191345643,
-0.056508068,
0.0083883498,
0.0004310644,
0.0252170023,
0.0328707919,
0.0031952537,
0.0266782008,
0.0447867252,
0.0265819896,
-0.0463850386,
-0.0208618976,
-0.0158057436,
-0.0200678259,
0.0369483866,
-0.0205718651,
0.0109586511,
-0.0763415173,
0.0384542048,
0.0330024362,
0.0080447113,
-0.0216548964,
-0.0123029193,
0.0566264354,
0.0297246817,
0.0823293924,
0.0773347467,
0.0284175109,
0.1014866605,
0.0390218459,
0.0200545657,
-0.0070632296,
-0.0136178564,
-0.0791973099,
-0.0348809436,
-0.0663626269,
0.0145731959,
0.0076127192,
0.0608798005,
0.0010332968,
-0.001416327,
-0.0460903309,
-0.0121485861,
-0.0240870062,
0.0065133567,
-0.0116614839,
-0.0104250414,
-0.0377636738,
0.0299896114,
-0.0110235997,
-0.0477484427,
0.0299892612,
0.0238682386,
-0.0114275496,
-0.0483431183,
-0.0011019773,
0.0279668309,
0.0218492243,
0.0120731238,
0.0041622734,
0.0447096154,
-0.0676686168,
-0.1137467474,
0.0360928625,
0.0462245755,
0.0151412422,
0.0831584558,
-0.0359034203,
-0.0225161966,
-0.0055012414,
-0.0118215028,
-0.0353845283,
0.0146407057,
0.0183509253,
0.0088727521,
0.0156170558,
-0.0602312572,
-0.0779943243,
-0.0220809542,
0.0262577254,
0.0577197038,
0.0089245941,
-0.0528133661,
-0.0140430937,
-0.0081766965,
0.0151839163,
0.0322075523,
0.0730382353,
-0.1037053391,
0.0125923455,
0.0143236397,
0.0302509014,
0.0080822501,
-0.0524891801,
0.016208468,
0.0110579701,
0.0135560678,
0.0175856352,
0.0536113344,
0.0248654969,
-0.0212618597,
-0.054355789,
0.0260046404,
-0.0057955468,
-0.050679747,
0.0182261374,
-0.0251932014,
0.0263837744,
0.0556096174,
0.0123668499,
0.0042023938,
0.0133408708,
0.0150442887,
0.0286467988,
0.0022955372,
0.0124711068,
0.0342005603,
-0.0636104047,
0.0123741962,
-0.0299650822,
-0.0735520422,
-0.0176343918,
0.0041445275,
0.0213731509,
-0.0044782064,
0.037596263,
0.0129406024,
-0.0659858361,
-0.0336981304,
0.0466586426,
0.0125774592,
-0.0624438114,
-0.0042874445,
-0.0395934246,
0.0440197997,
0.0695521384,
0.0146332616,
-0.0209453814,
-0.0182705689,
0.05100004,
-0.0248748884,
-0.0037841236,
0.0221607052,
0.0486473851,
0.0342692696,
-0.0080041112,
-0.0355380699,
0.0437992774,
0.0011630873,
-0.0125398925,
-0.0489670783,
0.0462597162,
-0.003348955,
-0.0409747995,
0.0164443254,
-0.0164993964,
0.0141396942,
0.0119152302,
0.0203586332,
-0.0634976178,
-0.0228089374,
-0.0328399949,
0.0392468609,
0.0553452484,
-0.0476686247,
0.0314715803,
0.0127981054,
-0.0641699135,
-0.0497315191,
0.0527621135,
0.0034220554,
0.0480229296,
0.0240897499,
-0.0490220562,
0.0101213669,
-0.0085760988,
-0.0280230176,
0.0389943346,
0.0230105072,
-0.0174402837,
-0.0827693343,
0.0089669954,
0.0312470105,
-0.0494791418,
-0.042163346,
-0.0232804678,
-0.1095902622,
-0.0422942042,
0.047004845,
0.0211807545,
0.0275515728,
0.0068092882,
0.0239536632,
-0.0497346632,
0.1084442288,
0.0436300747,
0.0501182042,
0.0293415952,
0.0718077198,
-0.0055286256,
-0.0325229242,
0.0067935623,
-0.0186458491,
-0.0053124484,
0.0106887668,
0.006887198,
-0.0215799306,
0.0670664534,
0.040728353,
-0.0874105319,
0.0749904066,
-0.0260442942,
-0.0361432731,
-0.0468482561,
0.0010666853,
0.0337049775,
0.0222531632,
-0.0062407814,
-0.0840207189,
0.0354932621,
0.0656068027,
0.0198443551,
0.0209763311,
0.0588553809,
0.0070377402,
-0.0031243297,
0.0017233036,
0.0417126715,
0.0104206568,
0.0291339196,
0.0765719041,
0.0255878642,
-0.0074718902,
-0.0268399566,
0.0176376719,
0.0263767187,
0.0366465524,
0.0491128862,
0.0153388102,
0.0060802139,
0.0065990146,
-0.0544003621,
0.0211294536,
-0.0494652204,
-0.0577777512,
0.0182534177,
-0.0281394254,
0.0197325796,
-0.0472151414,
0.0020996477,
-0.0106090456,
0.0055165691,
-0.0276493244,
-0.0106799453,
0.0217721146,
0.0110453777,
0.0326231681,
-0.024771845,
0.0491016097,
0.009133284,
0.0695715249,
-0.0711478442,
0.031387303,
-0.009811121,
-0.0112277977,
0.0184874646,
-0.0244542696,
-0.0197448675,
-0.02177836,
-0.0246632937,
-0.055765029,
0.0077015664,
-0.008736169,
0.0031772491,
-0.059377335,
0.0340952873,
0.0273962505,
-0.0272537582,
-0.0332321115,
0.0156755317,
-0.0454502366,
0.041098386,
0.0371954069,
-0.0306204986,
-5.584705335e-33,
-0.0277159382,
-0.0883693472,
0.020527998,
0.0185566675,
-0.0148047153,
-0.0005932231,
-0.0387377031,
0.039772518,
-0.027381083,
0.0050202049,
0.0178910885,
0.0366811827,
0.0060002925,
0.0313894823,
0.0108863302,
-0.0341522396,
0.025199974,
-0.0223102942,
-0.0117807128,
-0.0168722719,
0.0301011149,
0.0123826312,
0.0660115033,
-0.041255407,
-0.0361038335,
0.006001852,
-0.0209947303,
0.0288200658,
-0.0385892466,
0.0127337324,
-0.0168276522,
0.0379399508,
-0.0033112695,
-0.0648547336,
-0.0312082805,
0.0149342744,
-0.0664850548,
0.0020801397,
-0.0187314469,
-0.0003885553,
-0.0237700939,
-0.0008874856,
0.049830474,
-0.0117650675,
-0.0058230804,
-0.0499885418,
0.0536178984,
-0.030379314,
-0.0016038709,
-0.0233423673,
-0.0874866694,
-0.0127147464,
-0.0144225871,
-0.0303182583,
0.0086571183,
0.0529961176,
-0.0034913844,
0.0039194967,
-0.0728188902,
0.0177093446,
-0.0181104206,
0.0157546289,
0.0621526539,
0.0747735426,
0.0270217005,
0.0419211909,
0.031393718,
0.0422348492,
0.0129309976,
-0.0865399763,
0.0978339389,
0.025521066,
0.0113001503,
-0.0209736954,
0.0177838188,
0.002597776,
-0.0174162444,
0.0078142928,
-0.0173204131,
0.0509233773,
0.0151744802,
0.0130621744,
0.0452348813,
0.0018744818,
-0.0289528146,
-0.0413958132,
0.0037243762,
0.0119127631,
0.0074091638,
0.0141592147,
-0.0081023248,
-0.0143412622,
-0.0010535198,
-0.0154666109,
-0.0011065456,
-0.0090209804,
-0.0196822807,
0.049132403,
0.0107362363,
-0.0015461653,
-0.0500712842,
-0.0456887595,
-0.0203550998,
0.0686142445,
0.0405337587,
-0.0096610673,
0.0520729013,
0.0391234159,
-0.0403243713,
0.0098002097,
-0.0200529303,
0.0112905586,
0.0343538672,
0.0275633018,
0.0208162498,
-0.0263908394,
-0.0097831758,
-0.0178604927,
0.0137602361,
-0.0184214246,
-0.056483198,
0.09624625,
-0.0261661299,
0.002499056,
-0.0075934366,
-0.0220197123,
-0.0589247122,
0.0620381534,
0.053082101,
-0.0530080348,
-0.0252361968,
-0.0083314106,
2.41e-7,
0.0451639742,
0.0428816862,
0.0397401676,
0.0113823451,
0.0442389064,
0.0227492582,
-0.0294230543,
-0.012230685,
0.0388167687,
-0.0011996874,
0.0178875886,
-0.0144143961,
0.0554494597,
0.0181643125,
-0.0363275632,
-0.0484392047,
-0.0751195475,
0.0248055756,
-0.0263024531,
0.0281710923,
-0.0062817251,
0.038244728,
-0.0128675727,
-0.0267514642,
-0.0027222929,
0.0159791913,
0.0510503389,
-0.0147522427,
0.0842250809,
-0.0048268707,
-0.0722541586,
0.0029345546,
0.0060575088,
0.0276735146,
-0.0084546739,
0.0417378061,
-0.0375574678,
0.0003300066,
0.0230395757,
-0.05305814,
0.0170716345,
0.0619827919,
-0.0088500762,
-0.0083134267,
-0.0087824836,
-0.1060831994,
-0.0172109585,
-0.0322663411,
-0.0008654931,
-0.0149778714,
0.0188245401,
-0.0158343203,
-0.0086957244,
0.0046110624,
-0.0195723455,
0.0035228422,
0.0056186295,
-0.0601322502,
-0.0185294077,
0.0359495543,
-0.0531724654,
-0.0644390061,
-0.0303813349,
0.0088008428,
0.0368749723,
0.0242253449,
0.0197033975,
1.915062935e-34,
-0.0222552083,
-0.0183200873,
0.0459148288,
0.0055988636,
0.0043328721,
-0.0101587968,
0.0422005504,
0.0029831594,
0.0449750647,
-0.0333327167,
0.0217622221
] |
http://arxiv.org/abs/2402.05952v1 | http://arxiv.org/pdf/2402.05952v1 | Advancing Graph Representation Learning with Large Language Models: A Comprehensive Survey of Techniques | 2024-02-04T05:51:14 | 2024-02-04T05:51:14 | Qiheng Mao, Zemin Liu, Chenghao Liu, Zhuo Li, Jianling Sun | cs.LG | cs.LG, cs.AI, cs.CL | The integration of Large Language Models (LLMs) with Graph Representation
Learning (GRL) marks a significant evolution in analyzing complex data
structures. This collaboration harnesses the sophisticated linguistic
capabilities of LLMs to improve the contextual understanding and adaptability
of graph models, thereby broadening the scope and potential of GRL. Despite a
growing body of research dedicated to integrating LLMs into the graph domain, a
comprehensive review that deeply analyzes the core components and operations
within these models is notably lacking. Our survey fills this gap by proposing
a novel taxonomy that breaks down these models into primary components and
operation techniques from a novel technical perspective. We further dissect
recent literature into two primary components including knowledge extractors
and organizers, and two operation techniques including integration and training
stratigies, shedding light on effective model design and training strategies.
Additionally, we identify and explore potential future research avenues in this
nascent yet underexplored field, proposing paths for continued progress. | Advancing Graph Representation Learning with Large Language Models: A Comprehensive Survey of Techniques: The integration of Large Language Models (LLMs) with Graph Representation
Learning (GRL) marks a significant evolution in analyzing complex data
structures. This collaboration harnesses the sophisticated linguistic
capabilities of LLMs to improve the contextual understanding and adaptability
of graph models, thereby broadening the scope and potential of GRL. Despite a
growing body of research dedicated to integrating LLMs into the graph domain, a
comprehensive review that deeply analyzes the core components and operations
within these models is notably lacking. Our survey fills this gap by proposing
a novel taxonomy that breaks down these models into primary components and
operation techniques from a novel technical perspective. We further dissect
recent literature into two primary components including knowledge extractors
and organizers, and two operation techniques including integration and training
stratigies, shedding light on effective model design and training strategies.
Additionally, we identify and explore potential future research avenues in this
nascent yet underexplored field, proposing paths for continued progress. | [
0.0327694677,
0.0269102659,
-0.0056805448,
0.0443791375,
-0.0274986941,
0.0188968182,
-0.0241676793,
-0.0235622115,
0.0758073926,
-0.0615169331,
-0.0150590902,
-0.0353235416,
-0.0171561558,
0.0287418868,
0.041024588,
-0.0099968081,
-0.007057278,
-0.0277940091,
0.0168234296,
-0.0300585125,
-0.0133355316,
0.0289771389,
0.0629779696,
0.0037844914,
-0.0080279056,
-0.0065945094,
0.0120418156,
0.0068856352,
-0.0310932938,
-0.0265265629,
-0.044607617,
0.0109275989,
0.0650915354,
0.0412698947,
0.0000017079,
-0.0281593129,
-0.0377134718,
-0.0004778135,
0.0005170194,
-0.0190754868,
0.0576969907,
-0.0664494932,
0.0172173176,
-0.0100924727,
-0.0423823595,
-0.0304089487,
0.0646877736,
0.0163098183,
0.0421645083,
0.058949478,
-0.0244747065,
-0.0335600413,
0.0364509858,
-0.014530031,
-0.0240295026,
-0.0251293648,
0.0602454916,
0.05255685,
0.0096208332,
-0.0040196585,
-0.0635659397,
0.023671627,
-0.018410055,
-0.0247532371,
0.0426896624,
0.0424386151,
0.015639171,
0.0231983997,
-0.0097981226,
0.085625276,
0.0010684115,
-0.0127583975,
-0.0000576028,
0.0074604349,
0.0063934238,
0.0373028405,
-0.0675543621,
-0.013893716,
0.0183394104,
0.0327186435,
-0.0116197066,
-0.0538515374,
0.0870889947,
-0.0144793019,
-0.0275537204,
0.0590216219,
-0.0026589141,
-0.0182259716,
0.0068350863,
-0.0140666431,
-0.0344628878,
-0.0081403544,
-0.0271317232,
0.0295093395,
-0.003972311,
0.0256501008,
0.0122651188,
-0.001319148,
-0.0138383107,
-0.0287039243,
0.0496541001,
-0.0387416482,
0.0215766151,
0.0600070134,
-0.0354129449,
0.048721198,
-0.036887344,
0.0401679315,
-0.0476826988,
-0.0295489859,
-0.0587008968,
-0.0076873475,
-0.0155879194,
0.0252130292,
0.0359258913,
-0.0495057516,
-0.0192843582,
0.0428224243,
-0.038580332,
0.0324783847,
0.0079086879,
-0.0048795589,
0.0303311646,
-0.0251916982,
0.0104797166,
0.0212442856,
0.0364671722,
0.0150352847,
-0.0467059463,
-0.0029096638,
0.0125798462,
0.0101481751,
0.0361167043,
-0.0410158262,
0.0049785096,
0.0706024468,
0.0428026021,
-0.0108447531,
0.0051972182,
0.0133949947,
-0.0251910742,
-0.059745606,
0.0198360514,
-0.0967512354,
0.0388601944,
0.0373227149,
0.0106691709,
-0.0429769382,
0.0184522271,
-0.0404947996,
-0.0625102744,
0.0626292974,
-0.0087567242,
0.0316629671,
-0.024230605,
-0.054821264,
-0.0052129482,
0.0481920913,
0.0209943708,
0.0083803833,
0.0129746292,
-0.0208771098,
-0.0475884937,
0.041605588,
0.0154734161,
0.0241630226,
-0.0096937055,
0.0040021758,
0.0314858668,
-0.0049403706,
0.006489309,
0.0717533678,
-0.0290663652,
-0.0258016996,
0.0322767086,
0.0534473769,
-0.0002889608,
0.0061920704,
0.0482937098,
0.0071165408,
-0.0260277037,
-0.0059577134,
-0.0571818501,
-0.0264601354,
-0.0040493412,
0.0114964256,
-0.0246929638,
0.0050746216,
-0.0030654976,
0.0074773189,
-0.0100884307,
-0.0293904841,
0.0717260987,
0.0076619084,
0.0882277489,
-0.0103301015,
-0.01394442,
0.0138176484,
0.0376615413,
-0.0123942392,
0.0374126434,
0.0113781057,
-0.033188723,
0.0603615083,
0.0262576342,
-0.0435885526,
0.0052091992,
0.0007057346,
-0.0052465233,
-0.0334647074,
0.0347991697,
-0.0529600605,
-0.0320978127,
-0.0285042487,
-0.0279315077,
-0.0137243625,
-0.0104895523,
0.0209179372,
0.0052789268,
-0.0179870334,
0.010926703,
-0.0754732564,
-0.0613487028,
-0.0031563144,
0.0047867489,
-0.00042654,
-0.0167546533,
-0.0024229321,
-0.0039811567,
0.0173702762,
-0.0032274548,
0.0095228553,
0.0267873183,
-0.026008293,
-0.0153798675,
-0.0134162288,
-0.0048283017,
-0.0251740478,
0.0828119591,
0.0035156268,
-0.0206076037,
0.0039014751,
-0.0448495522,
-0.0240672268,
0.0443475135,
-0.0566157587,
0.0632841364,
-0.0011885933,
0.0099489884,
-0.0101399403,
-0.01902554,
-0.022012325,
0.0719343871,
0.0017824068,
-0.0106777623,
0.0128114223,
-0.0042736894,
-0.0075845523,
-0.038288001,
-0.016428547,
0.04395134,
0.0151328547,
0.0418991446,
0.0399980135,
-0.0363039859,
-0.0375598483,
-0.0002754224,
-0.0840703323,
0.0012195964,
-0.0040714894,
0.0172956977,
-0.0065142848,
-0.0050145318,
0.0183797255,
-0.0139806578,
-0.0620638765,
0.0155919045,
0.0214924384,
0.0338229351,
0.0048326752,
-0.0109936642,
-0.0462120287,
-0.0134111075,
-0.0619770028,
-0.0576904044,
0.0959332138,
0.0059268298,
-0.0622637048,
-0.0028928348,
0.029226847,
-0.0434973016,
0.0294261761,
-0.0266505703,
0.0303993281,
-0.0245275777,
-0.0366998985,
0.0391363874,
0.0333038718,
-0.035158325,
0.0449431874,
0.0622229539,
-0.0196354873,
-0.0116065452,
-0.0188848488,
-0.0616593733,
-0.0089327963,
-0.0634193644,
-0.017534636,
-0.0309262071,
0.002524761,
0.0044733877,
-0.0195239857,
-0.0390651859,
-0.0125051178,
-0.0484802164,
0.0058854083,
-0.02176922,
-0.0269326363,
0.0206742864,
-0.0151008731,
-0.0536449403,
-0.0455123298,
0.0627953261,
-0.0172539279,
0.0131193474,
0.043808762,
-0.0340293758,
-0.0823567733,
0.0692386478,
0.0643572435,
0.0425375961,
-0.0041063963,
-0.020258354,
-0.0339766219,
0.011915978,
0.0205069892,
-0.0377875492,
0.0380362123,
0.0162003711,
0.0531499833,
0.019761445,
-0.0253113452,
0.0000359605,
-0.0084851114,
-0.0379183367,
0.0528172813,
0.0259999279,
-0.0373503789,
0.0029500755,
0.0286372881,
-0.015096793,
-0.0301201139,
-0.0026832905,
0.0158381518,
-0.0332695358,
0.0319276601,
0.0380531624,
-0.0365872532,
0.032196112,
-0.0331693627,
0.0013346171,
0.0372900665,
0.0208435282,
-0.0388416983,
-0.0406119078,
0.019122418,
0.0132272802,
-0.0095429812,
-0.0084732743,
-0.0141761974,
0.023488624,
-0.0076946653,
0.0148198474,
0.0762512088,
0.0156676099,
-0.0489263795,
0.0078215143,
0.0202363841,
0.0262309704,
0.0320937522,
0.014213156,
-0.0801636651,
0.0106653562,
-0.0175235905,
-0.0021710086,
0.0360651761,
-0.0606468916,
-0.0110874837,
-0.103262417,
0.0244779419,
0.0698375702,
-0.0036418841,
-0.053544797,
-0.0615476854,
0.0056359512,
-0.0072958739,
-0.0088180834,
-0.0243581273,
-0.0558426827,
-0.0260037947,
0.0185078997,
-0.0342912525,
-0.0281190593,
0.0173271373,
-0.0239414517,
0.0026095763,
0.0013909488,
0.0357528403,
0.0115628866,
0.0336600356,
0.0157558434,
0.022090042,
-0.0333068185,
-0.011630876,
0.0177310612,
-0.0109935831,
0.0012662136,
-0.0204311814,
0.016744433,
-0.0295120999,
-0.015573117,
-0.0520052537,
0.1256402135,
0.0142216133,
-0.0256139636,
0.0175882988,
-0.0335565247,
-0.0022383572,
0.0117490385,
0.0464610122,
0.0025203568,
-0.0419122577,
-0.0489904284,
0.0132864062,
0.0054730182,
-0.0401789807,
0.0345679112,
-0.0403297581,
0.0396616571,
-0.0220711343,
-0.0061178594,
0.0326614603,
-0.0282066129,
0.0153815998,
0.0361733064,
0.0525946245,
-0.0581189208,
-0.0659020916,
0.010692806,
-0.0523846336,
0.0498000979,
-0.0007995368,
-0.0338353924,
0.0189909078,
-0.0012936246,
-0.0236642342,
0.0088871,
-0.0728272572,
-0.102086097,
-0.0389403738,
-0.0058777961,
0.007437577,
0.0379229672,
-0.0148427784,
-0.038289886,
0.0371660441,
0.0280640386,
-0.0079460004,
-0.0076074288,
0.0149445152,
-0.0091123981,
-0.0080894213,
0.0527217612,
-0.0678305477,
0.0141945165,
0.0332314186,
0.0254910327,
-0.0394904502,
-0.0313424878,
0.0444486216,
0.0798682794,
0.0156819895,
0.0311368704,
-0.0028434857,
-0.050274618,
-0.0414827466,
-0.033377517,
-0.0578237958,
-0.0066682277,
-0.0259694941,
0.0422067642,
0.0013199294,
0.0059583937,
-0.030161893,
0.0456655584,
0.0468114167,
-0.0418861508,
0.011081622,
0.0133592244,
-0.0735170022,
0.0616589561,
0.0169373751,
-0.0029428129,
0.0210460722,
0.0029514441,
0.0051619578,
0.0108367763,
-0.0272355285,
0.0069698873,
-0.0201823488,
-0.0694283396,
-0.006685033,
-0.0353195369,
-0.0440281965,
0.0137766367,
-0.0021764366,
0.0718957186,
-0.0006908009,
0.0599864945,
-0.0142688984,
-0.0274709389,
-0.0291776005,
0.0227113664,
-0.0125627536,
-0.057282269,
0.0583387651,
0.0295073483,
-0.0071148034,
0.0305911452,
0.0396343432,
0.0741762072,
0.0203992091,
-0.0050994926,
-0.0406153463,
-0.0293639936,
-0.0301110465,
0.0001364166,
0.0110492054,
-0.0259378552,
0.0340511575,
-0.0593591295,
-0.0964988023,
0.0138043407,
-0.0261110514,
-0.0099476343,
0.0045282333,
0.0003359917,
0.0326357819,
0.019699486,
-0.0074017486,
-0.0678319633,
0.0375341028,
-0.0092335381,
-0.0152195562,
0.0152940247,
-5.624093212e-33,
-0.0061401762,
-0.0602484532,
0.0208364632,
0.0118647907,
-0.0454674363,
0.0327523388,
-0.0178478286,
0.0057915826,
-0.0215789862,
0.0165792685,
0.016338801,
-0.0144384746,
-0.019464355,
0.0108237695,
-0.0352637917,
0.0097089503,
0.0299108457,
0.026879305,
0.0216730256,
0.0445830636,
0.0485813431,
0.0180680193,
0.0361098424,
-0.0767118558,
0.0230519343,
-0.0221369974,
0.016234478,
0.0403748862,
0.0492839366,
0.054586973,
-0.0213675816,
0.0443467759,
0.038035512,
0.0078655956,
-0.0113364654,
0.0311739445,
-0.0630531833,
0.0202135686,
0.0081346948,
0.0105436128,
-0.0052015483,
-0.0330827944,
0.0417554267,
-0.0067846468,
0.0320365876,
-0.0535090603,
0.0483323559,
-0.0043967743,
-0.0602946579,
0.030809056,
-0.0565646328,
0.0359889418,
-0.0067900768,
0.0216199551,
0.0653995872,
-0.0188535992,
0.0200001188,
0.0245226156,
-0.045987051,
0.0464914031,
-0.0191907026,
0.0012278372,
0.0347473212,
0.0571279898,
0.0354229882,
0.0472569205,
-0.0171127561,
-0.0069745919,
-0.0044260505,
0.0049274606,
0.0022589227,
0.0737184659,
-0.0147675602,
0.0088689458,
-0.048310671,
0.0130243171,
-0.0897376239,
-0.025409298,
0.011467889,
0.0196065735,
0.0071835862,
0.0322281644,
-0.0033333804,
-0.0632081181,
-0.0385820232,
-0.0113706132,
-0.0095891654,
-0.0169319864,
-0.0207200646,
-0.0114814518,
-0.0218132641,
0.0393621437,
0.0178747606,
-0.0863259509,
-0.0779879764,
0.0809301287,
-0.0017963025,
0.0529425628,
-0.008684502,
0.034039963,
-0.0770835653,
-0.051922109,
-0.0172946174,
-0.0597574562,
-0.0030367314,
0.0056724269,
0.0554877482,
0.0259788167,
-0.0221087802,
0.0034143222,
-0.0294737369,
-0.0471619926,
-0.0194850042,
0.0021579189,
0.0505818762,
0.0665708482,
0.0159134325,
0.0456997231,
0.0680226237,
0.0172837209,
0.0121260844,
0.0346108228,
-0.015015671,
0.0117465779,
-0.0184919443,
-0.0125927683,
-0.0414716713,
0.0306827296,
0.0581386983,
-0.04332624,
-0.0297748744,
-0.046176821,
2.482e-7,
0.071412757,
0.0531962179,
0.0333066396,
0.0354011618,
0.0302276853,
0.0203505084,
-0.0383029282,
0.0443395488,
0.0174649619,
-0.0057763397,
0.0469913967,
0.0716131851,
-0.0376254208,
0.0346653685,
-0.0418267734,
0.070249036,
-0.0690342039,
-0.0149536701,
-0.0377709977,
0.031049367,
0.0197521877,
0.0719171166,
0.0190536175,
-0.0026943737,
-0.0363485515,
-0.0810557529,
0.0939239264,
-0.0317030102,
-0.0224720426,
0.0193826742,
-0.0155998208,
0.0104058431,
-0.0187911931,
0.022209892,
0.0271277931,
0.0440895557,
0.031609226,
0.0240214057,
-0.0205975473,
0.0362400711,
0.0045777326,
-0.0337428674,
-0.0132252481,
-0.0301051997,
0.0582575053,
-0.0764712542,
-0.0743331388,
-0.0337834135,
-0.0186803285,
-0.0174653083,
-0.0057526631,
0.0086716143,
-0.0144581851,
0.0294194147,
0.0149613656,
-0.0198213831,
0.02837025,
-0.031567432,
0.0306642763,
0.0300870035,
-0.0170490686,
-0.0595154651,
0.0080507211,
0.0256982502,
0.0047254474,
-0.0052169398,
0.0619789213,
2.350292129e-34,
0.0441021509,
0.0153725976,
-0.0085534696,
0.0195803531,
0.0362300575,
-0.0482581854,
-0.040917363,
0.0305940658,
0.0344352089,
-0.1253087819,
0.0056535043
] |
http://arxiv.org/abs/2402.05122v1 | http://arxiv.org/pdf/2402.05122v1 | History of generative Artificial Intelligence (AI) chatbots: past, present, and future development | 2024-02-04T05:01:38 | 2024-02-04T05:01:38 | Md. Al-Amin, Mohammad Shazed Ali, Abdus Salam, Arif Khan, Ashraf Ali, Ahsan Ullah, Md Nur Alam, Shamsul Kabir Chowdhury | cs.GL | cs.GL, cs.CL, cs.HC | This research provides an in-depth comprehensive review of the progress of
chatbot technology over time, from the initial basic systems relying on rules
to today's advanced conversational bots powered by artificial intelligence.
Spanning many decades, the paper explores the major milestones, innovations,
and paradigm shifts that have driven the evolution of chatbots. Looking back at
the very basic statistical model in 1906 via the early chatbots, such as ELIZA
and ALICE in the 1960s and 1970s, the study traces key innovations leading to
today's advanced conversational agents, such as ChatGPT and Google Bard. The
study synthesizes insights from academic literature and industry sources to
highlight crucial milestones, including the introduction of Turing tests,
influential projects such as CALO, and recent transformer-based models. Tracing
the path forward, the paper highlights how natural language processing and
machine learning have been integrated into modern chatbots for more
sophisticated capabilities. This chronological survey of the chatbot landscape
provides a holistic reference to understand the technological and historical
factors propelling conversational AI. By synthesizing learnings from this
historical analysis, the research offers important context about the
developmental trajectory of chatbots and their immense future potential across
various field of application which could be the potential take ways for the
respective research community and stakeholders. | History of generative Artificial Intelligence (AI) chatbots: past, present, and future development: This research provides an in-depth comprehensive review of the progress of
chatbot technology over time, from the initial basic systems relying on rules
to today's advanced conversational bots powered by artificial intelligence.
Spanning many decades, the paper explores the major milestones, innovations,
and paradigm shifts that have driven the evolution of chatbots. Looking back at
the very basic statistical model in 1906 via the early chatbots, such as ELIZA
and ALICE in the 1960s and 1970s, the study traces key innovations leading to
today's advanced conversational agents, such as ChatGPT and Google Bard. The
study synthesizes insights from academic literature and industry sources to
highlight crucial milestones, including the introduction of Turing tests,
influential projects such as CALO, and recent transformer-based models. Tracing
the path forward, the paper highlights how natural language processing and
machine learning have been integrated into modern chatbots for more
sophisticated capabilities. This chronological survey of the chatbot landscape
provides a holistic reference to understand the technological and historical
factors propelling conversational AI. By synthesizing learnings from this
historical analysis, the research offers important context about the
developmental trajectory of chatbots and their immense future potential across
various field of application which could be the potential take ways for the
respective research community and stakeholders. | [
0.1012767479,
-0.0032923846,
-0.0531870499,
-0.0310094655,
-0.0458823144,
-0.019664837,
0.0259582642,
0.019896375,
0.0606071092,
-0.0411509648,
0.0266621225,
0.0107542416,
-0.0331898369,
0.0584421754,
0.0053841174,
-0.0580433086,
0.0024868627,
0.0081390031,
0.00187168,
0.0219102371,
-0.0076143206,
0.0167059246,
-0.0269257594,
0.0843256712,
-0.0671313778,
-0.0042571593,
-0.0277873725,
-0.0162012838,
-0.0091952831,
-0.0009508787,
-0.007288693,
0.0158681031,
0.0109348986,
0.0487013124,
0.0000018375,
-0.0641253218,
0.0086806389,
0.0168583319,
-0.0503514335,
-0.076611042,
0.0501840971,
0.0389219336,
-0.0328190513,
0.0398322418,
-0.06885124,
0.0357033238,
0.0636419356,
0.019484967,
0.0258744266,
0.0747133568,
-0.0237439182,
-0.0079551628,
0.0599656627,
0.0128137479,
0.0137414131,
-0.0423388779,
-0.022603875,
0.013788674,
0.0391659215,
-0.0223146211,
-0.0297123641,
0.0213947371,
0.0333286338,
-0.0353140868,
-0.0150270304,
0.0399995595,
-0.0537138134,
-0.0767128766,
-0.0357881524,
0.0348408557,
0.0278076548,
-0.0034058555,
0.0114987297,
0.0175884441,
0.0251565073,
0.0057819639,
-0.003431346,
0.0206434056,
-0.0121160047,
-0.0211684816,
-0.0023974718,
-0.0200206619,
-0.025717767,
0.0014117085,
-0.0178663507,
0.0189534295,
0.0014748831,
-0.0243835617,
-0.0077773225,
0.0058744121,
0.0300261769,
-0.0680111423,
0.0025062771,
0.0560145453,
-0.0117668156,
0.004653478,
-0.0188493729,
0.0048276987,
-0.0118420599,
-0.0299652647,
-0.0327566676,
0.0353979021,
-0.0038247933,
0.0265245531,
0.0054543992,
0.0889473706,
-0.0024335966,
-0.0079393638,
-0.015926268,
-0.0210958086,
-0.0770324916,
-0.0320498273,
-0.0445014052,
0.0401123576,
0.0124894278,
-0.0225726739,
-0.0166870542,
0.0826626644,
0.0005600029,
-0.0100730574,
-0.0205915496,
0.0102990037,
-0.0192630701,
0.0236021336,
-0.053217236,
-0.0217384342,
-0.0351647809,
0.0015917179,
0.0288054869,
-0.0945383683,
0.0106805963,
0.0182076655,
0.0007671249,
0.000233764,
0.0058185672,
0.0884993225,
0.0023857455,
-0.0645637959,
0.0291591268,
-0.0102728223,
-0.011954559,
-0.070284307,
0.0182470046,
-0.0280591063,
0.0279363059,
0.0169873014,
0.0031927142,
-0.0037686874,
0.0079093464,
0.0069660023,
-0.0100984937,
0.0124794701,
-0.0121819563,
0.0150784263,
0.0436204374,
-0.0013528996,
-0.0253343154,
0.0465248078,
0.0489808843,
0.0122918338,
0.0245744884,
-0.0092130508,
0.082418099,
0.0116542028,
-0.0399528667,
-0.0611326359,
0.0305076595,
-0.0194519926,
-0.0319463313,
0.0218430087,
-0.021407431,
0.0570365228,
-0.0463470183,
0.0192323811,
0.0504460931,
0.0722583607,
-0.0052482067,
0.0674744621,
0.074991934,
0.0185223445,
0.0370008908,
-0.0122835301,
-0.0139387613,
0.0070809782,
-0.0289339684,
-0.0320894495,
-0.0239194855,
0.0094974022,
-0.024590455,
-0.0587341972,
0.0415693931,
-0.0272150952,
0.0465573668,
-0.0258912314,
0.043094635,
0.0101870038,
-0.0317599215,
-0.0258028433,
-0.0042443932,
-0.0482949167,
-0.0684362873,
-0.0977274552,
0.0043217987,
0.0921789557,
0.0662334114,
0.0005741174,
-0.0227121003,
-0.007027695,
-0.0339878798,
0.0692396313,
0.0077156345,
0.0211790465,
-0.0149167106,
-0.003145104,
0.00938424,
0.0059992019,
-0.0219874457,
-0.001863216,
-0.0143023329,
0.0198285002,
-0.0270513203,
0.0536059663,
0.0187293291,
0.0213542879,
-0.0082651414,
-0.0026527226,
-0.0290004704,
-0.0106700137,
0.0476494543,
0.0345897228,
-0.016672913,
0.0027605486,
0.0102793667,
0.0030107999,
-0.021907052,
0.0659862012,
0.0014664754,
0.005076915,
0.0245456062,
-0.0734542236,
-0.0102075692,
0.0431081168,
0.0071342704,
-0.0241678525,
0.0220236126,
-0.0894432515,
0.0593535081,
-0.0312678181,
0.030526232,
-0.0366269909,
0.0354589224,
-0.0078726728,
0.0417636558,
0.0202209912,
-0.0433650948,
0.0049055065,
0.0171665419,
0.000908667,
-0.0322873183,
-0.0056947521,
0.0363428034,
-0.002116679,
0.0148167098,
0.0430348031,
0.0041556675,
-0.0402239524,
-0.0317024104,
-0.0966454893,
-0.0321160667,
0.0247053951,
-0.0315884948,
-0.0019169063,
-0.0092780581,
0.0016804429,
-0.0106580229,
-0.0262384061,
0.058671277,
0.0151365995,
0.0017782315,
0.0687888786,
-0.0229838174,
-0.0795702189,
-0.048028294,
0.0141586382,
-0.0149686821,
-0.0043025957,
0.0794488713,
-0.0158297345,
-0.0428691246,
0.0268323775,
-0.028781129,
0.0454643592,
-0.0304905511,
-0.030406883,
0.0527257286,
-0.0035090717,
0.0400440618,
0.0472418815,
0.0561803803,
0.0893796384,
0.0107821794,
0.0176969655,
-0.0416904353,
0.035786096,
-0.0247300901,
0.0019662366,
-0.0162430946,
-0.0278854407,
0.0133477794,
0.0372612439,
-0.013175833,
0.0245280489,
-0.0410935581,
0.0632449389,
0.0308435913,
-0.0791581944,
0.0141720772,
-0.0046976921,
-0.0335896909,
0.0409112237,
-0.0032168459,
-0.044439543,
0.0650783256,
-0.0038912932,
-0.0174264833,
-0.0238615777,
-0.0373486653,
0.0129919611,
0.0699331388,
0.0093353335,
0.0012282648,
0.0246297922,
-0.0265455451,
-0.0474536717,
0.0004207258,
0.0187388361,
-0.0553959869,
-0.0290066563,
-0.01627459,
0.0120849116,
0.0194686949,
-0.0241915677,
-0.0533206612,
-0.035623014,
-0.0310070887,
0.0107163442,
0.0630265996,
-0.0473320819,
-0.0626820475,
-0.0000432595,
-0.0037740124,
0.0422583744,
0.0401650891,
-0.0317820087,
-0.0303005949,
-0.0508467518,
0.001891078,
0.0437912941,
0.0054341936,
0.0559329018,
0.0652673393,
0.038251996,
0.0119651593,
-0.0076182182,
-0.02643588,
-0.0308061205,
0.0115076629,
-0.0175835267,
-0.0178361069,
0.0216726623,
0.0130117815,
-0.0312822834,
-0.0017891739,
0.0750072077,
0.0087912548,
-0.019039724,
0.0044770218,
-0.0155802732,
-0.0400709882,
0.0025602812,
0.0609874465,
-0.0819261447,
0.0091473032,
0.0086399605,
0.0047919112,
-0.0660278946,
-0.0297779683,
0.0024224911,
-0.1066313013,
0.0551438294,
-0.0504745319,
-0.0474208854,
-0.0507978387,
-0.0202641934,
-0.0409083813,
0.0201709364,
0.0148953358,
-0.010829417,
0.0140062394,
-0.0322655961,
0.0035199854,
-0.0768190697,
0.0124685122,
0.036387071,
-0.0383258723,
0.0717415214,
0.0823228136,
0.0787998885,
0.0069674444,
0.0318715237,
-0.0217240639,
0.0717162192,
0.033941634,
0.0008927139,
0.0648501739,
-0.0000221465,
-0.0122290216,
-0.0408046879,
-0.0135362213,
0.0098153744,
0.0197691824,
0.0165457074,
0.029644778,
-0.0634835809,
-0.0263763499,
-0.0433383435,
0.0346738026,
0.0222076345,
0.0288520865,
-0.0044671269,
-0.0515616499,
-0.0597116426,
-0.0332783423,
0.0484897606,
-0.0028533025,
-0.0576545373,
-0.0402258523,
0.0085236328,
-0.0573539212,
-0.096981369,
-0.039015159,
0.0138487024,
0.0205918737,
0.0250470079,
-0.0316289067,
0.028287299,
0.0292850621,
-0.0106436322,
-0.0286094584,
-0.0037477312,
-0.0340193957,
-0.0414887667,
-0.0571002066,
0.0073712477,
-0.0860912874,
0.0033893841,
0.0160844345,
-0.0083127078,
-0.0265685804,
0.023209393,
0.0528481938,
0.0310490299,
-0.0020706865,
-0.0317151919,
-0.0145935845,
0.0219772998,
0.0542540252,
-0.0195691343,
-0.0020939168,
0.0384165198,
0.0387449563,
-0.0069707097,
0.0060417685,
-0.0429256298,
0.0059529818,
-0.0443833023,
-0.0088683767,
-0.0134879462,
-0.0344663598,
0.053055197,
0.0238593668,
-0.0007750546,
0.0025900961,
0.0208312124,
0.0139350444,
-0.0082653938,
-0.0588269532,
0.0172533095,
-0.0477378629,
-0.0309571065,
-0.0292720851,
0.0293902773,
0.0413574912,
0.0025539764,
-0.0591128357,
0.0213028416,
0.0137437144,
0.0179924499,
0.0364320539,
0.0704623908,
-0.0109696239,
-0.0056535015,
0.0088349096,
0.0401048139,
-0.0055601168,
0.0504491217,
0.0291195232,
0.0686807781,
-0.0145377498,
-0.0250006448,
-0.0112664057,
0.0428133309,
-0.1102825999,
-0.1045179144,
0.0385437012,
0.0656684637,
0.0253149699,
0.0487813763,
0.0230433811,
-0.0423362851,
0.0089850258,
-0.000723802,
0.0308148786,
0.0076900483,
-0.0063427412,
-0.0436496511,
0.0050415765,
-0.0122495834,
-0.017035624,
0.0347649865,
-0.0268444661,
0.0487959795,
-0.018655723,
-0.0312612206,
-0.0335645899,
-0.0481971279,
-0.0026319188,
-0.040810205,
0.0372787192,
-0.012366035,
-0.0301306248,
-0.0323486142,
-0.021137666,
0.0246385001,
0.012936784,
-0.0341555253,
0.0533907339,
0.0515915565,
-0.0036668279,
-0.0438707918,
-0.0515595712,
0.0083723906,
-0.026016036,
0.0194528867,
-0.0029557685,
-5.609592387e-33,
0.0074482169,
-0.0205803961,
0.0556490161,
0.0531655662,
-0.0275012217,
-0.0266714506,
-0.0335516743,
0.0173261557,
-0.0402063318,
0.0066902055,
-0.0623122528,
0.0289686918,
0.0074356953,
-0.0304687545,
0.0178530477,
-0.0156289879,
0.041063685,
-0.0234972406,
-0.0179932695,
0.0303354971,
0.0168119017,
-0.0033473943,
0.0125640342,
-0.0463110693,
0.0283751488,
-0.0002644983,
-0.0134348208,
-0.0142003698,
0.0035127557,
0.0141993137,
0.001607862,
0.0176600721,
0.0290649571,
-0.0184146706,
-0.0019651356,
0.0388530791,
-0.0268468391,
-0.0177549981,
-0.0239272583,
-0.003476117,
-0.0159628894,
-0.0251041111,
0.0150287468,
-0.0201067217,
-0.0415466279,
0.0172430389,
0.0333387852,
-0.0049308292,
0.0428661518,
-0.0002132988,
-0.0705379918,
-0.0078574866,
-0.0415450372,
-0.0639551878,
-0.0008528472,
0.0540608875,
-0.0050236192,
-0.0238028616,
-0.0086473618,
0.0718859881,
-0.0148542579,
-0.0559085384,
0.041428335,
0.0067096371,
0.0355080515,
0.0505290888,
0.0943030193,
0.0468376763,
-0.0479108952,
-0.0182422139,
0.012727215,
0.0483863764,
0.0268056709,
0.0177372508,
-0.0210884847,
-0.0309105311,
-0.0380427092,
0.0080333864,
0.086705476,
0.0205291715,
0.0175402798,
-0.0078921197,
-0.0931228697,
0.0065126554,
0.0178870726,
0.0081284009,
-0.0024244005,
-0.0433681495,
0.0049922974,
0.0008833222,
0.0248470232,
-0.049231682,
0.0226044673,
-0.0810859129,
0.0385443941,
-0.0039692721,
-0.035304796,
0.0331564434,
0.0084345564,
-0.0293275397,
-0.0232050102,
-0.0560289621,
0.0451977476,
0.0280452538,
0.0393739864,
0.0166805349,
-0.0438823104,
0.052177906,
0.0049491515,
-0.0115521718,
-0.0114993574,
0.033319097,
0.0075722574,
-0.0396808423,
-0.0059221392,
0.0323430859,
-0.024103431,
-0.0022306114,
0.0245455708,
0.0497621931,
-0.051958099,
0.0137982043,
-0.0066285012,
-0.0186198726,
-0.0335912742,
-0.0159154031,
-0.0542107709,
0.0455772392,
0.0182733033,
-0.0494999997,
-0.0286509432,
-0.0282713659,
2.66e-7,
0.001738538,
0.0445427783,
0.0471003242,
0.0100041572,
0.01070817,
0.0179254692,
0.0218680911,
0.0072048963,
0.0196786951,
0.0527422875,
0.0127975913,
0.0253734514,
0.0152423391,
0.0044241864,
-0.0736040846,
-0.0636457279,
-0.0417260863,
-0.031381499,
-0.0207546186,
0.0274007116,
0.0666849762,
0.0599629357,
0.0003288176,
-0.0364173502,
-0.0033538048,
-0.0139677403,
0.001316333,
-0.0475282818,
0.0140148848,
0.0270811506,
0.0094953487,
0.0366281271,
0.0083658462,
0.0388695449,
0.0218588002,
0.0243844669,
-0.004396583,
0.0156965815,
0.0076636868,
0.0013902833,
0.0343520269,
0.0271670148,
0.0059215599,
-0.0312293991,
0.0455129333,
-0.0404210947,
-0.0496456064,
-0.0260448456,
-0.0234019719,
0.0006997884,
-0.0035320609,
-0.0218053479,
-0.0165959075,
0.007328039,
-0.0234103743,
0.0086711012,
0.0148325143,
-0.0338172615,
0.0349394754,
-0.0230542496,
-0.0490537547,
-0.0918397233,
-0.0163192283,
-0.0179959852,
0.0479105897,
0.0174722001,
0.0361573473,
1.721403914e-34,
-0.0066048224,
0.0055533755,
0.0235468559,
0.0567420796,
0.0263345521,
0.0378535502,
0.0307572279,
0.010949838,
0.0315613151,
-0.0338221714,
-0.029939549
] |
http://arxiv.org/abs/2402.02330v2 | http://arxiv.org/pdf/2402.02330v2 | Enhance Reasoning for Large Language Models in the Game Werewolf | 2024-02-04T03:47:10 | 2024-03-29T09:01:56 | Shuang Wu, Liwen Zhu, Tao Yang, Shiwei Xu, Qiang Fu, Yang Wei, Haobo Fu | cs.AI | cs.AI, cs.CL | This paper presents an innovative framework that integrates Large Language
Models (LLMs) with an external Thinker module to enhance the reasoning
capabilities of LLM-based agents. Unlike augmenting LLMs with prompt
engineering, Thinker directly harnesses knowledge from databases and employs
various optimization techniques. The framework forms a reasoning hierarchy
where LLMs handle intuitive System-1 tasks such as natural language processing,
while the Thinker focuses on cognitive System-2 tasks that require complex
logical analysis and domain-specific knowledge. Our framework is presented
using a 9-player Werewolf game that demands dual-system reasoning. We introduce
a communication protocol between LLMs and the Thinker, and train the Thinker
using data from 18800 human sessions and reinforcement learning. Experiments
demonstrate the framework's effectiveness in deductive reasoning, speech
generation, and online game evaluation. Additionally, we fine-tune a 6B LLM to
surpass GPT4 when integrated with the Thinker. This paper also contributes the
largest dataset for social deduction games to date. | Enhance Reasoning for Large Language Models in the Game Werewolf: This paper presents an innovative framework that integrates Large Language
Models (LLMs) with an external Thinker module to enhance the reasoning
capabilities of LLM-based agents. Unlike augmenting LLMs with prompt
engineering, Thinker directly harnesses knowledge from databases and employs
various optimization techniques. The framework forms a reasoning hierarchy
where LLMs handle intuitive System-1 tasks such as natural language processing,
while the Thinker focuses on cognitive System-2 tasks that require complex
logical analysis and domain-specific knowledge. Our framework is presented
using a 9-player Werewolf game that demands dual-system reasoning. We introduce
a communication protocol between LLMs and the Thinker, and train the Thinker
using data from 18800 human sessions and reinforcement learning. Experiments
demonstrate the framework's effectiveness in deductive reasoning, speech
generation, and online game evaluation. Additionally, we fine-tune a 6B LLM to
surpass GPT4 when integrated with the Thinker. This paper also contributes the
largest dataset for social deduction games to date. | [
0.0461617,
0.0325700082,
-0.0253625158,
0.0749412477,
-0.0442884155,
0.0131601533,
0.0037155501,
-0.0321586207,
0.039966315,
-0.0184217393,
-0.0112685841,
0.0126738977,
-0.0109043065,
0.0381158926,
0.0252125952,
0.0037969246,
0.0308193713,
-0.0646068379,
0.0412993431,
-0.0142870219,
-0.0291433968,
0.0085590407,
-0.0202739295,
0.0258557089,
0.0075697135,
-0.0197069664,
-0.0366649702,
0.008950837,
0.0339401402,
-0.0409102142,
-0.0545541681,
-0.0093934024,
0.0546351895,
0.034623459,
0.0000020056,
-0.0578652844,
-0.0009131858,
-0.0144836102,
-0.0263105128,
-0.0210753046,
-0.0180975907,
0.0340770297,
0.0055138455,
0.0224097334,
-0.0421133302,
0.0132833393,
0.0381753296,
0.0357814245,
0.0776135847,
0.0651275739,
0.0087150754,
-0.0527811795,
0.0075947386,
-0.0064281244,
0.0316138379,
-0.0473311245,
0.0202150084,
0.0367832147,
0.0628843009,
-0.0267423075,
0.0031046439,
0.0061177518,
0.0430414341,
-0.0016795313,
-0.0629070625,
0.0438688435,
-0.021081375,
0.0145982765,
-0.020742856,
0.0456903912,
0.0079251193,
-0.0199826192,
0.0003531571,
0.0292184502,
0.010388881,
0.0070706476,
-0.0058977162,
0.084139213,
-0.0265262462,
0.0003026441,
0.0099754091,
-0.0010395746,
0.0493994839,
0.0197180919,
-0.0597764477,
0.1360329688,
-0.0069966689,
-0.0163010173,
0.0032773921,
-0.0077250898,
-0.0556223132,
-0.057231307,
-0.0045270147,
0.0577194877,
-0.0766954646,
0.0308799092,
-0.0103598302,
0.014501689,
0.002547543,
0.0044997125,
0.0107063865,
0.0261774287,
0.0285654925,
0.0665235147,
-0.0080650533,
0.0313484371,
-0.0147259682,
0.0073502301,
0.0082457056,
-0.0195703655,
-0.0586421378,
-0.0204917323,
-0.0395635813,
0.0325307883,
-0.0208388027,
-0.0134349279,
-0.0130814854,
0.0434074365,
-0.0133080529,
0.0336718149,
-0.0106939971,
0.0103201037,
-0.0118097477,
-0.0324859172,
-0.0558422171,
0.0064504896,
-0.0633539706,
0.0192703735,
0.0414515622,
-0.0863757059,
-0.0100504328,
-0.0299123898,
0.0066104089,
-0.01144394,
0.0280526411,
0.0930680856,
-0.0060968497,
-0.0147599159,
-0.031273257,
-0.0363977849,
0.0082624909,
-0.1077230573,
-0.045331832,
-0.0788605958,
0.0135556655,
0.0033675984,
0.001128116,
-0.0255408324,
0.0213423334,
0.0011730832,
0.0074751624,
0.108899802,
-0.0143492799,
0.0085176416,
0.0130920419,
0.0192067493,
-0.0064763599,
0.0714823604,
0.0543605313,
0.0394979604,
-0.0078552067,
0.0311519671,
0.0359309539,
0.0307849441,
-0.0065621836,
-0.0257897265,
-0.0266213585,
0.0103506092,
-0.029141672,
0.0471295342,
-0.0213216636,
0.058232531,
-0.0373143628,
-0.0068156095,
0.0782776549,
0.0646796301,
-0.0051355474,
0.0278134774,
0.0568068884,
0.0224239733,
-0.0211818982,
-0.0382335223,
-0.032470677,
0.0333426073,
-0.0379040278,
-0.0023658979,
-0.0275439769,
0.0279750861,
0.0003195817,
-0.0234788358,
0.0171043929,
-0.0063989023,
0.0621159375,
0.0482583083,
0.0240718424,
-0.0000805125,
-0.0341905057,
0.0309887175,
0.0208435953,
-0.047381971,
-0.0245933384,
-0.063152574,
0.0512013845,
0.0674137846,
0.0672198832,
-0.0383670516,
-0.0369125567,
-0.0126611367,
-0.001178932,
0.0726598799,
-0.0013937787,
-0.0035314835,
0.0108996648,
-0.0340750478,
-0.0226794556,
-0.023813406,
0.0121379392,
0.0579232499,
-0.0062190811,
0.0113299228,
-0.0120874811,
0.0048424071,
0.0093828812,
0.005202055,
-0.0237947702,
-0.0182105862,
-0.0237455536,
0.008436393,
0.0314886533,
-0.0015324827,
-0.0093681719,
0.0087495539,
0.0242728535,
-0.0203138702,
-0.0216700267,
0.009165681,
-0.0018513723,
-0.0408778079,
0.0466224328,
-0.0173311401,
-0.014949128,
0.0184285231,
-0.0023574652,
-0.0047392868,
0.0248002317,
-0.069113344,
0.0586884357,
-0.0245055165,
0.0030750579,
-0.0207301453,
-0.0164600406,
-0.0174796674,
0.0815641433,
0.017376978,
-0.0411411263,
-0.0107327215,
0.0001683161,
0.0181664079,
-0.0742394403,
0.0353788249,
0.0065393262,
-0.0035981094,
-0.0008495583,
0.0387735032,
0.0025387562,
-0.0575967878,
-0.0386000872,
-0.0713679567,
0.0301049147,
-0.0056933798,
0.0157610737,
-0.0136206783,
0.0095421243,
-0.0066514146,
0.0108277723,
-0.0392283313,
-0.0032372423,
0.0137703605,
0.0100652203,
0.0202952716,
0.0362112671,
-0.0446780622,
0.0179491378,
0.0132047888,
-0.0237409063,
0.0173537862,
0.0033144548,
0.0070980573,
-0.0436584763,
0.0313760042,
-0.0826811716,
0.0310779177,
-0.0166692957,
-0.0603960939,
0.0534344874,
-0.0188753773,
0.0403665304,
0.0865925923,
0.0093409922,
0.0086900983,
0.0619890727,
-0.0476008803,
-0.0441551767,
0.0432830714,
-0.0004426553,
-0.0119431308,
-0.0646432787,
-0.021819938,
0.0114526926,
0.0472657233,
-0.0192758311,
-0.0043285829,
-0.0401367433,
0.0537555218,
0.0365573093,
-0.0569868013,
-0.0248842128,
-0.0216542166,
-0.003006058,
0.0284375455,
-0.0224845875,
-0.0007700751,
0.0967424139,
-0.0199627746,
-0.001911878,
-0.0130052622,
-0.0122700129,
-0.0404335707,
0.0827498585,
0.0084398212,
-0.0020307442,
-0.0266530737,
-0.0147323702,
-0.0383377299,
-0.0451268256,
0.0613847636,
-0.0150154689,
0.0219616052,
-0.067135185,
0.0005728258,
0.0391021594,
-0.0294182785,
-0.0438557044,
0.0161782578,
-0.0927745253,
-0.0166942403,
-0.0102795698,
-0.0272042416,
-0.067451857,
-0.0192651823,
0.0640394241,
-0.0042420235,
-0.0032173153,
-0.0122875292,
-0.0298754759,
-0.0230708476,
0.0563221052,
0.0295254663,
0.0403028429,
-0.0461261384,
0.0550513007,
0.0462514982,
-0.0139374798,
-0.0454643518,
-0.0390129015,
0.025151439,
-0.0054770634,
-0.0158716645,
0.0094590168,
0.038450852,
0.0282788686,
-0.0839820057,
-0.00721303,
0.1115141287,
0.0227821153,
0.0033572624,
0.0039143516,
-0.0240984838,
0.0270242989,
0.0153282946,
0.0303913765,
-0.0894619301,
-0.0663193837,
0.0073090252,
-0.0256185997,
-0.0235542227,
-0.0575064272,
-0.0229591187,
-0.1084455252,
0.0161589216,
0.004121019,
-0.0050812461,
0.0074982909,
0.0185360909,
-0.0058698067,
-0.0212298427,
-0.0332541838,
0.0097348997,
0.0217025205,
-0.0046880962,
-0.0204494502,
0.0168093685,
0.023211807,
0.0340670161,
-0.054203067,
0.048474513,
0.0500311777,
0.0606622137,
0.0032388316,
-0.0565338321,
0.0273054447,
0.0264730435,
-0.0045497254,
-0.0133529417,
0.0005585211,
-0.0122272307,
-0.0011265444,
-0.0302063357,
0.0326102413,
0.00728916,
-0.0535464399,
0.0195334423,
0.0759387985,
-0.0150906863,
-0.0406183377,
0.0140367625,
0.0183212273,
0.0074918671,
0.0907530859,
0.0037068808,
-0.0185446013,
-0.0410954878,
-0.0254449788,
-0.0704718903,
-0.016238628,
-0.0667870417,
-0.0261418223,
0.0344426483,
-0.0038119284,
-0.0531194881,
-0.0096698897,
0.0332635604,
0.0336831249,
-0.0058609759,
0.0233701151,
0.0150234727,
0.0496935844,
-0.0251445509,
0.0092800716,
-0.0571972467,
0.007939945,
0.0016220292,
0.0073077129,
-0.0016155037,
-0.055866003,
-0.0075233676,
0.0616877005,
-0.023349043,
-0.0519593582,
-0.0084911603,
0.0322813429,
0.0304745492,
0.0328911915,
-0.0104897907,
-0.0587045662,
0.0493877642,
0.0360841192,
-0.047105711,
-0.0089296615,
0.0066647111,
0.0478354767,
0.0508028828,
0.0415853411,
-0.0446567573,
0.0124120573,
0.0199329946,
0.0338154063,
-0.0226415731,
0.0104672657,
0.0216002017,
0.0344109796,
-0.0035567353,
-0.0127830943,
-0.0180064682,
-0.0076580248,
0.028277548,
-0.0788829699,
-0.0111471526,
-0.0498887599,
0.0080231922,
0.0077656428,
0.0090643261,
0.0162073094,
-0.030484641,
0.0150941703,
0.0060688439,
0.0357172862,
0.00306037,
0.0475937352,
0.0129849995,
0.0390118286,
-0.0224805605,
0.0082531106,
0.0172745418,
0.0056130807,
0.0495519601,
0.0246488675,
0.0627981871,
0.0488045849,
-0.0490508676,
-0.0252464674,
0.0159982871,
-0.0521134213,
-0.0820454359,
0.0891860798,
0.0461692959,
0.0371164456,
0.0603296869,
0.0031132947,
-0.0491003804,
0.0021654968,
0.0087533267,
0.0307295583,
0.0291965976,
-0.0695947036,
-0.0018044258,
0.0506230108,
-0.0440117009,
0.0045032483,
0.0196939297,
-0.0089800954,
0.0341887809,
0.0151308049,
-0.0230378713,
-0.0487063527,
-0.0125115141,
-0.037393529,
0.0191276595,
-0.0167550426,
0.0205033533,
-0.0761171132,
-0.0434363782,
-0.0261699818,
0.0670864657,
-0.0022568167,
0.0200632233,
0.0348067395,
0.003266908,
0.0029504306,
-0.0744383559,
-0.0239732806,
0.0385225564,
-0.0306625459,
-0.0201895181,
0.0024386665,
-5.82936024e-33,
-0.0281394515,
-0.0365600064,
0.0690812692,
0.0078040813,
-0.034642905,
-0.0331528969,
-0.0632511005,
0.0017723782,
0.0045539564,
0.024913881,
0.0073203342,
0.0379634574,
0.0062794061,
-0.0003945123,
-0.0148831094,
-0.0280430838,
0.0052616498,
0.0275377221,
-0.0435006618,
0.0220082924,
0.0326451361,
-0.0017880963,
0.0239163972,
-0.0333197601,
-0.0337333567,
-0.0020291894,
-0.0278276429,
0.0550170839,
-0.043461252,
0.0152185597,
-0.0172996037,
-0.0135098184,
0.0016332193,
-0.0301800054,
0.0110577578,
0.0202978905,
-0.0328338444,
-0.0100947907,
0.0259572621,
-0.0190573167,
-0.0325492099,
-0.044769451,
0.0142663894,
-0.0054414701,
0.0044745533,
-0.0068958355,
0.0116021307,
-0.0482227579,
0.0355491191,
0.0054213023,
-0.0439633876,
0.0031383892,
-0.0337757841,
-0.0195560716,
0.0071506132,
0.0222052522,
-0.0302982852,
-0.0013194403,
-0.0195491463,
0.0561454669,
0.0243154764,
-0.0429674685,
0.0662437454,
-0.0008015075,
0.038377244,
0.0470629893,
0.0197521113,
-0.0073305694,
-0.0878723636,
-0.0163590349,
0.0039954488,
0.0962199047,
0.0604865626,
0.0404280722,
-0.0245567467,
0.0302808378,
-0.0888010561,
0.0059132734,
0.0787132606,
0.05879388,
0.0368564539,
0.0061660362,
-0.0272096917,
0.0044230493,
-0.0190808009,
0.0253274236,
-0.0209046397,
-0.0446654111,
0.013388209,
-0.0181178488,
-0.0168313421,
-0.0311821345,
0.0214182753,
-0.0494029559,
-0.0048736027,
0.0325895809,
-0.0069198986,
-0.0181158967,
-0.0123341549,
-0.0052173352,
-0.0224065613,
-0.0013029505,
0.0345317647,
-0.0063229501,
0.0199885219,
0.0207206309,
-0.0141593115,
0.0224135071,
-0.020711137,
-0.0124283675,
-0.0571105033,
0.0289332885,
0.0000580902,
0.0098283077,
0.0362403318,
-0.0201819036,
-0.0140258614,
-0.0200978369,
-0.0201784372,
-0.0325527675,
-0.0162061639,
0.0545175485,
-0.0230027605,
0.0088716689,
-0.0186704993,
-0.0212817676,
-0.0690267831,
-0.0034151815,
-0.0145464279,
-0.0308163501,
-0.013325789,
-0.024432905,
2.877e-7,
-0.0203249976,
0.0542247631,
-0.0141898952,
0.042799823,
0.022515934,
0.0567013323,
-0.0359779745,
0.0156504847,
0.0241770167,
-0.0702247471,
0.0008197901,
0.0706363469,
0.0409637056,
0.0267340802,
-0.0750942975,
-0.0025767516,
-0.0116961086,
-0.0431245491,
-0.0628552139,
-0.0144014219,
0.0292147119,
0.0473965444,
0.0546552464,
-0.011928726,
0.039532382,
-0.0325520821,
0.0235533472,
-0.0431037992,
-0.0156745929,
0.0031436963,
0.0320257284,
0.0070258551,
-0.0047535137,
0.0185933802,
0.0110590542,
0.0098319203,
-0.0569004007,
0.0426881798,
0.0011041284,
-0.0181304943,
-0.0272573382,
-0.0169271398,
-0.0136389397,
-0.0399237573,
0.0631560907,
-0.0430132039,
-0.0623513423,
0.0211694725,
-0.0025995765,
-0.0502256937,
-0.0558121838,
-0.0200735591,
-0.0124929994,
-0.0099061746,
-0.0001297551,
-0.0267549921,
0.0084132468,
-0.0668035671,
-0.0118951686,
-0.0003318137,
-0.0567125827,
-0.0525343828,
-0.0355035588,
0.0100687612,
0.029094765,
-0.0700994283,
0.0025449891,
2.508330457e-34,
0.0206482802,
0.0095589086,
0.007887952,
0.0134820743,
0.06456393,
0.0295427907,
-0.0426085256,
0.0198416431,
0.067175597,
-0.080065608,
-0.0058370451
] |
http://arxiv.org/abs/2402.02318v1 | http://arxiv.org/pdf/2402.02318v1 | Diversity Measurement and Subset Selection for Instruction Tuning Datasets | 2024-02-04T02:09:43 | 2024-02-04T02:09:43 | Peiqi Wang, Yikang Shen, Zhen Guo, Matthew Stallone, Yoon Kim, Polina Golland, Rameswar Panda | cs.LG | cs.LG, cs.CL | We aim to select data subsets for the fine-tuning of large language models to
more effectively follow instructions. Prior work has emphasized the importance
of diversity in dataset curation but relied on heuristics such as the number of
tasks. In this paper, we use determinantal point processes to capture the
diversity and quality of instruction tuning datasets for subset selection. We
propose to measure dataset diversity with log determinant distance that is the
distance between the dataset of interest and a maximally diverse reference
dataset. Our experiments demonstrate that the proposed diversity measure in the
normalized weight gradient space is correlated with downstream
instruction-following performance. Consequently, it can be used to inform when
data selection is the most helpful and to analyze dataset curation strategies.
We demonstrate the utility of our approach on various instruction tuning
datasets. | Diversity Measurement and Subset Selection for Instruction Tuning Datasets: We aim to select data subsets for the fine-tuning of large language models to
more effectively follow instructions. Prior work has emphasized the importance
of diversity in dataset curation but relied on heuristics such as the number of
tasks. In this paper, we use determinantal point processes to capture the
diversity and quality of instruction tuning datasets for subset selection. We
propose to measure dataset diversity with log determinant distance that is the
distance between the dataset of interest and a maximally diverse reference
dataset. Our experiments demonstrate that the proposed diversity measure in the
normalized weight gradient space is correlated with downstream
instruction-following performance. Consequently, it can be used to inform when
data selection is the most helpful and to analyze dataset curation strategies.
We demonstrate the utility of our approach on various instruction tuning
datasets. | [
-0.005480085,
0.0444767252,
-0.0279761404,
0.0197097901,
-0.0284090191,
0.0459053107,
0.0369389951,
0.0190460794,
-0.01368271,
-0.0099514155,
-0.0315877497,
-0.0361599997,
0.0247932505,
-0.0311728287,
0.0408089869,
-0.036918737,
0.0477952659,
-0.0746383741,
-0.0464855693,
-0.0336693898,
-0.0299726184,
-0.0225370899,
-0.0074536176,
0.0512345694,
0.0281237885,
-0.0124139069,
0.0170878749,
0.0233663805,
0.0091316067,
-0.0167936366,
0.02336598,
0.0683062449,
0.0363720022,
-0.0008715201,
0.0000018789,
-0.0364070311,
0.0163890999,
0.006117722,
-0.0405115746,
-0.0002493233,
-0.0310118757,
0.0197211504,
0.0030158332,
-0.0054097758,
0.0212363824,
0.0119764283,
0.0641109869,
-0.0136448704,
0.0232741795,
0.0777806863,
0.0065234089,
-0.0170999821,
-0.0136080179,
-0.0245327819,
0.0298528541,
0.024624981,
0.0100441771,
0.0407592915,
0.0354341306,
0.0311607253,
-0.041626025,
0.0061989375,
0.0102778953,
0.027081782,
0.021983346,
0.0185496267,
-0.0183225535,
0.0098531367,
-0.0031766025,
0.0697890371,
-0.0123587754,
-0.0027879612,
0.0039010867,
-0.0341300406,
0.0120980283,
0.0033394152,
-0.0462885238,
0.0151220737,
-0.0182798374,
-0.0076408936,
0.0254425108,
-0.034645021,
0.0181211624,
-0.0437074043,
-0.0061068363,
0.043485079,
0.0298222806,
-0.065485321,
0.0005844251,
0.0138789359,
-0.0279825293,
-0.0372024477,
-0.0332160667,
-0.0117243668,
-0.0375298746,
0.0154786706,
-0.0642890856,
-0.0268254168,
0.0047490378,
-0.0082979072,
0.1071029678,
0.042326767,
-0.0305432025,
0.0204200372,
0.0195448846,
0.0631416738,
-0.0084797805,
0.0080950996,
-0.0360751897,
0.0395053811,
-0.0522111468,
0.0382316224,
-0.0040302081,
0.0276021175,
-0.022875594,
0.0361636877,
-0.0297264997,
0.0485559553,
0.0035861314,
-0.0437789187,
0.0524657965,
-0.0367930457,
-0.0689127818,
0.0253685787,
-0.0135600828,
0.0272575282,
-0.0677808449,
0.0214335434,
-0.0009122901,
-0.0477332473,
-0.0185825434,
-0.0018095501,
-0.0013649298,
-0.0651465058,
0.0360196754,
0.0467738882,
-0.0060357749,
-0.0458009839,
-0.0488902703,
0.0120681552,
-0.0138660576,
-0.0629150271,
-0.0068547083,
-0.0795527026,
0.0055233403,
0.0241084397,
-0.0041465862,
-0.0320022218,
-0.02735232,
-0.0013433298,
-0.0027940921,
0.0177704357,
0.0387737229,
0.0180015452,
0.0243816245,
-0.0418477058,
-0.031005783,
0.0884236991,
0.0032716088,
0.004339816,
0.040709082,
-0.0320401229,
-0.0317546874,
0.0203249622,
-0.0178349856,
0.0262581408,
-0.0101270424,
0.0291368943,
0.0122455815,
-0.00336417,
0.0080028996,
0.0049050087,
-0.0440258011,
0.0082271686,
0.0564706028,
0.0730499029,
0.0316726342,
-0.0373316146,
0.0527667366,
-0.0024392786,
-0.004266711,
0.0373536609,
-0.0362929627,
0.0260594077,
0.0059029926,
-0.0022403405,
0.0227032509,
0.0841685161,
-0.0356946513,
-0.023922421,
0.0033943378,
0.0226133317,
-0.0033053376,
0.0013474498,
0.0607963949,
-0.0648859367,
-0.0384946354,
0.0171120726,
0.0488921925,
-0.0296853296,
-0.0287853945,
-0.0395388603,
0.0298789293,
0.0530640408,
0.0049778917,
-0.0222834628,
-0.0294474568,
-0.007772774,
0.0868640766,
-0.0446710996,
0.0412850156,
0.0070564826,
0.0382725969,
-0.0574851483,
-0.0668644011,
0.0078965444,
0.0162178837,
0.0634612292,
0.0064952392,
-0.0224369951,
-0.0189460982,
-0.0034581232,
-0.0790108889,
0.0479476452,
-0.0231914744,
-0.0070568733,
0.0378921404,
0.0051289545,
0.0542433672,
-0.0047209775,
-0.0302653592,
-0.0160851069,
0.0098469583,
-0.0075847143,
0.0629325211,
-0.0331282355,
-0.0203431826,
-0.0175584424,
0.001805235,
-0.0634847358,
-0.0121280262,
0.0231493711,
-0.0135645708,
-0.0484640934,
0.0274772029,
-0.0244173612,
-0.0553454384,
0.0093424851,
0.0447437838,
-0.0292986501,
0.0315666273,
-0.0674858466,
0.0581908897,
-0.0215481147,
0.0382154547,
0.0365448892,
0.0153086334,
0.0047266698,
-0.0635536984,
-0.00720172,
-0.0116552999,
-0.0140617918,
0.0312557928,
0.0252537057,
0.0263558943,
0.0039546019,
-0.0429635495,
-0.0620470271,
0.0178306736,
-0.048425559,
0.0353046879,
0.0328778438,
0.0056685153,
-0.0045365598,
0.0032159088,
0.0087083708,
-0.014161502,
-0.0204632767,
0.0448692366,
0.0030136281,
0.0107214795,
-0.049533911,
-0.0224299114,
-0.0556213632,
0.0214632358,
0.0092962421,
-0.0070653004,
0.0224678535,
-0.0666681007,
0.0299995095,
0.0126467058,
0.0359658413,
-0.0184721909,
-0.0087062847,
0.0559451841,
0.0233622808,
0.0470502265,
0.1144862324,
-0.0331963599,
-0.0114848688,
0.0458688512,
-0.0026989835,
-0.0480113998,
0.0196921546,
-0.0209251605,
0.0425163731,
-0.056058947,
-0.0528999306,
0.0174950808,
0.0493162386,
0.0310382769,
-0.027323015,
-0.0726481527,
-0.0325794555,
-0.0403874256,
-0.0295317434,
-0.0317771547,
-0.0015277828,
-0.0263492931,
0.0013512051,
-0.0337289982,
-0.0879199058,
0.0558522008,
0.0138009544,
0.0181922242,
-0.0644831657,
-0.013845359,
-0.017584892,
0.0036399397,
0.0009053258,
-0.017717028,
-0.0125426613,
-0.0020874338,
-0.0626644269,
0.0434577242,
0.0532124043,
0.0057403343,
0.0985270962,
-0.076610364,
0.018732477,
-0.022905929,
-0.0015794443,
-0.0060765762,
-0.0132626388,
0.0721005201,
-0.0244237687,
0.0795174241,
-0.0509201884,
-0.0157811139,
0.000074944,
0.025028633,
0.0699312463,
-0.045542758,
0.0037169938,
-0.0076966551,
-0.0611821786,
0.0338732041,
0.0082523245,
-0.0172880981,
-0.0820128173,
-0.0031628371,
-0.0404016264,
0.038328018,
-0.0384863578,
-0.0061508203,
0.0481420755,
0.0215905178,
-0.0258319546,
-0.0430278331,
0.064381659,
0.1085975915,
-0.0144664543,
-0.0270985831,
0.0672692657,
0.0172762647,
-0.0573118739,
-0.0001908069,
0.0126536358,
0.0453103296,
0.026636878,
0.0208730754,
-0.0140125537,
-0.011675097,
-0.0311778877,
0.0457086451,
-0.0104673021,
-0.005588355,
-0.0011043048,
-0.0585324429,
0.0209108274,
0.0021578055,
-0.0637623966,
-0.0596203506,
-0.0233709887,
0.0080375262,
-0.046123568,
0.0277647674,
-0.0044866395,
-0.034964256,
0.0044394224,
0.0640181676,
-0.0171477608,
-0.1014907807,
0.0350262709,
-0.0063230516,
0.0377256088,
0.0538420044,
-0.0146424221,
-0.0051297098,
-0.0126857813,
0.034927357,
-0.0234396625,
-0.0022635346,
0.0039658849,
0.0480919741,
0.0280613992,
-0.0233397819,
0.0396150462,
0.0460237488,
-0.0109446887,
0.0407980829,
0.0256345402,
0.0524422117,
-0.040641699,
0.0370410532,
0.0115651777,
-0.063196443,
0.0382830128,
0.0274378471,
0.0419911034,
-0.0597659536,
0.0229717419,
-0.0327490121,
-0.0247103088,
-0.0735655278,
0.0331119858,
0.0185460448,
-0.0673160851,
-0.0229744408,
-0.0515978746,
-0.0191992614,
0.0504618362,
0.0454783998,
0.0021669012,
0.0192626659,
-0.0008286931,
0.0050946972,
0.0132235745,
0.0419953391,
0.0005053492,
0.0229653679,
-0.1010603607,
-0.0864050612,
-0.0410507321,
0.0020094635,
-0.0153566552,
-0.0105904024,
-0.1245109886,
-0.0780460387,
0.0342763625,
0.0511245355,
0.0185201764,
0.0341949277,
0.012808064,
-0.0707360879,
-0.0071071307,
0.0109433336,
0.0146589931,
0.02755633,
0.0417183004,
-0.0740109235,
-0.0255926698,
0.003218123,
-0.0376722999,
0.0553705767,
0.0331329815,
0.0000992045,
-0.0336210206,
-0.0099826446,
-0.0231336113,
0.0039197118,
0.0074731042,
-0.0187877491,
-0.0401135348,
-0.0714755058,
0.0382782333,
0.0072928648,
0.0343572795,
-0.0483545028,
-0.0290944409,
0.021842571,
-0.0186861884,
-0.0119153764,
-0.0035575004,
-0.0002697328,
0.0034616056,
-0.1014040634,
0.0459425747,
0.0159586407,
0.0622704253,
0.0399820358,
0.031282939,
0.0026850789,
0.0421203747,
-0.0033976301,
0.0093167415,
0.0267740153,
0.0460820235,
-0.0022687756,
0.0475701243,
0.013784226,
0.019681992,
-0.0687267706,
0.0008879126,
-0.0360501148,
-0.016868636,
0.0435163304,
-0.0366417356,
0.0587625131,
0.0046512135,
0.0208704304,
-0.0487552732,
0.0247925092,
0.0159129295,
0.0125710955,
-0.0020691245,
0.039722193,
0.0146521283,
-0.0217302069,
0.0833837688,
-0.0014007177,
0.0394952521,
-0.0154348621,
0.0223825406,
-0.0482091792,
-0.0245421138,
0.0228909981,
-0.0019544,
-0.0166913029,
0.0026889567,
-0.036420498,
-0.0402124263,
-0.0422970541,
0.0079759443,
0.0407883674,
0.0008176106,
0.0326870568,
0.0308919121,
-0.0524787828,
0.015955355,
-0.0241707191,
-0.0341728479,
0.0071081202,
-0.0168298129,
-0.0501739606,
-6.277440319e-33,
-0.0276063588,
0.0040184637,
-0.0104309684,
-0.0365894511,
0.0228627361,
-0.0002772003,
-0.003630613,
-0.0032019657,
-0.031664297,
-0.0118817007,
-0.0110098654,
0.0142868804,
0.0125655439,
0.008637323,
0.033420708,
-0.0043111676,
0.0277396739,
-0.0195360202,
-0.0390052721,
0.028891217,
0.04909667,
-0.0110392682,
0.0559076816,
0.017641183,
0.0139627,
-0.0160168763,
0.0221141614,
-0.016435381,
-0.0059588435,
0.0819089487,
0.0245037749,
0.0543450937,
0.0180047117,
-0.0631822124,
-0.0591540076,
-0.0589537285,
-0.0084341848,
-0.0808811635,
0.0088939816,
-0.0038372627,
0.003075602,
-0.0266782045,
0.0101692192,
0.0098557081,
-0.0538298748,
-0.0413234122,
0.0518373512,
-0.0150308982,
-0.0445611961,
0.0133409826,
-0.0311295055,
0.0262167249,
-0.0289158728,
-0.0206954293,
0.0149214026,
0.0169395637,
0.0221565161,
-0.0360286832,
0.0262974333,
0.0434101634,
0.0651785955,
0.0669554174,
0.0432794541,
0.0436677337,
0.0391341001,
0.0245334227,
0.0085775517,
0.0354691111,
0.066167213,
-0.0084374892,
0.0319930315,
0.0327033475,
0.0135517856,
-0.0043869079,
0.030884346,
0.0016137881,
-0.0790944248,
0.0115276333,
-0.0087093776,
-0.0093935272,
0.0172367413,
0.00135595,
-0.0639582127,
-0.0090352707,
-0.0014285366,
0.0139226625,
-0.0008765237,
-0.0016168274,
0.0546627529,
-0.0020092244,
-0.0560371503,
0.0062951902,
0.0684547722,
-0.0116107306,
-0.0372932106,
0.0075893379,
-0.0210287794,
0.0187923741,
0.0454639085,
-0.0117189297,
-0.0593503416,
-0.0712350905,
0.0415649302,
0.0188330151,
0.0409715138,
-0.0127411159,
0.0043814359,
0.0205525048,
-0.0461348891,
-0.0051852283,
-0.02584734,
0.0320377946,
-0.0049424865,
0.051602561,
0.0295356978,
-0.0193641894,
-0.0052926303,
0.0925175995,
0.0038744637,
-0.0585341975,
-0.0300714113,
-0.0232964735,
-0.0125937397,
0.0384531021,
-0.0061899503,
0.0274392664,
-0.0223622508,
0.0051951292,
0.0414027497,
-0.0087742945,
-0.0259931944,
0.0042045447,
2.598e-7,
0.0122678084,
0.0457165092,
0.0146363406,
0.0353361107,
0.0012385702,
0.0036899,
-0.0454555415,
0.0010674623,
0.0074824202,
-0.0192561857,
0.0184609964,
0.0284437705,
0.0345159955,
0.0159519762,
-0.0440969281,
0.0187779441,
-0.0223973673,
0.0693667978,
-0.0658441857,
-0.0053200726,
-0.0083204713,
0.0698312074,
-0.000837605,
0.006153923,
-0.0038298622,
0.0262145232,
-0.0112681445,
-0.0411595702,
0.0106615648,
0.0330881104,
0.0480146036,
-0.0286366604,
-0.0172221344,
0.0309070237,
-0.019924378,
0.0090369023,
-0.0167504326,
0.0457026586,
-0.0348582864,
0.0231010765,
0.017452823,
0.000829837,
-0.0413640887,
-0.0293335989,
0.0519420691,
-0.0107643493,
-0.0001286366,
-0.0389758684,
0.0048275259,
0.0075480035,
0.0433160178,
0.0001689971,
0.0219601542,
0.0160623249,
0.0238533542,
-0.0487184599,
0.0193181727,
-0.0178626142,
0.0156942178,
0.0184838772,
-0.0407080501,
-0.0745349377,
-0.0660057962,
0.0100666266,
0.0082419692,
-0.0010458237,
-0.0179605372,
2.67010626e-34,
0.0285484139,
-0.0256780814,
0.0526016951,
-0.0126438979,
0.0050831558,
-0.0378583744,
-0.0471646972,
0.0095831091,
0.0028040872,
-0.0803423896,
-0.037479043
] |
http://arxiv.org/abs/2402.02315v1 | http://arxiv.org/pdf/2402.02315v1 | A Survey of Large Language Models in Finance (FinLLMs) | 2024-02-04T02:06:57 | 2024-02-04T02:06:57 | Jean Lee, Nicholas Stevens, Soyeon Caren Han, Minseok Song | cs.CL | cs.CL, q-fin.GN | Large Language Models (LLMs) have shown remarkable capabilities across a wide
variety of Natural Language Processing (NLP) tasks and have attracted attention
from multiple domains, including financial services. Despite the extensive
research into general-domain LLMs, and their immense potential in finance,
Financial LLM (FinLLM) research remains limited. This survey provides a
comprehensive overview of FinLLMs, including their history, techniques,
performance, and opportunities and challenges. Firstly, we present a
chronological overview of general-domain Pre-trained Language Models (PLMs)
through to current FinLLMs, including the GPT-series, selected open-source
LLMs, and financial LMs. Secondly, we compare five techniques used across
financial PLMs and FinLLMs, including training methods, training data, and
fine-tuning methods. Thirdly, we summarize the performance evaluations of six
benchmark tasks and datasets. In addition, we provide eight advanced financial
NLP tasks and datasets for developing more sophisticated FinLLMs. Finally, we
discuss the opportunities and the challenges facing FinLLMs, such as
hallucination, privacy, and efficiency. To support AI research in finance, we
compile a collection of accessible datasets and evaluation benchmarks on
GitHub. | A Survey of Large Language Models in Finance (FinLLMs): Large Language Models (LLMs) have shown remarkable capabilities across a wide
variety of Natural Language Processing (NLP) tasks and have attracted attention
from multiple domains, including financial services. Despite the extensive
research into general-domain LLMs, and their immense potential in finance,
Financial LLM (FinLLM) research remains limited. This survey provides a
comprehensive overview of FinLLMs, including their history, techniques,
performance, and opportunities and challenges. Firstly, we present a
chronological overview of general-domain Pre-trained Language Models (PLMs)
through to current FinLLMs, including the GPT-series, selected open-source
LLMs, and financial LMs. Secondly, we compare five techniques used across
financial PLMs and FinLLMs, including training methods, training data, and
fine-tuning methods. Thirdly, we summarize the performance evaluations of six
benchmark tasks and datasets. In addition, we provide eight advanced financial
NLP tasks and datasets for developing more sophisticated FinLLMs. Finally, we
discuss the opportunities and the challenges facing FinLLMs, such as
hallucination, privacy, and efficiency. To support AI research in finance, we
compile a collection of accessible datasets and evaluation benchmarks on
GitHub. | [
0.0209108144,
0.0935105979,
0.0041194311,
0.0426856093,
-0.002551941,
0.0162582658,
0.0476167649,
0.0447017513,
0.0177836046,
-0.0276469756,
-0.0754164755,
-0.0593379848,
-0.0045319404,
0.0294870678,
0.0500968732,
-0.0258461013,
0.0196394995,
-0.0322120562,
0.0384504758,
-0.0099993479,
0.0232187118,
0.0159866214,
0.0142295966,
0.0410999097,
-0.0112413224,
-0.018337708,
-0.0039233938,
-0.0221399143,
-0.0107308887,
0.0017474443,
-0.0095188301,
0.0334349126,
0.0428748876,
0.0657866299,
0.000001992,
-0.0586084612,
-0.0269115698,
0.0040087979,
-0.0308188684,
0.0266719349,
0.0226455722,
-0.0509255417,
0.0553537458,
0.0067620804,
-0.0574557334,
0.0075363079,
0.0567045361,
0.0913532451,
0.0589770265,
0.0918250009,
-0.0153025547,
-0.0670133382,
0.0192150716,
-0.0856195837,
-0.014512863,
-0.0873748064,
0.0391546562,
0.0102077983,
-0.0460823923,
-0.0070885457,
-0.0211335383,
0.0309799071,
-0.0082711251,
0.0175882019,
-0.0003184512,
0.0086301425,
-0.0520345904,
0.0092661045,
-0.0515884943,
0.0666294098,
-0.0160891842,
0.0443733707,
0.0122403987,
-0.0408380628,
-0.0231658183,
0.0158031769,
-0.0360619165,
0.0637767613,
0.0159702394,
0.0159681235,
0.0125870099,
0.0026326792,
0.0467942059,
0.0095729902,
-0.0474653691,
0.0391262732,
0.0020826592,
-0.0216262061,
0.0428937934,
0.0163783468,
-0.0312448852,
-0.0294637401,
0.0276428219,
-0.0088026235,
0.0366579741,
0.0076005147,
-0.019738812,
-0.0312480293,
0.0042626248,
-0.084619306,
0.0432644635,
0.0160532389,
-0.0160766616,
0.0646014065,
-0.0372985527,
0.0497314706,
-0.0135243842,
-0.0133007448,
-0.0067659039,
0.0068646283,
-0.0456853397,
-0.0200334825,
-0.0539259724,
0.0801643282,
-0.0452419668,
-0.0608140081,
-0.0689554811,
-0.0353309922,
-0.0099335946,
0.018302653,
0.0160169005,
-0.0317123532,
0.0012822342,
-0.0495959669,
-0.0243389625,
-0.0220115948,
-0.0290231444,
0.0238868892,
-0.0197090954,
-0.030935511,
0.0095789097,
0.0419048965,
-0.0401815027,
-0.0299148038,
0.0134878075,
0.0744561329,
0.0714588538,
-0.0583671927,
-0.0417975783,
-0.0381142646,
0.0002897357,
-0.0384049602,
-0.0263078827,
-0.0291923806,
0.0307257976,
-0.0275736898,
0.0198227167,
-0.0325181223,
0.0186620355,
-0.0035547426,
-0.0192576628,
0.0987005159,
0.0123850042,
-0.0211877152,
0.0007019427,
-0.0430957749,
0.0197344497,
0.0710779428,
0.0002029585,
0.0293330494,
0.0510943905,
-0.0059248344,
0.0130264545,
0.0575994924,
0.0034440553,
-0.0505972914,
0.0027370588,
0.0206825938,
0.0109126111,
-0.0021154319,
0.0433896706,
0.0395091847,
-0.0250029191,
-0.0255763661,
0.0599126704,
0.0559675992,
0.0684810281,
0.0108496621,
0.0049603404,
0.0478440486,
0.0451531783,
-0.0312850028,
-0.0138387298,
-0.0292156637,
-0.0220305491,
0.0152978636,
0.0039238785,
0.0145181799,
-0.0348718874,
-0.0223185122,
-0.0175992027,
-0.0404281467,
-0.0194855556,
0.0014854822,
0.0286121089,
-0.0307486691,
-0.0472272001,
0.0718256459,
0.0339562297,
0.004329714,
-0.005420479,
-0.0369023271,
0.0332330912,
0.0970013589,
0.0307544898,
-0.0321225971,
0.0294733774,
-0.0247788262,
-0.0088231172,
0.0752920434,
0.0261078607,
0.0162810218,
0.0095072351,
-0.0577051044,
-0.0052309562,
0.0078855781,
0.0215345789,
0.030663766,
0.0102013778,
-0.0077882791,
0.0293013994,
-0.0142414877,
-0.0287822131,
0.03207241,
0.0002977391,
-0.0437067896,
0.0054319724,
-0.035332337,
-0.0030377132,
0.0168606713,
-0.0147303334,
0.0093826633,
-0.0011835735,
-0.0105954763,
-0.0165534671,
0.0017772665,
-0.0538626127,
-0.0019958678,
0.0074954159,
-0.0072346665,
0.0064721662,
0.0231845658,
-0.0545461588,
-0.0159440171,
-0.0405671336,
0.0079165073,
0.0300609395,
-0.0755690262,
0.0070282146,
0.0250432529,
0.0431913175,
0.0006543507,
0.068817623,
-0.0246234592,
0.0140250502,
-0.0207441095,
0.0263294559,
-0.0093189692,
-0.0358294994,
0.0314483494,
-0.0101099517,
-0.0226557907,
0.0329720974,
0.0250315033,
0.0062656337,
-0.0050092735,
-0.0202903058,
-0.059096735,
-0.0215749778,
-0.017256476,
0.0317813531,
-0.0091933273,
-0.0029416487,
-0.0200917739,
-0.0241515599,
-0.0248863399,
0.0046408842,
-0.0092298221,
-0.0054218625,
0.0397655666,
0.0006909811,
-0.0078771887,
-0.0553104281,
-0.0045109075,
0.0201227497,
0.038084887,
0.0016626267,
-0.0436327159,
-0.0313151851,
0.0051633986,
-0.0044043092,
0.020766424,
-0.0159914903,
0.0406945162,
-0.0070062676,
-0.028867824,
0.024743963,
0.0096589895,
0.0072498773,
0.0636703372,
-0.0344306752,
-0.0168782864,
-0.0214934312,
0.0093768397,
-0.0074580288,
-0.0027751292,
-0.0387415923,
0.0238518007,
0.0457214229,
0.0627033189,
-0.0349758938,
0.0005050357,
-0.0764709413,
0.0352836996,
-0.010269599,
0.0088559203,
-0.0314535126,
0.0072638523,
-0.0285818744,
0.0376554318,
0.0027590133,
-0.038446933,
0.0406509079,
-0.0339733362,
-0.0135210585,
-0.0150695471,
-0.0207131524,
-0.0814306512,
0.0578230433,
-0.0177965276,
0.0126287984,
-0.0079671927,
-0.0660613179,
0.0022641199,
-0.0087840045,
0.0196354426,
-0.0136752585,
0.0378209651,
-0.0315475501,
0.0467004366,
0.0015866217,
-0.0035085953,
0.000259312,
-0.1178584248,
-0.0459646769,
0.0087722605,
0.0395509191,
-0.0544610806,
-0.0402432941,
0.0291205253,
-0.0572355166,
-0.0085099665,
0.0025104394,
0.002158904,
0.0058128661,
0.022937458,
0.0846418738,
0.0138870468,
0.0396505408,
-0.0464667082,
0.033628054,
0.0092505049,
0.0162308291,
0.0145943658,
-0.0010759495,
-0.0219627544,
0.0010844833,
-0.0719544291,
-0.0083637387,
0.0060165627,
-0.0042539244,
-0.0510823168,
0.0112396153,
0.1219486371,
0.0124488873,
-0.0102348495,
0.003778554,
0.0233330037,
0.0010635452,
-0.0410237871,
0.0153998872,
-0.1180000901,
-0.0322577506,
-0.0213857722,
-0.0135900155,
0.0402499065,
-0.0440910794,
-0.0203302708,
-0.1259044856,
0.0274286661,
-0.0081239184,
0.0103021059,
-0.0226327311,
-0.0015888864,
-0.0388085954,
-0.0450458042,
0.0239102244,
0.0006791644,
-0.0227046497,
-0.0009064482,
-0.0004696897,
-0.0524708256,
-0.0204209182,
-0.0011756027,
-0.0320572108,
-0.0163639225,
0.0444074571,
-0.0080328984,
-0.0132388864,
-0.0204282254,
-0.0237359386,
0.0074020801,
-0.0032911301,
-0.0046646139,
0.0376772657,
-0.0254661068,
-0.0344620086,
0.0065984153,
0.031707149,
-0.0324943401,
0.0347023197,
0.0079272063,
-0.0124063045,
-0.0395242646,
-0.0348324217,
0.0196239073,
0.0373093486,
-0.0153055638,
0.0779875293,
0.0249445569,
-0.0445184223,
-0.033857096,
-0.0328144953,
0.0166006107,
-0.0398642197,
-0.0190861691,
-0.0018849563,
0.0393917747,
0.0088886935,
-0.0842167586,
-0.0595244616,
0.0320971794,
0.0123476172,
0.0088523533,
0.0268979631,
0.0507205576,
-0.0319297276,
-0.0297811162,
0.0497934818,
-0.0891906396,
0.0256216004,
-0.0440971032,
0.0214180946,
0.0314563923,
-0.0106379585,
-0.0498518646,
0.0545380563,
-0.090749681,
-0.0318679512,
0.0120110204,
0.0252478346,
0.0168006625,
0.001092059,
-0.0314753689,
-0.0044292086,
0.0336308517,
0.0214982089,
-0.0627292022,
-0.016215805,
0.0559146963,
0.006288297,
-0.0200972036,
0.0437460169,
-0.0290909931,
0.0221502818,
-0.0160564315,
0.0316455737,
-0.0348874666,
-0.0014816771,
-0.0085824691,
-0.0038628527,
0.0436897352,
0.0047147269,
0.0066759274,
-0.0331431963,
-0.0178475659,
-0.0178784542,
0.010261382,
-0.0366531424,
-0.0356586389,
0.0672248453,
-0.0350199006,
-0.0206227209,
0.0139578786,
-0.0246948358,
-0.0133877322,
-0.0182971321,
0.0265005492,
-0.0032974691,
0.0019298638,
0.0212351885,
-0.028122237,
0.0205242168,
0.0415431596,
0.0393264145,
0.0161500629,
0.0097280731,
0.0045515397,
0.0053998921,
0.0065053222,
-0.0297516715,
-0.0082194395,
0.0088302614,
-0.0553017445,
-0.0159451142,
0.0457501486,
0.0480075255,
-0.0411246456,
0.0119486758,
-0.0504604355,
0.0072251759,
-0.0017322723,
0.0216081496,
0.0243758187,
-0.0445799865,
0.0161135253,
-0.0007167635,
-0.0584816188,
-0.0135596022,
0.1368926615,
0.0353104025,
0.0115072029,
-0.0027564459,
0.0065955338,
-0.0556058511,
-0.0519238077,
0.0214611888,
0.0240131989,
-0.031089833,
0.0306885876,
-0.0686012059,
-0.0500911064,
0.0135103939,
-0.0409885645,
-0.0120386323,
-0.0069893096,
0.0362091511,
0.0527949035,
-0.0064396448,
0.0323409997,
-0.008119694,
-0.0087480759,
0.0069638803,
-0.0180136506,
-0.0752304718,
-6.289415668e-33,
-0.0303021725,
-0.0184801798,
0.0364914946,
0.004243453,
0.0155001525,
0.021457145,
0.0022793598,
-0.0111788055,
0.018169459,
0.0010142397,
0.0110814357,
0.0351377763,
0.0196808651,
0.0432828143,
-0.0124089243,
0.0059949979,
0.0459067188,
-0.0272111017,
-0.0030058718,
0.0002303975,
0.011087996,
0.0709029585,
0.0707162693,
-0.0608817227,
0.028433172,
-0.0026888938,
-0.0476079471,
0.0113324933,
0.0352052413,
0.0432700105,
-0.0350922905,
0.0342679955,
0.0166424625,
-0.0888480917,
-0.03024837,
-0.0027188533,
-0.016505165,
-0.0305306111,
0.0214270726,
0.0392264239,
0.0155814607,
-0.0209359247,
0.0363697186,
-0.0264372807,
-0.0096922256,
0.009013338,
0.0033364964,
-0.0076991878,
0.0226563625,
-0.0225873664,
-0.0744549856,
0.0563911349,
0.0140934559,
0.0209328718,
0.06629581,
0.0325533077,
-0.0119859967,
0.0334438644,
-0.0310821403,
0.0181354433,
0.0182487611,
-0.0039309124,
0.0143130356,
0.0081774471,
-0.0091748945,
0.0573253185,
0.0005763263,
0.0024711599,
-0.0186961386,
0.0140092503,
0.0228705332,
0.0773600414,
0.0511089414,
0.0166192371,
0.0140937977,
-0.0175425578,
-0.0158829652,
0.0108842952,
-0.0161471516,
0.0192599073,
0.0239312928,
0.0203296393,
-0.0301736724,
-0.0353715979,
-0.0232767519,
0.0076324893,
-0.0173660591,
-0.0047304113,
0.0129852537,
0.03003872,
-0.0183404367,
0.0364419408,
0.0220011771,
-0.0514233857,
-0.0702125281,
0.0086409571,
0.0263367295,
0.0028038097,
0.0011538907,
0.0168442167,
-0.048221413,
-0.0526913069,
-0.0026591984,
-0.0207924657,
-0.0102565605,
-0.0410557762,
0.0273713917,
0.000828861,
-0.0481845252,
-0.0087395096,
0.0201741401,
0.0059141656,
0.0237137377,
0.0513318665,
0.0274873488,
0.02553601,
-0.0052695526,
0.022835901,
0.0055301525,
0.0014748741,
0.0082258573,
-0.0180905238,
0.011053537,
0.0632820129,
0.0254083928,
0.0062459246,
-0.0177335981,
0.0250921343,
0.0638163611,
-0.0194474831,
-0.0280493572,
0.0313872285,
2.579e-7,
0.0457703173,
0.0641231164,
0.0188305005,
0.0221127737,
-0.0005221469,
0.0270870365,
-0.0171718039,
0.0025920728,
-0.0199997295,
0.0174435228,
0.0106574306,
0.0308717564,
0.089985244,
-0.0287589356,
-0.0868975148,
-0.0484563895,
-0.0548234209,
0.0385368355,
-0.0992160812,
0.0010021839,
0.0091268541,
0.1211114153,
-0.0212566461,
-0.0366409384,
0.0063963733,
-0.1064640433,
0.0710397512,
-0.0575249307,
-0.0439531431,
0.0259180069,
0.0001803792,
0.0544324517,
-0.0030869253,
0.0138904918,
0.0030594256,
0.0232845098,
0.0366756022,
0.0533477999,
0.0324865803,
-0.0081897806,
-0.0041349074,
0.0195984263,
-0.0256488007,
-0.0347483009,
0.0457159169,
-0.0617168844,
-0.0458513498,
-0.0475302562,
-0.0275730323,
-0.0248244945,
0.0564539917,
0.0008726554,
-0.0096253324,
0.0149050197,
0.0284037422,
-0.0203050748,
-0.00202301,
-0.0435339808,
-0.0057212966,
0.0129925758,
-0.0370494127,
0.0464700982,
-0.0445706211,
-0.0029354044,
0.01881692,
-0.0531965345,
0.0223227683,
2.317306426e-34,
0.0696049556,
-0.0276499167,
-0.0242446475,
0.0769416541,
0.0741944686,
0.0203878861,
-0.0200704746,
0.0617422499,
-0.0195512772,
-0.1075211763,
0.0140401563
] |
http://arxiv.org/abs/2402.02314v3 | http://arxiv.org/pdf/2402.02314v3 | Selecting Large Language Model to Fine-tune via Rectified Scaling Law | 2024-02-04T01:55:00 | 2024-05-28T16:16:42 | Haowei Lin, Baizhou Huang, Haotian Ye, Qinyu Chen, Zihao Wang, Sujian Li, Jianzhu Ma, Xiaojun Wan, James Zou, Yitao Liang | cs.LG | cs.LG, cs.AI, cs.CL | The ever-growing ecosystem of LLMs has posed a challenge in selecting the
most appropriate pre-trained model to fine-tune amidst a sea of options. Given
constrained resources, fine-tuning all models and making selections afterward
is unrealistic. In this work, we formulate this resource-constrained selection
task into predicting fine-tuning performance and illustrate its natural
connection with Scaling Law. Unlike pre-training, we find that the fine-tuning
scaling curve includes not just the well-known "power phase" but also the
previously unobserved "pre-power phase". We also explain why existing Scaling
Law fails to capture this phase transition phenomenon both theoretically and
empirically. To address this, we introduce the concept of "pre-learned data
size" into our Rectified Scaling Law, which overcomes theoretical limitations
and fits experimental results much better. By leveraging our law, we propose a
novel LLM selection algorithm that selects the near-optimal model with hundreds
of times less resource consumption, while other methods may provide negatively
correlated selection. The project page is available at
rectified-scaling-law.github.io. | Selecting Large Language Model to Fine-tune via Rectified Scaling Law: The ever-growing ecosystem of LLMs has posed a challenge in selecting the
most appropriate pre-trained model to fine-tune amidst a sea of options. Given
constrained resources, fine-tuning all models and making selections afterward
is unrealistic. In this work, we formulate this resource-constrained selection
task into predicting fine-tuning performance and illustrate its natural
connection with Scaling Law. Unlike pre-training, we find that the fine-tuning
scaling curve includes not just the well-known "power phase" but also the
previously unobserved "pre-power phase". We also explain why existing Scaling
Law fails to capture this phase transition phenomenon both theoretically and
empirically. To address this, we introduce the concept of "pre-learned data
size" into our Rectified Scaling Law, which overcomes theoretical limitations
and fits experimental results much better. By leveraging our law, we propose a
novel LLM selection algorithm that selects the near-optimal model with hundreds
of times less resource consumption, while other methods may provide negatively
correlated selection. The project page is available at
rectified-scaling-law.github.io. | [
-0.0106701441,
0.038150128,
-0.0178449061,
0.0279020295,
-0.0080691772,
0.0227954481,
0.0039291983,
0.042705901,
-0.026856035,
-0.0497951582,
-0.0644625947,
-0.0280426163,
0.008877432,
0.0113888374,
0.0354157984,
-0.0143699814,
0.0370958038,
-0.0506825671,
-0.0220716391,
-0.0125278598,
-0.0126889972,
0.0063219774,
0.0229389016,
0.0206456669,
0.0345170684,
0.0247872714,
0.0185201075,
0.0213158429,
-0.0051912684,
-0.0159190074,
0.0224049762,
0.0504929759,
0.0066584414,
0.0160203055,
0.0000020283,
-0.0729988143,
-0.0136752231,
-0.0221256651,
-0.0266801044,
0.0283561777,
0.0330060311,
-0.0045413389,
-0.0238596462,
0.025706768,
-0.0365282334,
-0.0043392754,
0.0419993736,
0.0108984075,
0.016011402,
0.0712433159,
0.0116433892,
-0.0336121172,
0.0092602335,
-0.0605996661,
-0.0600882918,
-0.0579139106,
0.0357483812,
0.0105917742,
-0.0374139398,
0.0116189066,
-0.0354330316,
0.0097107347,
-0.0307020601,
-0.0051722657,
-0.003574729,
0.0168449618,
-0.0277454592,
-0.0093465345,
-0.0027924089,
0.0863067284,
-0.0609763749,
-0.0286780484,
0.0264945477,
-0.0397881754,
-0.0188216772,
0.010573186,
-0.045062393,
-0.00040474,
0.0466492884,
0.0394590385,
0.0183310322,
-0.0205449164,
0.0736670047,
0.0290817134,
-0.036100775,
0.0308077186,
0.0185054094,
-0.0170074906,
-0.0011990088,
-0.0218281914,
-0.0344170816,
-0.0464150272,
-0.0266648345,
0.0091119939,
-0.0397334434,
-0.007747822,
-0.0378237069,
-0.0146160014,
0.0124844238,
0.0125985919,
0.067539908,
-0.009675839,
0.009099436,
0.0484520011,
-0.0288051683,
0.0254892129,
-0.0058162832,
0.0157360807,
-0.0378606357,
0.0181685388,
-0.0095944377,
0.041278664,
-0.0370310843,
0.0120001631,
-0.0118044727,
-0.001586378,
-0.037025664,
-0.024104517,
0.0291828774,
-0.0158105344,
0.0391856097,
-0.0424684621,
-0.0339834355,
0.0064076381,
0.0251672138,
0.0577388592,
-0.0374414995,
0.0103554623,
-0.0237842891,
-0.0638780296,
-0.0147847226,
0.0114919869,
-0.0064651184,
-0.0237740725,
0.0364129506,
0.0440374911,
0.0064071948,
-0.0378830321,
-0.0864684284,
0.0089434301,
-0.0147904446,
-0.0689767748,
-0.0094034327,
-0.057040412,
-0.0042640446,
0.0019554398,
0.0266121179,
-0.0256134402,
-0.0054866625,
-0.0209375378,
-0.0427731723,
0.0487074777,
-0.0295904055,
0.0262881834,
-0.0162672698,
-0.02055059,
0.0222170055,
0.0821994469,
0.0575992912,
0.007287995,
0.0438878648,
-0.0389730111,
-0.0032110352,
0.0607904829,
0.0052264258,
0.0448187254,
-0.0252328794,
-0.0220317338,
0.0112696141,
0.0177306272,
0.0193473417,
0.0131475767,
-0.0263736024,
-0.0091593852,
0.0447683334,
0.0416785739,
0.0593758747,
0.0664537698,
-0.0024680791,
-0.0114545459,
0.0002040393,
0.0051540672,
-0.0624171793,
0.020316001,
-0.0526867881,
0.0088562369,
0.0222695973,
0.0426067449,
-0.0319911987,
-0.0421926752,
-0.0288381036,
-0.0167105403,
-0.0263738297,
-0.0077625117,
0.0483928472,
0.0098574366,
-0.0266327485,
0.0755070522,
0.0459967703,
-0.0327434577,
-0.0122686718,
-0.0547910854,
-0.0011197256,
0.036720667,
0.0484814942,
-0.0488107204,
-0.0088248095,
-0.0095615108,
0.0389653631,
0.0024503993,
0.0681567937,
-0.0026679551,
0.0018177857,
-0.0892025679,
-0.0369427577,
0.0058557061,
0.0238106549,
-0.0266097523,
0.0244195629,
0.0216707438,
0.0077313296,
-0.0602188148,
-0.0337734483,
0.0322547294,
-0.039953135,
-0.0335118957,
0.0151946452,
-0.0069521228,
0.0232458543,
-0.0337483361,
-0.0281621218,
-0.0128849782,
0.006195162,
-0.0294184573,
0.0168703496,
-0.0061855013,
-0.0338142626,
0.0092034107,
0.0213413145,
0.0156067628,
0.0255431589,
0.0119515602,
-0.062894918,
-0.0510609709,
0.0109224264,
-0.0005464677,
0.030538192,
-0.0349266231,
0.0370211266,
-0.0435421951,
0.0380939543,
-0.0680435449,
0.034451779,
-0.0047366833,
0.0088756159,
0.0313352756,
0.0033295809,
0.0238287523,
-0.0030037363,
-0.0466722772,
0.044106625,
0.001111176,
0.0180363376,
0.0085973917,
0.0109237097,
-0.0222409237,
-0.0461821817,
-0.0300732926,
0.0090380805,
-0.0556214154,
0.0287197269,
0.0174226668,
0.012467755,
-0.0051076296,
-0.005675056,
0.0118691381,
0.0186105855,
-0.0146415522,
0.021977853,
0.0362617224,
-0.0317828842,
-0.0303938463,
-0.0718721449,
-0.0508228652,
0.0183747429,
0.0322150476,
0.0038059524,
0.0100258598,
-0.0643521696,
0.0790658742,
-0.0035706977,
0.0064117876,
-0.0274920017,
-0.0674623847,
-0.0252902694,
-0.0373455472,
-0.0028367287,
0.0959077999,
0.0099279964,
-0.0137607483,
0.0749213845,
0.0175180193,
-0.0489113368,
-0.0037137659,
0.0089940354,
-0.0519897975,
-0.0688107237,
0.0031100155,
0.0068326672,
0.0755550116,
-0.029020261,
-0.0040223212,
-0.038533669,
-0.013888659,
-0.0155732213,
0.0195635241,
0.0155596305,
0.0555511527,
-0.0167114232,
0.0689891428,
0.0220370237,
-0.0882887617,
0.0373948105,
-0.020526519,
0.0104678115,
-0.0169846956,
-0.0158509985,
-0.0555006191,
0.0329941586,
0.0043131574,
-0.0109732533,
0.0164512433,
-0.0175258107,
-0.0646467805,
-0.0449798964,
0.0356418267,
-0.0223584622,
0.0790968463,
-0.0506504141,
0.0310702585,
-0.0310303252,
-0.0631023869,
-0.0045055519,
-0.0067998702,
-0.0406251699,
0.0077597201,
0.0239683315,
-0.0407342874,
-0.0362338983,
-0.0127937598,
-0.0124675129,
0.0330134295,
-0.0109366802,
0.0025947278,
-0.0240829512,
0.031531021,
0.0453144051,
0.0598662868,
0.0647254512,
-0.0319972225,
0.0070247664,
0.002888364,
0.0203646626,
0.0279734079,
-0.0530655719,
0.0062752413,
0.0475148857,
-0.0571016297,
-0.0646460503,
0.0435637012,
0.0481363982,
-0.0451954007,
0.009103098,
0.1231099591,
-0.0190345403,
-0.0416987799,
0.0144760599,
0.0282264743,
0.0445726402,
0.0374147482,
0.0184252616,
-0.0550416596,
-0.0373955444,
0.0159635879,
0.0233224444,
0.021130817,
-0.0224887561,
0.015788842,
-0.0901489183,
0.041379828,
0.0359245203,
0.0177936964,
0.0088821305,
-0.032686241,
0.0396438353,
-0.0639855266,
0.0315788612,
0.0132035036,
-0.0589411445,
-0.0113968207,
-0.0035488561,
-0.0027003435,
-0.0495855734,
0.0493751727,
-0.0366695113,
0.0339268856,
0.0460566767,
0.0383878052,
-0.020483844,
-0.0343062617,
0.0725045502,
-0.0024624285,
0.0121620195,
-0.0465034731,
0.0332558155,
0.0465585552,
-0.0116619766,
0.0291905627,
0.0242882483,
-0.013951662,
0.0146684404,
-0.0005983323,
0.0430291332,
-0.0132444697,
-0.0414779969,
0.019126961,
-0.0214163195,
-0.0141797541,
0.0639471188,
0.0512973592,
-0.0284600146,
-0.0607814454,
-0.0115060313,
-0.000774726,
-0.0379250422,
-0.0624710098,
0.0127555067,
0.018866457,
0.0258967672,
-0.0414896049,
0.0405027159,
0.0118813086,
0.02946903,
0.0134034744,
0.0357032381,
0.0347672291,
0.0096671665,
-0.0120474864,
0.0200526938,
-0.0489329025,
0.0073429947,
-0.0889377967,
-0.0183292236,
0.0666903257,
-0.0147594484,
-0.0039142598,
-0.0335280113,
-0.0895785764,
-0.0631162226,
0.011015526,
0.039992746,
0.0243552439,
0.0302484091,
0.0026786318,
-0.0910178125,
0.031391222,
0.0396476947,
0.0055914796,
0.0530927256,
0.0353376977,
-0.0075014904,
-0.0324292742,
0.0590867922,
-0.0342824869,
0.0232667383,
0.0512263589,
0.0253606904,
-0.0297089014,
-0.0655218661,
-0.0386988223,
0.0363778956,
0.0308925826,
-0.0326874144,
-0.0183128621,
-0.0209337249,
0.0194132607,
0.0650913268,
0.0161080584,
-0.034511108,
0.0105023552,
-0.0028437402,
-0.0242790058,
-0.0061920574,
-0.0130993295,
-0.0057828128,
-0.011967795,
-0.0556267425,
0.0576980785,
-0.0022389488,
-0.0017832372,
0.0689256191,
-0.0103626391,
-0.0050955485,
0.0728362128,
0.0097503727,
0.0461906306,
-0.0167770665,
0.0196403153,
0.0068280273,
0.0442473479,
-0.0698109865,
0.0323150642,
-0.0509916656,
-0.0700682998,
-0.0133963879,
0.0216561053,
0.0709280819,
-0.0472291857,
0.0699148402,
0.0094536301,
-0.0124111418,
-0.0326637551,
0.0251586419,
0.0028462622,
0.0234241821,
0.0176700037,
0.0243824888,
-0.0408771485,
0.044637993,
0.0733557791,
0.0043207323,
0.0440721996,
0.028625695,
0.0560243838,
-0.002420377,
-0.0498092361,
-0.0402752236,
-0.0165996645,
-0.0053154668,
-0.0056822905,
-0.0450999364,
-0.058590278,
-0.0144003425,
-0.0031689622,
-0.0119501678,
-0.0310226493,
0.0321108624,
0.0538230762,
-0.0066856979,
0.0093929777,
-0.0178559851,
-0.0220110472,
0.0308272112,
-0.0156135075,
-0.0027507702,
-6.219353266e-33,
0.0070214393,
0.0241430067,
0.0014576092,
0.021596279,
-0.0039725876,
-0.00122708,
-0.0201535281,
-0.0130149303,
-0.0543599017,
-0.0132884448,
-0.0292367246,
-0.0033875541,
0.01423149,
0.0376826376,
0.0064671803,
-0.0206984598,
0.0672103837,
-0.0178944115,
-0.0284046009,
0.0362078324,
0.0024035478,
0.0286610909,
0.0531563871,
-0.0410366915,
-0.0504734926,
0.0137563506,
-0.0202133339,
0.0262607653,
-0.037800312,
0.0676752999,
0.0327715501,
0.0201366823,
0.0177181866,
-0.0535040163,
-0.0397954099,
-0.0384211577,
-0.0523609482,
-0.0343834981,
0.0047754645,
0.0015366103,
-0.0109174047,
-0.0151726799,
0.0698017776,
-0.0159056839,
-0.0185731947,
-0.0421206057,
0.0260803569,
-0.0255789757,
-0.0395908542,
-0.0249370318,
-0.0276753679,
0.0327253528,
0.0001679961,
-0.0133533869,
0.049266845,
-0.0392783657,
-0.022976784,
0.0546788946,
-0.0252796579,
-0.0057293186,
0.0652452707,
0.0832388923,
0.0515431315,
0.0406716876,
-0.0115741258,
0.0415873118,
-0.0145001812,
0.0224569924,
0.0198355261,
-0.0267858189,
0.0594535694,
0.0874612033,
0.0161691159,
0.011693378,
0.0248189177,
-0.0320217311,
-0.088783361,
-0.0164844263,
0.0392843038,
0.0304774977,
-0.0180161614,
0.0363489836,
-0.0006437136,
-0.0303887464,
-0.0643964335,
-0.0309757106,
-0.0237407535,
-0.0006787845,
0.0098567288,
0.0045386627,
-0.0151385032,
-0.0408574864,
0.0289663542,
-0.0534100756,
-0.0239944365,
0.0253723916,
-0.0024766643,
0.018543195,
-0.0063760215,
0.0066786325,
-0.0439460911,
-0.0671967566,
-0.0050313124,
-0.0112010483,
0.0274293162,
-0.0179523472,
0.0060480358,
0.0156932976,
-0.0561443157,
-0.0097594559,
-0.0063851262,
-0.012331143,
0.0389023945,
0.0482015796,
0.0269459132,
0.003433408,
0.0256069638,
0.044866167,
0.0101966485,
-0.0286343731,
-0.0199748799,
0.0287761111,
0.0390328839,
0.0418305323,
-0.0333659798,
-0.0004543586,
-0.0493366532,
0.0135016115,
0.0555549376,
-0.0215552971,
-0.0219837278,
0.0258886237,
2.65e-7,
0.0383319035,
0.051549878,
-0.0190536957,
0.0109419087,
0.0293441825,
0.0460259765,
-0.0234678295,
-0.0112015,
0.0012086375,
0.0001135924,
0.0133968666,
0.0302423667,
0.0936386809,
0.0043269135,
-0.0770388395,
0.0290360022,
-0.0676404387,
0.0310785864,
-0.0237190984,
-0.0458580181,
-0.0372337922,
0.0998190194,
0.0105200661,
-0.0279852841,
-0.0092718042,
-0.0345479958,
0.0316855013,
-0.0579529479,
0.0347835049,
0.068796955,
-0.0051636882,
0.0378255509,
-0.016984826,
0.0139707848,
-0.0305125229,
0.0358187407,
-0.0310244057,
0.0914117321,
0.0165511966,
0.0150233749,
0.014014611,
-0.0207158402,
0.0034225646,
-0.0428374857,
0.0862501934,
-0.0410832949,
-0.0020086032,
-0.048590187,
0.0348276459,
-0.0465046689,
0.0315974504,
-0.0423130356,
0.0031132274,
0.0656010062,
0.0218464248,
-0.0257928614,
-0.0348309614,
-0.0762976184,
-0.0038187006,
0.0403558314,
-0.0289660562,
-0.0372962318,
-0.0396574363,
-0.0327812433,
0.0361807197,
-0.0260138791,
0.0260282885,
2.485258166e-34,
0.0376027711,
-0.0439799204,
0.0297090542,
0.0086743366,
0.0134794442,
0.002771277,
-0.0179030579,
0.0033042284,
-0.0043641073,
-0.1023809686,
0.0027903947
] |
http://arxiv.org/abs/2402.02309v1 | http://arxiv.org/pdf/2402.02309v1 | Jailbreaking Attack against Multimodal Large Language Model | 2024-02-04T01:29:24 | 2024-02-04T01:29:24 | Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, Rong Jin | cs.LG | cs.LG, cs.CL, cs.CR, cs.CV | This paper focuses on jailbreaking attacks against multi-modal large language
models (MLLMs), seeking to elicit MLLMs to generate objectionable responses to
harmful user queries. A maximum likelihood-based algorithm is proposed to find
an \emph{image Jailbreaking Prompt} (imgJP), enabling jailbreaks against MLLMs
across multiple unseen prompts and images (i.e., data-universal property). Our
approach exhibits strong model-transferability, as the generated imgJP can be
transferred to jailbreak various models, including MiniGPT-v2, LLaVA,
InstructBLIP, and mPLUG-Owl2, in a black-box manner. Moreover, we reveal a
connection between MLLM-jailbreaks and LLM-jailbreaks. As a result, we
introduce a construction-based method to harness our approach for
LLM-jailbreaks, demonstrating greater efficiency than current state-of-the-art
methods. The code is available here. \textbf{Warning: some content generated by
language models may be offensive to some readers.} | Jailbreaking Attack against Multimodal Large Language Model: This paper focuses on jailbreaking attacks against multi-modal large language
models (MLLMs), seeking to elicit MLLMs to generate objectionable responses to
harmful user queries. A maximum likelihood-based algorithm is proposed to find
an \emph{image Jailbreaking Prompt} (imgJP), enabling jailbreaks against MLLMs
across multiple unseen prompts and images (i.e., data-universal property). Our
approach exhibits strong model-transferability, as the generated imgJP can be
transferred to jailbreak various models, including MiniGPT-v2, LLaVA,
InstructBLIP, and mPLUG-Owl2, in a black-box manner. Moreover, we reveal a
connection between MLLM-jailbreaks and LLM-jailbreaks. As a result, we
introduce a construction-based method to harness our approach for
LLM-jailbreaks, demonstrating greater efficiency than current state-of-the-art
methods. The code is available here. \textbf{Warning: some content generated by
language models may be offensive to some readers.} | [
0.0504380688,
0.090904519,
-0.0136630349,
0.0713371858,
-0.0517972372,
-0.0053888122,
0.0432343297,
0.0074541443,
0.044838883,
-0.0396637619,
-0.0198409669,
-0.0108740553,
-0.0147586456,
0.0018682769,
0.0549667701,
-0.0179948453,
0.0149655389,
-0.0486906543,
-0.0396523289,
-0.0223875586,
0.0151522802,
-0.0071159885,
0.0166470036,
0.0242412761,
-0.0070830546,
-0.0256528892,
-0.0535615236,
0.0449031144,
-0.0079516647,
-0.0327028073,
-0.0427732728,
0.0278021786,
-0.0225921255,
0.0473311283,
0.0000022375,
-0.0266268551,
-0.0475673005,
0.0331248827,
-0.0178896785,
0.0308505539,
0.0217274521,
-0.0304236021,
-0.0310117062,
-0.0064461897,
-0.0103667183,
0.0183024462,
0.0507497787,
0.0718516558,
0.0951045677,
0.0642358065,
-0.0041739112,
-0.0092130639,
0.0215597339,
0.0017104788,
0.0516536273,
-0.0309571587,
0.0135331685,
-0.0206753332,
0.022555491,
0.0097696735,
0.0138474638,
0.0113726659,
0.0165457875,
-0.0124523528,
-0.0781921521,
-0.0186668988,
-0.0077265478,
-0.0334535167,
-0.0102700414,
0.0524853729,
-0.0182263236,
0.0027277346,
0.0169168673,
0.0305402428,
0.0382988453,
0.0693452209,
-0.0238835271,
0.0599918328,
-0.0185618382,
-0.0142500382,
0.0007847825,
-0.0388155691,
0.0496437661,
-0.0050131627,
-0.0199063998,
0.034588486,
0.0529974066,
-0.0223305654,
-0.0162821673,
-0.0568718985,
-0.0034739911,
-0.0380978771,
-0.006883075,
-0.0060010022,
0.0184599981,
0.0118343737,
-0.0773254037,
-0.0173847042,
0.0242573712,
-0.0339660756,
0.017429525,
0.0034360164,
-0.0116214948,
0.038812276,
0.0099924104,
-0.025603747,
-0.0553190112,
0.0091827577,
0.0000127805,
0.0074014533,
0.0102490736,
-0.0157484598,
-0.0070661805,
0.0276464205,
-0.041790165,
-0.0012438206,
-0.018482592,
0.0562642254,
-0.0232732408,
0.0119928541,
-0.0819787681,
0.0120995706,
-0.0021215072,
0.0208274201,
0.0439850055,
-0.031119464,
0.0048439559,
-0.0405818708,
0.00255844,
-0.0661273748,
0.0063760448,
0.0210007988,
-0.0470471978,
-0.0179257132,
-0.0067806994,
0.1306649894,
0.0243097674,
-0.0125419479,
-0.0823811144,
-0.1056818441,
-0.0259763617,
-0.0435282327,
0.0175138414,
-0.0890852436,
-0.0231792759,
-0.0389196426,
0.0069201137,
-0.0310408473,
-0.0063502304,
-0.0047662705,
-0.0343463011,
0.0810780823,
0.0136961471,
-0.0368804336,
0.0045617232,
0.0434319116,
0.0241807848,
0.0816499665,
-0.0425011851,
0.0710251629,
0.0723334178,
-0.0297198687,
0.0441925302,
0.07187213,
-0.0154982973,
0.0405648723,
-0.0425724462,
-0.0169505905,
0.061773438,
-0.0230660848,
-0.0450172201,
0.0260263123,
-0.035436634,
-0.0221237801,
0.0828196481,
0.0532544702,
-0.0250037964,
0.0296213217,
-0.0266263243,
-0.0273612887,
0.0019047522,
0.0243874006,
-0.0656902492,
0.0082928808,
-0.107747972,
-0.017307736,
-0.0235052053,
0.0348254666,
-0.0089623127,
0.0198590104,
0.0255969465,
-0.019363232,
-0.0445784256,
0.0096313823,
0.009418346,
0.020606935,
0.013949981,
0.0556020588,
0.0191424415,
-0.0282462221,
-0.0361407064,
-0.0577117614,
-0.0159223005,
0.0114227151,
0.0526471213,
0.0160875786,
0.0223417785,
-0.0111315539,
-0.006040615,
0.0224188771,
-0.0317397639,
0.0112122856,
0.0159564298,
0.0048521929,
-0.0081372578,
-0.0217447076,
0.0238959417,
0.0078518838,
-0.0282156523,
-0.0106947999,
-0.0106325559,
-0.0355464332,
0.0170319248,
-0.0008091711,
0.0105525861,
-0.0231305454,
-0.0010887421,
0.0400714763,
0.0173135325,
0.0094668381,
-0.0219862089,
0.0067452388,
0.0658284649,
-0.0000935444,
0.0340687931,
-0.0042260308,
-0.0174783617,
-0.0077149929,
0.0363987498,
-0.0778272524,
0.0170950778,
0.0588220023,
-0.015851479,
0.0009929207,
0.0444825888,
-0.02686744,
0.0565463975,
-0.0004237531,
0.0301075075,
0.0147659536,
-0.0040300353,
-0.0351751335,
-0.0083083222,
-0.0274031963,
0.0066680131,
0.015471654,
-0.0060684448,
0.0154516213,
0.0040700054,
-0.0600204095,
0.0436159968,
-0.0284835659,
0.0215730425,
-0.0032465961,
0.02582011,
-0.0098210732,
-0.040088892,
-0.0639568642,
-0.0292452611,
-0.0139953326,
0.0094863046,
0.0374337584,
0.0565078072,
-0.0322646424,
-0.0228434969,
-0.0152611174,
-0.0231420361,
0.0008443774,
0.0299236365,
-0.0032567373,
-0.0104338275,
-0.0510271452,
0.0082172574,
-0.0340760276,
0.0220161229,
0.0728389844,
-0.0409416854,
-0.0229367334,
-0.0180988181,
-0.0141376704,
-0.0163096562,
0.0252390169,
-0.0286953188,
-0.0382987224,
0.042558603,
-0.0306520779,
-0.0104426695,
0.0683053955,
0.0152920084,
0.0575647578,
0.0441121981,
-0.0051691229,
-0.0313289836,
-0.0351064689,
0.0446724035,
-0.0453096367,
-0.028418798,
-0.0810890049,
0.0316994116,
0.0143387923,
0.0353716463,
0.0185460392,
-0.0016151941,
0.0503630154,
0.0175952241,
-0.0479839109,
-0.011480676,
0.0298135914,
-0.0263805687,
0.0000160145,
0.0068735108,
-0.0865781233,
0.0416405685,
-0.0460157581,
0.0241202433,
0.0310404468,
-0.0491473824,
0.0125092091,
0.0568605103,
0.0004352331,
0.0522330627,
0.0592128187,
-0.0237016529,
-0.0206829682,
0.0071599749,
0.0504756868,
-0.0123551069,
0.018055534,
-0.0539998412,
-0.017134171,
-0.0143757807,
-0.0365068726,
-0.0299211144,
-0.0707016885,
-0.0028458731,
-0.062562421,
-0.0013603654,
-0.0439590365,
-0.0100688525,
-0.0358748958,
0.0252110325,
-0.0317964666,
-0.0229896139,
-0.0047991779,
-0.0272403043,
0.0459706113,
0.0378809087,
-0.0341630355,
-0.017143216,
-0.0144344186,
0.038312342,
0.0975410864,
0.046924118,
0.0016178174,
0.00681506,
0.0490530543,
0.0287401695,
0.0499276184,
-0.020439012,
0.0169422124,
0.0558742583,
-0.0138655808,
0.0010895734,
0.0269065686,
0.014475368,
-0.0273039341,
0.0078138113,
0.0335332565,
0.0379940458,
0.0079751061,
0.027797645,
-0.0846443474,
-0.0445538051,
0.0087113818,
0.066878058,
-0.040089339,
-0.0301606413,
-0.0312522165,
-0.0894864649,
0.0145011907,
0.029830493,
0.0233606137,
0.0158355869,
-0.0127513278,
-0.0191724021,
0.0372503884,
0.003798957,
-0.0221260879,
-0.0163682941,
0.0140266232,
-0.0059607476,
0.0138619114,
0.0083630094,
-0.0101347063,
-0.0243053436,
0.045527257,
0.0477024093,
-0.0203438122,
-0.0392062888,
0.0153491385,
0.0534510389,
0.027885234,
-0.0185657348,
-0.0006093441,
-0.000076524,
0.0207784772,
-0.0204361398,
-0.0200897511,
0.0561862253,
0.0240471307,
-0.0215732846,
0.030342564,
0.0390625633,
-0.0016896382,
0.0033575615,
-0.0229502637,
-0.0408910289,
0.01437495,
0.0333222784,
0.02310008,
-0.044780314,
-0.0289237294,
-0.097330451,
-0.0408491157,
-0.0012563361,
-0.0142821427,
0.0292582735,
0.0138895875,
0.0079103503,
-0.015970625,
0.0345633887,
-0.0105002895,
0.080506362,
0.0334287882,
0.0036566379,
-0.0397006646,
-0.0050793481,
-0.0220643803,
-0.0182106495,
-0.0182304829,
0.0262652487,
-0.0773278698,
0.0453226231,
0.0179853383,
-0.0366682261,
0.0099036302,
0.0080566416,
-0.0641035438,
-0.0812083036,
0.0140096806,
0.0626523942,
0.0261636339,
-0.0028231335,
-0.0083241491,
-0.071148321,
0.0227702148,
0.0972905606,
-0.0565448366,
0.0322900563,
0.0677121431,
-0.0021643138,
-0.0075275097,
0.0217576865,
-0.0287791789,
-0.0195042416,
0.0806641132,
-0.0187314302,
-0.0434199683,
0.0094200531,
0.0177014396,
0.0688707829,
-0.0345083326,
-0.0200683624,
-0.013821464,
0.0466577746,
0.0255365558,
-0.0033761906,
-0.031476859,
-0.0193707701,
0.0245516412,
0.0067525418,
-0.0238474254,
0.052741684,
-0.0164214764,
0.029825775,
-0.0126013765,
-0.0338327251,
0.0280880146,
0.0371557213,
-0.0007468914,
0.0745937079,
0.0081874328,
-0.0108664678,
0.0671337619,
0.0184992142,
0.0342130028,
-0.0103563434,
0.0656014606,
0.0098237693,
-0.0309065543,
-0.0218143985,
0.0593290329,
-0.0418663509,
-0.0560691804,
0.0127078434,
0.0867288485,
0.0188544318,
0.0124533251,
-0.0169469379,
-0.0064616115,
0.0007521795,
-0.0484750941,
-0.0087600118,
0.0172989871,
-0.0494591855,
0.0328744352,
-0.0064911679,
0.0081367949,
0.0346353017,
0.0296813715,
0.0319838487,
0.0528731458,
0.0288079903,
-0.0113832355,
-0.0692169741,
0.0105857486,
0.0709217712,
0.0261162817,
-0.0380908027,
-0.0364852361,
-0.0472024232,
-0.0650333539,
-0.0506685972,
-0.0076290951,
0.021880297,
0.0128816664,
-0.0449974127,
0.0445376858,
0.0265528746,
0.0194509141,
-0.0619650483,
0.0371076129,
0.0131889954,
-0.058096271,
0.0534760989,
-6.517429246e-33,
0.0273600481,
-0.0686291233,
-0.054082863,
-0.0370026864,
0.0269525778,
-0.0047286502,
0.0140214292,
0.0329969898,
0.00780519,
-0.0415722542,
-0.0180178229,
-0.0343013518,
0.0009614737,
0.0143576916,
-0.0473075286,
-0.0203191154,
-0.0105117792,
0.0249134339,
-0.0251429155,
0.0150433872,
0.0574003831,
0.0074279266,
0.0071579781,
-0.0258394368,
0.0384776779,
-0.0574735776,
-0.0597702302,
0.0335904695,
-0.0352846272,
0.0768998638,
0.0018929834,
0.0257300269,
-0.0011616673,
0.0008241458,
0.0002603699,
0.0586801805,
-0.0103596803,
0.008889812,
0.0278766751,
0.0095595624,
0.0012761607,
-0.0501356497,
0.0156227872,
-0.0264103785,
0.0124169001,
-0.0146076698,
-0.0151465079,
-0.0015602709,
-0.0010875447,
0.0117979813,
-0.0028824999,
-0.0014799548,
-0.042856507,
0.004129108,
0.0171110313,
-0.0158637501,
0.015299513,
-0.0148360552,
-0.0053781918,
0.0388325527,
0.0497172326,
-0.0191336032,
0.0482612923,
0.0140647646,
0.0006956054,
0.0613893606,
-0.0030100823,
-0.0349656641,
0.0237962455,
-0.0157791469,
-0.0057255239,
0.0788912773,
0.0354068801,
0.0813695565,
0.0406288505,
-0.0015335222,
-0.0262442585,
-0.0067645549,
0.018477397,
0.0091376817,
0.0288907774,
-0.0207053833,
-0.0252042729,
0.0006776419,
-0.026284514,
-0.0405038521,
-0.030302301,
-0.0419906974,
0.0278567132,
0.0095334984,
-0.0533387363,
-0.0373037085,
0.0345972143,
-0.0621499494,
-0.0420563184,
-0.0126187522,
0.0002575352,
-0.0080846502,
-0.0229797363,
0.011532384,
0.0427204967,
0.029194463,
0.0217284467,
-0.0761715621,
0.018703565,
0.0114748608,
0.0114336675,
0.0256286822,
-0.0692062601,
-0.0250246543,
-0.0240976233,
0.0265804287,
0.0077893357,
-0.0159735512,
0.0414741188,
-0.0519437678,
-0.0094160493,
-0.0333234295,
0.0395652056,
-0.0021980056,
-0.0469723418,
0.0072235642,
-0.0297040623,
0.0204028711,
-0.0025266549,
0.0019949356,
-0.0100538619,
-0.0127135431,
0.0687116385,
-0.0622748137,
-0.0050788135,
0.0010599374,
3.087e-7,
-0.0401873924,
-0.0293808598,
0.0166012309,
-0.0013548482,
0.0259132348,
0.088437818,
-0.0179412048,
0.0517028458,
-0.0120000821,
-0.0708651245,
-0.0522583723,
-0.0367198996,
0.0607289225,
0.0186252687,
-0.0678882524,
0.0421019979,
-0.0034200153,
-0.0007987866,
-0.0650963411,
0.0110568581,
0.0415140465,
0.0685865209,
0.0025330905,
-0.0490724631,
0.0264678504,
-0.0514091589,
-0.00515074,
-0.0351511352,
-0.0328676067,
-0.0329288654,
0.0192988198,
0.0123177795,
0.0219734628,
0.0181502867,
-0.0246414896,
-0.0572379753,
-0.0258895252,
0.0278857015,
0.0376839153,
-0.0124721965,
-0.0277516786,
-0.0438708737,
-0.0225297436,
-0.0444090217,
0.0814623609,
0.0629483461,
-0.0441928878,
-0.0275905877,
0.006174692,
-0.078337647,
0.0062674149,
-0.0657849833,
-0.0065784757,
0.0131157152,
0.0246527977,
-0.0045325342,
0.0212872233,
-0.0395841971,
-0.0068681622,
-0.0111967409,
-0.0135377077,
-0.0334046148,
-0.022811586,
-0.000320783,
0.0490879528,
-0.0450993702,
0.0037142034,
3.128337376e-34,
-0.0108479299,
-0.0283377208,
0.0325751342,
0.0337998383,
0.0269842017,
0.0053574713,
0.0183769092,
0.0097255427,
0.0022064196,
-0.0705643892,
-0.0249256194
] |
http://arxiv.org/abs/2402.05121v1 | http://arxiv.org/pdf/2402.05121v1 | Large Language Model for Table Processing: A Survey | 2024-02-04T00:47:53 | 2024-02-04T00:47:53 | Weizheng Lu, Jiaming Zhang, Jing Zhang, Yueguo Chen | cs.AI | cs.AI, cs.CL | Tables, typically two-dimensional and structured to store large amounts of
data, are essential in daily activities like database queries, spreadsheet
calculations, and generating reports from web tables. Automating these
table-centric tasks with Large Language Models (LLMs) offers significant public
benefits, garnering interest from academia and industry. This survey provides
an extensive overview of table tasks, encompassing not only the traditional
areas like table question answering (Table QA) and fact verification, but also
newly emphasized aspects such as table manipulation and advanced table data
analysis. Additionally, it goes beyond the early strategies of pre-training and
fine-tuning small language models, to include recent paradigms in LLM usage.
The focus here is particularly on instruction-tuning, prompting, and
agent-based approaches within the realm of LLMs. Finally, we highlight several
challenges, ranging from private deployment and efficient inference to the
development of extensive benchmarks for table manipulation and advanced data
analysis. | Large Language Model for Table Processing: A Survey: Tables, typically two-dimensional and structured to store large amounts of
data, are essential in daily activities like database queries, spreadsheet
calculations, and generating reports from web tables. Automating these
table-centric tasks with Large Language Models (LLMs) offers significant public
benefits, garnering interest from academia and industry. This survey provides
an extensive overview of table tasks, encompassing not only the traditional
areas like table question answering (Table QA) and fact verification, but also
newly emphasized aspects such as table manipulation and advanced table data
analysis. Additionally, it goes beyond the early strategies of pre-training and
fine-tuning small language models, to include recent paradigms in LLM usage.
The focus here is particularly on instruction-tuning, prompting, and
agent-based approaches within the realm of LLMs. Finally, we highlight several
challenges, ranging from private deployment and efficient inference to the
development of extensive benchmarks for table manipulation and advanced data
analysis. | [
0.0239972826,
0.0939287245,
-0.0140218139,
0.0298786573,
-0.0625495166,
0.036837738,
-0.0144844092,
0.0224169213,
0.0096402057,
-0.025826605,
-0.0186367817,
-0.0271065198,
-0.0483414456,
0.0199023504,
0.0209196024,
0.0165338479,
0.0449437313,
-0.0162827782,
-0.0361343808,
0.0174894612,
-0.0234176125,
-0.0445100591,
-0.0131352153,
-0.0173588134,
0.0032835065,
-0.0060729277,
-0.038962651,
-0.0219021887,
-0.0452871621,
0.0197754521,
-0.01762647,
0.0766049325,
0.020569481,
0.0435476713,
0.0000019324,
-0.0164420437,
-0.0488770045,
0.0144694168,
-0.0272293836,
0.0126653053,
0.0212171879,
-0.0325206108,
-0.0415636301,
-0.0081473058,
-0.0546381958,
-0.0231562667,
0.0228596739,
0.0297111329,
0.0678332672,
0.0581861474,
0.0055536767,
-0.0326779,
-0.0125043569,
0.0043919305,
0.0504608899,
-0.0656310245,
0.0418264866,
0.0107149091,
0.0135119874,
-0.0275124833,
0.0124310888,
0.030100368,
0.0298602488,
-0.0268112887,
-0.0279362332,
0.0264378786,
-0.0302428883,
0.0421105772,
-0.0121735772,
0.0548976958,
0.0205373038,
-0.0052155419,
0.0203775149,
0.0227476899,
0.0257110465,
-0.0208249781,
-0.0406248793,
0.0205745697,
0.0118482225,
-0.0212875474,
0.0138854263,
-0.0474920124,
-0.0198234916,
0.0082896613,
-0.0497148968,
0.0610290952,
-0.0122444937,
-0.054097753,
-0.0198772326,
-0.0253081061,
0.0403771661,
-0.0586641543,
0.0220485441,
0.0847673789,
-0.0164132938,
0.0279327314,
0.0155423135,
-0.0209626313,
0.0387370773,
-0.0585369579,
0.0354267061,
0.0315901078,
0.0042188582,
0.0414941646,
-0.0212707091,
0.0186717939,
-0.0109239537,
0.034888003,
-0.0161239337,
-0.0121274628,
-0.0340272449,
0.0242637061,
-0.0478784256,
0.0240596198,
-0.0565688722,
0.0149386032,
-0.0001424745,
0.0312861428,
-0.0183720458,
0.0958235338,
-0.0070524076,
-0.0003851016,
-0.0018697463,
-0.0010437168,
-0.0534515791,
0.0095679136,
-0.0272116158,
-0.0229778029,
-0.0272903237,
-0.0969967544,
-0.0147878313,
-0.0050747055,
0.0218931958,
-0.0286846254,
0.0677592829,
0.0843206272,
-0.0349696614,
0.0111022256,
-0.0482037403,
-0.016688304,
-0.0080082454,
-0.069292441,
-0.0065438971,
-0.0572440773,
0.0389927141,
0.0020880755,
-0.0201843921,
-0.0370116532,
0.014776079,
0.0552008525,
0.015736917,
0.0541052744,
0.0206917282,
0.0244708974,
-0.0035252133,
-0.0311446562,
-0.0124165742,
0.0612293556,
0.0289628115,
0.0285247993,
-0.00995088,
-0.0264599212,
-0.0029607331,
0.0572957024,
0.0095588956,
0.0025188103,
-0.0177222993,
0.0812387019,
0.0374065675,
0.0275746919,
-0.0220833104,
0.0490505844,
-0.0017912522,
0.0257213488,
0.0468311422,
0.0447917394,
-0.0039292541,
0.0717469975,
0.0088302055,
0.0049466565,
0.0461125374,
0.0003599407,
-0.0186182875,
0.0270638317,
0.0153805148,
0.0128349122,
-0.0049061733,
0.0270792861,
-0.0163875539,
-0.0449295789,
0.0177035984,
0.0066234721,
0.000300545,
0.0051160101,
0.0667951033,
0.0149301691,
-0.0350115709,
-0.0177174117,
0.0054223584,
-0.0571872368,
-0.0345924273,
-0.0461375192,
0.0226306077,
0.0610866919,
0.0363995321,
-0.0762177706,
-0.0161047969,
-0.0096616,
0.0262870807,
0.0109669268,
0.0510663167,
-0.0528154932,
-0.0485341735,
0.0037471657,
0.002002605,
0.0535753667,
0.0269563273,
-0.0231371392,
0.0456346422,
-0.0132469637,
-0.0145185627,
-0.0400030315,
-0.055149246,
0.0272359233,
0.0379584022,
-0.0322930738,
-0.0093869483,
0.027198365,
0.0390918627,
0.0281361081,
-0.0148652876,
0.0241514854,
0.0497320481,
-0.0079248073,
0.0212520696,
-0.036917612,
0.0127251549,
0.0321780518,
0.0544674881,
0.0478239469,
-0.082860209,
-0.0147116659,
-0.0366272144,
-0.0147189926,
0.0156425908,
-0.027234491,
-0.037478812,
-0.0385455415,
0.0211050343,
-0.0025385423,
-0.0103503885,
-0.0067731673,
0.0712384135,
-0.0037892156,
0.0166215654,
0.002639408,
-0.0856070817,
0.0231780559,
-0.046485506,
0.0133835282,
-0.0069485176,
-0.0178483725,
0.053776633,
-0.0063709021,
0.023776086,
-0.0575329475,
-0.0206647143,
-0.0764400885,
0.0462609157,
-0.0381413624,
0.038148433,
0.016174024,
0.0086852154,
-0.0251602549,
-0.036581073,
-0.0629841089,
-0.0083559677,
0.0035608879,
0.0157712419,
0.0700857118,
0.0051885084,
-0.0431309454,
-0.0283703282,
-0.0231412034,
-0.0351416357,
0.0048428462,
-0.0076308521,
-0.0282454491,
0.002808176,
-0.0179572944,
-0.0106711863,
0.0513497926,
0.0259716883,
-0.0118390014,
-0.0074057486,
-0.0356017314,
0.0397212878,
0.0797906891,
-0.026228331,
0.031921912,
0.0318235233,
-0.0176297761,
-0.0132816192,
0.0348533988,
-0.0107537936,
-0.0053477022,
0.0008398025,
-0.0564959086,
0.0283140857,
0.0366986133,
-0.0472500771,
-0.0064266743,
0.0064222459,
0.0091181211,
0.0262360256,
0.0056615039,
-0.0484239571,
0.0659796447,
-0.0101221344,
0.0050705723,
-0.0190597512,
0.0117608681,
0.0248222984,
-0.0416147821,
0.008203893,
-0.0263726339,
-0.0249293949,
-0.0603069216,
0.0875826627,
0.0265398212,
0.0108798668,
-0.0493447147,
-0.0217248034,
-0.0286167786,
0.0000927147,
0.014168378,
-0.046010986,
0.1123829037,
-0.054828994,
0.0205050446,
-0.0171201397,
-0.0321071558,
-0.0184664782,
-0.0351276956,
-0.0484621525,
0.0257206801,
0.0131177958,
-0.0249184482,
-0.0392111354,
-0.0330084041,
0.0556249805,
0.0496409014,
0.0067573283,
-0.006341218,
-0.0407142825,
0.0164688993,
0.0199889168,
0.0322311036,
-0.0557207987,
-0.02228266,
0.0304597598,
0.0715376213,
0.0143738193,
-0.0456679948,
-0.0206975657,
0.0194304455,
0.0068518221,
-0.0625628009,
0.0336040743,
0.0399401635,
-0.0429635122,
-0.0689458027,
-0.0190173723,
0.0718939304,
0.0267598685,
-0.0021804604,
-0.0106982514,
0.007871707,
0.0003774162,
0.0223095622,
0.0661732778,
-0.1104787663,
-0.0148448301,
-0.0072142575,
-0.0216055587,
-0.0009675184,
-0.0501176752,
-0.0308563225,
-0.1202817857,
0.0593535192,
-0.0359046496,
-0.0381390937,
-0.082105957,
-0.0582676418,
0.0143171428,
-0.0259557758,
0.0203136299,
0.0744095445,
-0.0263971519,
-0.0060993549,
0.0191529561,
-0.0390030965,
-0.0441876724,
0.0024147853,
-0.0627411678,
0.0370779671,
0.0452969,
0.0224476382,
-0.0357027315,
-0.0010830648,
-0.0033629446,
0.0292005464,
-0.0003061605,
-0.0109862164,
0.0461106002,
0.0067206589,
-0.0073856572,
-0.019899033,
0.0004016802,
-0.0167286713,
0.0132033136,
0.0458912402,
0.0899795964,
-0.0213455055,
-0.0062516755,
0.0385267399,
-0.0300204828,
-0.006263799,
0.0696869269,
0.0268669967,
0.0011675898,
-0.0615431145,
-0.0300599467,
-0.0894843712,
-0.0435989834,
0.0237144083,
0.0290003978,
-0.0326987803,
-0.0138633773,
-0.0350847133,
-0.0042644064,
0.0975430384,
0.0382819884,
-0.0018555233,
0.046912875,
0.0050339745,
-0.0126123559,
-0.0516468808,
-0.0424070135,
-0.0822287649,
0.0413777344,
-0.0234024543,
-0.0512537658,
0.0042574704,
-0.0143767567,
-0.0267318059,
-0.0110106282,
-0.0662429929,
-0.0216957368,
0.0108520081,
0.0263449401,
0.0242128149,
0.0097600734,
-0.0043964456,
-0.0058614803,
-0.0299018752,
-0.0003468448,
-0.0836258754,
-0.0276276283,
-0.0152920661,
-0.0497909859,
0.0152801862,
0.0311800744,
-0.022807762,
-0.0046182433,
-0.0168193094,
0.0236107446,
-0.022357028,
0.0134765571,
0.0429863073,
0.0421538539,
-0.0371160433,
0.0653045624,
-0.0638678297,
-0.0028027457,
0.0501489602,
-0.0612387285,
-0.0498379916,
0.0117385341,
-0.0101088006,
0.0075709801,
0.0102643119,
-0.0021150571,
0.0065642735,
0.0222206935,
0.0126082767,
-0.0163546149,
0.0253765583,
0.0170671567,
-0.009864416,
0.0327871852,
-0.0206541773,
-0.0122806747,
-0.0239959341,
0.0407453291,
0.0340622514,
0.0151684098,
0.0103794904,
0.0277914144,
-0.0408814028,
-0.0514798909,
0.027633369,
-0.0628544763,
-0.0434173197,
0.0139399106,
0.0223726742,
0.0402319171,
0.0056618094,
-0.007813435,
-0.041234009,
-0.028777549,
0.0212125238,
-0.017742265,
0.033750575,
-0.004575097,
0.0192477889,
0.0087927189,
0.0038655985,
-0.0233939495,
0.0785828307,
0.0412663668,
0.0237316452,
0.0072079394,
-0.0097743878,
-0.0711613446,
-0.0098092118,
-0.0173172224,
-0.0294375606,
-0.0327802561,
-0.0365467668,
-0.0348178856,
-0.0523009375,
0.0098952949,
0.0707479566,
-0.0296120904,
-0.0177429244,
0.0220029205,
0.059985172,
0.0712954476,
0.007609467,
-0.0547503233,
0.0413000174,
0.0037768437,
-0.0410493091,
0.0045420714,
-5.984738185e-33,
-0.0183824617,
-0.0200588852,
0.0462478884,
-0.0161529444,
0.0172683895,
-0.0318483151,
-0.0453927815,
-0.0343056247,
-0.0042426614,
-0.0016397347,
-0.0453313366,
0.0367612727,
0.031710837,
-0.025442645,
0.0223899838,
-0.0602507852,
0.0387847163,
-0.025060365,
0.014213087,
-0.0155180851,
-0.0220330898,
0.0195739157,
0.0062632607,
-0.0045905472,
0.037066251,
0.0013142206,
-0.0127200559,
0.0106845126,
-0.0230761655,
0.0119344601,
-0.0424502045,
0.0255556535,
0.0241014808,
-0.0403159373,
-0.0239095408,
0.0140953017,
-0.14342843,
-0.0210645441,
-0.0116838384,
0.0296783,
0.0024987618,
-0.0244329125,
0.0173752476,
-0.0033541266,
-0.0296049248,
0.0240775421,
0.01701748,
-0.0361941867,
0.0227661654,
0.0247759279,
-0.0221221205,
0.0171435997,
0.0137154311,
0.01848731,
0.0759514943,
-0.0095079066,
0.0251843799,
-0.0569124855,
-0.0187576078,
0.0368338861,
0.0538859479,
0.0143844988,
0.0547645055,
-0.0268050227,
0.0376614667,
0.013970213,
0.0384802967,
0.0358793326,
-0.0262958277,
0.0191868134,
-0.0426840521,
0.0591966063,
0.0247008987,
0.001772853,
0.0084158629,
0.0252575111,
-0.0681663007,
0.0635813549,
0.0316325165,
0.0350136682,
-0.0170055386,
-0.0321221016,
-0.0372970663,
-0.0068043447,
-0.0048933825,
0.050835494,
-0.0324296132,
0.0000915305,
0.0115230205,
-0.0016837366,
0.0183409285,
0.0016096792,
0.0007672759,
-0.002283639,
-0.05051177,
0.0246227682,
0.0166899841,
0.0092205312,
0.0147533203,
0.0010096162,
-0.0337618068,
-0.0179932024,
-0.0196604524,
-0.0253082905,
0.014708939,
0.0576263182,
0.0241775941,
0.030188527,
-0.0717899427,
0.0028903508,
-0.0432174951,
0.0607461073,
0.0216501541,
0.0276070703,
-0.001569039,
0.0154458582,
0.02869609,
0.0637616292,
0.0294331089,
-0.0056730653,
-0.0048594507,
0.013698739,
0.0012481725,
0.0020679862,
0.0152883362,
0.0062634507,
-0.0487198867,
0.0302331969,
0.0000602782,
-0.0370928049,
0.029617738,
0.0018200368,
2.518e-7,
0.0562781803,
0.0289736651,
0.0513443202,
0.0584150404,
-0.0016407161,
0.0087455697,
-0.0261283498,
-0.013568949,
-0.0393327586,
-0.0064806947,
0.0506103039,
0.0130469091,
0.043894276,
0.0334261842,
-0.0363034159,
-0.0271101166,
-0.0407525636,
-0.0049061649,
-0.0634691119,
0.0158318821,
0.0237244293,
0.0815282539,
-0.0652638897,
-0.0073681329,
-0.0159798171,
-0.0376039371,
0.0170428138,
-0.0511965677,
0.030014623,
0.0391138904,
0.0417455733,
-0.0120687727,
0.0110599427,
0.0438052677,
-0.0114993211,
-0.000529752,
-0.0314934067,
0.0647940561,
0.0001017339,
0.0770532563,
0.0068701669,
-0.0920376554,
0.0242447797,
-0.0514121167,
0.1014524847,
-0.0413000397,
-0.0073250616,
0.0710228309,
-0.009229783,
-0.0013668867,
-0.0214326736,
0.0040170578,
-0.0176269468,
-0.0005661058,
0.0153849991,
-0.0035550131,
0.0184885729,
-0.0291867163,
-0.009975424,
0.0064324387,
-0.022595549,
-0.0343537629,
-0.0076487688,
-0.0144686494,
-0.0220376682,
-0.042730134,
-0.013166693,
2.526109121e-34,
0.0520113595,
-0.0015038907,
0.0249581672,
0.0526894368,
0.0645266697,
-0.02224271,
-0.0177271869,
0.014078957,
0.0167140607,
-0.0928212628,
-0.0335140787
] |
http://arxiv.org/abs/2402.02302v1 | http://arxiv.org/pdf/2402.02302v1 | Predicting positive transfer for improved low-resource speech recognition using acoustic pseudo-tokens | 2024-02-03T23:54:03 | 2024-02-03T23:54:03 | Nay San, Georgios Paraskevopoulos, Aryaman Arora, Xiluo He, Prabhjot Kaur, Oliver Adams, Dan Jurafsky | eess.AS | eess.AS, cs.CL | While massively multilingual speech models like wav2vec 2.0 XLSR-128 can be
directly fine-tuned for automatic speech recognition (ASR), downstream
performance can still be relatively poor on languages that are
under-represented in the pre-training data. Continued pre-training on 70-200
hours of untranscribed speech in these languages can help -- but what about
languages without that much recorded data? For such cases, we show that
supplementing the target language with data from a similar, higher-resource
'donor' language can help. For example, continued pre-training on only 10 hours
of low-resource Punjabi supplemented with 60 hours of donor Hindi is almost as
good as continued pretraining on 70 hours of Punjabi. By contrast, sourcing
data from less similar donors like Bengali does not improve ASR performance. To
inform donor language selection, we propose a novel similarity metric based on
the sequence distribution of induced acoustic units: the Acoustic Token
Distribution Similarity (ATDS). Across a set of typologically different target
languages (Punjabi, Galician, Iban, Setswana), we show that the ATDS between
the target language and its candidate donors precisely predicts target language
ASR performance. | Predicting positive transfer for improved low-resource speech recognition using acoustic pseudo-tokens: While massively multilingual speech models like wav2vec 2.0 XLSR-128 can be
directly fine-tuned for automatic speech recognition (ASR), downstream
performance can still be relatively poor on languages that are
under-represented in the pre-training data. Continued pre-training on 70-200
hours of untranscribed speech in these languages can help -- but what about
languages without that much recorded data? For such cases, we show that
supplementing the target language with data from a similar, higher-resource
'donor' language can help. For example, continued pre-training on only 10 hours
of low-resource Punjabi supplemented with 60 hours of donor Hindi is almost as
good as continued pretraining on 70 hours of Punjabi. By contrast, sourcing
data from less similar donors like Bengali does not improve ASR performance. To
inform donor language selection, we propose a novel similarity metric based on
the sequence distribution of induced acoustic units: the Acoustic Token
Distribution Similarity (ATDS). Across a set of typologically different target
languages (Punjabi, Galician, Iban, Setswana), we show that the ATDS between
the target language and its candidate donors precisely predicts target language
ASR performance. | [
-0.0141227413,
0.0803386122,
-0.0276975315,
0.0037792979,
-0.0557433069,
-0.0079666767,
0.0173331238,
0.0143425185,
-0.0707732588,
0.0025363585,
-0.0528638177,
-0.0387179926,
0.0315943249,
0.0088720657,
0.0351598002,
-0.0665815547,
0.0263934657,
-0.0649951771,
0.0338314474,
-0.0101752328,
-0.0174561404,
-0.002728713,
-0.0277386252,
0.0499868914,
-0.0368360467,
0.0412211046,
0.0249927621,
0.018164495,
0.0076665557,
-0.0369101614,
-0.0619340092,
0.018126281,
0.0001205258,
0.0053677433,
0.0000017674,
-0.0605247281,
-0.0010146574,
0.0012246153,
-0.0268696733,
0.0259977244,
0.0395163,
0.0240604058,
-0.004231357,
0.0425528511,
0.0167838428,
-0.0144306468,
0.0594491139,
0.0243722517,
-0.0052087083,
0.0443178229,
-0.0239145793,
0.0010469521,
0.0010896552,
0.0099947499,
-0.0104746642,
-0.0262928624,
0.0100928536,
-0.0400062539,
0.0167375803,
-0.0420205966,
-0.032514248,
0.0202348623,
-0.0076220725,
-0.0069559398,
0.0263462514,
0.0091153141,
-0.1174261197,
-0.0438955836,
0.004635192,
0.0341538489,
-0.0276739597,
-0.0393422917,
-0.0001604566,
0.0071950126,
-0.008275101,
-0.0123519199,
-0.0296902694,
-0.0509822927,
0.0387772247,
0.017941894,
0.0408989526,
0.0143240392,
0.0252038669,
-0.022251565,
-0.0093559455,
-0.0101661729,
0.0273015033,
-0.0403664298,
-0.0062456364,
-0.0270949192,
-0.039975252,
-0.0197832417,
0.0388560854,
-0.0180873144,
0.0092199724,
-0.0094155148,
0.0073790061,
-0.0427808538,
0.0005837581,
0.0087169139,
0.0056732483,
0.06029284,
-0.0939579606,
-0.0017977806,
-0.0574115627,
0.0447423905,
0.0058874362,
0.0222953632,
-0.0534570105,
-0.0298722647,
-0.0238723848,
-0.0142164314,
-0.045443289,
0.018864112,
0.0017565425,
0.0006352177,
-0.0328804664,
-0.0047364109,
-0.0189725738,
-0.0291046519,
0.029241506,
-0.0235679001,
-0.0382483266,
-0.0079573169,
0.0242860336,
0.0347220786,
-0.0154947443,
0.0491388924,
-0.0143758599,
-0.0983402282,
0.0074987533,
0.029796036,
0.0145233516,
-0.0072225993,
0.010444778,
0.08453428,
-0.0018935828,
-0.0175558552,
-0.026860293,
0.035409756,
0.00668862,
-0.0337774903,
-0.0326318517,
-0.0336255282,
0.0152551085,
0.0356312618,
0.0121711455,
-0.0021049448,
-0.0348871313,
-0.0317765176,
0.0201522838,
0.0457121693,
0.0053308406,
0.0339438803,
0.0466999933,
-0.0157392584,
-0.0114500718,
0.0323282443,
0.0162574127,
-0.036653161,
0.0317334644,
-0.0570203811,
-0.0393960066,
0.0519980565,
0.024034705,
0.0373061635,
-0.0235827919,
-0.0035890071,
0.034895055,
0.0063943272,
0.0366925299,
0.0233233925,
-0.026143048,
-0.0027479888,
0.0398248062,
0.1013920754,
0.0071699447,
0.0182317793,
0.0179498829,
-0.0053496473,
-0.0358892046,
-0.0061951829,
-0.0232388172,
0.0031517472,
0.0411746576,
0.0047979252,
0.0648645461,
0.0285061691,
-0.0349004082,
0.0347752497,
-0.0171591341,
-0.0267979167,
-0.0180763006,
-0.0104647949,
0.066285871,
-0.0135836508,
-0.0465318523,
0.0524577051,
0.037124496,
-0.0366039723,
0.0575600117,
-0.0744918287,
0.0157058667,
0.0083527528,
-0.0206974819,
-0.0369291678,
0.0217875633,
0.0056145028,
-0.0101695424,
0.0110433176,
0.009723898,
0.016643187,
0.05183281,
-0.040242292,
-0.0070105987,
0.0514993705,
-0.0190844815,
0.064218834,
-0.0486413799,
0.0583194792,
0.0009796107,
-0.0782189816,
-0.006991012,
-0.0028044146,
-0.0066939266,
0.014125783,
0.0635739192,
0.0240669325,
0.0766585097,
0.0373151116,
-0.0068006143,
-0.0088601653,
0.0301559642,
-0.0556646734,
-0.0026328596,
0.0714485049,
-0.0679666027,
0.0187752824,
0.0124200135,
0.0136578484,
0.0697595254,
0.0233486425,
0.0176718179,
-0.0692271963,
0.0806102604,
0.0097389799,
0.1097155884,
-0.0284678116,
0.0593301728,
0.0003115885,
0.0164471995,
0.0361556783,
-0.0159799829,
0.035689652,
0.013613414,
0.0063714646,
0.0019948997,
0.0554163232,
-0.0529646091,
0.0123255383,
0.0378297195,
-0.0202817675,
-0.0035107662,
0.0290791634,
-0.0032600558,
-0.0096999044,
0.0070412075,
-0.0883206874,
-0.0226296335,
-0.0110701509,
0.0287790839,
-0.0238828473,
0.0323109105,
0.0027618543,
-0.0113215828,
-0.0349102616,
-0.0516387895,
-0.0008535931,
0.0182310361,
-0.0327169299,
-0.0159787741,
-0.0153905936,
0.0061689597,
-0.0276122056,
0.047342293,
0.0259758849,
-0.0130640212,
-0.0122809401,
-0.054888878,
-0.0324580073,
0.0089580491,
-0.0149792833,
0.002952659,
-0.0571652092,
-0.0358290114,
0.0150624244,
0.0160640199,
0.0225021411,
0.0051076156,
-0.0642465204,
0.0082051381,
0.0140981367,
-0.0360073149,
-0.027952591,
0.0225906,
0.0207584221,
-0.0148883108,
0.001065956,
-0.0092094988,
0.0026441845,
-0.004148663,
-0.0088185081,
-0.0735544339,
-0.0783384442,
-0.0469062179,
-0.0341599844,
-0.0117153581,
-0.0231954642,
-0.0380619131,
0.0184605792,
-0.0138778668,
-0.040337313,
0.0298302025,
-0.0699354708,
-0.0212277733,
-0.0329766273,
-0.0340675153,
-0.0169270393,
-0.0112507483,
0.003771489,
0.0065872259,
-0.0506473668,
-0.0477582961,
-0.0062938421,
0.0052657961,
-0.0059446367,
-0.0287763476,
-0.0139846113,
-0.0079478286,
0.019952476,
-0.0249749869,
-0.0491298921,
-0.0008733052,
-0.0065997881,
0.0282427575,
0.0435667969,
0.023567915,
-0.0088601653,
0.0014432166,
0.0437514372,
0.0144411484,
0.053444352,
0.0113052567,
0.0275887009,
0.0079720672,
0.0256778914,
0.0010672787,
0.0361559279,
-0.0141566331,
-0.030020643,
-0.0047270418,
-0.047752507,
0.1019918695,
-0.0193363875,
-0.0057627177,
0.0080659753,
0.0557700954,
0.0062809289,
-0.0195612274,
0.0280578695,
0.0692700371,
0.0362025946,
0.0063947011,
0.044358287,
-0.0157578401,
-0.0226093885,
0.025843896,
-0.0476791374,
0.0322975181,
0.0078667859,
-0.0340050422,
-0.0377958305,
-0.0048944517,
0.0538422428,
0.033070609,
-0.0221000444,
0.0120274993,
-0.0160556082,
-0.105660215,
-0.0064677261,
-0.0017863081,
-0.0838107914,
-0.0165279284,
-0.0663704649,
-0.0103409616,
-0.0399375074,
0.0549413376,
0.0062687322,
-0.0158785153,
-0.0304462425,
0.0747458786,
-0.016863035,
-0.0507150106,
-0.0049716346,
-0.0410987139,
0.0676036477,
0.039253369,
-0.0123242289,
-0.0283404142,
-0.031975463,
0.0096247699,
-0.0246499889,
-0.0020487588,
-0.01735829,
0.0162997544,
0.0327146575,
-0.0202719029,
0.0318278186,
0.0371241458,
-0.0068557742,
-0.0120404642,
0.0322373994,
0.0881830528,
0.0285632517,
-0.0111755859,
0.0280555338,
-0.0283080302,
0.0031956083,
0.0523631573,
0.010525207,
-0.0461913683,
0.0402639806,
0.0529850535,
0.0333619937,
-0.0131271323,
0.0175335016,
-0.0026577273,
-0.0663418397,
-0.0737087429,
-0.0477250926,
0.0202007536,
-0.0013328218,
0.0937474146,
0.0542167239,
0.0847083479,
-0.0344794579,
0.0080097523,
0.0017075645,
0.0080868965,
-0.0119172698,
0.0464310907,
-0.0772016793,
-0.0369125381,
0.0080617052,
-0.0480573885,
-0.0310955662,
-0.0427903906,
-0.0221841391,
-0.0548569635,
0.0137738148,
0.0032577671,
0.0184002463,
0.0419548452,
0.0162167605,
-0.0584495328,
0.0685185343,
0.0278949887,
-0.0799372345,
0.0003006129,
0.0516372062,
-0.0343213044,
-0.004811991,
0.0344103053,
-0.0416298024,
0.0077387532,
0.0087982323,
0.0030300836,
0.0000374538,
0.0122270845,
-0.0173646901,
0.057323724,
0.0551487729,
-0.0566642992,
-0.0310103334,
-0.0108517669,
0.0421309993,
0.0089275083,
0.0596822239,
-0.0201659743,
-0.0414642692,
0.0155624868,
0.0324665792,
-0.0250509195,
-0.0043479786,
-0.0242522601,
0.0138273221,
0.0270674918,
0.0727526769,
0.0046443609,
0.0739790127,
0.032746017,
0.0039887442,
0.0008921401,
-0.0380856991,
0.0117015522,
0.0027690006,
-0.0601344854,
-0.0020118351,
0.0105405906,
0.0464109704,
0.0737476796,
0.0192484185,
0.0144239059,
-0.0135465376,
0.0162601005,
0.0215577334,
0.0205727629,
-0.0384808704,
0.0563132502,
0.0034364185,
0.0226610694,
-0.0507882275,
0.0082484493,
-0.0508082174,
0.0015686598,
-0.0068512745,
-0.0053312844,
-0.0244438611,
0.0538085327,
0.0401572958,
0.0010766967,
0.0121452622,
-0.0702048242,
0.0443134308,
0.0155081097,
-0.0158932749,
0.0189876854,
0.0171674155,
-0.0160498917,
0.0212963205,
-0.0333755463,
-0.0845361203,
-0.0200662334,
-0.0445404798,
0.0240691137,
0.0199587662,
0.0125772497,
0.0346520096,
-0.0438802689,
-0.0098493816,
0.0009219792,
-0.0277149547,
0.0308754668,
-0.0029822474,
0.0562133826,
-5.856434079e-33,
0.0250999946,
0.0213964302,
-0.0264143217,
-0.0141220549,
-0.0241213739,
0.0239872504,
-0.0034569739,
-0.0256668553,
0.0186015666,
-0.0388211273,
-0.0141485222,
-0.0302537922,
0.0306484979,
0.0272156503,
0.0226659793,
0.0213986114,
0.0845117196,
0.0393246114,
0.0028350377,
0.007935876,
0.0548349768,
0.0213905275,
0.0452403836,
-0.0274606142,
-0.0406989641,
-0.0116746565,
0.0609540083,
-0.0136942538,
0.0575669147,
0.0146099562,
0.0056506977,
0.044415094,
0.0008649383,
-0.044249624,
-0.0045075347,
-0.0336228162,
-0.0773319751,
-0.0228666272,
0.0557849817,
0.0580075197,
0.0190514103,
-0.0221870653,
0.014749283,
0.0107198199,
-0.0570869669,
0.03674509,
-0.0018370989,
-0.004658448,
-0.0437289663,
-0.0308301169,
-0.0528106838,
0.0552354418,
-0.0202736873,
0.0398282222,
0.0463846922,
0.0588191971,
-0.0227066036,
-0.0378639996,
-0.022183869,
0.0139874164,
-0.0145509569,
0.0772583112,
0.0489600562,
0.0388106629,
-0.005054065,
-0.0078295302,
0.0739729106,
-0.0969926119,
0.0744176805,
-0.0354158133,
0.0587404482,
-0.012376816,
0.0048487517,
0.0088212863,
0.0609888248,
-0.0964370146,
-0.0320601016,
-0.0050339131,
-0.0071515203,
0.0085352324,
0.0240582656,
0.0471179895,
-0.0151976347,
-0.0035744,
-0.0223621801,
-0.0447319634,
-0.0348520614,
-0.0425642394,
0.0735099465,
-0.0103924647,
0.0103295473,
0.0141370054,
0.0785498396,
0.0280681625,
-0.0083219074,
0.024228951,
-0.022688292,
-0.0093936427,
-0.0343877934,
0.0117724389,
-0.0754702762,
-0.07615906,
0.0485787615,
0.0234038811,
0.0249287374,
-0.0560543612,
0.0353373624,
0.0163733829,
-0.0729549676,
0.0042054765,
0.0183946807,
0.0195023138,
0.0288547166,
-0.0032724456,
0.0510322377,
0.022687437,
0.0032090298,
-0.0039418181,
0.0070917234,
-0.0299219675,
-0.0112233218,
0.0255304631,
-0.006078246,
0.0205261856,
-0.0668958053,
0.0165862311,
-0.0205378886,
-0.0097377198,
0.0022557329,
-0.0343619138,
-0.0337499082,
0.010372987,
2.682e-7,
-0.0739264414,
0.0137061914,
0.0192701537,
-0.0750974193,
-0.0207441617,
0.0056662285,
-0.0725991428,
-0.0040149544,
0.010539839,
0.0462220609,
0.0117215812,
-0.0099768881,
0.0564976037,
0.0324392878,
-0.0916031748,
-0.0050515747,
-0.071527876,
0.0080090901,
-0.0570907593,
0.0135735869,
-0.0194608178,
0.0790157169,
0.0079345843,
-0.0407605805,
-0.0025567194,
-0.0494538955,
0.0073753782,
-0.0322658494,
0.0016473588,
-0.0303399395,
0.012838508,
-0.0219045673,
0.0012399397,
0.0486231036,
-0.0484439991,
0.0220452305,
0.0319064334,
0.0115023935,
-0.0388908386,
-0.0257790647,
-0.0045084539,
0.0506509319,
-0.0314639993,
-0.0217441358,
0.0447724275,
0.0371722467,
-0.0468283333,
-0.0231282171,
-0.0652787536,
0.0136746727,
0.0142523153,
-0.0049373414,
-0.0037411328,
0.0434804447,
0.016449431,
-0.013626066,
-0.022179449,
0.0005390832,
0.0389868803,
0.0083763022,
-0.0317063257,
-0.0547039546,
-0.0322949402,
0.0594719984,
0.0204972606,
-0.0016719998,
0.0350026302,
1.904616762e-34,
0.0507786162,
-0.0156431142,
0.0342241563,
-0.0019034778,
0.0037245199,
-0.0439183302,
-0.0094532752,
0.0389690585,
-0.014696613,
-0.020473456,
-0.017510809
] |
http://arxiv.org/abs/2402.02289v1 | http://arxiv.org/pdf/2402.02289v1 | SemPool: Simple, robust, and interpretable KG pooling for enhancing language models | 2024-02-03T23:03:51 | 2024-02-03T23:03:51 | Costas Mavromatis, Petros Karypis, George Karypis | cs.CL | cs.CL, cs.AI | Knowledge Graph (KG) powered question answering (QA) performs complex
reasoning over language semantics as well as knowledge facts. Graph Neural
Networks (GNNs) learn to aggregate information from the underlying KG, which is
combined with Language Models (LMs) for effective reasoning with the given
question. However, GNN-based methods for QA rely on the graph information of
the candidate answer nodes, which limits their effectiveness in more
challenging settings where critical answer information is not included in the
KG. We propose a simple graph pooling approach that learns useful semantics of
the KG that can aid the LM's reasoning and that its effectiveness is robust
under graph perturbations. Our method, termed SemPool, represents KG facts with
pre-trained LMs, learns to aggregate their semantic information, and fuses it
at different layers of the LM. Our experimental results show that SemPool
outperforms state-of-the-art GNN-based methods by 2.27% accuracy points on
average when answer information is missing from the KG. In addition, SemPool
offers interpretability on what type of graph information is fused at different
LM layers. | SemPool: Simple, robust, and interpretable KG pooling for enhancing language models: Knowledge Graph (KG) powered question answering (QA) performs complex
reasoning over language semantics as well as knowledge facts. Graph Neural
Networks (GNNs) learn to aggregate information from the underlying KG, which is
combined with Language Models (LMs) for effective reasoning with the given
question. However, GNN-based methods for QA rely on the graph information of
the candidate answer nodes, which limits their effectiveness in more
challenging settings where critical answer information is not included in the
KG. We propose a simple graph pooling approach that learns useful semantics of
the KG that can aid the LM's reasoning and that its effectiveness is robust
under graph perturbations. Our method, termed SemPool, represents KG facts with
pre-trained LMs, learns to aggregate their semantic information, and fuses it
at different layers of the LM. Our experimental results show that SemPool
outperforms state-of-the-art GNN-based methods by 2.27% accuracy points on
average when answer information is missing from the KG. In addition, SemPool
offers interpretability on what type of graph information is fused at different
LM layers. | [
0.0356193222,
0.0128040146,
0.0299470797,
0.039959114,
-0.0250626877,
0.0020082109,
-0.0289152674,
-0.0297554173,
0.0560426563,
-0.0332175419,
0.0064194356,
0.02102088,
-0.0023576752,
0.0374923311,
0.0283717383,
-0.0068745469,
0.0173365939,
-0.0185462199,
0.0186234675,
-0.0225137584,
-0.0043538576,
0.0633911639,
-0.0042026932,
0.013998135,
-0.0201389696,
-0.0060789809,
-0.00939677,
0.0202914774,
-0.0291547496,
0.0033825026,
-0.0057212347,
0.0108641647,
0.0190021843,
0.0170280114,
0.0000019873,
-0.0196342487,
-0.0016687351,
-0.0063680741,
-0.0613025725,
0.0186273921,
-0.0215435885,
0.0008367244,
0.0354616493,
0.006534636,
-0.0471172482,
-0.0262589585,
0.0164199397,
0.0113162398,
0.0661319643,
0.0264442805,
-0.001774329,
-0.0124080563,
-0.0052132895,
-0.0095164459,
-0.0113290651,
-0.079610303,
0.0288490243,
0.0467859618,
0.0170348212,
-0.0275596268,
-0.0393820181,
0.0266782269,
0.0422827564,
-0.0236808602,
0.028711129,
0.0474792048,
-0.0227705054,
0.0411538854,
0.0114034554,
0.129431352,
-0.0170443021,
-0.0430310108,
-0.0284452401,
-0.029445013,
0.0090683755,
0.029778,
-0.0567453764,
-0.0182339121,
0.0234186444,
0.0076923389,
-0.0635272712,
-0.023843253,
0.0742585137,
0.0179393776,
-0.0273472294,
0.0860896781,
0.0010119337,
0.0183002558,
0.009055797,
0.0181535967,
-0.0627890304,
-0.0360848866,
-0.023084458,
0.0867773518,
-0.0021865091,
0.0374331474,
-0.0174549818,
0.0021686461,
0.0178177878,
-0.0488514602,
-0.0042290771,
0.0030716681,
-0.0015426115,
0.0925876722,
-0.0250401814,
0.0242494829,
-0.0279888436,
0.0812712833,
-0.0597898737,
-0.0587303005,
-0.0390230045,
-0.0314021707,
-0.0415141769,
0.0328949057,
-0.0102740377,
-0.0573464073,
-0.0398916639,
0.0533871837,
-0.0174672101,
-0.006548929,
0.0159365423,
0.0798746496,
0.0360222682,
-0.02922998,
-0.0186148342,
0.0622649118,
-0.0468324758,
0.0310541075,
-0.0022343011,
-0.0224932898,
-0.0353636295,
-0.0054217982,
0.0540596694,
-0.0130345104,
0.0156172346,
0.0663568452,
0.024658097,
0.0263020117,
-0.0106469123,
-0.0305597447,
-0.0049801865,
-0.027119441,
0.0089228274,
-0.0813493356,
0.0729843751,
0.024851324,
-0.036812108,
-0.0486551896,
0.0167225674,
-0.0043537295,
-0.0411749817,
0.0457073152,
-0.0413366929,
-0.0048592412,
0.0164114591,
-0.0053220764,
-0.0236410964,
0.0658797994,
0.0294631608,
0.0101793772,
0.0421162434,
-0.0017039591,
-0.0185694396,
0.013535413,
0.0625605434,
-0.0051049371,
0.036736127,
0.0204742067,
-0.033921089,
-0.0153412502,
-0.0015342045,
0.0237260405,
-0.0224958267,
-0.0442609824,
0.0233802013,
0.0509648658,
-0.0058967965,
0.0118735908,
0.0151516488,
0.0163035858,
-0.0124646313,
-0.0063166809,
-0.0394273736,
0.0208950378,
-0.0412946045,
0.0197004415,
-0.0123783741,
-0.009937874,
-0.0278938971,
-0.015548015,
-0.0037926889,
-0.0196359977,
0.0765627474,
0.0552664697,
0.0512921177,
0.0207975265,
-0.0195670817,
0.0744036138,
0.0334278755,
0.0092552798,
0.0160334446,
-0.0105619794,
0.0056879963,
0.0475097559,
0.0336096622,
0.0127843395,
-0.0559155457,
0.0159411989,
0.014068203,
-0.0159321483,
0.0183758605,
-0.0181316733,
0.0407646708,
-0.0583579391,
-0.0308817327,
-0.0622974113,
-0.0183397718,
0.0189118106,
0.0229575466,
-0.0120229349,
0.0265246425,
-0.0869118497,
-0.0065802666,
-0.0018171755,
0.0228642952,
-0.0080456054,
0.0171768107,
-0.0158444569,
-0.0222877525,
0.0156336892,
0.0410139412,
0.0463050529,
0.017905144,
-0.0022550924,
-0.0563426279,
0.0036098994,
-0.0536293611,
-0.0696806088,
0.0339137651,
0.0139795309,
-0.0114424005,
-0.0140983667,
-0.0784179047,
-0.0469867438,
0.0535859875,
-0.0592355058,
0.0563641042,
-0.0128682116,
-0.0335953757,
0.0275793634,
0.0010109319,
-0.0051085982,
0.0674527213,
-0.0228128526,
-0.0025327955,
-0.01342298,
0.0469825305,
0.0101779373,
-0.0417618714,
0.025097359,
0.006940783,
-0.0052126846,
-0.0068862564,
0.0031742833,
-0.0202953778,
-0.068000488,
0.0439968072,
-0.1141120195,
0.0071784542,
0.0046993215,
0.0174350794,
0.0102704735,
0.0188228786,
0.0257573277,
-0.0396118052,
-0.078347072,
0.0267345756,
0.0363626517,
0.0137863066,
0.0255631451,
0.0193588007,
0.0244439431,
-0.0295380391,
-0.0597319715,
-0.0217501167,
0.101187177,
0.0256651342,
-0.0104415659,
0.0121848127,
0.001393638,
-0.0807741359,
0.0299624093,
-0.0118051283,
-0.0062694699,
-0.0090078255,
-0.0420645773,
-0.0094567668,
0.0262110271,
-0.0296021756,
0.0042524259,
0.0327804536,
-0.0170871671,
-0.0581052452,
0.0116947768,
-0.0576577596,
0.0047443029,
-0.0184548944,
0.0109804496,
-0.0012726322,
0.0127336625,
0.0174468905,
-0.02513762,
-0.0467585996,
0.0167373158,
0.0172666535,
0.0179570485,
-0.0383145586,
-0.0324701667,
0.0186356325,
-0.017283855,
-0.0487759598,
0.0049213511,
0.0891251713,
0.0091299973,
-0.0260310043,
-0.0025753377,
-0.0428170487,
-0.0797610506,
0.0350759104,
0.0023396469,
-0.0006230372,
-0.0430622883,
-0.002319887,
-0.0098541817,
-0.0465358905,
0.0124275433,
-0.0328188762,
0.0285448246,
-0.003744887,
0.0368256755,
0.0159661118,
0.0047354177,
-0.0382006466,
-0.0042843213,
-0.0488337539,
0.0438492224,
-0.0132175302,
0.0062145023,
0.0343311615,
-0.0251210984,
-0.0020142801,
-0.0054134415,
-0.0281752851,
0.0205898583,
-0.0096698506,
0.0592820905,
0.0586381182,
0.0463131033,
0.0308571681,
0.0038809483,
0.029193854,
0.0773432925,
0.0247783028,
-0.035466902,
-0.0240210965,
0.0173429921,
-0.0138835004,
0.0008600068,
0.0121951206,
-0.019730635,
0.0599856637,
-0.0670992509,
0.0201017894,
0.087078616,
0.0204396341,
-0.0124243731,
-0.008178019,
-0.0046445741,
0.0026645812,
0.0017005178,
0.0227738451,
-0.0835564733,
-0.0284932815,
-0.0248240866,
0.030830387,
0.0342322327,
-0.0297725704,
-0.0219752658,
-0.0961390063,
-0.0170130897,
0.0155024445,
-0.025555905,
-0.032951124,
-0.0522803962,
-0.0147795593,
-0.0184028726,
-0.0076789055,
-0.0098088169,
-0.0084574046,
0.0047758911,
0.0215955283,
-0.0157344248,
-0.0071406206,
0.0557865724,
-0.0178284813,
-0.0093771499,
-0.0010239208,
0.038733881,
0.0126843499,
-0.0178739391,
0.0619671457,
-0.0352998376,
-0.0229246132,
-0.0102779279,
0.0193022452,
0.0122474702,
-0.0095079886,
-0.0427566804,
0.0283759031,
-0.0531439334,
-0.0298543572,
-0.032319285,
0.0697552487,
0.0024318751,
-0.045779746,
0.0089507382,
0.0158768874,
-0.0289392564,
0.0273385178,
0.0584077947,
0.0223824047,
-0.0416954346,
-0.0459364355,
-0.00113381,
-0.0604112074,
-0.0249003526,
0.0454123206,
0.0145195415,
0.0365607813,
-0.0589285158,
-0.0470366068,
0.039429646,
0.0184057951,
-0.0222640652,
0.0137938615,
0.0329533443,
-0.0121436706,
-0.0669824854,
0.0517004654,
-0.0532023907,
0.0196355525,
0.0110167721,
-0.0165120419,
0.0081532989,
0.0049504084,
-0.0242490806,
0.0127284974,
-0.0722179934,
-0.0952915326,
-0.0376383103,
0.0255118236,
0.0173203126,
0.0496891476,
-0.0559595041,
-0.0098957438,
0.0446061827,
-0.0010169165,
-0.0409247354,
0.0039426307,
0.0248234402,
-0.0181925427,
-0.0188020859,
0.02636512,
-0.1042574048,
-0.0031547465,
0.0180964768,
0.0396098047,
-0.0654637516,
-0.0338050835,
-0.0048990683,
0.0262591932,
-0.0216335654,
0.0497803986,
0.0179532543,
-0.0271829255,
-0.0049241367,
-0.0115447026,
0.0043359022,
-0.0391190797,
-0.0244355965,
0.0069412892,
-0.0386114605,
0.0019059416,
-0.0547875501,
0.0213411953,
0.021916518,
-0.0150860362,
-0.0399356671,
0.0438628085,
-0.0369600505,
0.0193433948,
0.0083408151,
-0.0396054387,
0.0471476316,
-0.0041379794,
0.040167354,
-0.0087803742,
0.0352520645,
0.0240323506,
-0.0410784967,
-0.0558100715,
0.001863174,
-0.0845466033,
-0.0737521127,
-0.0189341605,
0.0160387438,
0.0529084243,
0.0090498142,
0.0111891609,
-0.0047634747,
-0.0207719449,
-0.0201755669,
0.0185513031,
0.0137629919,
-0.0539682508,
0.0538279302,
0.0385884792,
0.0010580213,
0.0251697619,
0.0499711893,
0.0539381169,
0.035408482,
-0.0024384877,
-0.0241345875,
-0.018333938,
-0.0383554101,
-0.0291919988,
0.0668017268,
-0.0054817745,
0.0267258603,
-0.0344880372,
-0.0963654742,
0.000792992,
-0.0052125785,
-0.005037488,
0.0193198659,
0.024269376,
0.0196069907,
0.0471519455,
-0.0171214119,
-0.0296755824,
0.0678640604,
-0.0413809158,
0.010998318,
0.0186878871,
-6.372656831e-33,
-0.0185237937,
-0.0661604032,
0.0077889641,
0.0072150049,
-0.0291372053,
0.0077403262,
-0.0488342345,
-0.0058841682,
-0.0595540963,
0.0073368661,
0.0102774594,
0.0128054181,
-0.0098120617,
0.0451513045,
-0.0230883379,
-0.0453946292,
-0.005930556,
0.0420536809,
0.0111367917,
0.036871925,
0.0437400192,
0.0350569263,
0.0399297178,
-0.0547059849,
0.0201172512,
-0.0434869453,
0.0356908366,
0.0165197197,
-0.0013056956,
0.0461924002,
-0.0075047286,
0.0648980215,
0.0361936353,
0.0026191764,
-0.0064991815,
0.0125427721,
-0.1090759933,
-0.0511443987,
0.0426118858,
0.027995592,
-0.0461891182,
-0.0147129185,
0.0434650034,
-0.0061098132,
0.0766130462,
-0.0467724167,
0.0306846146,
0.0137870871,
-0.0509979352,
-0.0027276219,
-0.0033566798,
0.0262926333,
-0.0131989401,
-0.011550596,
-0.02295986,
-0.0143711753,
0.0004366919,
-0.0066585583,
-0.0531890467,
-0.0113293221,
-0.0195063837,
-0.0107837338,
0.0306201242,
0.0345742851,
0.011649278,
0.0409606621,
-0.0409215763,
-0.0389433429,
-0.0794978738,
0.0181905814,
0.021475317,
0.0438932739,
0.0451418571,
-0.0050080493,
-0.0143293338,
0.0219542459,
-0.0785506889,
0.0033835962,
0.0164532084,
0.0292747412,
-0.0301040243,
0.0211925991,
0.0098702405,
-0.0547441766,
-0.0353835598,
0.0205813237,
-0.044325389,
-0.033624582,
-0.0031056942,
-0.0109770438,
-0.0446096994,
0.0456371084,
-0.0050723865,
-0.0628833324,
0.0024088356,
0.1142823622,
0.0263971519,
0.0099284304,
-0.0040890523,
0.0431133956,
-0.0308058616,
0.0508816987,
0.0414801203,
-0.0287945308,
-0.0170115586,
0.0498964861,
0.020325521,
-0.0026865848,
-0.0047499966,
-0.013178288,
-0.0650710762,
-0.0159451496,
-0.0429914407,
0.0047705942,
0.0321516469,
0.0796855092,
-0.0169130303,
0.0177200362,
0.0182871316,
0.0622676127,
0.0103312926,
0.0354494266,
0.016438365,
0.0372180566,
-0.047465574,
-0.0252862051,
-0.0166156758,
0.0127376504,
-0.0159952864,
-0.0467244349,
-0.009305886,
0.0168205239,
2.659e-7,
0.0935485736,
-0.0054045017,
0.0448092036,
0.030076718,
0.0026343844,
0.0162274856,
0.0046465541,
-0.0030997293,
0.0407467149,
-0.0191902034,
0.0105344849,
0.0819552019,
-0.0218268447,
0.0474228784,
-0.0478432775,
0.0333644636,
-0.0561327264,
-0.03451588,
-0.0218053311,
-0.0072734645,
0.0426521823,
0.0535023324,
0.0594050549,
-0.0212897453,
-0.0259520188,
-0.080007799,
0.0636118799,
0.0151984142,
0.0012097267,
0.0050260341,
0.0098344358,
0.0191872027,
-0.0117005259,
0.0222662371,
-0.022180263,
0.0178800337,
0.0288083926,
0.0390683152,
-0.0131780598,
-0.0568584763,
-0.0242344011,
-0.0407997221,
0.044970233,
-0.0432753861,
0.111926131,
-0.0466917343,
-0.025317505,
-0.0284996685,
-0.0166431218,
-0.0264571235,
0.0289415605,
0.0005102997,
-0.001132932,
0.0094148489,
0.0212054737,
-0.0452745892,
0.0152406842,
-0.0944739953,
-0.0243094359,
0.0092804795,
-0.0208244696,
-0.0044699148,
0.0007175319,
0.0061004148,
0.0249944422,
0.0008466004,
0.0286959466,
2.210515749e-34,
0.0147672649,
0.0174085759,
0.0120458314,
0.0262869038,
0.0140938098,
-0.005794548,
-0.0084096473,
0.0138087515,
0.0504868999,
-0.0836851969,
0.0033259369
] |
http://arxiv.org/abs/2402.02285v1 | http://arxiv.org/pdf/2402.02285v1 | SynthDST: Synthetic Data is All You Need for Few-Shot Dialog State Tracking | 2024-02-03T22:49:00 | 2024-02-03T22:49:00 | Atharva Kulkarni, Bo-Hsiang Tseng, Joel Ruben Antony Moniz, Dhivya Piraviperumal, Hong Yu, Shruti Bhargava | cs.CL | cs.CL, cs.AI, cs.LG | In-context learning with Large Language Models (LLMs) has emerged as a
promising avenue of research in Dialog State Tracking (DST). However, the
best-performing in-context learning methods involve retrieving and adding
similar examples to the prompt, requiring access to labeled training data.
Procuring such training data for a wide range of domains and applications is
time-consuming, expensive, and, at times, infeasible. While zero-shot learning
requires no training data, it significantly lags behind the few-shot setup.
Thus, `\textit{Can we efficiently generate synthetic data for any dialogue
schema to enable few-shot prompting?}' Addressing this question, we propose
\method, a data generation framework tailored for DST, utilizing LLMs. Our
approach only requires the dialogue schema and a few hand-crafted dialogue
templates to synthesize natural, coherent, and free-flowing dialogues with DST
annotations. Few-shot learning using data from {\method} results in $4-5%$
improvement in Joint Goal Accuracy over the zero-shot baseline on MultiWOZ 2.1
and 2.4. Remarkably, our few-shot learning approach recovers nearly $98%$ of
the performance compared to the few-shot setup using human-annotated training
data. Our synthetic data and code can be accessed at
https://github.com/apple/ml-synthdst | SynthDST: Synthetic Data is All You Need for Few-Shot Dialog State Tracking: In-context learning with Large Language Models (LLMs) has emerged as a
promising avenue of research in Dialog State Tracking (DST). However, the
best-performing in-context learning methods involve retrieving and adding
similar examples to the prompt, requiring access to labeled training data.
Procuring such training data for a wide range of domains and applications is
time-consuming, expensive, and, at times, infeasible. While zero-shot learning
requires no training data, it significantly lags behind the few-shot setup.
Thus, `\textit{Can we efficiently generate synthetic data for any dialogue
schema to enable few-shot prompting?}' Addressing this question, we propose
\method, a data generation framework tailored for DST, utilizing LLMs. Our
approach only requires the dialogue schema and a few hand-crafted dialogue
templates to synthesize natural, coherent, and free-flowing dialogues with DST
annotations. Few-shot learning using data from {\method} results in $4-5%$
improvement in Joint Goal Accuracy over the zero-shot baseline on MultiWOZ 2.1
and 2.4. Remarkably, our few-shot learning approach recovers nearly $98%$ of
the performance compared to the few-shot setup using human-annotated training
data. Our synthetic data and code can be accessed at
https://github.com/apple/ml-synthdst | [
0.007902734,
0.0716445222,
0.0121640069,
0.0215595905,
-0.0018338383,
0.0074342871,
0.0345663466,
-0.0052561886,
0.0199294407,
-0.0661083087,
-0.0283184517,
-0.0604676381,
-0.0324333087,
0.0655877888,
0.0204032287,
-0.0688060671,
0.0474678315,
-0.040000312,
-0.0093220295,
0.0059362403,
0.0323108286,
0.0129512306,
-0.057896819,
0.0528850965,
-0.019026814,
-0.0347271785,
-0.0533602498,
0.0228558742,
-0.0032887978,
-0.032474339,
-0.0138731468,
0.0458431132,
0.0432662517,
0.0357866548,
0.0000018715,
-0.0498385876,
0.0058339913,
-0.0237757359,
-0.0598762743,
-0.0179587528,
-0.0003730333,
-0.0058038021,
-0.0309712701,
0.0453358069,
-0.0305416528,
-0.0096562989,
0.0306236614,
0.013631667,
0.0815896988,
0.0889050663,
-0.0126371393,
-0.0216738954,
0.0154569233,
-0.0186796375,
0.0269186459,
-0.0522294454,
-0.0238204934,
-0.0074815857,
0.0024949647,
0.0110686235,
0.006619175,
0.0186065212,
0.010469771,
-0.0467570387,
-0.0230021942,
0.0142957177,
-0.0471753851,
-0.0286790114,
-0.036949873,
0.1006585434,
0.0242444277,
0.030495977,
0.0090044299,
0.0006031597,
0.0172053874,
-0.0379387215,
-0.0721691623,
-0.0115232002,
-0.0150079811,
-0.005691722,
-0.0332756601,
0.0079568485,
-0.0180264004,
-0.0183016695,
-0.0031323999,
0.0055049495,
0.0038802521,
-0.0128901927,
-0.0278786868,
-0.0102211693,
0.0075313374,
-0.084523119,
-0.0273722261,
0.018572757,
-0.0765006691,
0.0202936511,
-0.0760406703,
0.00136133,
-0.0038131839,
-0.0090194168,
0.0346287042,
0.0563951619,
-0.0178815126,
0.0421734452,
0.0198316593,
0.052339118,
0.006459902,
-0.0146545731,
-0.0157025922,
-0.0056295679,
-0.0594144687,
0.0168502498,
-0.0664614886,
0.0061133332,
-0.0534662157,
-0.055547934,
-0.0292364433,
0.0448826142,
-0.0002131226,
0.0389091671,
-0.0290118996,
0.0057310532,
-0.0293756276,
-0.002489408,
0.0014450725,
0.0389032662,
-0.0287539754,
0.0007953179,
0.0493641309,
-0.0634828806,
-0.0240272526,
0.0335598178,
0.043726977,
0.0024180159,
-0.0123986714,
0.0647161677,
-0.0161106624,
-0.0388183035,
-0.0206039492,
-0.0282264706,
-0.0224074963,
-0.048426874,
-0.0109605985,
-0.0370546766,
0.0331774279,
-0.0217581578,
-0.0116066588,
-0.0222305693,
0.0019124126,
0.0245173629,
0.0136605985,
0.0416168943,
-0.0345464125,
-0.0115005104,
0.0044437349,
-0.0015023135,
-0.020917749,
0.0820241868,
0.0367547236,
0.0212800503,
0.0326061025,
-0.0072550513,
0.0386145487,
0.0364617631,
0.0396537073,
0.0261388738,
-0.0343037732,
0.0326336175,
-0.005738982,
0.020783836,
0.0319765061,
0.0287954267,
-0.0608724505,
0.0152427629,
0.0708548427,
0.05156992,
0.0320253074,
-0.0171764996,
0.0335626155,
0.0497411303,
0.0131058367,
-0.0199756511,
-0.0276109278,
0.043095462,
-0.0003108876,
0.0072955163,
0.0341339931,
0.0058480687,
-0.0499846637,
-0.0216240864,
0.0212628301,
-0.0004368983,
-0.0374380797,
0.0045917295,
0.0238255169,
0.0274610594,
-0.0718146637,
0.0301967356,
0.0047802604,
-0.0416626334,
-0.0563044213,
-0.0119524971,
0.0111050857,
0.0805602595,
0.0254884623,
-0.0168383513,
-0.0534740202,
-0.0698475316,
0.0246772207,
0.0590921454,
0.0242563672,
-0.0393007286,
0.0713534281,
-0.0535558239,
-0.0074327313,
-0.0310578551,
0.0108116763,
0.0182275046,
0.0085330922,
-0.0178700984,
0.0088613238,
0.0250183512,
0.0636646301,
0.0099968286,
0.0011303485,
-0.0429263972,
0.0149749359,
-0.001393831,
0.0485567264,
0.033590205,
-0.004981989,
-0.021818107,
0.0064955801,
-0.004742343,
0.0245831385,
0.023238875,
-0.0699374676,
-0.0844433308,
0.0176879428,
0.037537083,
-0.0163301788,
-0.0106629971,
-0.0487551652,
-0.062718153,
-0.0115394574,
-0.033250995,
0.021810459,
-0.0331915282,
-0.0021260735,
0.0454007238,
0.010499578,
-0.0219589379,
0.0210456867,
0.0066723428,
-0.0290922541,
0.0524961203,
0.0386023521,
0.0098673981,
-0.0313787982,
0.0203542989,
0.0322680138,
-0.0243786685,
0.004980125,
-0.0038649472,
0.0130199175,
-0.0658617988,
0.0019329926,
-0.0847109556,
0.0138000548,
-0.0259869359,
-0.0069200066,
0.015823964,
0.0146247456,
-0.027081063,
-0.0225417446,
-0.0531386323,
-0.0156308282,
0.0314070322,
-0.0208204258,
0.0152031844,
0.0250616521,
-0.023338329,
-0.0490161963,
0.0360217579,
0.0030342678,
0.0217482876,
-0.0289576128,
-0.0142656043,
-0.0613539964,
0.0281147882,
-0.0255384147,
0.0514551625,
-0.0261883643,
-0.0257083979,
0.0415871516,
0.0137759726,
0.0345921628,
0.0384272262,
0.0073127421,
0.0283589475,
-0.0021443891,
-0.0231193397,
-0.072043322,
0.0510828421,
0.0021473304,
0.0094223069,
-0.0322176963,
0.0194644071,
0.0154735036,
0.0308228601,
0.0157815181,
0.0515022315,
-0.0578383617,
0.0401313901,
0.044420816,
-0.0255585797,
-0.0115637314,
0.0006429997,
-0.05156729,
0.0337172449,
0.0046829493,
-0.0463124216,
0.0561230779,
-0.051109761,
-0.0545675345,
-0.0187563766,
-0.0163433813,
-0.0072967368,
0.0505454578,
-0.0269746128,
-0.0008857119,
0.0243424103,
-0.0299775414,
-0.0222759657,
-0.0179870334,
0.0214422885,
-0.0271564964,
0.0431950353,
-0.0433327369,
0.0181762762,
-0.021161804,
-0.0468719192,
-0.06545908,
-0.0336464271,
-0.0379208848,
-0.0077289045,
-0.0257617012,
-0.0235939045,
0.0041342378,
0.0112857688,
0.010409655,
0.0805641785,
0.0089332461,
-0.0135973003,
-0.0351462439,
0.0473797098,
0.0768949613,
-0.0196217597,
-0.0284030084,
-0.0158467852,
0.1211402044,
0.0449013673,
0.0712081417,
0.0209847223,
-0.0221202727,
-0.0144549627,
0.0054983241,
0.0085397754,
0.0057744118,
0.0199868418,
0.0758604556,
-0.0362499245,
-0.0137370741,
0.1222131327,
0.0107388832,
-0.0447521098,
-0.0158431102,
-0.0035892953,
0.0285290051,
0.0311065707,
0.0233759377,
-0.1115681678,
-0.0614732876,
0.0199745689,
0.0346853808,
-0.0095286313,
-0.0236270297,
-0.0299241208,
-0.1129920036,
0.0368312076,
0.0014427364,
-0.0609797239,
-0.0068466975,
-0.0117443372,
0.012071725,
0.0225315932,
0.0192343518,
0.0227581114,
0.0034903686,
-0.0109351445,
0.0093121985,
-0.0010063209,
-0.0265146494,
0.0289420001,
-0.0139360214,
0.0986319557,
0.0656716302,
0.0319277272,
0.0359288901,
0.0188291762,
0.0318358839,
-0.0086713564,
0.0001002841,
0.0111871986,
0.0510621034,
-0.0274302755,
-0.0246058442,
-0.0112652555,
0.0646452904,
-0.009795988,
-0.0179293528,
0.0070024333,
0.0119821271,
-0.0552622341,
0.0298060346,
-0.0177621506,
0.0161124784,
0.0257004667,
0.0461847,
0.0512375645,
-0.0601671152,
-0.0038024881,
-0.0226383135,
-0.0167501941,
-0.0519759282,
0.0109379897,
-0.0194298998,
0.0125788748,
0.0161094368,
-0.0889931619,
-0.0470004827,
0.055066254,
0.0716225728,
-0.0428395495,
-0.0131291244,
-0.0227409415,
0.0084315958,
0.0182973556,
0.0244211666,
-0.0342582837,
0.0119591625,
-0.0610581823,
-0.0200967453,
0.0533685796,
-0.0018124958,
0.0017070642,
0.0441187657,
-0.0650998801,
-0.0612293482,
-0.0262897667,
0.0674843565,
0.0230982825,
0.078331843,
0.0150246425,
0.00550461,
-0.0035502736,
0.0500981249,
-0.0094229104,
-0.0171712786,
0.0554264784,
0.0654937252,
-0.0309350435,
0.0254092999,
-0.0236894954,
0.0207630619,
-0.0199832227,
0.0319688506,
-0.0645079911,
-0.0501665585,
0.0144991716,
-0.0242774542,
0.0633556098,
-0.0075720991,
-0.0033847357,
-0.0122701591,
0.0388100743,
-0.0283045415,
-0.0163360089,
-0.0930742249,
0.0008112385,
-0.0238745213,
-0.0465164483,
0.0079494836,
-0.0212782919,
-0.0449207909,
0.0432889275,
-0.0197676234,
-0.0354983732,
0.0066088825,
0.0170553755,
0.0102682011,
-0.0154324854,
-0.01914379,
0.0567365251,
0.0189622901,
0.033302784,
-0.0125084799,
0.1123753041,
0.0105923424,
-0.0219444707,
0.0170628112,
0.0241665486,
-0.0649417639,
-0.0448701903,
-0.0412868038,
0.0550899245,
0.0457976386,
0.0111298282,
0.0497249626,
-0.0149021819,
0.0100060413,
-0.0059444802,
0.0733963624,
-0.0072181667,
-0.0412448235,
0.015033355,
0.034487579,
-0.0097705321,
-0.0031672076,
0.0738881081,
0.0124293193,
0.0392317623,
0.0085089216,
0.0459191054,
-0.0721336454,
-0.0186313391,
0.0195566341,
0.027693212,
0.0049892953,
0.0074510137,
-0.0483907945,
-0.0358954258,
0.0052763587,
0.0025078601,
-0.0168521907,
-0.0172673874,
0.0586171709,
0.0174250826,
-0.0231986772,
0.0068050819,
-0.0235322751,
-0.0097274119,
0.0114702098,
-0.0330852158,
-0.0081297103,
-5.436774147e-33,
-0.0480495989,
-0.034372054,
0.0389698818,
-0.0122001441,
-0.026359586,
-0.0328243822,
-0.053569831,
0.0139808469,
-0.02056114,
-0.0150541076,
-0.0193133708,
0.0186828692,
0.016877966,
0.0379233062,
-0.0320516825,
-0.0187867023,
0.045982331,
-0.0136571825,
-0.0284043718,
0.0406293385,
0.0350488871,
0.0127257258,
0.0214524716,
-0.0366579667,
0.0012320863,
-0.0138628883,
-0.0174557157,
0.0262466241,
0.0114867706,
0.0340655185,
-0.0036557044,
0.0112644173,
-0.0089174276,
-0.0419079736,
-0.0144421868,
0.0345747732,
-0.0097922431,
-0.0120465811,
0.0070087127,
0.0000489547,
-0.0051270919,
0.0046973974,
-0.0190304387,
-0.0327075049,
0.0219732895,
0.00612634,
-0.0067375242,
-0.0172689445,
-0.041482579,
-0.0333713517,
-0.0542993993,
0.0257069319,
-0.0238946546,
-0.0902017727,
-0.0118394429,
0.1117794961,
0.042893894,
-0.0208678972,
-0.0177580379,
0.0142591773,
-0.0262324959,
-0.0113023492,
0.0434081741,
0.0330284052,
0.0181186851,
0.0904982686,
0.0586340092,
0.0128288297,
-0.0026946547,
-0.0293436721,
-0.0042268867,
0.031253498,
0.0360099673,
0.0609608777,
0.015492226,
0.0099025732,
-0.0374012217,
0.0125914905,
0.0225603282,
0.011507093,
-0.0186093077,
0.0229858849,
-0.0204709079,
0.0003503581,
-0.019786207,
-0.0065788049,
-0.0041203941,
-0.0066857985,
0.0388524607,
0.0375425331,
-0.0179875307,
-0.0207283236,
0.0858232751,
-0.0107533028,
0.0128589747,
0.0161998197,
-0.0359236188,
0.0249962173,
-0.0286010504,
0.0018764075,
-0.0338585488,
-0.028879201,
0.0637874827,
0.0334812216,
0.0330886655,
0.0226053111,
0.0184934177,
0.0084942896,
-0.061164055,
0.0096941534,
-0.0502671413,
0.0678892657,
-0.0357779935,
0.0085655823,
0.0198412351,
-0.0219735354,
-0.0289256368,
0.0894920155,
-0.0227579307,
-0.0546509214,
-0.0181219522,
-0.0239823088,
-0.0104985852,
0.0121653164,
-0.0367876031,
-0.002162012,
-0.0177749265,
-0.0348106176,
0.033337038,
-0.037252143,
0.0036408929,
-0.0075669903,
2.557e-7,
-0.0272685345,
0.0225730501,
0.0322389826,
0.0184945781,
0.0271595325,
0.0672443733,
-0.0243060961,
0.0233665351,
0.0295797084,
-0.0328157209,
0.0084558027,
0.0050450517,
0.0664941519,
0.0325137228,
-0.1242411658,
-0.0239432119,
-0.0633067861,
-0.0144627485,
-0.0567605384,
0.0158530436,
0.0345884003,
0.0771894306,
0.0010146699,
0.0010748256,
0.0583572239,
-0.0153296916,
0.0559477322,
-0.0138907311,
-0.010904952,
0.0173578635,
0.0022135116,
-0.009038073,
-0.0237447228,
-0.0188609045,
-0.0112109873,
0.0204711296,
-0.021416137,
0.0979879498,
-0.0003747948,
-0.0420801453,
-0.0132433986,
-0.0054872432,
0.0056641251,
-0.0203714315,
0.0386682488,
-0.054635521,
-0.0088036628,
-0.0543142669,
-0.0234600976,
-0.0275084618,
0.0164163187,
-0.002496148,
-0.0063127126,
0.0154808071,
0.0132530378,
-0.01263397,
-0.002045478,
-0.0740373582,
-0.0371936411,
-0.027284652,
-0.0421893857,
-0.0354176722,
-0.0300900284,
-0.0217814371,
0.0028710335,
-0.0098804552,
-0.0183137842,
2.109694379e-34,
0.0396840833,
-0.0099167731,
0.0406630114,
0.0621868521,
0.0077068796,
-0.0037271292,
0.0090339109,
-0.0048112948,
0.0592548773,
-0.0514862984,
-0.0133162318
] |
http://arxiv.org/abs/2402.02262v1 | http://arxiv.org/pdf/2402.02262v1 | Data Quality Matters: Suicide Intention Detection on Social Media Posts Using a RoBERTa-CNN Model | 2024-02-03T20:58:09 | 2024-02-03T20:58:09 | Emily Lin, Jian Sun, Hsingyu Chen, Mohammad H. Mahoor | cs.CL | cs.CL, cs.AI | Suicide remains a global health concern for the field of health, which
urgently needs innovative approaches for early detection and intervention. In
this paper, we focus on identifying suicidal intentions in SuicideWatch Reddit
posts and present a novel approach to suicide detection using the cutting-edge
RoBERTa-CNN model, a variant of RoBERTa (Robustly optimized BERT approach).
RoBERTa is used for various Natural Language Processing (NLP) tasks, including
text classification and sentiment analysis. The effectiveness of the RoBERTa
lies in its ability to capture textual information and form semantic
relationships within texts. By adding the Convolution Neural Network (CNN)
layer to the original model, the RoBERTa enhances its ability to capture
important patterns from heavy datasets. To evaluate the RoBERTa-CNN, we
experimented on the Suicide and Depression Detection dataset and obtained solid
results. For example, RoBERTa-CNN achieves 98% mean accuracy with the standard
deviation (STD) of 0.0009. It also reaches over 97.5% mean AUC value with an
STD of 0.0013. In the meanwhile, RoBERTa-CNN outperforms competitive methods,
demonstrating the robustness and ability to capture nuanced linguistic patterns
for suicidal intentions. Therefore, RoBERTa-CNN can detect suicide intention on
text data very well. | Data Quality Matters: Suicide Intention Detection on Social Media Posts Using a RoBERTa-CNN Model: Suicide remains a global health concern for the field of health, which
urgently needs innovative approaches for early detection and intervention. In
this paper, we focus on identifying suicidal intentions in SuicideWatch Reddit
posts and present a novel approach to suicide detection using the cutting-edge
RoBERTa-CNN model, a variant of RoBERTa (Robustly optimized BERT approach).
RoBERTa is used for various Natural Language Processing (NLP) tasks, including
text classification and sentiment analysis. The effectiveness of the RoBERTa
lies in its ability to capture textual information and form semantic
relationships within texts. By adding the Convolution Neural Network (CNN)
layer to the original model, the RoBERTa enhances its ability to capture
important patterns from heavy datasets. To evaluate the RoBERTa-CNN, we
experimented on the Suicide and Depression Detection dataset and obtained solid
results. For example, RoBERTa-CNN achieves 98% mean accuracy with the standard
deviation (STD) of 0.0009. It also reaches over 97.5% mean AUC value with an
STD of 0.0013. In the meanwhile, RoBERTa-CNN outperforms competitive methods,
demonstrating the robustness and ability to capture nuanced linguistic patterns
for suicidal intentions. Therefore, RoBERTa-CNN can detect suicide intention on
text data very well. | [
0.0779502541,
0.094859302,
-0.0018154788,
-0.0074847401,
-0.0083680255,
0.0270642694,
-0.013258039,
-0.0068434109,
0.0175024122,
-0.013402313,
-0.003602101,
0.0316160694,
-0.0639525726,
-0.0001273582,
-0.0760323107,
-0.0192239601,
0.0061741536,
0.0427341275,
0.0576271638,
0.0050948076,
0.0249504205,
0.0450014733,
-0.0114038335,
0.0534211732,
-0.0115216058,
-0.0144736776,
0.0255269799,
0.0253401399,
0.0222363919,
-0.0483526103,
0.0410263799,
0.045498427,
-0.0286081862,
0.0261677168,
0.0000021197,
-0.0112898145,
-0.0125271743,
-0.0146694407,
-0.0268865973,
-0.0572213754,
0.0274011828,
-0.0734748691,
0.0101219807,
0.0015942033,
-0.0609116033,
-0.0338653475,
0.0204494279,
0.0578973331,
0.035985209,
0.0928993374,
-0.0244124141,
-0.0401821397,
0.0633997768,
-0.007238924,
-0.0299425907,
0.0072166845,
0.0338934734,
0.0023081515,
-0.0074441787,
0.002205675,
-0.0332405306,
0.0112678735,
0.0110598067,
0.0188077707,
-0.0115368571,
-0.0167898778,
-0.0265177786,
-0.061002098,
0.0211676806,
0.0500642508,
-0.0397335328,
0.0040725335,
0.0380760767,
-0.0181382168,
-0.0271591153,
-0.0371754244,
0.0115047414,
-0.0235211123,
-0.0264902022,
0.0314035825,
0.023536047,
0.0318423249,
0.007328792,
0.0183844622,
0.0591109209,
-0.0198238436,
-0.0105866585,
0.0054052565,
0.0029611283,
0.0011914668,
-0.0473313928,
-0.0349954404,
0.0203826986,
0.01112282,
-0.0550298691,
0.0127855521,
-0.0673603565,
-0.0741476491,
0.0215615798,
-0.0437780432,
0.066002503,
0.0381170586,
-0.0278927293,
-0.0058802534,
0.0078572277,
-0.0121182334,
0.0191233605,
-0.0177779868,
-0.0402371399,
0.0278221238,
-0.0619528927,
0.0297917016,
-0.0314907841,
0.0052811061,
0.0554628484,
-0.0209863558,
-0.0803885087,
0.0214341003,
-0.004055724,
-0.0021536776,
-0.0126314405,
-0.0030287632,
-0.0258498993,
0.0036371322,
0.0264662039,
0.023629833,
-0.0432866067,
0.0130905863,
-0.0199701935,
-0.0321081392,
0.0355124511,
0.0186874717,
0.0334024914,
-0.0241213292,
0.0225776099,
0.0500389151,
0.0427900739,
-0.0250683222,
-0.0470294803,
-0.0361595154,
-0.0483253524,
-0.0418151245,
-0.0030621914,
-0.0142701929,
0.0022300682,
-0.0183356293,
-0.0194701459,
-0.0368348733,
0.0414049216,
-0.0287264287,
-0.0001258868,
0.0619135574,
-0.0276999585,
-0.0162558388,
0.0428663976,
-0.0003352529,
-0.020821996,
0.0580983162,
-0.0071136379,
0.075987719,
0.0241018515,
-0.0100957351,
0.0036491202,
-0.0162897222,
0.0185799152,
-0.0188010726,
-0.0303696804,
0.0022315406,
-0.0974856913,
0.0278955903,
-0.0041817217,
0.0010310139,
0.0084328912,
-0.0072161043,
0.0557360649,
0.0386027321,
-0.0061612469,
0.0538631044,
0.0252706762,
0.0551291592,
-0.0334360823,
-0.0590043031,
-0.0356449671,
0.0134726577,
-0.0142211178,
-0.0064020548,
0.015408827,
-0.013150502,
0.0016905742,
-0.0498870276,
-0.0266810153,
-0.0204642043,
-0.0175971482,
-0.0124379853,
0.0048294947,
-0.003700773,
-0.0259695109,
-0.01176567,
0.0167930704,
-0.0022813948,
-0.0326383077,
-0.0372015424,
0.0371297002,
0.0288927332,
-0.0015753398,
-0.0386485793,
-0.0375626162,
-0.0189366955,
0.0328008235,
-0.000261728,
0.0115610799,
0.0343333557,
0.0234675538,
-0.0341243148,
-0.0479874499,
0.034916807,
-0.0066300291,
-0.0395063907,
0.0115877185,
0.0526351631,
-0.0227182768,
0.0035166857,
-0.0361232944,
0.0364172533,
0.0540167727,
-0.0274675675,
0.0149995768,
0.0376416333,
0.0332605764,
0.0051131789,
0.000658926,
-0.0188481528,
0.0007770464,
-0.0209272392,
-0.0621994697,
-0.037942186,
-0.008751356,
-0.0455119573,
-0.0143850669,
-0.0706914067,
-0.010705783,
0.0970408469,
-0.0383364148,
-0.0282892734,
0.0193256568,
-0.052271869,
0.0043850658,
-0.0224376787,
-0.0501084812,
-0.0254806764,
0.0136765838,
0.0260079317,
0.0697516054,
-0.0347560793,
0.0179358833,
-0.0143978307,
-0.0410542376,
-0.002769425,
-0.0261040889,
-0.0677964091,
-0.0672639459,
-0.0254999399,
0.0042068455,
0.0294670854,
0.0086250938,
-0.0095535954,
-0.0234040339,
-0.1011186391,
0.0349483043,
-0.0495765992,
-0.0031354169,
-0.0150004476,
-0.0012326429,
-0.0217549037,
-0.029971445,
-0.0268058833,
0.0229899753,
-0.019120343,
0.0061183991,
0.0214742105,
-0.0240636393,
0.0148676205,
-0.0262768697,
-0.0267581362,
0.0113878082,
-0.0018204605,
0.0071598706,
-0.0492984429,
-0.0092021795,
0.0055939993,
-0.0262622871,
-0.0046134787,
-0.0001052905,
-0.0735381767,
-0.0107266968,
-0.0064980388,
0.0284488741,
0.0323648266,
0.024396779,
0.0346075259,
0.0041900137,
0.0828562602,
-0.0150019797,
0.0347362645,
-0.0664469227,
0.0112239718,
-0.0473456606,
-0.0241504796,
-0.0133096594,
0.0304068513,
0.0136283888,
-0.0267428085,
-0.0264499336,
-0.0486063622,
-0.0363048427,
0.0011303242,
0.035610076,
-0.0116166435,
0.0298520811,
-0.0150603913,
0.0016232945,
-0.0473101139,
0.0421707779,
0.0509827845,
0.0047659315,
-0.0087781316,
-0.0263385084,
-0.0351980999,
0.0319144204,
0.0312497262,
0.0129020987,
-0.0054006432,
0.0052863336,
0.0460375622,
-0.0182789378,
-0.0347125158,
-0.0154313026,
0.0088396017,
-0.0081816204,
0.0296881981,
0.0103033818,
-0.0290144905,
-0.0554945953,
-0.0365731157,
-0.076921232,
-0.0417463705,
-0.0417135656,
-0.0047125877,
0.0038270592,
-0.0253853612,
-0.0280449465,
0.0299364906,
-0.0270906035,
-0.0340381302,
-0.0136501305,
0.0052363323,
0.0806210861,
-0.0887775868,
0.0415054224,
0.0051489277,
0.049742423,
0.0509285443,
-0.0793065056,
0.1017545313,
-0.0492005982,
0.0193754956,
-0.0143373609,
0.0163599066,
-0.0020513812,
0.0005660229,
0.0336982496,
-0.0089903567,
-0.0007865633,
0.0747685879,
-0.0034131138,
-0.0574458018,
0.0122509701,
0.0254960507,
0.0137429172,
0.0048809978,
-0.0100578088,
-0.0064377524,
-0.0667390749,
-0.0251857154,
0.037940897,
-0.0452875011,
-0.0276309084,
-0.0015124895,
-0.0264601074,
0.0624032021,
-0.0607962757,
-0.0003030242,
0.0434201621,
-0.0294021256,
0.0284761935,
-0.0466483943,
0.0149640106,
-0.0700362921,
-0.015539417,
0.0053966185,
-0.0138431275,
-0.093618378,
-0.0062157391,
-0.0055051865,
-0.0211781338,
0.0859188959,
-0.0069348812,
0.061182145,
-0.0159580521,
-0.0310926046,
0.0255514849,
0.0363444351,
-0.0070948093,
-0.0084995823,
0.1039696559,
0.0053994334,
-0.036798995,
-0.0782578215,
0.0706023872,
-0.0585800782,
-0.0286882482,
-0.0264211204,
0.023596324,
-0.0365121774,
-0.0363476649,
-0.0543316267,
0.0506317392,
0.0217335112,
0.0555923097,
0.0263098255,
-0.0675941706,
-0.0162924491,
0.0129425693,
-0.0211203508,
-0.0699732974,
-0.0189113542,
-0.0005120061,
-0.0377206914,
-0.0555692613,
-0.027240634,
-0.0275806673,
0.0027941344,
-0.0411051437,
-0.0263330415,
-0.0181092769,
0.0314970724,
0.0189313814,
-0.1029619426,
0.0431477241,
0.026171634,
0.0170407407,
-0.0259988829,
0.0304484833,
0.0313203521,
-0.0212738533,
-0.0154038984,
-0.0019425154,
-0.0692038834,
0.0190964639,
0.046880763,
0.0282346159,
-0.0002956325,
-0.017312862,
-0.0170414411,
-0.0476936884,
0.0435486808,
-0.0136111798,
0.0177934375,
-0.0210571215,
0.0308631528,
0.0114274835,
0.0090248045,
0.0483124964,
0.0158975013,
0.0781968459,
0.0340570398,
-0.0118660033,
-0.0509621203,
0.0536645912,
-0.0145880552,
-0.0112682702,
0.0157971084,
0.0074541033,
-0.022340335,
-0.0018972552,
-0.0188537277,
0.0455586985,
0.0007462116,
-0.0285016075,
-0.0234142747,
0.076827012,
0.0220782999,
-0.0479178987,
-0.022502061,
-0.0081204865,
-0.0418261625,
-0.0164463222,
0.0082270131,
0.0601116195,
0.0652436316,
0.0361547954,
0.0186135694,
-0.0585238039,
0.0560239591,
-0.0081353569,
0.0077571091,
-0.0235385485,
0.03054858,
0.0117507977,
-0.0103383036,
0.081947878,
0.0378682353,
-0.0609608442,
-0.0560090914,
-0.0354369469,
-0.0008564188,
0.0307890605,
-0.0054105744,
0.0409438163,
-0.0175420158,
0.0366780348,
-0.0005767833,
0.0581143945,
-0.0144051993,
-0.0247368179,
-0.0051513696,
-0.0084887985,
-0.0125541501,
-0.0201354809,
0.0060591721,
0.0225027688,
0.0234010015,
0.0322378017,
-0.0043418356,
-0.0180836525,
0.0351604261,
0.0449869707,
0.0434159413,
-0.0213043876,
0.0179260354,
0.0179707874,
-0.0423374102,
0.0670741946,
-0.0373882763,
-0.0186310094,
-0.0047341678,
0.0216504838,
0.0117130112,
0.0138570555,
-0.0271546151,
-0.067014344,
0.0075987447,
-0.0214501172,
-0.0302487146,
-0.0299904998,
-5.65351657e-33,
0.0477351137,
0.0021111146,
-0.0505437367,
0.0229208227,
0.0087026292,
0.0445492938,
0.0086684693,
-0.000631915,
-0.0213997532,
0.053347528,
0.0190240871,
0.0947350562,
-0.0006072586,
0.0534396842,
0.0217518974,
-0.0222011898,
-0.0025603557,
0.0053161988,
0.0233753528,
0.0138541227,
-0.0041616582,
0.0755667016,
0.0759057477,
-0.0367106721,
0.0199946985,
0.0176095888,
-0.0134484656,
0.0218161643,
0.0588161238,
0.0427179076,
-0.0443983413,
0.0528924242,
0.0522618368,
0.0218165442,
-0.0150310798,
0.0327921622,
-0.0452746935,
-0.0446504913,
-0.0200479757,
0.0298997667,
-0.0640683621,
0.0049784207,
0.0287229344,
-0.0028024747,
-0.0183170941,
-0.0297634602,
0.0401391499,
0.0118591841,
0.0262020398,
-0.0049202107,
-0.0195027906,
0.0050744694,
-0.0216888525,
-0.0135102002,
0.0219104383,
0.0054259989,
0.0117297415,
0.0150638046,
-0.0735131651,
-0.0116087552,
-0.0830911547,
-0.0008890061,
-0.0298975725,
-0.0466982834,
-0.0502351373,
0.0791664645,
0.0196342431,
-0.0169462617,
-0.0152550153,
-0.0024712342,
0.0210785531,
0.0836420432,
0.0788012147,
0.0244164485,
0.0489183813,
-0.0127692372,
-0.0598938763,
0.0199010745,
0.0466872975,
-0.0058051855,
0.0353659205,
-0.0193219464,
-0.0346951373,
0.0280907229,
0.020929113,
-0.0507242791,
-0.0264771618,
0.0143755842,
-0.0068463176,
0.0094571291,
-0.0194681697,
-0.0033233517,
-0.0156151205,
-0.0527292527,
0.0167706776,
0.0260827169,
0.0210399777,
0.0305961426,
0.0191106293,
0.0107149091,
-0.0025447893,
0.0272872262,
0.0048124311,
-0.0011407873,
0.03604329,
-0.0045657051,
0.0685624033,
0.0071849157,
-0.034123566,
-0.0249266475,
-0.0044298177,
0.0269792434,
-0.0113843801,
0.0373775847,
-0.0160172917,
0.0211138111,
-0.0004393887,
0.0520857163,
0.0355885737,
0.0774271935,
-0.0482360609,
-0.0223564971,
0.0062545626,
0.0044250046,
0.0092699779,
-0.071373187,
-0.024536917,
0.0622339062,
-0.0074844654,
-0.080788523,
0.0089710886,
0.003620737,
2.767e-7,
0.0663043484,
0.0367790535,
0.00474957,
0.0194790661,
-0.0211040042,
0.0271275807,
0.0419703871,
0.0599693432,
0.0235995073,
0.0087937396,
0.0209171921,
0.0333311409,
0.040332336,
0.0622326098,
-0.0642596036,
-0.0280044451,
-0.0176197775,
0.0126814954,
-0.0625377372,
0.0149956495,
0.0973943099,
0.0787956566,
0.0518686101,
-0.0524516441,
0.0263375398,
-0.0079833614,
0.0216187667,
-0.027537616,
0.0491846353,
0.0305273235,
0.0271757096,
-0.019302927,
0.0070015909,
-0.0371582806,
-0.0038794063,
-0.0509836078,
0.0064784428,
0.0762739554,
-0.0350299887,
-0.0299796686,
0.0119138993,
0.0367709175,
0.0261119697,
-0.0210947022,
0.0682694688,
-0.0343416445,
-0.0052500917,
-0.0816707909,
-0.0157200899,
0.0027748698,
0.0670730621,
0.0023527471,
-0.0104597993,
0.0362736173,
0.0313559324,
-0.0050621075,
-0.0289302915,
-0.0819983855,
-0.0331452787,
-0.0025398366,
-0.0082001947,
0.0061321943,
-0.0645726249,
-0.0137784882,
0.0529826097,
-0.0445056185,
0.0078401323,
2.031019369e-34,
0.0152940797,
0.0288999956,
0.0287658833,
-0.0283661354,
-0.0245502945,
0.0344127826,
0.008635154,
-0.0022338219,
-0.0186051354,
-0.0640187114,
0.00576908
] |
http://arxiv.org/abs/2402.02255v1 | http://arxiv.org/pdf/2402.02255v1 | Frequency Explains the Inverse Correlation of Large Language Models' Size, Training Data Amount, and Surprisal's Fit to Reading Times | 2024-02-03T20:22:54 | 2024-02-03T20:22:54 | Byung-Doh Oh, Shisen Yue, William Schuler | cs.CL | cs.CL, cs.LG | Recent studies have shown that as Transformer-based language models become
larger and are trained on very large amounts of data, the fit of their
surprisal estimates to naturalistic human reading times degrades. The current
work presents a series of analyses showing that word frequency is a key
explanatory factor underlying these two trends. First, residual errors from
four language model families on four corpora show that the inverse correlation
between model size and fit to reading times is the strongest on the subset of
least frequent words, which is driven by excessively accurate predictions of
larger model variants. Additionally, training dynamics reveal that during later
training steps, all model variants learn to predict rare words and that larger
model variants do so more accurately, which explains the detrimental effect of
both training data amount and model size on fit to reading times. Finally, a
feature attribution analysis demonstrates that larger model variants are able
to accurately predict rare words based on both an effectively longer context
window size as well as stronger local associations compared to smaller model
variants. Taken together, these results indicate that Transformer-based
language models' surprisal estimates diverge from human-like expectations due
to the superhumanly complex associations they learn for predicting rare words. | Frequency Explains the Inverse Correlation of Large Language Models' Size, Training Data Amount, and Surprisal's Fit to Reading Times: Recent studies have shown that as Transformer-based language models become
larger and are trained on very large amounts of data, the fit of their
surprisal estimates to naturalistic human reading times degrades. The current
work presents a series of analyses showing that word frequency is a key
explanatory factor underlying these two trends. First, residual errors from
four language model families on four corpora show that the inverse correlation
between model size and fit to reading times is the strongest on the subset of
least frequent words, which is driven by excessively accurate predictions of
larger model variants. Additionally, training dynamics reveal that during later
training steps, all model variants learn to predict rare words and that larger
model variants do so more accurately, which explains the detrimental effect of
both training data amount and model size on fit to reading times. Finally, a
feature attribution analysis demonstrates that larger model variants are able
to accurately predict rare words based on both an effectively longer context
window size as well as stronger local associations compared to smaller model
variants. Taken together, these results indicate that Transformer-based
language models' surprisal estimates diverge from human-like expectations due
to the superhumanly complex associations they learn for predicting rare words. | [
0.0243426915,
0.0428009033,
-0.0224817805,
0.0223373882,
-0.009085075,
0.0296109505,
0.0505052619,
-0.0016341867,
-0.0643556789,
-0.0146280099,
-0.0414645299,
-0.0541573204,
0.0654624403,
0.0165779088,
0.0174956303,
-0.0329977721,
0.0345299765,
-0.0311541744,
-0.0195315499,
-0.0163556058,
-0.0227506105,
-0.0084934672,
0.0216468628,
0.0317101665,
0.0083421646,
-0.0026734895,
0.0225247387,
-0.0296353176,
0.0046280865,
0.0090198079,
0.0078563699,
0.0746814087,
0.0126826717,
0.0694401562,
0.0000016522,
-0.0576224811,
-0.0038830999,
0.0268502217,
-0.0138431797,
0.0314108133,
0.0627575144,
-0.0435963348,
0.0012726665,
0.0288136546,
-0.0048504709,
-0.0248318054,
0.0472458377,
0.0248274803,
-0.0598848872,
0.045284193,
-0.0127546284,
-0.0327373035,
0.0488268398,
-0.0233059879,
0.0314627029,
-0.0321890935,
0.0088298675,
-0.002184591,
-0.0181824677,
-0.0126401838,
-0.0727504343,
0.0539513528,
0.0190082118,
0.044607643,
0.0138577512,
0.0511747301,
0.009340167,
-0.0437653586,
-0.0019643079,
0.0121247154,
-0.000199019,
-0.0375504158,
-0.0340633579,
0.0056792232,
-0.0210215598,
0.0695591271,
-0.0069882334,
0.022756651,
0.0036505216,
0.0093535772,
0.0212807134,
0.0068313871,
0.0255672596,
-0.0096952803,
-0.057550218,
0.0321528278,
0.025913395,
-0.0461531878,
0.0169256944,
0.0208953805,
-0.037125349,
0.0184667241,
-0.0081446432,
-0.0225614458,
0.0246601421,
0.0344048589,
-0.0394654945,
-0.0405778512,
0.0401445888,
-0.0146797597,
0.0323693268,
0.0444764122,
-0.0214840434,
-0.0005276573,
-0.0259655807,
0.0557830408,
-0.005995376,
0.0135971233,
-0.0429126695,
-0.0297406986,
-0.015687814,
-0.0087377392,
0.0031047862,
-0.0445198566,
-0.0433334112,
-0.0203544311,
-0.0781438798,
0.0028141923,
0.0583822243,
-0.0340279154,
0.0379930846,
0.0111555271,
-0.0562023893,
-0.0015222543,
-0.0303659961,
0.0338066593,
0.0046419343,
0.0259596985,
-0.0463493653,
-0.054866977,
0.0238728672,
0.0008124466,
0.0209608972,
-0.0615533926,
-0.0006444992,
0.0364907794,
-0.0373062156,
-0.0600782409,
-0.0311833508,
0.0243559331,
-0.0467456914,
-0.0538538806,
-0.0133834248,
-0.0595637821,
-0.0064451015,
0.0642644688,
0.0002562102,
-0.0282648429,
-0.034075439,
0.0101799667,
-0.047952991,
0.0379515961,
0.0020648169,
0.0582088269,
0.0321798846,
-0.0034854051,
0.0081635825,
0.0663043186,
0.045233421,
-0.0177333131,
-0.0043457467,
-0.0342481434,
-0.0251802579,
0.038172096,
-0.008337005,
0.022545632,
-0.011451086,
-0.0239190813,
0.0052656396,
-0.0185186397,
0.0724661574,
0.0391882099,
-0.0088528618,
-0.0576176234,
0.0830872357,
0.0653066337,
0.0429311767,
0.0536325462,
0.019514095,
0.0056909565,
0.0047617271,
0.0059331786,
-0.0445868373,
-0.0040113083,
-0.0157952104,
0.0195342321,
-0.0227832943,
0.048175659,
-0.0050216764,
-0.0285163037,
-0.0067632091,
-0.0001206675,
0.0365609229,
-0.0026309176,
0.0645068288,
-0.0153905721,
-0.026450308,
0.082648173,
0.0467696302,
-0.0094146663,
-0.0367756113,
-0.017618278,
-0.0082159396,
0.033212319,
-0.0080953734,
0.0093715414,
0.0185962189,
0.0166136939,
-0.0089360047,
0.0419041701,
0.0231281873,
0.0230775997,
0.0175458733,
-0.0204659719,
0.008315335,
0.0370572619,
0.0173296873,
0.0116451876,
0.0265872907,
0.0401399694,
0.0177723169,
-0.0603778958,
-0.0637758896,
0.0330108367,
-0.1207606792,
-0.024360735,
-0.0399742201,
0.0126713831,
-0.0034210056,
-0.0370758995,
-0.0041709435,
0.0037820777,
-0.0023283947,
-0.0004458437,
0.0344662666,
0.0248813573,
-0.0606174916,
0.0162972547,
0.0193443075,
-0.017927492,
0.0306745488,
0.0501047373,
0.0359097421,
-0.0158922244,
-0.0144812828,
-0.0517591424,
0.0737740919,
-0.0193712972,
0.037362922,
-0.0738978982,
0.0431621782,
-0.0318568982,
0.0253116209,
-0.030034164,
0.0218951553,
-0.0180284474,
-0.0227288902,
0.0049697394,
0.0028109911,
-0.0453121513,
0.0055627506,
-0.0207724646,
-0.0068676732,
0.0201731194,
-0.0347096585,
-0.0229989812,
-0.0380554497,
-0.0797855929,
-0.0335255228,
-0.0136449011,
-0.0114652533,
-0.0002804238,
-0.005824388,
-0.0214854628,
0.022262685,
-0.0051330482,
-0.0085186921,
-0.0071497797,
0.0526081249,
0.002234512,
-0.0024002406,
-0.045103278,
-0.0504020602,
-0.03130088,
-0.0315604173,
0.0702567846,
0.0153155401,
-0.0141573502,
-0.0635984465,
0.0676593482,
0.0014206539,
0.0094552394,
-0.0228096563,
-0.0480390228,
0.0190659203,
0.0322976075,
0.0589289404,
0.0480822623,
0.0291537661,
0.0278066229,
0.0622798838,
0.0182378199,
-0.0371181592,
-0.0255448595,
-0.0505156107,
-0.04829197,
-0.0701445267,
-0.0758210644,
-0.035571415,
0.0930860713,
0.0082789762,
-0.0040950305,
-0.044527445,
0.0193062052,
-0.016562419,
-0.0322295837,
0.0108695012,
0.0157987587,
-0.0250587165,
0.0602571182,
-0.0153663494,
-0.0811619461,
0.0355080664,
-0.0118975043,
0.0061725862,
-0.0446318164,
-0.0088671669,
-0.0240164921,
0.041559305,
0.017690843,
0.0221835971,
-0.003633548,
-0.0242671445,
-0.0702065453,
-0.0086454246,
0.0444547348,
-0.0141711468,
0.0990353599,
-0.0032756794,
-0.0003267532,
0.0043630437,
-0.0312698819,
-0.0340264551,
-0.0444462113,
-0.0273370314,
-0.0227965936,
0.0296741985,
-0.0525496267,
-0.0720746815,
-0.0224702787,
0.0174044874,
0.0389327146,
-0.011974345,
-0.0349530503,
0.0156360827,
0.0144549254,
0.0479651168,
0.0065788492,
0.0398685262,
-0.0688729584,
0.0180806201,
0.0141707547,
0.0008131987,
0.0289605279,
-0.0330832638,
-0.0271126255,
0.0133464318,
-0.0141676664,
-0.0082366839,
0.0186192598,
0.0605303198,
-0.0278407447,
0.0011084911,
0.107501395,
0.0107207531,
-0.0284285415,
0.0337377377,
-0.0255743656,
0.084468022,
0.005259369,
0.0121458797,
-0.0257063061,
0.0055319639,
-0.0127692875,
0.0284626205,
-0.0031206044,
-0.0242797881,
-0.0240866132,
-0.0803232267,
-0.0165261328,
0.0112480652,
-0.0458012782,
0.0011545647,
-0.0229142793,
-0.009813522,
-0.0409072153,
0.0514228828,
0.0128752571,
-0.0449162312,
-0.0066625201,
0.0389736481,
-0.0104642082,
-0.0384417661,
0.0396901444,
-0.0223175772,
0.0171124954,
0.010538497,
0.0244492292,
-0.0453767926,
-0.035273347,
0.0049074246,
0.0232762676,
0.03208334,
0.0225140043,
0.066522643,
0.0230863914,
0.0136583438,
-0.0168162547,
0.0033789971,
-0.014904825,
-0.005229237,
-0.0358478539,
0.0498973988,
0.0328263864,
-0.0518086702,
-0.0245595034,
-0.0186241455,
-0.0238056518,
0.0643739328,
0.0158113725,
-0.0471492484,
-0.0310149044,
0.0325221568,
-0.0076763821,
0.0190813206,
-0.0584043525,
0.0563940927,
0.0075552687,
-0.0717549473,
-0.0474997722,
-0.0319608785,
-0.0063638263,
0.0276473854,
0.0272640102,
0.0301154647,
0.0069469945,
-0.0188596547,
-0.0520486012,
0.0566593707,
0.0282743331,
-0.0062058759,
-0.0795231983,
-0.0137648582,
0.0158565771,
-0.0266526751,
-0.0326283462,
-0.0367554575,
-0.1144165397,
-0.0428928956,
0.0531734452,
-0.0079246741,
0.0348192155,
0.0438637063,
0.0450707301,
-0.0532850251,
0.0688764676,
0.0334733464,
-0.0385281704,
0.0443489887,
0.0064470121,
-0.0425328538,
0.0024851479,
0.0120273139,
-0.0073708291,
0.0273931045,
0.0507850386,
-0.0047748848,
0.0034295288,
-0.0090286257,
-0.0107412264,
-0.0048058406,
0.0407908037,
-0.032674294,
-0.0095845517,
-0.0242685359,
-0.0112481574,
0.0036043508,
0.0539758466,
-0.0059658452,
-0.0277674515,
0.0528982431,
0.0224886891,
-0.0335078128,
-0.0112112435,
-0.0005583632,
-0.0034483573,
-0.0489167832,
0.0423303097,
-0.0134695675,
-0.0093508409,
0.0504163243,
0.0488161519,
-0.0091329096,
0.0090113739,
-0.032169342,
0.0519628301,
0.0349811278,
0.0415570252,
-0.0188215096,
0.0071540922,
-0.0395924263,
-0.0138932755,
-0.07001964,
-0.0268194024,
-0.0209997296,
-0.029326316,
-0.0024087587,
-0.0637931004,
0.0495918877,
-0.0370492972,
0.0219641048,
-0.042763792,
0.0059265126,
-0.0225878283,
0.0267014205,
0.0095022945,
-0.0021238246,
-0.0167705342,
0.019930087,
0.0955065861,
0.0618428513,
0.006710561,
0.0108209755,
-0.0399738736,
-0.0141729703,
-0.0088291233,
0.0299134292,
-0.0256511718,
-0.0401366204,
-0.0119492589,
-0.0805461183,
-0.0656762123,
0.0273833647,
-0.0349535197,
0.021960916,
-0.0149079431,
-0.0124230478,
0.0619058982,
-0.0359251387,
-0.0200371798,
-0.0425606444,
0.0079483632,
0.0013467281,
-0.0080218492,
0.0157804471,
-5.821311409e-33,
0.0062171463,
-0.0192788951,
0.0130418595,
-0.0028408985,
0.0149497036,
0.01282134,
-0.0109163588,
0.0263021514,
-0.0365685411,
0.0277791247,
-0.0152750006,
0.0124948788,
0.0182201043,
0.0107408687,
0.0094284182,
0.0124717392,
0.0391657278,
0.0223035589,
-0.0035833023,
-0.0243231952,
0.0434995443,
0.0474985987,
0.0649366453,
-0.0683624744,
-0.0013919336,
0.0201521851,
0.0267325751,
-0.020417124,
0.0394629575,
0.0625759959,
-0.0338586904,
0.0817049071,
0.0226003714,
-0.055126179,
-0.016705649,
0.0253365189,
-0.064030014,
-0.0129536148,
0.0394861326,
0.0223045778,
0.0294279568,
0.0089904899,
0.0586774088,
-0.0196290556,
-0.0387058891,
-0.0592055693,
0.0525475331,
-0.0358593501,
-0.0562850609,
-0.0237684529,
-0.0762798339,
0.0354058444,
0.0231092572,
0.0193542652,
0.0210828539,
0.0543116666,
0.0096823154,
0.0284193736,
-0.0398635417,
-0.0020904762,
0.0659869537,
0.0524750277,
0.0360863619,
0.0363031328,
0.0111516053,
0.0561782569,
0.0640785173,
0.0347062051,
0.0064654755,
-0.0119465478,
0.0939950347,
0.0228747614,
0.0188528895,
0.0586082861,
0.081516251,
-0.0481765047,
-0.0229194108,
0.0179446638,
-0.0111665437,
0.0108393366,
0.0675380975,
-0.0096579744,
-0.0740645975,
-0.0333149768,
-0.0467601977,
0.0231182128,
-0.0286576245,
-0.0383816436,
0.0236561466,
-0.0051503479,
0.0083693936,
-0.0112448921,
0.0312682539,
-0.0491053946,
-0.0356033742,
0.0169945788,
-0.014915525,
0.0317536145,
-0.0318392739,
0.01463906,
-0.0367853567,
-0.0239994638,
-0.0192969069,
0.043904677,
0.0134175252,
0.0051306817,
0.0317900032,
0.0160070993,
-0.0896833763,
-0.0089149391,
-0.019016223,
-0.0504483171,
0.0302665681,
0.0090395799,
0.0422331057,
-0.0089224251,
-0.0167336054,
0.0132050086,
0.0282023475,
0.0012189364,
-0.0453553945,
0.0103925383,
-0.0035003917,
0.0285346955,
-0.0177486315,
-0.0434164219,
-0.0166620612,
0.0166016892,
0.0767926127,
-0.0341805145,
-0.0077433921,
0.0099099688,
2.438e-7,
0.0393220186,
0.0724240988,
-0.0334018692,
-0.0438592136,
0.015202825,
0.0061149066,
0.0309981648,
0.0195041727,
-0.0037210092,
-0.0088322498,
0.0118063716,
0.0140447374,
0.0543972105,
0.0220784061,
-0.0440428406,
-0.0617800206,
-0.0798159987,
0.057162337,
-0.0413731076,
0.0356894769,
-0.0115236621,
0.0812405348,
-0.007953058,
-0.0511597171,
-0.0026167114,
-0.0172989499,
0.0464223139,
-0.0588088483,
0.0052459296,
0.015079271,
0.0075500808,
0.0112627866,
0.0514079854,
0.0149472449,
-0.0180234946,
0.0410960875,
-0.0229350962,
0.0305758025,
-0.0140880756,
0.0031782638,
0.0213801675,
0.0136661613,
-0.0334081016,
-0.0894040912,
0.0342207737,
-0.0436206497,
-0.0403670296,
0.0006598916,
-0.0076407501,
0.0213381927,
0.0153264031,
-0.0077652489,
0.017264355,
0.0168800019,
0.0294684619,
-0.0406525172,
-0.0082602575,
-0.0487041846,
-0.0375197642,
0.0341152363,
-0.0411168896,
-0.0616518669,
-0.055829946,
-0.0362743065,
0.0386968628,
-0.0126673291,
0.0218055118,
2.158038651e-34,
0.0194257535,
0.0059614126,
0.0243585166,
0.007147498,
0.0204274915,
-0.0212084856,
0.0740711018,
0.0166533925,
0.0171163138,
-0.1021599621,
-0.0165409781
] |
http://arxiv.org/abs/2402.02244v3 | http://arxiv.org/pdf/2402.02244v3 | Beyond the Limits: A Survey of Techniques to Extend the Context Length in Large Language Models | 2024-02-03T19:20:02 | 2024-05-29T13:38:25 | Xindi Wang, Mahsa Salmani, Parsa Omidi, Xiangyu Ren, Mehdi Rezagholizadeh, Armaghan Eshaghi | cs.CL | cs.CL, cs.LG | Recently, large language models (LLMs) have shown remarkable capabilities
including understanding context, engaging in logical reasoning, and generating
responses. However, this is achieved at the expense of stringent computational
and memory requirements, hindering their ability to effectively support long
input sequences. This survey provides an inclusive review of the recent
techniques and methods devised to extend the sequence length in LLMs, thereby
enhancing their capacity for long-context understanding. In particular, we
review and categorize a wide range of techniques including architectural
modifications, such as modified positional encoding and altered attention
mechanisms, which are designed to enhance the processing of longer sequences
while avoiding a proportional increase in computational requirements. The
diverse methodologies investigated in this study can be leveraged across
different phases of LLMs, i.e., training, fine-tuning and inference. This
enables LLMs to efficiently process extended sequences. The limitations of the
current methodologies is discussed in the last section along with the
suggestions for future research directions, underscoring the importance of
sequence length in the continued advancement of LLMs. | Beyond the Limits: A Survey of Techniques to Extend the Context Length in Large Language Models: Recently, large language models (LLMs) have shown remarkable capabilities
including understanding context, engaging in logical reasoning, and generating
responses. However, this is achieved at the expense of stringent computational
and memory requirements, hindering their ability to effectively support long
input sequences. This survey provides an inclusive review of the recent
techniques and methods devised to extend the sequence length in LLMs, thereby
enhancing their capacity for long-context understanding. In particular, we
review and categorize a wide range of techniques including architectural
modifications, such as modified positional encoding and altered attention
mechanisms, which are designed to enhance the processing of longer sequences
while avoiding a proportional increase in computational requirements. The
diverse methodologies investigated in this study can be leveraged across
different phases of LLMs, i.e., training, fine-tuning and inference. This
enables LLMs to efficiently process extended sequences. The limitations of the
current methodologies is discussed in the last section along with the
suggestions for future research directions, underscoring the importance of
sequence length in the continued advancement of LLMs. | [
0.0449519306,
0.0602720417,
-0.0276284833,
0.0448924005,
-0.058331687,
0.0043807342,
0.0201016739,
-0.0313725173,
-0.0296085533,
-0.0394460075,
-0.0063617802,
-0.0746207088,
0.0153400842,
0.026926009,
0.0391459465,
-0.0531158112,
0.0430706404,
-0.04195638,
-0.0023858785,
0.0033479324,
-0.0123915086,
-0.010607007,
0.0307956822,
0.0185594037,
0.0203797556,
-0.0173413269,
-0.0674941838,
-0.0213225633,
0.0097378166,
-0.0545692109,
-0.0228735525,
0.0426007509,
0.0457656607,
0.0548930541,
0.0000019894,
-0.031692367,
0.0050634132,
-0.006194172,
-0.0224212836,
0.0337748379,
0.0002351653,
-0.014861702,
-0.0016700121,
-0.0104538631,
-0.03300827,
0.0154286334,
0.0613205247,
0.0026623677,
-0.0018511458,
0.0762549192,
-0.0021143924,
-0.0600178018,
0.0111522125,
-0.0632762983,
0.0207490586,
-0.0517749675,
0.0356465392,
-0.0112941647,
0.0857117921,
0.0344970673,
-0.0535736755,
0.0202251505,
0.0603859089,
0.0206259917,
-0.0172684584,
0.0448726937,
0.0110426098,
-0.021019714,
-0.005284796,
0.0396705084,
-0.0449029729,
-0.0377747826,
-0.0228021219,
-0.0001962168,
0.0076319743,
-0.0585275777,
-0.0401639678,
0.0209740903,
0.0123622557,
0.0166469812,
0.0372163393,
0.0018895372,
0.0305639375,
0.0055851871,
-0.0829284564,
0.0385244377,
0.0242009237,
-0.0400734134,
-0.0456690155,
-0.02579212,
-0.014066156,
-0.0572384261,
0.0069743306,
0.0745486468,
-0.027028488,
0.046592705,
-0.0339010432,
-0.0178733487,
0.0300269388,
-0.0403092094,
0.0305923931,
0.0581593178,
-0.0128260758,
0.0384936593,
-0.0496645533,
0.041768834,
-0.0444572866,
-0.0110593382,
-0.0522561483,
0.0252441447,
-0.0054816986,
0.0102254516,
-0.0372067057,
-0.0087533258,
-0.0692451298,
-0.0428844728,
-0.0338003188,
0.0689501092,
0.0393090397,
-0.0060518002,
-0.0157396067,
0.0141272694,
-0.0155297387,
-0.0164726581,
0.0051062419,
0.0302757788,
-0.0456267633,
-0.0000292238,
0.017520329,
-0.0369552374,
-0.0158352945,
0.0239962805,
0.0163605791,
-0.0456436537,
-0.0087698828,
0.0723930672,
0.0242677275,
-0.0638688654,
-0.0210106261,
0.0065712151,
-0.0545998923,
-0.0719981566,
0.0275167841,
-0.0542379469,
0.0051680375,
-0.0069394247,
-0.0281905197,
-0.009180991,
0.0280736051,
0.006912448,
-0.0169730056,
0.0723219216,
0.0027388397,
-0.0072073811,
-0.0134306904,
0.0136734964,
0.0344139785,
0.0663371459,
0.0262454618,
0.0131726889,
0.010165791,
-0.0197327659,
-0.0321368724,
0.0257186089,
0.0160507672,
-0.0147955976,
-0.030000288,
0.0201102942,
0.0193516091,
0.0328360423,
0.0175789725,
0.0292479694,
-0.0092272013,
-0.0476669483,
0.0367443711,
0.0457141697,
0.0346710943,
0.0118869478,
0.0743684098,
-0.0431213044,
-0.0225750636,
0.0320462622,
-0.0527085848,
0.0397815034,
-0.0018909253,
0.0525870435,
0.0270418506,
0.0454830527,
-0.023579523,
-0.0305715278,
-0.0345280953,
-0.0029823137,
0.0309440382,
0.0381381847,
0.036070168,
-0.017578911,
-0.0448546968,
0.0296829697,
0.0199597813,
-0.025104953,
-0.0164885297,
-0.0404917412,
0.0361746512,
0.0698989257,
-0.0027181732,
-0.0415553749,
-0.0340854675,
0.0167225637,
0.0048860377,
0.0666837543,
0.0069373921,
0.0220918451,
-0.0106299883,
-0.0345286131,
-0.0132511687,
0.0028796692,
-0.0048991092,
0.0512947477,
0.04305657,
0.0044654883,
0.0333658829,
-0.0028791777,
0.0209196676,
0.0275643822,
-0.0465174802,
-0.0398441888,
-0.0382631831,
0.0191297308,
0.0515833721,
0.007005366,
0.0183678437,
0.0365116447,
0.0526586734,
-0.0433222093,
0.0168910362,
-0.0138676409,
-0.1105239466,
-0.0497435667,
0.0500368774,
0.0316659324,
0.0049000299,
-0.0134567432,
-0.0228800923,
-0.036241889,
0.0273546502,
-0.0642792508,
0.0674418956,
-0.0336947739,
0.0276186112,
0.0202540066,
0.0543731861,
-0.0562934093,
0.064256236,
-0.0303983092,
0.0068116831,
0.04937068,
0.028063627,
-0.0180016924,
-0.0452231281,
-0.0053812563,
0.040033225,
0.0016291223,
-0.0147382421,
-0.0118959295,
-0.0219992083,
-0.0419418439,
-0.0289185755,
-0.0718259066,
-0.0079075405,
-0.0208127592,
0.0321044736,
-0.0052719237,
0.0144792031,
-0.0033592891,
0.0077437395,
-0.0442869812,
0.0029476287,
0.0114568044,
0.0302226823,
0.0049895821,
0.0213676076,
-0.0733884275,
-0.0321976915,
-0.038329754,
0.0006380759,
0.0629113913,
-0.0407520458,
-0.040611852,
-0.0709444433,
0.0039473241,
-0.0212554391,
0.0258510262,
-0.0357375555,
-0.0178842284,
0.0038312529,
-0.0380439013,
0.0588099472,
0.0679294989,
0.0160400849,
0.0329167433,
0.0589439571,
-0.0207377989,
0.0013709605,
0.0233462881,
-0.0367544107,
-0.0089870123,
-0.040570803,
0.0041775387,
-0.0343376845,
0.0686917454,
-0.048232343,
-0.0113609424,
-0.0582100786,
0.000647583,
0.0203482825,
-0.0451375619,
0.006753779,
-0.0321750753,
-0.0520282127,
0.0465840697,
-0.0416794568,
-0.0357008614,
0.0586108156,
-0.0520030521,
-0.0006070266,
-0.0424996205,
-0.0033695386,
-0.0455529504,
0.0890101269,
0.0299157426,
0.0252527408,
-0.0079110553,
-0.0046833917,
-0.0286696311,
0.0123255942,
-0.0509229042,
-0.0152967758,
0.0492334366,
-0.0297332443,
-0.0026544502,
-0.0113906702,
-0.0042653424,
-0.046150282,
0.011902988,
-0.0159758441,
0.0239899904,
0.009478162,
-0.0068710749,
-0.0172927305,
0.053406924,
0.0101662641,
0.0456691161,
-0.0054536294,
0.0504205823,
-0.0362329558,
0.025180757,
0.0367105603,
0.0217866171,
0.0203298181,
-0.0220173672,
0.0194716267,
0.0478905179,
0.0255156066,
-0.0181845389,
-0.0246239323,
0.0063964264,
0.0128145758,
-0.0420971699,
-0.0200842489,
0.0478487425,
0.0393227413,
-0.1017553732,
-0.0296762437,
0.1104957834,
0.0220983457,
-0.0256731305,
0.0341942906,
-0.0185637716,
0.040226955,
0.0174021907,
0.0081903515,
-0.0594624504,
-0.0297346599,
0.028071437,
-0.034853138,
-0.0190512408,
-0.0002179963,
-0.0162802059,
-0.1270027459,
0.0058581871,
0.0291690603,
-0.0783408582,
-0.0545318983,
-0.0310033318,
0.0035802149,
0.0734529421,
0.0314838886,
0.0660119131,
-0.0110092377,
-0.0160784703,
0.0051833582,
0.0050563617,
-0.0400627851,
0.0266137794,
-0.0079424148,
0.0454802662,
0.04843181,
0.0069539538,
0.0324529186,
-0.0316842198,
0.0276838467,
-0.021214895,
0.060934756,
-0.0000880338,
0.0071165059,
-0.0172299966,
-0.0117392428,
0.0003929769,
0.0241683554,
-0.0155583043,
-0.0371600017,
-0.0034157475,
0.089043729,
-0.0138189187,
-0.015014559,
-0.0211860724,
0.0124505777,
-0.0167588592,
0.0599517673,
0.0797352567,
-0.0462178849,
-0.0409642942,
-0.0049039214,
-0.044736132,
-0.0409635566,
0.0158136878,
0.0001484313,
0.0023473748,
-0.0307165887,
-0.0326220766,
-0.0525443219,
0.0185478739,
0.0393067412,
0.0045229662,
0.0187123865,
0.0062174788,
-0.0273473896,
-0.0662052557,
0.0142701603,
-0.0253796596,
0.0312558301,
-0.0365418233,
0.0210818797,
-0.0183860809,
-0.0266320184,
0.0040074023,
0.04387445,
-0.0776947811,
-0.0871362388,
0.0040033543,
0.0396268256,
0.0195206944,
0.0549903288,
0.0166578107,
-0.022054607,
0.0332188047,
0.0006941975,
0.005411943,
0.0332332514,
0.0296982508,
0.0186669789,
0.0061089606,
0.0278583486,
-0.0125704845,
0.0220302697,
0.0319793634,
0.0661680847,
-0.0326810703,
-0.0304298997,
0.0054388768,
-0.0206093099,
0.014932924,
0.0068191849,
0.007705343,
0.013106578,
0.0680881366,
-0.055238612,
0.010034305,
-0.0081976755,
-0.0151748788,
0.0112692416,
0.0001914718,
0.0154311601,
-0.0419223197,
0.0025953618,
-0.0146229547,
-0.0117393248,
0.0852139518,
-0.0010534959,
0.0012726298,
0.035541907,
0.0120347235,
-0.0326521508,
0.0063550533,
-0.0244086217,
0.0477233827,
-0.0306238718,
0.0542335249,
0.0251409523,
0.0012092886,
-0.0453134738,
0.0257912464,
-0.0852765515,
0.0109576648,
-0.0032075674,
0.0284283329,
0.0612754077,
-0.0146843093,
0.0090464931,
-0.0044782991,
-0.0223221406,
-0.0141303334,
0.0280058384,
0.0001789369,
-0.0620735288,
0.0052195881,
0.0470134653,
-0.0191790834,
-0.0160601474,
0.0420586877,
0.0182534251,
0.0295994356,
0.0165666025,
0.0008545391,
-0.0197442118,
-0.0453933477,
-0.0061462773,
0.0222437847,
-0.0407512225,
0.0050236727,
-0.04416731,
-0.0937517807,
0.0046898653,
0.0086320108,
0.0056122988,
-0.0148392329,
0.0102355117,
0.0354716405,
0.0008944503,
-0.01110138,
0.006465679,
0.0254495032,
-0.0036786117,
-0.0255537257,
-0.014687594,
-5.954555898e-33,
-0.0224257559,
-0.0189696047,
0.040773429,
-0.0493808836,
-0.072060965,
0.0070581459,
-0.06251017,
0.0028463353,
0.0182879381,
-0.0218397956,
-0.0163883846,
-0.0230404045,
-0.0011412494,
0.0323170461,
-0.0062788087,
-0.0125132408,
0.017744977,
0.0358058587,
0.0239050463,
0.0143961608,
0.014519087,
0.0265592989,
0.0330171995,
0.0018503814,
0.0086049745,
0.02058479,
0.0292134192,
0.0150596574,
-0.0456383899,
0.0736139938,
-0.0227477346,
0.0104923481,
0.0393340811,
-0.0455797166,
-0.0272449851,
0.0418023057,
-0.0704320297,
0.017824959,
0.0375659019,
-0.0056636804,
-0.0176915806,
-0.0639691353,
0.0449905992,
0.0086168451,
0.0238987301,
-0.0161513407,
0.0022904889,
-0.0601947792,
-0.0502443835,
0.0032940733,
-0.0762517527,
0.0252631884,
0.0207596347,
0.0122789098,
0.0227545742,
0.0097543374,
-0.0127355112,
0.0199799128,
-0.0109977368,
-0.0117493542,
0.0363160335,
0.0761826336,
0.045846954,
0.0636088625,
0.0298810899,
0.0085105477,
0.0101121143,
-0.0314205289,
-0.0515939146,
-0.0216065794,
0.0136522632,
0.0154752024,
0.0373978987,
0.0025901408,
0.0441550128,
-0.0015609462,
-0.0571841225,
-0.0471984595,
0.0060048285,
0.0540247522,
0.046913866,
0.0328427181,
-0.0190998465,
-0.0012253998,
-0.0202679038,
0.0068718125,
-0.0406578332,
-0.0160782896,
0.0488943942,
0.0178935155,
-0.0022148439,
0.0102072572,
0.0590173826,
-0.0163421743,
-0.0012650219,
0.0327590182,
0.0240540896,
0.0035948537,
0.0026722862,
0.0385237001,
-0.0352954604,
-0.0336515494,
0.0357951894,
-0.012928186,
-0.0211343579,
-0.0336477831,
0.0358859003,
0.0143590476,
-0.0517215766,
-0.0080660935,
-0.0618472435,
0.0011389862,
0.0022931783,
0.0484232083,
0.0654788464,
0.0024257281,
0.0012316343,
0.0514221378,
0.0127569186,
-0.0305925589,
-0.0145608392,
-0.0118917115,
-0.0009423752,
0.0323851891,
-0.0427426472,
0.0118893031,
-0.0309166946,
0.0353169851,
0.0312335398,
-0.0499921367,
-0.0155300256,
0.0422073565,
2.676e-7,
0.038767267,
0.0405566096,
0.0643627122,
0.017631853,
0.0352463052,
-0.0091024088,
0.0162431486,
0.0203516986,
0.0149315298,
-0.016539529,
0.0099906046,
0.0360764191,
0.0259467121,
-0.0164155625,
-0.0730065778,
-0.0152495131,
-0.0470063388,
0.0242991447,
-0.0612499155,
0.0149997864,
0.0586097054,
0.1354162395,
-0.0105872396,
-0.006252429,
0.0188890547,
-0.0369971767,
0.0288087446,
-0.0219673477,
0.0181028079,
-0.0153384823,
-0.0298189484,
-0.0135055184,
-0.002705219,
-0.0053665205,
0.0139089832,
0.0236183852,
-0.040815305,
0.0571885556,
-0.0159986634,
-0.0315742157,
0.0113944896,
-0.0044784294,
-0.0327349529,
-0.0627859831,
0.0605473854,
-0.0680021569,
-0.0308202412,
-0.0234103445,
-0.0139273303,
-0.0153577011,
-0.0392990187,
0.0037152783,
0.0089096809,
0.0175198112,
0.015386167,
-0.0196452029,
0.0096431328,
-0.1252270043,
-0.0340384245,
0.0097980676,
-0.0587026477,
0.000185912,
-0.0810723677,
-0.0043829591,
0.0534051619,
-0.0432320088,
-0.0307243522,
2.889388532e-34,
0.0271295886,
-0.0327518806,
0.0100139491,
0.0151557429,
0.0193620659,
-0.0344709009,
-0.023777714,
0.0240710061,
0.0519237518,
-0.1186248884,
-0.0544312596
] |
http://arxiv.org/abs/2402.02243v1 | http://arxiv.org/pdf/2402.02243v1 | Language Writ Large: LLMs, ChatGPT, Grounding, Meaning and Understanding | 2024-02-03T19:19:34 | 2024-02-03T19:19:34 | Stevan Harnad | cs.CL | cs.CL, q-bio.NC | Apart from what (little) OpenAI may be concealing from us, we all know
(roughly) how ChatGPT works (its huge text database, its statistics, its vector
representations, and their huge number of parameters, its next-word training,
and so on). But none of us can say (hand on heart) that we are not surprised by
what ChatGPT has proved to be able to do with these resources. This has even
driven some of us to conclude that ChatGPT actually understands. It is not true
that it understands. But it is also not true that we understand how it can do
what it can do. I will suggest some hunches about benign biases: convergent
constraints that emerge at LLM scale that may be helping ChatGPT do so much
better than we would have expected. These biases are inherent in the nature of
language itself, at LLM scale, and they are closely linked to what it is that
ChatGPT lacks, which is direct sensorimotor grounding to connect its words to
their referents and its propositions to their meanings. These convergent biases
are related to (1) the parasitism of indirect verbal grounding on direct
sensorimotor grounding, (2) the circularity of verbal definition, (3) the
mirroring of language production and comprehension, (4) iconicity in
propositions at LLM scale, (5) computational counterparts of human categorical
perception in category learning by neural nets, and perhaps also (6) a
conjecture by Chomsky about the laws of thought. The exposition will be in the
form of a dialogue with ChatGPT-4. | Language Writ Large: LLMs, ChatGPT, Grounding, Meaning and Understanding: Apart from what (little) OpenAI may be concealing from us, we all know
(roughly) how ChatGPT works (its huge text database, its statistics, its vector
representations, and their huge number of parameters, its next-word training,
and so on). But none of us can say (hand on heart) that we are not surprised by
what ChatGPT has proved to be able to do with these resources. This has even
driven some of us to conclude that ChatGPT actually understands. It is not true
that it understands. But it is also not true that we understand how it can do
what it can do. I will suggest some hunches about benign biases: convergent
constraints that emerge at LLM scale that may be helping ChatGPT do so much
better than we would have expected. These biases are inherent in the nature of
language itself, at LLM scale, and they are closely linked to what it is that
ChatGPT lacks, which is direct sensorimotor grounding to connect its words to
their referents and its propositions to their meanings. These convergent biases
are related to (1) the parasitism of indirect verbal grounding on direct
sensorimotor grounding, (2) the circularity of verbal definition, (3) the
mirroring of language production and comprehension, (4) iconicity in
propositions at LLM scale, (5) computational counterparts of human categorical
perception in category learning by neural nets, and perhaps also (6) a
conjecture by Chomsky about the laws of thought. The exposition will be in the
form of a dialogue with ChatGPT-4. | [
0.0090747904,
0.0147674317,
-0.0310745146,
0.0466051288,
-0.0520087294,
0.0309673082,
-0.0482681021,
0.0295764282,
0.0135045629,
-0.047667142,
-0.0714194551,
-0.0171867087,
0.0152182337,
0.0041268156,
0.0211263299,
-0.0165806413,
0.0460330471,
-0.0038535001,
-0.0260946117,
-0.0150810182,
-0.0299764667,
-0.0287339985,
-0.0179367159,
0.0388568491,
-0.019595189,
-0.0229099989,
-0.0285282712,
-0.0328249671,
-0.0441875681,
-0.003553835,
-0.0491817482,
-0.0146986246,
0.0192595646,
0.065161705,
0.0000021036,
-0.0505162887,
0.0272577573,
-0.0192286838,
-0.0280340575,
0.008936421,
0.01219997,
-0.0109584602,
0.0011397605,
-0.0055523203,
-0.0310918819,
0.0017328358,
0.0433081202,
0.0640500784,
0.0058536348,
0.0500402823,
-0.0021609953,
-0.0456740074,
0.0134673016,
0.0021611019,
0.0645975694,
-0.0290444456,
0.0264996216,
0.0220063813,
-0.0003067663,
-0.0155415749,
-0.0834362283,
-0.0009632877,
0.032476794,
0.0247588959,
-0.0118838651,
0.0690689385,
-0.0342499726,
-0.0312464517,
0.009014355,
0.0354929082,
0.048668541,
-0.045825962,
-0.0316962376,
0.0102710146,
-0.0032127586,
-0.0056026843,
-0.0478823483,
0.0514953546,
-0.0105661983,
0.059955921,
0.0195170902,
-0.001897428,
0.00592785,
0.0278989524,
-0.0188928209,
0.0624499731,
-0.0011678549,
-0.0334412269,
0.0423641168,
0.0542716868,
-0.0263283793,
-0.0554932356,
-0.0278157983,
0.0509341434,
-0.0013648223,
0.0255894307,
0.0065879696,
0.0596252456,
0.0113632772,
-0.0908984244,
0.0156251416,
0.0818796381,
-0.0737534985,
0.0219192691,
-0.0395756476,
0.0489261113,
-0.0455213375,
0.0722602233,
-0.0352926105,
-0.003802635,
-0.0420816243,
-0.0368925445,
-0.0289095286,
0.0526082069,
-0.0173036437,
-0.0252901949,
-0.0195776429,
0.0408702157,
0.0025749351,
-0.030384073,
-0.0277331378,
0.0030328587,
-0.044431705,
0.0213076193,
-0.0428349823,
0.0323080309,
-0.0129879862,
0.0269237868,
-0.0082988525,
-0.0763453916,
-0.0129435761,
0.0194496159,
-0.0318595804,
-0.0105059044,
-0.0132890614,
0.0955240875,
0.0261957925,
-0.0185298678,
-0.0667093769,
-0.022508733,
-0.0440446548,
-0.051693473,
0.0338251702,
-0.01051025,
-0.0374411941,
0.0350677855,
-0.0343586802,
-0.0319188014,
0.0048319409,
-0.0132477814,
-0.0634321868,
0.055055242,
0.0054478394,
-0.0124131711,
0.0052364818,
0.0379296876,
-0.0135057596,
0.061456617,
0.0052028536,
0.0191479474,
0.0205033515,
0.0075896811,
-0.000643192,
0.0010781434,
0.0187295899,
-0.0156962387,
-0.0013834679,
-0.0322107673,
0.0078248866,
0.0522321202,
0.0234812461,
0.0797805339,
-0.0169462785,
-0.0076808915,
0.058917243,
0.0638399646,
0.0028979953,
0.0431481376,
0.0152070997,
-0.0245082006,
0.0181910899,
0.0918731466,
-0.0498583987,
0.0159926303,
-0.0152615765,
0.0098047117,
-0.0353343114,
0.0019721072,
-0.039996095,
-0.0432862453,
-0.027477663,
0.0235769954,
0.0067746388,
-0.0542883798,
0.0176691692,
0.0076022441,
-0.0005846224,
0.03965896,
0.0521608628,
-0.0300187785,
0.0257222243,
-0.0876365975,
0.0077086999,
-0.0391281396,
0.0132845007,
0.0134030282,
-0.0210930388,
0.0232155155,
0.0385405906,
0.0772279948,
0.036261078,
0.0216826014,
-0.0263162479,
-0.0215451196,
0.013248344,
-0.0348170921,
0.0016295449,
0.0132652558,
-0.0580403842,
0.0076160864,
-0.0320416354,
-0.0202325527,
0.0134251164,
0.049755875,
-0.0049044606,
-0.0433429554,
-0.0376942232,
0.0193269961,
0.0569650456,
-0.0092025315,
-0.000961598,
0.0312594362,
0.032833416,
-0.0137452027,
-0.0257110465,
0.0737515166,
-0.0622248165,
-0.0031242047,
0.0280644931,
-0.034341339,
-0.0048217252,
-0.0351498276,
-0.0029295716,
-0.0306518134,
0.048043251,
-0.0209285393,
0.0922063664,
-0.0365166999,
-0.0117979869,
0.0039887661,
0.0101928785,
-0.0388283543,
0.0527056269,
0.0185188986,
-0.0339927003,
-0.0048093209,
0.0205600597,
0.0520771816,
-0.018349817,
-0.0122854086,
0.0256846827,
0.0062021753,
0.0115845958,
-0.028051639,
0.0131675806,
-0.0719066486,
-0.0277644657,
-0.0886625424,
-0.015968686,
0.0197090916,
0.0151356878,
0.0281273089,
-0.0232720859,
-0.0185632855,
0.0152407931,
-0.0096968887,
0.0355914943,
0.0268471539,
0.0353257582,
0.0533258729,
0.0045985542,
-0.0397456028,
0.0044136094,
0.0244516488,
-0.0394456834,
0.0294660814,
-0.0070196372,
-0.0254667737,
-0.0539414361,
0.0455230214,
-0.0490536503,
0.0068689478,
-0.014279468,
-0.0324880108,
0.0484209135,
0.0266160313,
-0.0012494525,
0.0522417165,
-0.0013134782,
0.0975750089,
0.0326629207,
-0.0251742173,
-0.0185323395,
-0.0293468498,
-0.0230815373,
-0.0767539665,
-0.0051172525,
0.0212099589,
-0.051597964,
0.0909881741,
0.0336638652,
-0.0170800518,
-0.0450167656,
0.0168084092,
-0.0221548658,
-0.0193383526,
-0.0351971574,
-0.019257186,
-0.0690460578,
0.0459320024,
-0.0320912525,
-0.0308850277,
0.0410501324,
-0.0275031887,
-0.0179035608,
-0.0356767252,
0.0219457336,
-0.052812174,
0.0906767696,
-0.0316512957,
0.0179100558,
0.0318018086,
-0.035683088,
-0.0741856918,
-0.0312111415,
-0.0116467401,
0.0228869971,
0.0352629833,
-0.0507238396,
0.0109385569,
-0.0119189546,
-0.0198627785,
-0.0571623333,
0.0202566907,
-0.0163273886,
-0.0061309482,
-0.0114014996,
0.0485931784,
0.0140577322,
0.0063379207,
0.0092949206,
-0.0457397811,
-0.0224779993,
-0.0232633166,
-0.0329918154,
-0.0030868787,
-0.0114893774,
0.0385432541,
0.0421264134,
-0.0024646274,
-0.0005059301,
0.0758002922,
-0.0226530079,
-0.0145707633,
-0.0166832116,
0.0446693525,
0.0042352164,
-0.039968621,
-0.0585739501,
0.0076504289,
0.0484749228,
-0.0321179107,
-0.0390111506,
0.0857656375,
-0.0057057636,
-0.0352462307,
0.0549818426,
-0.0453256108,
0.0038979833,
-0.0011484287,
0.0103087351,
-0.0901303217,
-0.0004094836,
0.0094016111,
-0.0180795398,
0.0060349084,
-0.0331086144,
-0.022852391,
-0.0760508031,
0.0368654765,
0.0108385496,
-0.0571256243,
-0.0075419485,
0.0108326552,
0.0122963283,
0.0233566687,
0.0730222911,
0.0349700674,
-0.0203791074,
0.012323048,
0.047178898,
-0.0067511364,
0.0045639467,
0.0121966908,
-0.035252206,
0.0428283289,
0.0021790271,
0.0443165787,
-0.0142485676,
-0.0469934493,
0.0934956148,
-0.0021896323,
0.0391443633,
-0.0270436648,
-0.0147691704,
-0.0048566908,
0.0019562896,
-0.0201921146,
0.0390835516,
-0.0202650931,
-0.0080273598,
-0.0157575328,
0.0673444644,
-0.0523641966,
-0.0295749269,
0.0105367508,
0.0258080456,
-0.0479372516,
0.0711432546,
0.0493957326,
-0.0044057006,
-0.0311360322,
-0.019067131,
0.0100866714,
0.0872638598,
-0.0808939263,
0.030564785,
0.0040499512,
-0.0339662172,
-0.0432410687,
-0.0434877798,
-0.0043462696,
0.0316228531,
0.0239364561,
0.0011070435,
-0.022963319,
-0.0192304924,
0.0052882568,
0.035410516,
-0.0248174705,
-0.0349939354,
-0.0318365321,
-0.0045492891,
0.0293831192,
-0.0705163106,
-0.0171624664,
-0.0469176359,
-0.0194096621,
-0.0517355651,
0.0385625549,
-0.0042959135,
0.0409593657,
0.0074604801,
0.0143982507,
0.0062795714,
0.0027619509,
0.0415959135,
-0.0064680013,
0.0854108706,
0.0215130635,
0.075583525,
0.0356658697,
0.0337117091,
-0.0207194407,
0.0102170734,
0.0171046872,
0.0151871657,
0.0262773596,
-0.0040758145,
0.0411115289,
0.0292855799,
0.0754836798,
-0.0032388407,
0.016968986,
-0.0305288769,
-0.0058171456,
0.0042056213,
0.01473419,
-0.0353393815,
-0.0294758324,
0.035120178,
0.0446952879,
-0.0408737101,
-0.0530977845,
-0.0366588496,
0.0527913943,
0.0274925828,
0.000153682,
0.001508908,
0.0026907939,
-0.0136012314,
0.0319475122,
-0.0314546786,
0.0258514676,
-0.0332581028,
0.0235740412,
-0.0091823013,
0.0540115163,
0.0488394834,
0.0207445845,
-0.0048220567,
0.0167513452,
-0.0589635037,
-0.0485680662,
0.0323290117,
0.0706683993,
0.0656544417,
-0.0072144433,
-0.0107668247,
-0.0674662963,
0.0071308454,
0.0250877105,
0.0182201769,
-0.0079085194,
-0.0254746471,
0.0396273844,
-0.0132495994,
-0.0207278337,
-0.0110786268,
0.0546017475,
-0.0276525244,
0.0519652963,
0.03037907,
0.0103214784,
-0.022256393,
-0.0380475484,
0.0279282611,
0.0049797618,
0.0114392312,
0.0124268737,
-0.0936594382,
-0.0448146313,
-0.0305287968,
0.0135892006,
0.0016993567,
-0.0477727428,
0.0236728024,
0.0526044443,
0.0183343161,
-0.0200711042,
0.02298229,
-0.0035230177,
-0.0314135626,
0.01242337,
0.0197321437,
-5.989865544e-33,
-0.0539752021,
-0.0216040537,
0.0444631465,
0.0693250895,
-0.0605845749,
-0.0043588984,
-0.0286267325,
0.0271069892,
-0.0329463482,
0.015375739,
-0.0243313834,
-0.0170194283,
0.0155836381,
-0.0002008098,
0.0137972394,
-0.0556326136,
-0.0095938127,
-0.0175480489,
-0.0010449706,
-0.02364989,
0.0319321752,
0.0072924751,
0.064108856,
-0.0635399967,
0.0409045741,
0.0125168832,
0.0016229376,
0.0183346067,
-0.0337311924,
0.0333613418,
-0.0440660268,
0.0525317229,
0.0332628265,
0.0064536915,
0.0040135346,
-0.0014809844,
-0.023928687,
-0.0657384321,
0.0764207318,
0.0427690968,
-0.0085522244,
-0.0200315882,
0.0384158157,
-0.017405672,
-0.0632358193,
0.0195423719,
0.0029156154,
-0.0373919643,
-0.0257402863,
0.0475448743,
-0.0707973242,
-0.0307486001,
0.0340963565,
-0.0201098714,
-0.0119291507,
-0.0789544955,
-0.0350978561,
-0.0218166169,
-0.0190291759,
-0.0042594485,
-0.0331245475,
0.0094286827,
0.0566964559,
0.0179096609,
0.0187415648,
0.073751092,
0.0613007843,
0.0043867873,
-0.0567443334,
-0.0358709395,
0.0199097749,
0.0308955871,
0.0153290164,
0.0075108474,
0.0263268128,
0.0179511402,
-0.056266468,
-0.0317796692,
0.0553799048,
0.0573299117,
0.0173286442,
0.0226971433,
-0.0262455028,
-0.0338253528,
0.0095648887,
-0.0288657732,
-0.0273757447,
-0.0418276042,
0.0216751229,
-0.0357957892,
0.0139734708,
-0.0220442042,
0.0214412939,
-0.0529582724,
0.0229941364,
0.0018761805,
-0.0244241413,
0.0038381258,
0.0035860918,
-0.0092920372,
-0.0697613135,
-0.0466828421,
-0.0325836726,
0.0249424148,
0.0069385669,
0.0363991298,
-0.0052856784,
0.0061035901,
0.0094561754,
-0.0099671409,
-0.0454007238,
-0.0444234014,
0.0384785309,
-0.0279054102,
0.0491415001,
-0.0112885144,
-0.0116520738,
0.0195416696,
0.0285905134,
0.0180239622,
-0.018925583,
-0.0041760439,
-0.0383959711,
0.0059540854,
-0.0465567186,
0.000230695,
0.0036402051,
0.017249899,
0.0369245335,
-0.0842008963,
-0.0214960352,
-0.001755517,
2.935e-7,
0.0603528991,
0.0391915664,
0.0352981277,
0.0223619789,
0.0363002308,
0.0114865685,
-0.0280131418,
0.0037600745,
0.0012535837,
-0.039807193,
-0.0051997844,
0.060999874,
0.0174684227,
0.0280783679,
-0.0342929214,
0.0085007558,
-0.0492560603,
0.0002622576,
-0.0226043798,
0.0349941514,
0.068320252,
0.0716927052,
-0.004293493,
-0.0224779025,
-0.0076196818,
-0.0431721658,
0.0662446469,
-0.0249722116,
0.0387439691,
-0.0167015716,
0.0362756215,
0.001264442,
0.0017873805,
0.0211502183,
0.0568959005,
0.0005936999,
-0.0144869573,
0.0104836868,
0.0044799433,
-0.0191206709,
-0.015393069,
-0.0239880886,
-0.002788414,
-0.0531212874,
0.0674128085,
-0.0847018585,
-0.0534629487,
0.0080360807,
0.0402099416,
-0.0406009369,
0.0110618817,
0.0220764056,
-0.0080780527,
-0.0189328846,
0.0187690258,
-0.0273400769,
-0.0348181576,
-0.080235675,
-0.0036459218,
-0.0132783642,
-0.0323587097,
-0.0671443567,
0.0097049251,
0.0234461986,
0.0709580779,
-0.0508295149,
-0.0639399663,
2.792412544e-34,
0.0121462224,
-0.029447509,
0.0530035384,
0.0163913015,
0.0164838862,
0.0009285879,
0.0835297257,
0.0177719574,
0.017723985,
-0.0397866927,
-0.025700042
] |
http://arxiv.org/abs/2402.02212v1 | http://arxiv.org/pdf/2402.02212v1 | A Data Generation Perspective to the Mechanism of In-Context Learning | 2024-02-03T17:13:03 | 2024-02-03T17:13:03 | Haitao Mao, Guangliang Liu, Yao Ma, Rongrong Wang, Jiliang Tang | cs.CL | cs.CL | In-Context Learning (ICL) empowers Large Language Models (LLMs) with the
capacity to learn in context, achieving downstream generalization without
gradient updates but with a few in-context examples. Despite the encouraging
empirical success, the underlying mechanism of ICL remains unclear, and
existing research offers various viewpoints of understanding. These studies
propose intuition-driven and ad-hoc technical solutions for interpreting ICL,
illustrating an ambiguous road map. In this paper, we leverage a data
generation perspective to reinterpret recent efforts and demonstrate the
potential broader usage of popular technical solutions, approaching a
systematic angle. For a conceptual definition, we rigorously adopt the terms of
skill learning and skill recognition. The difference between them is skill
learning can learn new data generation functions from in-context data. We also
provide a comprehensive study on the merits and weaknesses of different
solutions, and highlight the uniformity among them given the perspective of
data generation, establishing a technical foundation for future research to
incorporate the strengths of different lines of research. | A Data Generation Perspective to the Mechanism of In-Context Learning: In-Context Learning (ICL) empowers Large Language Models (LLMs) with the
capacity to learn in context, achieving downstream generalization without
gradient updates but with a few in-context examples. Despite the encouraging
empirical success, the underlying mechanism of ICL remains unclear, and
existing research offers various viewpoints of understanding. These studies
propose intuition-driven and ad-hoc technical solutions for interpreting ICL,
illustrating an ambiguous road map. In this paper, we leverage a data
generation perspective to reinterpret recent efforts and demonstrate the
potential broader usage of popular technical solutions, approaching a
systematic angle. For a conceptual definition, we rigorously adopt the terms of
skill learning and skill recognition. The difference between them is skill
learning can learn new data generation functions from in-context data. We also
provide a comprehensive study on the merits and weaknesses of different
solutions, and highlight the uniformity among them given the perspective of
data generation, establishing a technical foundation for future research to
incorporate the strengths of different lines of research. | [
0.0026206018,
0.0406487733,
-0.0215501077,
0.043426156,
-0.0174715016,
0.04024297,
0.0040426725,
-0.0319904834,
-0.0268981811,
-0.041248139,
-0.0718809664,
-0.0439188145,
0.0246361978,
-0.0045077964,
0.0512654968,
-0.0394194871,
0.0076228022,
-0.0359587446,
-0.018245507,
0.0015869986,
-0.0267844032,
-0.0033403186,
0.0107718799,
0.0421808362,
-0.0019397095,
-0.022077756,
0.0115360236,
0.0058931834,
-0.0164381061,
-0.017761156,
0.0263902172,
0.0186700895,
0.0480836518,
0.0475728028,
0.0000021456,
-0.0369067378,
-0.0031358094,
-0.0249012318,
-0.0387600139,
-0.0344807692,
0.0345558189,
-0.0488456935,
-0.0086446879,
0.0237272382,
-0.057618238,
-0.0393619537,
0.0256121866,
-0.0336047672,
0.0070502926,
0.0705294386,
-0.0075146696,
-0.0383784994,
0.0028487844,
-0.0370273702,
0.0130107328,
-0.044902429,
0.0241632275,
0.0009201598,
-0.0120783998,
-0.016544532,
-0.0390283614,
0.0485147201,
0.0415920392,
-0.0027768556,
0.006733303,
0.0501097329,
0.0122885359,
-0.005258746,
-0.0114190718,
0.0418043546,
-0.0252669696,
-0.0197678246,
-0.0248269737,
-0.0136440275,
0.0146706328,
0.0003024414,
-0.0498825833,
0.0351200476,
-0.0043925825,
0.0421122499,
0.0200686734,
-0.0155915171,
0.0623508729,
0.0026309884,
-0.0302413832,
0.0379862003,
-0.0155173093,
-0.0281421114,
0.0049122754,
0.0126319146,
-0.0256294906,
-0.0540419333,
-0.0305197053,
0.0370551124,
-0.0436538532,
0.0515361317,
-0.0490897186,
0.0033785335,
-0.0101930387,
-0.0245493427,
0.0224448945,
0.0225873683,
-0.021993747,
0.0162068959,
-0.0231560655,
0.019324597,
-0.0060007931,
0.0135321477,
-0.0446893908,
-0.0202477947,
-0.0477512106,
-0.0148808453,
-0.0274720974,
-0.0183771867,
-0.0276362076,
-0.0010273337,
-0.0104665281,
0.0751641244,
-0.0335303061,
0.017744679,
-0.0187934153,
-0.028340606,
-0.0520901568,
0.0223673508,
-0.0068049286,
-0.0113757784,
0.013313774,
0.0343552828,
-0.0050684372,
-0.0323943645,
0.0056590457,
0.0073775821,
0.0222705211,
-0.0066558383,
0.0224921219,
0.0486250073,
-0.023456065,
-0.0345142186,
-0.0403749086,
-0.0177869555,
-0.0164366551,
-0.0792953745,
-0.0157648493,
-0.0678902864,
-0.0058557405,
0.0301727448,
0.0112730274,
-0.0466361344,
-0.0204152558,
0.0168558098,
-0.0126648452,
0.0580359809,
-0.0189006161,
0.022549618,
0.0220081527,
0.0015554857,
0.021603331,
0.0844907686,
-0.0018416931,
-0.0306915659,
0.0189025998,
0.0048481221,
0.005777671,
0.0448876321,
0.0497050919,
0.0147669995,
-0.0342030898,
0.0037278163,
0.0315046348,
0.0350866467,
0.0323678404,
-0.009747074,
-0.0384439714,
0.0533433445,
0.0697929859,
0.0618057251,
0.0562961772,
0.0128951361,
0.057526432,
0.0020615852,
0.0024253032,
0.0237085968,
-0.0599767789,
-0.018556213,
0.0069638113,
0.0112570394,
-0.0078441175,
0.0001423235,
-0.034438584,
-0.0404132418,
-0.0019457862,
-0.0041987174,
-0.005498589,
0.0374607667,
0.0435483381,
0.0568554029,
-0.003782847,
-0.0193936862,
0.0254214257,
-0.0173736047,
0.0457600318,
-0.0537246689,
-0.0018620432,
0.0664168298,
0.0337392986,
-0.0257860087,
-0.0066556251,
-0.0162733756,
0.0492648222,
0.016551761,
-0.0001816497,
-0.0582359359,
0.0357042812,
-0.0264079459,
-0.0289985016,
-0.0219817068,
0.0191073027,
-0.0045691617,
0.0084448401,
-0.0253008604,
0.0030778404,
-0.0266594365,
-0.0262994859,
0.0429152288,
-0.0605132356,
-0.0250311326,
0.0210269447,
0.0178745948,
0.0161154661,
-0.0007547634,
-0.0053972625,
-0.0037216514,
0.0209139269,
-0.0360556915,
0.0194143578,
0.0569517799,
-0.10147468,
-0.0531110577,
0.0116865002,
0.0011229621,
-0.0066229035,
-0.0269619562,
-0.0278830212,
-0.0409296528,
0.0055842716,
-0.017013859,
0.0879560485,
-0.0414611287,
0.0184620917,
-0.0211734623,
0.0146647375,
-0.0184122212,
0.0973087177,
-0.0002728468,
0.00082279,
0.0482713766,
0.0883672759,
-0.0149797769,
-0.026462758,
0.0023485676,
-0.005697289,
0.0105655156,
0.0272039268,
0.0026585485,
0.0317685045,
-0.0474165976,
0.0091298912,
-0.0970629454,
-0.0057498091,
-0.0545872115,
-0.0161541328,
0.0545526072,
-0.0099358223,
-0.0032801726,
-0.0328173451,
-0.0134067899,
-0.0107550472,
0.0147643723,
0.0184352119,
0.0437074788,
0.0341076031,
-0.0188390762,
-0.0132528385,
-0.0000859663,
0.0057020467,
0.0565845929,
0.0136209615,
0.0077719283,
-0.0837901905,
0.0322854929,
0.0257032868,
0.0249541774,
-0.0283451807,
0.0162941311,
0.0243781712,
-0.0039354451,
0.0259549208,
0.0453235954,
-0.0194662269,
0.0663090721,
0.0427442268,
-0.00041069,
-0.0500770956,
-0.0058330507,
-0.0369269811,
-0.015844509,
-0.0728111267,
-0.0105791353,
0.0068557663,
0.0356705971,
0.020197669,
-0.0045898352,
-0.0356836244,
0.0128960451,
-0.0124108866,
-0.0251451768,
-0.067493625,
-0.0086983992,
-0.0356436409,
0.0277249832,
-0.0409387201,
-0.0989081189,
0.0309152305,
-0.0114132771,
-0.0036652246,
-0.0761732087,
-0.0305831973,
-0.0183486454,
0.058832299,
-0.0079470677,
0.0115525424,
-0.0176357962,
-0.0313616917,
-0.0614027642,
0.010179813,
0.0222935248,
0.0021527999,
0.0587000474,
-0.0761222467,
-0.0016937308,
0.0252405088,
-0.0152738281,
-0.0669659972,
-0.0095360242,
0.0020880962,
0.0320609026,
0.0125701055,
-0.0035469183,
0.0293653142,
-0.0197776593,
0.0269477442,
0.0625040159,
0.0216531921,
-0.0387018956,
-0.026468588,
0.027574176,
0.0293738022,
0.0026280677,
0.0101803942,
-0.042848099,
0.0219722874,
0.0330102816,
0.0771680251,
-0.0074898819,
-0.0281479172,
0.0192030836,
0.0395525582,
-0.0117222043,
-0.0151307983,
0.0399063304,
0.0227493327,
-0.0581941865,
-0.0407582149,
0.096253559,
-0.0049262363,
-0.0967567712,
0.0180217456,
0.0220909566,
-0.0151965981,
0.0464573726,
0.0076329638,
-0.066813454,
-0.0200947057,
0.0122419288,
0.0223291125,
-0.0097895619,
-0.0088286102,
-0.030309204,
-0.0912281498,
0.0421350151,
0.0029239045,
-0.0854140297,
-0.0427971818,
-0.0308462903,
0.0160869565,
0.0135653876,
0.0524496101,
0.0184099302,
-0.034661144,
-0.0082095237,
0.0465073735,
0.0070514157,
-0.1239565462,
0.0195319913,
-0.0449030101,
0.0488556288,
0.0748876408,
0.0242954753,
0.0146065084,
0.0065934202,
0.0468351729,
-0.0354044139,
0.0196556672,
-0.030444406,
0.0313151516,
-0.018976124,
-0.000225419,
0.0245539024,
0.0807177499,
-0.0361784101,
-0.0232842658,
0.0135363135,
0.0576107018,
-0.0174427852,
0.0087712873,
0.0159437954,
-0.0493519977,
0.0671415627,
0.0709219053,
0.0242535323,
-0.0772878975,
-0.0106070777,
0.0044732909,
0.0270152017,
0.0103357024,
0.0010762395,
0.0134926196,
-0.0176741611,
-0.0062280223,
-0.0324742086,
-0.0249987002,
0.0329920314,
0.049924098,
0.0225375425,
0.0215984099,
-0.0197305474,
-0.0130706094,
-0.0493641645,
0.0160826463,
-0.0192787759,
0.024336243,
-0.0281725973,
-0.0579591468,
0.0260884631,
-0.0647756904,
-0.0184636358,
0.0094555933,
-0.0826068297,
-0.0544150099,
0.0079203583,
0.0241069924,
0.0190979335,
0.1045649573,
0.0113259591,
-0.0453284234,
0.0896937177,
0.0321775414,
-0.0219808947,
0.0271416996,
0.035912279,
0.0137962829,
0.0063114683,
0.0390794165,
-0.0536262244,
-0.0070062703,
0.0604845211,
0.0173507873,
-0.0291683879,
0.0573428534,
0.0414209701,
-0.0658446774,
0.0163020883,
-0.0428627841,
-0.03134422,
-0.0617031232,
0.0035897389,
0.0306754913,
-0.0211284086,
-0.0407785401,
-0.025584463,
0.0257373638,
0.0187693052,
0.0080978759,
0.0071053952,
0.0173886772,
0.0265256725,
-0.0461360365,
0.0011532194,
0.0164786801,
0.04102429,
0.0405993424,
0.0228632241,
-0.0054373336,
0.0181752127,
-0.0267690774,
0.0242978241,
-0.0093034934,
0.0390823185,
0.0227970909,
-0.0220795125,
-0.0187945906,
0.0154548641,
-0.0969759971,
-0.012352922,
-0.0583262965,
0.0440420695,
0.0963540077,
-0.0352645107,
0.0331498832,
-0.0398460701,
-0.0033924151,
-0.0132431136,
0.0162135586,
-0.0124864159,
-0.0570305847,
0.0008193994,
0.0014510159,
-0.0226552375,
0.0089482786,
0.091972813,
-0.0002160485,
0.0612898245,
-0.0296113882,
0.0036795298,
-0.0280738976,
-0.0159606542,
0.0637406558,
0.0481262244,
-0.0166840255,
0.0216311924,
-0.0365401991,
-0.071752958,
-0.0446374863,
-0.0089695929,
-0.0236800145,
-0.0148985935,
0.0290581118,
0.0855162442,
-0.0118568279,
0.0023222698,
-0.0395112187,
0.0086788647,
-0.0054953755,
-0.0131619265,
-0.0413465798,
-6.306004832e-33,
-0.020025881,
-0.0452260748,
0.0321512111,
0.0181266461,
0.0177677982,
0.01961991,
-0.0249970518,
-0.0012011469,
-0.0079104593,
-0.0063234251,
-0.0108619463,
0.0198922846,
0.008226471,
0.0395120606,
0.0083248829,
-0.0113114165,
0.0503411368,
-0.017149752,
-0.0061963503,
0.0489472561,
0.0528878495,
0.0166433603,
0.0832112879,
-0.0619399436,
0.0454220846,
-0.0501648225,
-0.0055465749,
0.0401906893,
0.001806351,
0.0336546898,
-0.0152325472,
0.0419688784,
0.0408568718,
-0.0469699837,
-0.0451364666,
0.007090684,
-0.0935808942,
-0.0050354828,
0.0207962971,
0.0283835512,
-0.0404426269,
-0.023187425,
0.0642750487,
-0.0198418349,
-0.0279436167,
0.0080704149,
0.0464230403,
-0.028337203,
-0.0283330474,
-0.0352096669,
-0.0623402074,
0.0020475455,
-0.0057818568,
-0.0238472447,
0.0246408638,
0.0088160001,
-0.004226055,
-0.0200529192,
-0.0832929164,
0.0157894026,
-0.0040657204,
0.0347277299,
0.0860540494,
0.0521095917,
0.042600859,
0.0504478738,
0.045688618,
-0.0131607549,
0.0370244607,
-0.0729853958,
0.0726350918,
0.0176456608,
-0.0287814774,
0.0401065238,
0.0430067107,
-0.0180482715,
-0.0595311634,
-0.0047386317,
0.0186732989,
0.0031689235,
0.0224252138,
-0.0048882398,
-0.0230544321,
-0.0559283756,
-0.0400108099,
-0.0537238643,
-0.0037354014,
-0.0009039526,
0.0418508723,
0.0319484025,
-0.0030595839,
0.0049585556,
0.0104842717,
-0.0580111668,
-0.0235167556,
0.036443688,
-0.0114523517,
0.0070203803,
0.0251143984,
-0.0093118139,
-0.0674758106,
-0.0501554646,
-0.0239234585,
0.0262858048,
0.0066881133,
0.0016349244,
0.0677935481,
0.0549140312,
-0.0260856673,
0.0003060564,
-0.056281995,
0.0506829619,
-0.00425696,
-0.0098626912,
0.0348522402,
0.0073043942,
-0.024052063,
0.0243806709,
0.0335932523,
-0.0173594411,
-0.0105564157,
0.0215321053,
-0.029920198,
0.0208849758,
-0.0251867864,
-0.0228161793,
-0.0071163485,
0.0020509763,
0.0811694562,
-0.0274260752,
0.0020240964,
0.0007553629,
2.636e-7,
0.0347606651,
3.268e-7,
-0.0018001656,
-0.0035345638,
0.0593972988,
0.0136616807,
-0.0365039967,
-0.001401184,
0.0251025837,
0.0318802446,
0.0200295616,
0.0229789633,
0.0436865352,
0.028509682,
-0.0457379445,
0.0174892265,
-0.0540374219,
0.0321888477,
-0.0205067471,
0.0036727006,
-0.0183288362,
0.0614981577,
-0.0013450442,
-0.0397802331,
0.0074639576,
-0.0507784933,
0.0645279512,
-0.0660546198,
0.063197203,
0.0247893576,
-0.0350264907,
0.0428415462,
0.0273464546,
0.023167029,
-0.0124387825,
0.0402427316,
-0.0188185424,
0.0536185466,
0.0238575619,
-0.0144625437,
-0.0069556278,
0.0118703796,
0.0121198576,
-0.0855454654,
0.0606656857,
-0.1016718671,
-0.0038016154,
-0.0552993193,
0.0362557657,
-0.0335660838,
-0.0194594823,
0.0154225221,
-0.0233106203,
0.0071742749,
0.0256703962,
-0.0185054634,
-0.0314208046,
-0.0717682168,
-0.0250905324,
0.0180091411,
-0.0412414074,
-0.0528598353,
-0.0232872218,
-0.0008095873,
0.0362099372,
-0.0388977826,
0.0197656285,
2.330856754e-34,
-0.0065238066,
0.0136133563,
0.0318568461,
0.0314971469,
-0.0237567611,
-0.0159552488,
0.0062325867,
0.0106268404,
0.0591007024,
-0.0368356891,
-0.0033658589
] |
http://arxiv.org/abs/2402.05948v1 | http://arxiv.org/pdf/2402.05948v1 | DE$^3$-BERT: Distance-Enhanced Early Exiting for BERT based on Prototypical Networks | 2024-02-03T15:51:17 | 2024-02-03T15:51:17 | Jianing He, Qi Zhang, Weiping Ding, Duoqian Miao, Jun Zhao, Liang Hu, Longbing Cao | cs.LG | cs.LG, cs.CL | Early exiting has demonstrated its effectiveness in accelerating the
inference of pre-trained language models like BERT by dynamically adjusting the
number of layers executed. However, most existing early exiting methods only
consider local information from an individual test sample to determine their
exiting indicators, failing to leverage the global information offered by
sample population. This leads to suboptimal estimation of prediction
correctness, resulting in erroneous exiting decisions. To bridge the gap, we
explore the necessity of effectively combining both local and global
information to ensure reliable early exiting during inference. Purposefully, we
leverage prototypical networks to learn class prototypes and devise a distance
metric between samples and class prototypes. This enables us to utilize global
information for estimating the correctness of early predictions. On this basis,
we propose a novel Distance-Enhanced Early Exiting framework for BERT
(DE$^3$-BERT). DE$^3$-BERT implements a hybrid exiting strategy that
supplements classic entropy-based local information with distance-based global
information to enhance the estimation of prediction correctness for more
reliable early exiting decisions. Extensive experiments on the GLUE benchmark
demonstrate that DE$^3$-BERT consistently outperforms state-of-the-art models
under different speed-up ratios with minimal storage or computational overhead,
yielding a better trade-off between model performance and inference efficiency.
Additionally, an in-depth analysis further validates the generality and
interpretability of our method. | DE$^3$-BERT: Distance-Enhanced Early Exiting for BERT based on Prototypical Networks: Early exiting has demonstrated its effectiveness in accelerating the
inference of pre-trained language models like BERT by dynamically adjusting the
number of layers executed. However, most existing early exiting methods only
consider local information from an individual test sample to determine their
exiting indicators, failing to leverage the global information offered by
sample population. This leads to suboptimal estimation of prediction
correctness, resulting in erroneous exiting decisions. To bridge the gap, we
explore the necessity of effectively combining both local and global
information to ensure reliable early exiting during inference. Purposefully, we
leverage prototypical networks to learn class prototypes and devise a distance
metric between samples and class prototypes. This enables us to utilize global
information for estimating the correctness of early predictions. On this basis,
we propose a novel Distance-Enhanced Early Exiting framework for BERT
(DE$^3$-BERT). DE$^3$-BERT implements a hybrid exiting strategy that
supplements classic entropy-based local information with distance-based global
information to enhance the estimation of prediction correctness for more
reliable early exiting decisions. Extensive experiments on the GLUE benchmark
demonstrate that DE$^3$-BERT consistently outperforms state-of-the-art models
under different speed-up ratios with minimal storage or computational overhead,
yielding a better trade-off between model performance and inference efficiency.
Additionally, an in-depth analysis further validates the generality and
interpretability of our method. | [
0.0307813864,
0.0849562213,
0.0052605998,
0.0139991408,
-0.0150097283,
-0.0044215014,
0.0179554559,
-0.0178494249,
-0.0479942001,
0.0007766084,
-0.0242227968,
-0.0575884953,
0.0026626235,
-0.0034533287,
0.0458848663,
-0.0621216521,
0.0359912142,
-0.0175072718,
-0.0221583545,
-0.0133881131,
-0.0095525961,
-0.0146454014,
0.0067924405,
0.0824242756,
0.0316759981,
0.0095012356,
-0.0421714224,
0.0402567089,
-0.0050253081,
-0.0264951941,
0.0466337204,
0.0685499609,
0.0124459062,
0.0483637005,
0.0000021518,
-0.0553154871,
0.005419055,
0.0107157957,
-0.0222062077,
0.039020434,
-0.0562329032,
0.0429625474,
-0.0012721503,
0.0358458497,
0.0092131235,
-0.0437617078,
0.0845756158,
-0.0114128049,
0.0147618726,
0.0459774584,
-0.0079864273,
-0.0133239878,
0.0167541318,
-0.036597982,
0.0282213259,
0.0025322314,
0.0262719654,
-0.0833571851,
-0.0290339626,
-0.0940443724,
0.0018476931,
0.010250642,
0.0396475606,
0.0370284617,
-0.0389266014,
0.0660343692,
-0.0343575329,
-0.039534051,
-0.0206936095,
0.044927381,
-0.0534903631,
-0.0386072397,
0.0058220793,
-0.0197722949,
-0.0102491984,
-0.0349393189,
-0.005784105,
-0.0157562476,
-0.0055116951,
0.0083829332,
-0.0033086685,
0.0544315763,
0.019277459,
0.0132816695,
-0.0444855802,
-0.0068381294,
0.0486896224,
-0.0197346527,
-0.0035779162,
0.0267495755,
-0.0248274729,
-0.022372622,
0.0255797505,
-0.0108106146,
0.0363733396,
0.0168825574,
-0.0507674664,
-0.0050434386,
-0.0266173314,
-0.0465812609,
0.0195170511,
0.0433729142,
-0.0031614825,
0.0275673661,
-0.0140856421,
-0.0037669621,
-0.0245251488,
0.0317260064,
-0.0310268011,
-0.0022512346,
-0.0123911668,
0.0144518437,
-0.0591113381,
0.0348594002,
0.0509652048,
-0.0349315554,
-0.0507402495,
0.0209392961,
0.0273754541,
0.0002875028,
-0.0025443144,
0.0324475206,
-0.0255122706,
-0.0061407494,
-0.0139137898,
0.0332256891,
-0.0554950461,
-0.0043443074,
0.0030755955,
-0.0352283157,
-0.0297035798,
0.0356154926,
0.0173931643,
-0.028728012,
0.0395710953,
0.0608381592,
0.0237636026,
-0.0411263369,
-0.0667857602,
0.0353646465,
-0.0000049789,
-0.0620668046,
0.0123891179,
-0.0590321347,
0.0151112089,
0.0095521528,
-0.0186887924,
-0.0124848969,
-0.0157403238,
-0.0148229254,
-0.0042100218,
0.0373223871,
-0.0564781949,
0.0030323782,
-0.0065255715,
0.005596465,
-0.0029823114,
0.0597725771,
0.0305280909,
0.0260403957,
0.0314086638,
-0.0223150253,
-0.0074962741,
0.0828425139,
0.0481915735,
-0.0150866937,
-0.0582036749,
-0.0118801603,
-0.0005286701,
0.0117159486,
0.01697753,
0.0185755994,
-0.0397530794,
-0.002305642,
0.0769194141,
0.0652051494,
0.0149794994,
0.0359858163,
0.0340448096,
-0.01064335,
0.0164131839,
-0.0003991114,
-0.0515318178,
-0.0099262344,
-0.0168390553,
0.0195676032,
0.0414319821,
0.0591528639,
-0.0386062004,
-0.0286956206,
-0.0090200752,
-0.026764458,
-0.0529751815,
-0.0155185815,
0.0466107465,
-0.0015304163,
-0.0331139825,
0.0805112347,
0.0168881193,
-0.0298800692,
-0.0140323341,
-0.0463768244,
0.0305819847,
0.0715874806,
0.0296042655,
-0.0311891884,
-0.0808692798,
0.0345693827,
-0.0358932987,
0.0396799967,
0.0186939556,
0.0120773166,
0.078639105,
-0.0555196777,
-0.0448697321,
0.0591131039,
0.0047030691,
0.0138172368,
0.0630299896,
-0.0048292791,
-0.0081232199,
-0.1198177785,
0.0290997364,
0.0266495757,
-0.0079920664,
0.0087238392,
0.0162587557,
0.0504783057,
0.0339881778,
-0.0173329059,
-0.0350346752,
-0.0114191324,
0.0262386594,
-0.006121065,
-0.0000281886,
0.0265124254,
-0.0616400279,
-0.0741643608,
0.0034150742,
0.0033976154,
0.0010469009,
0.0167489648,
-0.0192264393,
-0.0191425476,
0.0199665409,
-0.0429599136,
0.1137879118,
-0.0345634297,
0.0239890944,
-0.0293982662,
-0.0073003815,
0.0090297069,
0.0631964654,
-0.0425005369,
0.0295236129,
0.0105013745,
0.013389919,
0.0158437509,
-0.0048006745,
-0.0518700853,
0.0038244084,
-0.0187644921,
-0.0174855366,
0.0331523009,
0.00904008,
-0.0413317308,
-0.015192112,
-0.0510943159,
0.0152469529,
-0.0050815521,
0.0533129238,
0.01428454,
0.0138144614,
-0.0184495989,
-0.0284814835,
-0.0004274272,
-0.0120162833,
-0.0630028695,
0.0085585732,
0.0223170612,
0.0305531211,
-0.0583692044,
-0.0322785601,
-0.0278766938,
-0.0046382095,
0.0911428481,
-0.0426542163,
-0.0335035101,
0.009596915,
-0.0349311531,
-0.0397199951,
0.0713483542,
0.0176444538,
-0.0469237417,
0.0043450808,
-0.058761403,
0.0478872098,
-0.008756144,
0.0024246718,
0.0047091739,
0.0115579786,
0.0089184316,
-0.064187102,
-0.0099662812,
-0.0487723053,
0.0420234017,
-0.0167415738,
-0.0061697802,
-0.0237961859,
0.080035679,
0.0031897947,
0.0043397606,
-0.0655156896,
-0.0058870283,
0.0349695906,
-0.0214479845,
0.0046133394,
-0.0060714995,
-0.0048299772,
-0.0050834171,
-0.0279389862,
-0.036670357,
0.1063449755,
-0.0389707349,
-0.0076192962,
-0.0009886156,
0.0150942672,
0.0001963378,
0.0557982698,
0.0172094014,
0.005984074,
0.0269490648,
-0.0487856939,
-0.0348502994,
-0.0026349898,
0.003188201,
-0.0062521622,
0.0065218876,
-0.0362554342,
-0.0049734851,
0.0159817636,
-0.0450405814,
-0.0161812659,
-0.019845536,
0.0040547163,
0.0001964671,
-0.0178298615,
-0.0418040343,
-0.0432193652,
0.0025148848,
-0.0162622519,
0.0913219228,
-0.0079630874,
0.0241758097,
-0.0039626569,
0.0703880414,
0.0015315925,
-0.0090160836,
0.0184350032,
-0.0446632653,
0.0435838625,
0.0261007957,
0.0630680621,
0.0190003701,
-0.0776690543,
-0.0107027851,
0.044240348,
-0.0129577201,
-0.0489340909,
0.0409019627,
0.000666305,
-0.0567073114,
-0.0281903427,
0.1338260323,
0.0031051205,
-0.0342933014,
0.0008170292,
0.0462808162,
0.0769075975,
0.002410423,
0.0177423488,
-0.0266732965,
-0.0394288003,
0.0074133929,
0.0432480834,
0.024823986,
-0.0288221464,
0.012866376,
-0.1215699613,
0.024092989,
-0.0040477575,
-0.0418012701,
-0.0147823272,
-0.0350569971,
0.0037049174,
-0.0027993945,
-0.0306377765,
0.0026922608,
0.0206398126,
0.0051453728,
0.0529864654,
-0.0474261418,
-0.0256013386,
0.0186827015,
-0.0506492779,
0.0240757875,
0.0423341393,
-0.007741503,
-0.0172082502,
-0.0379288569,
0.0415458381,
-0.0825924873,
-0.0225572418,
-0.0016134843,
0.0794244185,
-0.0185519438,
-0.0152010517,
-0.0302083995,
0.0112738591,
-0.0072539328,
-0.0064700609,
-0.0099224066,
0.0414256901,
0.0020190149,
-0.0673504099,
-0.0164025649,
0.0193328578,
-0.0259236619,
0.0575377047,
0.0424925834,
-0.09838406,
-0.0323007032,
0.0113701746,
0.0018316915,
-0.0567505211,
0.0525704734,
-0.0088527836,
-0.0243704207,
-0.0962870941,
-0.0219300576,
0.0014237883,
0.0276719462,
0.0783271343,
0.0035098838,
-0.0308907554,
-0.0223165043,
0.0373998396,
-0.04233649,
0.0189053733,
0.0505311042,
-0.0069202776,
-0.0536184423,
0.0127390549,
0.057442829,
-0.0104080308,
-0.0443691462,
0.0108646527,
-0.0267717112,
-0.0565533377,
0.0282539427,
0.0472064093,
0.0176360346,
0.0398898274,
-0.0182287041,
-0.0606432408,
0.0900834203,
0.0528143048,
0.0375959054,
-0.0068957959,
0.0527841151,
-0.0146528697,
0.0136149209,
0.026946418,
-0.0415118001,
0.0232902467,
0.0321916193,
0.0205401629,
-0.0422478542,
-0.0149730192,
0.0049526771,
-0.008297598,
0.0626217723,
-0.0319804922,
-0.0165299661,
-0.054124862,
0.0373759829,
0.041355513,
-0.0115531282,
-0.0489960425,
0.0173977558,
0.0447658673,
0.0463271663,
-0.0170619916,
0.0089365728,
0.0377159081,
-0.0083530927,
-0.0447989106,
-0.0070739137,
0.0420326851,
0.0567088351,
0.0307241753,
-0.01200746,
-0.0264349673,
0.0197739378,
-0.0037971644,
0.0110155884,
-0.0639255047,
0.0229213759,
0.055936791,
0.0076011652,
0.0131498212,
0.0354683958,
-0.04487684,
-0.0084701627,
-0.1016489789,
0.0660166368,
-0.0045741871,
0.034339536,
0.0145266708,
-0.0215069875,
0.0090572871,
0.0081109218,
0.0200772602,
-0.0313861482,
-0.0651426017,
0.0083897812,
0.0208120756,
-0.0018743249,
-0.0319237188,
0.0398112573,
0.020400703,
0.0759952888,
-0.006946546,
-0.0100955665,
-0.0272482727,
-0.0144834071,
0.057616353,
0.0179722812,
-0.0285750888,
0.0262595061,
-0.012999868,
-0.0625297055,
-0.0365735143,
0.0423796326,
0.0022390638,
0.0400321111,
0.0539170504,
0.0532323383,
-0.0105400784,
-0.0533304885,
0.0212921798,
-0.0024162282,
-0.0057873507,
-0.00254496,
0.02247045,
-6.878156136e-33,
-0.0311805513,
0.0102435639,
-0.0026464327,
-0.0205455702,
-0.0089060655,
0.0053829164,
-0.0390987955,
0.0112316143,
-0.0174847096,
-0.0168920942,
0.0117140841,
0.0253157672,
0.0181824975,
0.0247878656,
-0.010833879,
-0.012434477,
0.0505212061,
-0.0015135399,
-0.0185525417,
0.001887318,
-0.0258997921,
0.0549451895,
0.0860892162,
0.0096125575,
-0.0451526977,
0.0010640465,
-0.0465763584,
0.0116941351,
0.014726148,
0.0300022233,
-0.0209963061,
0.0228123348,
-0.0152849881,
0.0022525222,
-0.0092226537,
0.0011179344,
-0.0796880499,
-0.0426929332,
-0.0121402079,
0.0228386205,
-0.0186296571,
0.0224098433,
0.0448397845,
0.0050040558,
-0.0045818309,
0.014714838,
0.0381755568,
0.019752549,
-0.0193704832,
-0.0510304049,
-0.0263051242,
0.0352727212,
-0.0185493361,
-0.0295640435,
-0.0091896858,
0.0208932627,
-0.0040420275,
-0.0221851654,
-0.0526025929,
0.0035267442,
0.0334251747,
0.0035738095,
0.0164488368,
0.0405916385,
0.0093821324,
0.0540787838,
0.0562080443,
-0.0124897854,
0.0280522201,
-0.0459508263,
0.0879118145,
0.0286414735,
0.0147696044,
-0.0555402189,
0.0454326198,
0.0046198708,
-0.0595145747,
0.0341707096,
0.0383449756,
0.0033504141,
0.0179012939,
-0.0137308463,
-0.0324862115,
-0.0299083032,
-0.0281358436,
-0.0583576113,
-0.0312538929,
-0.0090760505,
0.0262663942,
0.0013779071,
-0.0308046062,
0.0025011185,
0.0160013363,
-0.0199864879,
0.0253899582,
0.007652353,
-0.0219236556,
0.0192538984,
-0.0001995314,
0.0381627679,
-0.0673227534,
-0.0617494658,
-0.0150329378,
0.0078734336,
0.0117136175,
-0.003088743,
0.0297072008,
0.0471192747,
-0.0702160448,
-0.001586984,
-0.047490824,
-0.001699895,
-0.0044177482,
-0.0136258584,
0.0351364128,
-0.0069633769,
-0.0071700588,
0.0097723054,
0.0299057011,
0.051664006,
-0.0066521894,
0.0104881218,
0.0340625942,
0.0407116376,
-0.011401834,
-0.023673933,
-0.0339293107,
-0.017327344,
0.0340069868,
-0.0535052642,
-0.0248122178,
0.0231954064,
2.928e-7,
0.0519435182,
0.0175271612,
0.0646036491,
-0.0037443105,
-0.0147388773,
0.020208383,
0.0082122004,
0.0113156922,
0.0002748528,
0.0477185622,
0.0602404624,
-0.0116414111,
0.0064564752,
0.0098277535,
-0.0937972367,
-0.0013473197,
-0.0713094324,
0.01255948,
-0.0452145077,
-0.001606026,
-0.0082527306,
0.0810365826,
0.0365352854,
-0.0235676449,
0.0160431806,
-0.0590257421,
0.0411102213,
-0.0177801475,
0.0194128361,
0.0196270254,
0.0368779898,
0.0058433474,
-0.0046622525,
-0.0057791616,
0.0072186696,
0.045459345,
-0.0056242635,
0.0094091538,
-0.0206085462,
-0.0351927318,
-0.0016414702,
-0.0019589323,
0.0023009379,
-0.0570007004,
0.05543378,
-0.0454500578,
-0.0222671833,
0.0320924148,
-0.0211785547,
-0.0310333241,
-0.0070710988,
-0.0313576013,
-0.025453208,
-0.0066353143,
0.0220291447,
-0.0573961139,
-0.0204049405,
-0.0724081919,
-0.0433859117,
0.0221681353,
-0.0451204777,
-0.0440732874,
-0.0412616543,
-0.0314551629,
0.0417046063,
-0.0841768011,
0.0072982255,
2.877345225e-34,
0.0447096787,
-0.0171401165,
0.0114131048,
0.0171626899,
0.0133027257,
-0.0071495753,
0.0542242303,
-0.013003015,
0.0217416082,
-0.0475428626,
-0.0521657579
] |
http://arxiv.org/abs/2402.02175v1 | http://arxiv.org/pdf/2402.02175v1 | Enhancing Complex Question Answering over Knowledge Graphs through Evidence Pattern Retrieval | 2024-02-03T14:54:13 | 2024-02-03T14:54:13 | Wentao Ding, Jinmao Li, Liangchuan Luo, Yuzhong Qu | cs.CL | cs.CL, cs.IR | Information retrieval (IR) methods for KGQA consist of two stages: subgraph
extraction and answer reasoning. We argue current subgraph extraction methods
underestimate the importance of structural dependencies among evidence facts.
We propose Evidence Pattern Retrieval (EPR) to explicitly model the structural
dependencies during subgraph extraction. We implement EPR by indexing the
atomic adjacency pattern of resource pairs. Given a question, we perform dense
retrieval to obtain atomic patterns formed by resource pairs. We then enumerate
their combinations to construct candidate evidence patterns. These evidence
patterns are scored using a neural model, and the best one is selected to
extract a subgraph for downstream answer reasoning. Experimental results
demonstrate that the EPR-based approach has significantly improved the F1
scores of IR-KGQA methods by over 10 points on ComplexWebQuestions and achieves
competitive performance on WebQuestionsSP. | Enhancing Complex Question Answering over Knowledge Graphs through Evidence Pattern Retrieval: Information retrieval (IR) methods for KGQA consist of two stages: subgraph
extraction and answer reasoning. We argue current subgraph extraction methods
underestimate the importance of structural dependencies among evidence facts.
We propose Evidence Pattern Retrieval (EPR) to explicitly model the structural
dependencies during subgraph extraction. We implement EPR by indexing the
atomic adjacency pattern of resource pairs. Given a question, we perform dense
retrieval to obtain atomic patterns formed by resource pairs. We then enumerate
their combinations to construct candidate evidence patterns. These evidence
patterns are scored using a neural model, and the best one is selected to
extract a subgraph for downstream answer reasoning. Experimental results
demonstrate that the EPR-based approach has significantly improved the F1
scores of IR-KGQA methods by over 10 points on ComplexWebQuestions and achieves
competitive performance on WebQuestionsSP. | [
0.0298213642,
0.0178084131,
-0.0060161287,
0.044648312,
-0.0564078651,
-0.0095212702,
-0.0486552753,
0.0341151506,
0.0343116261,
-0.0322781801,
0.0038526813,
0.0002724456,
-0.0415748209,
0.0252713822,
-0.0068481127,
-0.0202407092,
0.0348182656,
0.0216835793,
0.052535519,
0.0115964534,
0.0010959127,
0.0280826371,
-0.0459652655,
0.007946441,
-0.0374276489,
-0.0531960092,
-0.0112889959,
0.009260715,
0.0314808227,
-0.0411261581,
0.0166055225,
-0.0044223801,
-0.0031979275,
-0.0459570773,
0.0000016409,
-0.0354291499,
-0.0320244357,
0.0271093436,
-0.0427322835,
-0.0042802268,
-0.0540114604,
0.0486323349,
0.0482852906,
0.0127320196,
-0.0529343784,
-0.0071374751,
0.0194748864,
0.0752149895,
0.0298828948,
0.0229272954,
-0.0054990384,
-0.0350923464,
0.0046484442,
0.0344668254,
-0.0118615739,
-0.0086880177,
0.0420067124,
0.046939224,
0.0345858261,
0.022447139,
0.0030319872,
0.0059143361,
0.0101274764,
0.0276949015,
0.0067079426,
0.0131181628,
-0.0216425844,
0.0466231592,
0.0361247212,
0.1177110374,
0.0408957936,
-0.0089423684,
-0.0067182244,
-0.0068462146,
0.0134627204,
0.0888840705,
-0.0725870058,
0.0049118311,
-0.0492985584,
-0.0356251858,
0.0126224579,
0.0248313807,
0.023485478,
-0.0115416106,
-0.0098138694,
0.078118369,
-0.0065694479,
-0.0383721516,
0.0038672697,
-0.0098374821,
-0.0540938899,
0.0053698709,
0.024947159,
0.0433623716,
-0.000659347,
0.0386253186,
0.0460291468,
0.020263603,
0.0209336895,
-0.0320085809,
-0.0242950749,
0.0186605491,
-0.0093582552,
0.0641462058,
-0.0242861751,
0.0202789735,
-0.0205916259,
-0.0322646014,
-0.0413541682,
-0.0229771622,
-0.100451462,
-0.0309484918,
-0.0462752469,
0.0810461566,
-0.0374571681,
-0.047846552,
0.0232429281,
0.0251975022,
-0.0081922403,
-0.0448972546,
0.0229959823,
0.0665959269,
0.0180747081,
0.0123134973,
-0.0599541254,
0.0631212443,
-0.0486042351,
0.0171212442,
0.0132364472,
-0.0064329132,
0.0046880343,
0.0038308771,
0.0552654937,
-0.0268877987,
-0.0199825075,
0.0338708349,
0.0207870677,
-0.0049764868,
-0.0316080526,
-0.085788779,
-0.0025689667,
0.0143987676,
0.0204142556,
-0.0744785443,
0.033272285,
0.0094395205,
-0.0419053324,
0.0070640212,
0.0162147321,
0.0276361126,
-0.0409658253,
0.005810814,
-0.0153692719,
-0.0442753024,
0.0067415447,
0.0211651158,
-0.0323432274,
0.0394155793,
-0.0054887817,
0.0192228816,
0.055807896,
0.0137942173,
-0.038427975,
-0.0404449403,
0.0337281823,
-0.0154910116,
0.0282693878,
0.0550379977,
-0.0103403516,
-0.0032175125,
-0.02427903,
0.0307231601,
-0.0709719509,
-0.0333453342,
0.0029409225,
0.0523230061,
0.018381197,
-0.0136222867,
0.0013058616,
0.0192169733,
0.0261146519,
0.0339131765,
0.0135588171,
0.0032076398,
-0.0278055072,
0.0096096713,
-0.0082960874,
0.0102036819,
-0.0312835574,
-0.0608929284,
-0.0156373475,
0.0378353447,
0.0094314441,
0.0347558409,
0.0320589095,
-0.010413534,
-0.0252162088,
0.0190153401,
0.0383340754,
0.0130673889,
0.0019255417,
-0.0143929832,
0.0254167188,
0.0768407658,
0.0445140339,
-0.0338533372,
-0.0447105989,
0.0454335958,
-0.0321974754,
-0.0514554679,
-0.017704593,
-0.0445897803,
0.0034800807,
-0.0330390632,
-0.0406643599,
-0.0222770385,
-0.0311733875,
-0.0315018147,
0.0167357381,
-0.0055808332,
-0.0061851023,
-0.0480595678,
-0.0180527326,
-0.0029305336,
0.0249469262,
-0.0049951044,
0.0154835181,
0.0053878971,
0.0330091491,
0.0377928354,
0.0037132853,
0.0387683958,
0.0422529094,
-0.0168565586,
-0.0429455452,
-0.0142721916,
-0.0306060407,
-0.075155884,
0.033592023,
-0.0318570584,
0.0306155197,
-0.0360513218,
-0.0137952929,
-0.0540676154,
-0.0111119989,
-0.0598976649,
0.057766851,
0.0035511774,
-0.0645953789,
0.063970387,
-0.0473834276,
-0.0633860007,
0.0678931773,
0.0069629727,
-0.0019686371,
-0.025120303,
-0.0157624166,
0.0154621303,
-0.0253364481,
0.0397202075,
-0.0101721743,
-0.0318624042,
0.0171960127,
0.0509126186,
-0.0517835133,
-0.0585485175,
0.0200491138,
-0.0485318452,
0.0468180627,
0.0082946578,
0.0241705161,
-0.0066277203,
0.0061224792,
0.0222594291,
-0.0448444225,
-0.0519900434,
0.0355110057,
-0.0227694437,
0.0123491716,
0.0249951817,
0.0469294116,
0.0040128632,
-0.0197909679,
-0.0613106601,
-0.0160756204,
0.0636753142,
0.0014887005,
0.017387107,
0.0090124225,
0.0075707934,
-0.0411499254,
0.0136302505,
-0.0084720021,
0.008149459,
0.0152759003,
-0.0360809974,
-0.0142332632,
0.0538935363,
-0.0387840457,
-0.0073663373,
0.0148538304,
-0.0396684706,
-0.0794952661,
0.0180284474,
-0.0302352328,
0.0508976206,
0.0086312313,
-0.0210205317,
0.0168339517,
0.0034844649,
0.0267077927,
-0.0342848711,
-0.0176564809,
0.0019370444,
0.0163956881,
0.0290134996,
-0.0333384201,
-0.0124094142,
0.0480847731,
0.0114503521,
-0.0509311557,
0.0123711918,
0.081601508,
0.0166117474,
0.0473905914,
-0.0334261544,
0.0017741235,
-0.0137485331,
0.0441177264,
-0.0056064194,
0.0172542427,
0.0119617525,
-0.0132303983,
-0.0260684732,
-0.0126380334,
0.0392901562,
-0.0478897579,
-0.006837361,
0.0018441421,
0.034803845,
0.0095532862,
-0.0112932138,
-0.0343054309,
0.0168408249,
0.000835447,
0.0505053364,
0.0101429978,
-0.0136414841,
-0.0094166724,
0.021647593,
0.0315022096,
-0.0430671126,
-0.0412640348,
-0.0083552357,
-0.0090701841,
0.0793266073,
0.0597791001,
0.0187985729,
0.037479464,
0.0394210741,
0.0197438803,
0.0764139071,
-0.0260309838,
-0.0193727445,
0.0088056652,
-0.0183565896,
-0.011935005,
-0.0601798296,
-0.0126608517,
-0.009624227,
0.02732151,
-0.0597831085,
0.0103420634,
0.0593761131,
0.0501179844,
0.0063067661,
-0.0240876414,
-0.0106389159,
0.0112370504,
0.0096563492,
0.0302480124,
-0.0864545181,
-0.0201314948,
-0.0075750174,
0.0310092941,
-0.0080123767,
-0.027848646,
0.0537797622,
-0.0884822458,
0.0365751497,
-0.0052578761,
0.000135386,
-0.038632147,
-0.064897567,
0.0098741753,
0.0145792291,
-0.0360674709,
0.0183051713,
-0.036956463,
-0.0048021423,
0.0329370163,
-0.0085463487,
-0.0000950496,
0.0211106539,
-0.0185039844,
-0.0116061931,
-0.0055225678,
0.0020469176,
0.0030267539,
-0.0439297408,
-0.0530060679,
-0.0518187881,
0.0333848596,
-0.013034598,
0.0011170412,
0.0235663708,
-0.0051434515,
-0.0613830239,
-0.0173972379,
-0.0373267159,
-0.0252186097,
0.0208564587,
0.1030113921,
-0.0283295251,
0.0010319974,
-0.0307722073,
-0.0280680824,
-0.0326381475,
0.0504801534,
0.0340329669,
0.0556205213,
-0.0346478783,
-0.0130283032,
0.0244497415,
-0.0391461141,
-0.0107781449,
0.0360119157,
-0.0131680034,
0.0747698024,
-0.0873573273,
-0.0495105349,
0.0352811627,
0.0547045507,
-0.0576720387,
0.0073588789,
0.0444712751,
0.0105557786,
-0.0997116715,
0.0568292402,
-0.0168033056,
-0.0008310268,
-0.0307557695,
-0.0071113836,
-0.0024006348,
-0.0221584328,
-0.0288248081,
-0.0215057153,
0.0466169007,
-0.0290990528,
-0.0290641449,
0.0143632265,
-0.0005072783,
0.0329475068,
-0.0102509093,
-0.0236532893,
0.0280602332,
0.0000998933,
-0.0455121845,
-0.0687091127,
0.0099276165,
-0.0405898057,
-0.0307144132,
-0.0160574615,
-0.0864251181,
0.0132266991,
0.0046604518,
0.0968617722,
-0.0511933863,
-0.0076871729,
0.0553362221,
0.0526585393,
0.0089200865,
0.0541278981,
-0.0288422555,
-0.0269971676,
0.0031037296,
-0.0118795233,
0.0025023336,
-0.0108096292,
0.0192790274,
0.0052224388,
-0.0303242132,
0.0027701496,
-0.0229354389,
-0.0048432867,
-0.0021867233,
0.0103850709,
-0.0267789364,
0.066612795,
-0.0024230187,
0.0199302547,
-0.0254747272,
-0.0377718955,
-0.0040331241,
0.0175815262,
0.0476599708,
-0.0160873774,
0.038786862,
0.0388288051,
-0.0284451693,
-0.0597625338,
0.0065903743,
-0.0535123535,
-0.0334145837,
-0.0026042499,
-0.004991123,
0.0486402623,
0.0074343737,
-0.0231709946,
0.0350595079,
-0.033169698,
0.0033306349,
-0.0114213154,
0.0042292066,
0.0310473759,
0.0707109794,
0.070510827,
-0.0057276827,
0.0161134154,
0.0611756071,
0.0021039748,
0.0099269673,
-0.0142082041,
-0.0669370294,
-0.0063970257,
-0.0342717208,
0.0196832493,
0.0618498214,
-0.0418128744,
0.0188255198,
-0.0128720105,
-0.0496603176,
-0.0068225791,
0.022619931,
-0.0198722128,
-0.0029730871,
0.0536933057,
0.0317393616,
0.0621742979,
-0.0532249734,
-0.0237354208,
0.0178778004,
-0.0616290122,
-0.0157892946,
0.0136483051,
-6.188111565e-33,
-0.0105806571,
-0.0467475392,
0.0311994944,
0.0030669936,
-0.011170432,
0.0406461172,
-0.012361303,
-0.0301804282,
-0.0455723405,
0.0362221375,
0.0115017276,
-0.0097206216,
-0.0031995552,
0.0038847048,
-0.0076614739,
-0.0014894841,
-0.0080033597,
0.0249114726,
0.0029406345,
0.0327942818,
0.0580704957,
0.0152955903,
0.0163946822,
-0.0476024598,
0.024551427,
-0.0231542848,
0.0680186599,
0.0448921546,
-0.0123364534,
0.0893928185,
-0.0159243513,
0.0333312936,
0.038397681,
-0.0078754649,
0.0351108164,
0.0003110936,
-0.1105036736,
-0.0666963682,
0.0283385925,
-0.0218370352,
-0.0680458173,
-0.0262953527,
0.0272983648,
0.0065014306,
0.0818831176,
-0.0231538918,
0.0484226197,
0.0012716483,
-0.0491533168,
-0.0273435339,
-0.0298363455,
0.0340765901,
-0.0260176547,
0.0214514639,
-0.0552627556,
0.0221574884,
-0.0022120236,
-0.0116906036,
-0.0752444193,
-0.0184900481,
-0.0298096929,
0.0318249129,
-0.0265802518,
0.0279517565,
0.014649082,
0.0291344095,
0.0113098975,
-0.025816502,
-0.0140832663,
-0.0108239176,
-0.0210154317,
0.0651667565,
0.0210725218,
0.0072563365,
0.0548503958,
0.0231704153,
-0.1169825569,
0.0115139447,
0.0586251169,
0.0605339669,
-0.0319389589,
0.0087898551,
-0.0531384908,
-0.0374034382,
-0.0096127568,
0.016809931,
-0.0397600234,
-0.0201809853,
0.018315034,
-0.0056293015,
-0.002577797,
0.0374503322,
0.0224610437,
-0.0401595868,
0.0350907482,
0.0608269572,
0.0118482867,
0.039832063,
-0.0085593592,
0.0562836565,
-0.0343500935,
0.0144695276,
0.0190671738,
-0.0620591231,
-0.0002567958,
0.0360401385,
0.0102390759,
-0.0202695094,
0.0039142319,
-0.0301748347,
-0.0178106483,
-0.037905097,
-0.0446655974,
-0.0677931309,
0.0381266251,
0.0899285227,
0.0209917966,
0.0240247771,
0.0243026465,
0.037272837,
-0.0036820271,
-0.0054061469,
0.0099685118,
-0.0044717174,
0.0066372799,
0.0087128328,
-0.0376319773,
0.0186109357,
-0.0819856897,
-0.0721644014,
-0.0112750344,
0.0667841285,
2.554e-7,
0.05182226,
0.0174503103,
0.0089622624,
0.0096569555,
-0.0057034474,
0.0002403193,
0.0142111154,
-0.0329131596,
0.0393404588,
-0.0293186922,
0.0217735674,
0.0600683354,
-0.0169544276,
0.0559169538,
-0.0502695106,
-0.0274645369,
-0.0192414615,
-0.0877186283,
-0.0496065244,
-0.0103994543,
0.0745716691,
0.014800122,
-0.0053130924,
0.005664058,
-0.0269532762,
-0.0534623191,
0.0534951873,
0.0548650362,
0.0160169993,
-0.006724257,
0.0575840734,
-0.0230318122,
0.0096099712,
0.0423879996,
0.0051509268,
0.0463007949,
0.0148819052,
0.0159665998,
-0.0181355402,
-0.0446715057,
0.0064432672,
-0.0526718237,
0.0543751046,
-0.0020902692,
0.1020217314,
-0.0395774767,
-0.0163592752,
0.0298308562,
-0.0788854361,
-0.0193737876,
-0.0385706238,
0.0197550934,
-0.0017544216,
0.0132547747,
-0.0076439562,
-0.0063071526,
0.0067318105,
-0.0377088971,
0.0340992212,
0.034396071,
0.0087025659,
0.0021211824,
0.0309507698,
-0.0262686722,
0.0222227145,
0.0298153218,
-0.0000248696,
1.964972887e-34,
0.0218464956,
0.0206004512,
0.0291399658,
0.0352998488,
0.0358545631,
-0.0287379306,
-0.0811850652,
-0.0115790088,
0.0255547706,
-0.1067531854,
-0.0049784789
] |
http://arxiv.org/abs/2402.02145v1 | http://arxiv.org/pdf/2402.02145v1 | Analyzing Sentiment Polarity Reduction in News Presentation through Contextual Perturbation and Large Language Models | 2024-02-03T13:27:32 | 2024-02-03T13:27:32 | Alapan Kuila, Somnath Jena, Sudeshna Sarkar, Partha Pratim Chakrabarti | cs.CL | cs.CL | In today's media landscape, where news outlets play a pivotal role in shaping
public opinion, it is imperative to address the issue of sentiment manipulation
within news text. News writers often inject their own biases and emotional
language, which can distort the objectivity of reporting. This paper introduces
a novel approach to tackle this problem by reducing the polarity of latent
sentiments in news content. Drawing inspiration from adversarial attack-based
sentence perturbation techniques and a prompt based method using ChatGPT, we
employ transformation constraints to modify sentences while preserving their
core semantics. Using three perturbation methods: replacement, insertion, and
deletion coupled with a context-aware masked language model, we aim to maximize
the desired sentiment score for targeted news aspects through a beam search
algorithm. Our experiments and human evaluations demonstrate the effectiveness
of these two models in achieving reduced sentiment polarity with minimal
modifications while maintaining textual similarity, fluency, and grammatical
correctness. Comparative analysis confirms the competitive performance of the
adversarial attack based perturbation methods and prompt-based methods,
offering a promising solution to foster more objective news reporting and
combat emotional language bias in the media. | Analyzing Sentiment Polarity Reduction in News Presentation through Contextual Perturbation and Large Language Models: In today's media landscape, where news outlets play a pivotal role in shaping
public opinion, it is imperative to address the issue of sentiment manipulation
within news text. News writers often inject their own biases and emotional
language, which can distort the objectivity of reporting. This paper introduces
a novel approach to tackle this problem by reducing the polarity of latent
sentiments in news content. Drawing inspiration from adversarial attack-based
sentence perturbation techniques and a prompt based method using ChatGPT, we
employ transformation constraints to modify sentences while preserving their
core semantics. Using three perturbation methods: replacement, insertion, and
deletion coupled with a context-aware masked language model, we aim to maximize
the desired sentiment score for targeted news aspects through a beam search
algorithm. Our experiments and human evaluations demonstrate the effectiveness
of these two models in achieving reduced sentiment polarity with minimal
modifications while maintaining textual similarity, fluency, and grammatical
correctness. Comparative analysis confirms the competitive performance of the
adversarial attack based perturbation methods and prompt-based methods,
offering a promising solution to foster more objective news reporting and
combat emotional language bias in the media. | [
0.0570140518,
0.0691571683,
-0.0081135137,
0.0337638184,
-0.066470325,
-0.0195077714,
-0.0128279375,
-0.0413845889,
0.001962526,
-0.0439282134,
-0.0262300819,
0.0026945118,
-0.0233880877,
0.0151858944,
-0.024976952,
-0.0750592425,
0.0648883656,
0.0664152578,
0.040369682,
-0.0081891418,
0.069926627,
0.0093858596,
0.0016132311,
0.0109935999,
-0.0288774502,
0.0052983584,
0.0192733072,
0.0384179391,
0.0370449573,
0.0005218019,
0.0281770919,
0.0323559567,
-0.0204427652,
0.0153267141,
0.0000018955,
0.0101590035,
-0.0181983896,
-0.0202504843,
0.0522200018,
0.0206943229,
0.1347471029,
-0.0674446002,
-0.0147072412,
0.0229520183,
-0.0440850481,
-0.0258457437,
0.0442324281,
0.0615012757,
0.1248540431,
0.1208705157,
-0.0277850106,
0.0237843785,
-0.0287388973,
-0.0157650188,
-0.004199272,
-0.031785734,
0.0351238661,
-0.0355163589,
-0.0585011914,
0.0032613459,
-0.0030251446,
0.0411614478,
0.0190993827,
0.0108337179,
-0.0342324376,
-0.0124218911,
-0.0225210413,
0.0055601397,
-0.0065520676,
0.0162091646,
-0.0028774568,
-0.0412575863,
-0.0027112765,
-0.0055055027,
-0.0035653342,
0.0154979443,
-0.0295658503,
0.0009515297,
-0.0039879344,
-0.0037908601,
-0.0246047098,
-0.0196623765,
0.0177728087,
0.0367049128,
-0.0041872249,
0.0264737587,
0.0269719362,
-0.0501094013,
-0.023908997,
-0.007624127,
-0.0329218656,
-0.0012152804,
0.0550461076,
-0.0245004296,
-0.0015658459,
0.013999247,
-0.0535320602,
-0.0502980426,
-0.0211059637,
-0.0978203639,
-0.034529835,
0.0090300096,
-0.0087840315,
0.025083337,
0.0225246195,
-0.0004868065,
-0.0011776818,
0.0018381659,
-0.0475481972,
-0.0003932923,
-0.0114766425,
-0.0019395279,
-0.0030212002,
0.0129430005,
-0.0448494852,
0.0002803522,
-0.0507296696,
-0.0035547707,
-0.0232682917,
-0.0117063606,
-0.0237861313,
0.0183158722,
-0.0170347001,
-0.0109182047,
0.0177206229,
0.0410176069,
-0.0173481368,
0.0297512375,
-0.0283714123,
-0.002664482,
-0.0245391373,
0.07347624,
-0.0398063324,
-0.0201064218,
-0.0153827742,
0.0462333299,
0.0004469871,
0.0113735804,
-0.0168592017,
-0.0146022923,
-0.0388256051,
-0.0218043961,
0.0183183122,
-0.0263804328,
-0.0069247801,
-0.0251221433,
-0.0282881986,
-0.0277045146,
0.0054031126,
0.0110382307,
-0.0384565629,
0.0852484256,
0.040390905,
-0.0404440649,
0.0370461047,
0.0345097817,
0.0205704886,
0.0386798866,
-0.0064230007,
0.0878387168,
0.0483383238,
0.0030224074,
0.0177992731,
0.0185814425,
0.0171417929,
0.0138105964,
-0.006196036,
-0.011219169,
-0.002638214,
0.0135423057,
0.0058067916,
0.0672680438,
-0.0108588543,
-0.0380050018,
0.0849229619,
-0.0203380622,
-0.0569320135,
0.0272034798,
0.009947353,
0.0322287679,
-0.0137470327,
0.0426358953,
-0.0517058149,
0.0205029547,
-0.0653756708,
-0.0364775173,
-0.0032120585,
0.0259433314,
-0.0130801266,
-0.0251568332,
0.0132388035,
0.0031861039,
0.0551397875,
-0.0284621045,
-0.0062857489,
0.033241678,
-0.0601657853,
0.0243621618,
0.0287001785,
-0.0492465496,
0.0078098085,
-0.0315912552,
-0.0248105563,
-0.0049106339,
0.0548463091,
-0.0244781133,
0.008195295,
-0.0113864997,
0.0148175629,
0.0363440886,
0.0616665259,
0.0228654109,
-0.0410360917,
-0.0130790398,
0.0173180513,
-0.006748314,
0.0310590975,
0.0347997881,
-0.0081160115,
0.0327389576,
0.0048943395,
0.0300957728,
-0.006925182,
0.0316794515,
0.0344605595,
-0.0286320839,
0.0225003492,
0.0326513946,
0.035963282,
0.0267975852,
-0.0141439606,
0.0022401093,
0.017388612,
0.0017572663,
-0.0585694723,
0.0368800052,
-0.0425981767,
-0.0548192374,
-0.0050896439,
-0.0707964152,
-0.0064531411,
0.056374602,
-0.0442985818,
0.0169705711,
0.0691504925,
-0.0349182263,
0.0001456466,
-0.0146657648,
-0.0140764127,
-0.0108425487,
-0.0011838118,
-0.0579355992,
0.0450881049,
0.0025938717,
-0.0081378045,
-0.021462407,
-0.0309458729,
0.003412982,
0.0114386082,
-0.0324307457,
0.0022809985,
0.0248224642,
-0.0022181363,
0.057927411,
-0.0246658064,
-0.0041718106,
-0.0721084848,
-0.0508841574,
0.0006820899,
-0.013797475,
0.0202920288,
0.0637190416,
0.002350911,
-0.0132579599,
0.0425550044,
-0.0257173218,
-0.0497617386,
-0.0242300835,
0.0340658985,
0.0225275941,
0.0297146235,
0.0003028237,
-0.0435360111,
0.009081793,
-0.0463495404,
0.0173382405,
0.0109369364,
-0.0578202866,
-0.0241324473,
0.0542865768,
-0.0195158515,
0.0038395384,
-0.0281829666,
-0.0615828969,
-0.000939493,
0.0127662439,
0.0320523456,
0.0797120407,
0.0343249589,
0.069616355,
0.0220449679,
0.0028943894,
-0.0062180199,
-0.0073709795,
-0.0511313491,
0.0183465779,
0.0007800604,
-0.0331017189,
0.0558848865,
0.0127497744,
0.0526509285,
0.0016141937,
-0.0403882861,
0.0436216891,
-0.0245094765,
-0.0304850936,
0.0041589229,
-0.0165573824,
-0.008292404,
-0.0272378363,
-0.0341437906,
-0.0119723538,
0.062875092,
0.0159721002,
0.0394308753,
-0.0852908343,
-0.0125107225,
-0.0279925223,
0.0159531031,
-0.0254359283,
0.0276314653,
-0.035344746,
0.0127548408,
-0.0794988796,
-0.0609522872,
0.0678563118,
-0.0030735228,
-0.0008497169,
-0.0263646822,
-0.0092627853,
0.0247700866,
-0.0447437316,
-0.0153869977,
-0.0875752643,
0.0210841112,
-0.007293188,
0.0813839808,
-0.0149771003,
-0.0508880354,
0.0015919229,
0.0318684317,
-0.0306839813,
-0.0271717813,
0.0092366058,
0.0014272464,
0.0359717198,
0.0404584371,
-0.0586863831,
0.0344300419,
0.0164470449,
-0.0019502211,
0.0472133383,
-0.0345734209,
0.0131946728,
-0.0019997514,
-0.0191623904,
-0.0307277795,
-0.0356744193,
-0.0104523767,
-0.0608804077,
0.0317022875,
-0.027064221,
0.0393295474,
0.0761648864,
0.0405242555,
-0.0301868711,
0.006519177,
0.0192743428,
-0.0293794312,
0.0149645964,
0.0316885933,
-0.0675316527,
0.0138840647,
-0.0090145301,
0.0392783806,
-0.0078382706,
-0.0065301256,
-0.0521859527,
-0.1011330485,
-0.0345377624,
0.0118338522,
0.0113859903,
-0.0344402865,
-0.0406201556,
0.0067920787,
-0.0341913365,
-0.0083530955,
-0.0244153626,
0.0426680669,
0.0185681246,
-0.0390157551,
-0.0130514642,
-0.0257946234,
-0.0319198184,
-0.0217038766,
0.0291782953,
0.0359291621,
0.0616356879,
-0.0192118287,
-0.0086954329,
0.0511526912,
0.0298168939,
-0.0188865084,
-0.0331219099,
0.0376484133,
-0.0326204672,
-0.0108953351,
-0.0171822105,
0.0481591225,
0.0344907492,
-0.0069292937,
-0.0068149483,
0.0077416459,
0.0215929262,
-0.0322322063,
-0.0262378287,
0.0534572005,
0.087544702,
0.0873113796,
0.0138531551,
-0.0344928913,
-0.0005965207,
-0.0448831022,
0.0187094342,
-0.0263487045,
-0.064760156,
-0.0379172191,
-0.0006108656,
0.0078962641,
-0.0838465616,
-0.0467506535,
-0.0129384389,
0.0145757785,
0.0069338279,
-0.0137135983,
-0.0210820269,
-0.0185534731,
0.0007725239,
-0.000367589,
0.0057277684,
0.0157709345,
0.0152321579,
-0.0163441766,
0.0163642485,
-0.0075040893,
-0.027633518,
-0.0078231944,
-0.0224605463,
-0.0567577295,
0.037353605,
0.044300165,
0.007101723,
0.027767932,
-0.0228980053,
-0.0190508552,
0.0030418548,
0.1165477633,
-0.0347784571,
0.0454589725,
0.0350603648,
0.0372532494,
0.0217423793,
0.0254167859,
-0.0027641805,
0.0347753093,
0.0540231988,
-0.011617613,
-0.0250682309,
-0.0101943528,
0.0290095638,
0.0184026789,
0.0510483049,
0.0139596863,
-0.0279037692,
-0.0239903107,
-0.0112205734,
-0.0420607775,
-0.024036726,
0.0094194068,
0.0005816098,
-0.0186136793,
0.0195940677,
-0.0627301857,
0.0282701757,
-0.0374042578,
0.0230981484,
-0.0330951065,
0.0194894783,
0.0241818707,
0.0223492403,
-0.0138959996,
0.0339692645,
-0.0036475188,
0.0331581756,
0.007657364,
0.009714161,
-0.0144177023,
-0.0174437743,
0.0424332395,
0.0099986987,
-0.0062115635,
0.0158954319,
-0.0614769608,
0.0362608992,
-0.020306142,
0.0256047919,
-0.0027932192,
-0.0008428479,
0.0085376333,
-0.022343548,
-0.0180375725,
-0.0018741869,
-0.0167329572,
0.0261736158,
-0.0258161426,
0.0248001087,
0.0057277652,
-0.0218212027,
0.0220531635,
0.028336985,
-0.0616614856,
0.1041508466,
0.0481105708,
-0.0335108675,
-0.0407658033,
-0.0160173774,
0.081107229,
-0.0211549513,
0.0086607346,
-0.0011420177,
-0.0582792461,
-0.0364121161,
0.0343653262,
-0.0092348466,
0.0153707583,
-0.0051171505,
0.0052960101,
0.0654707476,
-0.0042622401,
0.0431506075,
-0.0156622846,
0.0253598318,
-0.0473194122,
-0.076917395,
-0.0011004569,
-6.405830016e-33,
-0.0147382617,
0.0069905231,
-0.0455356017,
0.0204922464,
0.0046959296,
0.0053771925,
-0.0165811852,
0.0146017205,
-0.0128256408,
0.0346590765,
0.0139741786,
0.0410183892,
-0.0053096288,
-0.0093611991,
-0.0616052411,
-0.0003074219,
0.071602799,
-0.0125350198,
-0.027154319,
0.0048902766,
0.0247387551,
0.017390959,
0.0375667214,
-0.1187697873,
0.0403928086,
0.0042460104,
-0.0902395025,
-0.0161915906,
0.025500983,
0.0170592796,
0.0099038389,
0.0314213298,
0.0048639784,
-0.0416884609,
-0.009475599,
0.0780551136,
0.0082353931,
-0.0297388248,
0.0585314631,
0.0152069973,
0.007440872,
-0.0486220308,
-0.0031474282,
-0.0082279565,
-0.0351268947,
0.029136423,
0.0107360501,
-0.0058819856,
0.0319822617,
-0.049144648,
-0.0459436215,
0.0094318865,
-0.0231780373,
-0.0087574674,
-0.0274096485,
0.0117035368,
0.0239839498,
0.0486088842,
-0.0552593023,
0.0216774084,
-0.0325642824,
0.0545684509,
-0.0105549451,
-0.0367607549,
0.0176739395,
0.0671636984,
0.024017334,
0.0196852218,
-0.0259158071,
0.0274127666,
0.0324974097,
-0.0148027614,
0.0802558586,
0.1098959148,
0.0183676779,
-0.0593033098,
-0.0697578937,
-0.0006088766,
0.0340199061,
0.0130174411,
0.0281963516,
0.0230273474,
-0.0215210132,
-0.0040833903,
-0.0164257325,
-0.0835772455,
-0.0417839959,
-0.037791457,
0.0129443742,
0.0145257246,
-0.0717033818,
-0.0053665624,
-0.0176079329,
-0.0518579297,
0.0190651584,
0.0394777805,
-0.0525076576,
-0.0038549695,
0.0015585072,
0.0202985238,
-0.0539254248,
0.0453544185,
0.0175087899,
0.0159066413,
0.0048191692,
-0.0133236721,
0.0588187799,
0.0231456775,
-0.0639831051,
-0.001849521,
0.0090312008,
0.0475317873,
-0.000805766,
0.0124299945,
-0.001372612,
0.0275380444,
-0.0102947801,
-0.0150378356,
0.0574849211,
0.0312841535,
-0.0077203833,
0.0032640502,
-0.0315916725,
0.0224633049,
-0.0391612388,
-0.0183549356,
0.0273484215,
-0.0135370521,
0.0006719935,
-0.0804546326,
-0.0269040931,
-0.0346991904,
2.797e-7,
0.0317273699,
0.0913450494,
-0.0039563156,
0.0143933436,
0.0177809652,
0.0328958742,
0.0164706595,
0.0353276506,
-0.0064760502,
-0.0550019965,
0.0133445226,
0.0098117273,
0.0083342073,
0.0022650412,
-0.0939650387,
0.035360232,
-0.0373966098,
0.0130556449,
-0.0132647967,
0.0210743807,
0.0882801116,
0.0759555399,
0.028174933,
0.0134744914,
-0.0345328972,
-0.0171185583,
0.0519265831,
-0.026414847,
-0.0496687628,
0.0115657877,
-0.0067546479,
-0.0323325917,
0.004709858,
-0.0311468281,
-0.0170424543,
-0.0273125302,
0.0053932518,
0.0362816304,
0.0354815461,
-0.0816702172,
-0.0256944876,
0.0207970925,
-0.0236524679,
-0.0087389071,
0.0777899027,
-0.0641977936,
-0.0199542046,
-0.100862287,
-0.0584398098,
0.001730739,
0.0515493602,
-0.0305085108,
-0.0719615296,
0.0076932237,
0.0116945235,
0.0620388463,
-0.0247222986,
-0.0488138385,
-0.002981435,
-0.0420707949,
-0.0332877003,
-0.0233039819,
0.0051107914,
0.0637602285,
0.0075098518,
-0.0170220286,
0.030301908,
2.572868311e-34,
0.0652826354,
-0.0187637806,
-0.0090486556,
-0.0379221812,
0.0321149565,
0.0478083119,
-0.0482549779,
0.0087921526,
0.0295894369,
-0.0545094013,
-0.0096596237
] |
http://arxiv.org/abs/2402.02144v1 | http://arxiv.org/pdf/2402.02144v1 | Probing Critical Learning Dynamics of PLMs for Hate Speech Detection | 2024-02-03T13:23:51 | 2024-02-03T13:23:51 | Sarah Masud, Mohammad Aflah Khan, Vikram Goyal, Md Shad Akhtar, Tanmoy Chakraborty | cs.CL | cs.CL | Despite the widespread adoption, there is a lack of research into how various
critical aspects of pretrained language models (PLMs) affect their performance
in hate speech detection. Through five research questions, our findings and
recommendations lay the groundwork for empirically investigating different
aspects of PLMs' use in hate speech detection. We deep dive into comparing
different pretrained models, evaluating their seed robustness, finetuning
settings, and the impact of pretraining data collection time. Our analysis
reveals early peaks for downstream tasks during pretraining, the limited
benefit of employing a more recent pretraining corpus, and the significance of
specific layers during finetuning. We further call into question the use of
domain-specific models and highlight the need for dynamic datasets for
benchmarking hate speech detection. | Probing Critical Learning Dynamics of PLMs for Hate Speech Detection: Despite the widespread adoption, there is a lack of research into how various
critical aspects of pretrained language models (PLMs) affect their performance
in hate speech detection. Through five research questions, our findings and
recommendations lay the groundwork for empirically investigating different
aspects of PLMs' use in hate speech detection. We deep dive into comparing
different pretrained models, evaluating their seed robustness, finetuning
settings, and the impact of pretraining data collection time. Our analysis
reveals early peaks for downstream tasks during pretraining, the limited
benefit of employing a more recent pretraining corpus, and the significance of
specific layers during finetuning. We further call into question the use of
domain-specific models and highlight the need for dynamic datasets for
benchmarking hate speech detection. | [
0.0309681352,
0.0456217602,
-0.0290762279,
0.0326431617,
0.0045931269,
0.0431915633,
0.0221292302,
0.0110955238,
-0.0406320803,
-0.0604299419,
-0.0292728562,
-0.0339656211,
-0.007891003,
0.0228339024,
-0.022991294,
-0.0299668834,
0.0308974814,
0.0205571726,
0.037770696,
-0.0152569348,
0.065914683,
0.0089576207,
0.003675079,
0.022271011,
-0.072256811,
0.0272975899,
0.0367683396,
-0.0015922256,
0.0321276709,
-0.0092331981,
0.0538699478,
0.0527394116,
0.0365630984,
0.0462761447,
0.000001856,
-0.0354964435,
0.0206182972,
-0.0136539144,
-0.0141528314,
-0.0131052341,
0.0713766217,
-0.023438096,
-0.0249735918,
0.0359182619,
-0.0416699462,
0.0132679585,
0.052074708,
0.0436199158,
0.0373502783,
0.0471237525,
-0.0059637162,
-0.0116291391,
0.0156558249,
0.0036792031,
-0.0406807512,
0.0186596066,
0.0420083664,
-0.0697177574,
-0.063128598,
-0.0142036304,
-0.0207324922,
-0.0107317502,
0.0262450706,
0.033233393,
-0.002849492,
-0.0161408968,
-0.0177036729,
-0.0396041982,
-0.0209583603,
0.0483643077,
-0.0639564469,
0.0029328035,
0.000011505,
-0.0138098178,
-0.0391226858,
-0.0501971021,
-0.0022442581,
-0.0022154937,
0.0009500739,
0.0194274448,
-0.0173902363,
0.0253672525,
0.0590341501,
0.0749751404,
0.0142852366,
0.0258435626,
-0.0077128238,
-0.0297778249,
0.0289856251,
-0.0019290479,
0.0124725588,
0.0148507403,
0.0151608326,
-0.0079413895,
0.0091895387,
0.0164734293,
-0.0706982762,
-0.0400052406,
0.0036090105,
-0.0605702959,
0.0738896355,
0.0385155641,
-0.0397728458,
-0.0042037242,
-0.0102921482,
0.0276577603,
0.0030269974,
0.0355150998,
0.0010773125,
0.0133182723,
-0.0075956015,
0.0143947722,
-0.0077420007,
-0.016326787,
0.0052702897,
0.0300168339,
-0.0340885036,
-0.0220517162,
-0.0226078071,
-0.0493878312,
-0.0413629822,
-0.0050155153,
-0.0187563319,
-0.0085811652,
-0.0218798984,
-0.0030154085,
-0.0274188109,
0.0292453747,
-0.0332613476,
-0.0991990492,
-0.0107223531,
0.0365346782,
-0.0367666632,
-0.0213736072,
0.0344975926,
0.0358254351,
0.0201781318,
-0.0436227359,
-0.0514736511,
-0.0181786455,
-0.0319387205,
-0.0691633448,
0.0157108232,
-0.0433731116,
0.0088508194,
-0.0174337439,
-0.0196728744,
-0.0527210236,
-0.001690397,
0.0092500737,
-0.052568417,
0.0817332417,
-0.0036190951,
-0.0142740244,
0.0262207482,
0.0221420694,
0.0141143296,
0.0764218792,
0.022723373,
0.0183169357,
0.0465098694,
-0.0290133748,
-0.0002025578,
0.0779156163,
0.0543677658,
0.0188843291,
-0.0510054193,
0.0065436307,
-0.0471087135,
-0.0455510281,
0.0246778894,
0.0135450326,
-0.0040357667,
-0.0587685406,
0.0688715279,
0.102089636,
-0.0354338512,
-0.0029948701,
0.056467697,
0.0099952864,
-0.0088926824,
-0.0161844473,
-0.0653431937,
-0.0632242039,
-0.032484021,
-0.0360304676,
-0.0070917304,
0.0825656876,
-0.0223029815,
-0.0422598422,
-0.0262783486,
-0.024245115,
0.0225633122,
-0.0273559019,
0.0427519083,
0.0246327594,
-0.0146076148,
0.0220912788,
0.0461512022,
-0.0163519513,
0.0120104216,
-0.0361013561,
-0.0080261817,
0.0638291538,
-0.0043181642,
-0.0031161727,
0.014896851,
-0.0560461059,
0.0358491689,
0.0543643981,
0.0173636079,
0.0072580422,
0.0371174142,
-0.0489622355,
-0.0247688517,
0.024512114,
0.0339824371,
0.0141068147,
0.077384986,
0.0114436299,
0.0218265001,
0.0049400679,
-0.0526527017,
0.0336307436,
-0.0030673726,
-0.0213257689,
-0.0214067604,
-0.0061153667,
0.0413531326,
0.0150959902,
-0.0565609708,
-0.0503714532,
0.0212403219,
-0.0021014006,
-0.006963741,
0.0181331038,
-0.0585275814,
0.0192650296,
0.0170624703,
-0.0344711207,
-0.0105949612,
0.0602211244,
-0.0010897269,
0.0152374478,
0.0519026294,
0.0032238758,
0.0029255075,
-0.0389818214,
-0.0251678117,
-0.071835123,
0.0531023666,
-0.0108173247,
0.0369203985,
-0.0049059442,
-0.0251607243,
0.0298486073,
-0.0325554535,
0.0310395397,
0.0218475703,
-0.0634384081,
-0.0399182998,
-0.0273012202,
0.024849046,
-0.024458956,
0.0068752286,
-0.0155577753,
0.027728986,
-0.0689168125,
-0.0117098577,
-0.045288451,
-0.0388781652,
-0.003602932,
-0.0264430288,
-0.0387466624,
-0.0060154181,
0.0136762634,
-0.0266692862,
0.06394016,
0.0264978893,
-0.0175509341,
-0.0596236475,
-0.0081049101,
-0.0297140069,
-0.0139222704,
0.0219569746,
0.003202816,
0.000219311,
-0.0502540395,
0.0500984006,
-0.0104012853,
-0.0151603976,
0.0309412237,
-0.0208151098,
-0.0620025322,
0.0336295478,
0.0362265073,
-0.0076211561,
0.0362334773,
0.065216206,
0.1063638255,
0.0093531469,
-0.001854329,
-0.0266751889,
-0.0024275039,
-0.0395505652,
0.010382819,
-0.0550277308,
-0.0203265175,
-0.0044889897,
0.116463311,
0.0340913571,
0.0070051723,
-0.0496274307,
0.0350397192,
-0.0586609356,
-0.0364076234,
-0.0054555163,
-0.0660687536,
-0.0365157463,
-0.0064805266,
-0.0526701249,
-0.0987000987,
0.0519195944,
-0.0199949816,
0.020596467,
-0.011044452,
-0.0336491168,
0.0093211671,
0.0362154879,
0.0168764722,
0.0133647379,
-0.013955215,
0.0163374897,
0.0120345708,
-0.0039641443,
0.0108168311,
-0.0345137827,
-0.0174471587,
-0.0222973544,
0.0272433423,
0.0009071357,
-0.0278653521,
-0.0689020157,
-0.0707434043,
-0.0581080392,
-0.0430489145,
0.0336251929,
0.0009906924,
0.0037719759,
0.0195733756,
-0.0082633728,
-0.0025315334,
0.000766195,
0.0182615947,
0.0313753001,
-0.030067021,
-0.0009172899,
0.0026912873,
0.0231676009,
-0.0178983025,
-0.0212519895,
0.0245941691,
0.0100691896,
0.0015142306,
-0.02675361,
0.0248049442,
0.0235212483,
0.029873807,
0.0133094117,
-0.0125004305,
0.0169277117,
-0.0229168478,
0.0026724567,
0.093533583,
-0.0006448066,
-0.0125611378,
-0.0050250161,
0.0285556093,
0.0135451118,
0.0172820557,
-0.0340431817,
-0.0814561173,
-0.0157324281,
-0.0148315169,
0.0427150801,
-0.0625443161,
-0.0126267998,
-0.0200638492,
-0.067518115,
0.0061038369,
-0.0202729162,
-0.0500386134,
0.0090011004,
0.0066855797,
0.0271106213,
0.0016086943,
0.0353462137,
-0.0407464243,
0.0234931447,
-0.0075552599,
0.0314344727,
-0.0656219423,
-0.0617011376,
0.0293591619,
-0.0348753482,
0.0428891405,
0.075463213,
0.0066506607,
-0.0154482983,
-0.0178052913,
0.0571746901,
-0.0017133037,
0.0026781042,
-0.0655714869,
0.0556783564,
-0.0095762396,
-0.0466413684,
0.0022549706,
0.0418135747,
-0.015747996,
0.0135963736,
0.0125984075,
0.0198891405,
0.0275080763,
-0.0149039635,
-0.0642302707,
0.042076949,
-0.0161725283,
0.074349165,
0.0334216692,
-0.0622501485,
0.0244483426,
-0.0420806855,
0.0031563577,
-0.0565980747,
-0.0987394676,
-0.0072298036,
-0.0403439626,
0.0241050683,
0.0016108641,
-0.0263830405,
0.0239473712,
0.0275708009,
0.0297596268,
-0.009239018,
-0.0128504122,
-0.0084004859,
-0.0758760497,
0.0584173016,
0.0054363273,
0.0461363159,
-0.046172671,
-0.0275358353,
0.0076717576,
-0.0169535782,
-0.0254066009,
0.0034536228,
-0.0726636574,
-0.0516235605,
0.096178554,
0.0619066767,
0.0222822875,
0.0217244402,
-0.0018713915,
-0.0591429137,
0.0948400199,
0.0327616595,
-0.0158808772,
0.0123463944,
0.0580293424,
-0.0613186248,
-0.013982892,
0.0107352221,
0.0171085279,
0.0774608478,
0.0648120567,
-0.0591476597,
-0.0332870372,
-0.0168725997,
-0.0325162709,
-0.0256762803,
0.0822904557,
-0.0067529269,
-0.064764984,
-0.0322775207,
-0.0059209042,
0.0205545053,
-0.0329621024,
-0.0363677926,
-0.0295102503,
0.0064861709,
0.0042063198,
-0.0456473306,
-0.0132072791,
-0.0454911888,
0.0291912444,
-0.0195791665,
0.0330040865,
0.0396188796,
0.0238135979,
0.0576549657,
0.0380471349,
-0.0142242787,
0.0758373663,
-0.0136881769,
-0.0402234606,
0.0114382077,
0.0110160196,
0.0067215641,
0.0277668908,
0.039890632,
-0.0187168829,
0.0236033071,
-0.0402093604,
-0.0721188709,
-0.0328234732,
-0.0171871744,
-0.0532497391,
0.0272532441,
0.007916593,
-0.0094847903,
-0.0185349584,
-0.013026177,
-0.0159890819,
-0.0252832156,
0.0283400286,
0.0227209888,
-0.0438513868,
0.0262715593,
0.0500538126,
0.0262153745,
0.0498507358,
0.0496401079,
-0.0221509933,
-0.0152175371,
0.0072718649,
0.0437673591,
0.0250436049,
0.003715226,
0.02530591,
-0.0515701808,
-0.0350294113,
-0.0021074468,
-0.0377412215,
0.0319560356,
0.0063496591,
0.0379711054,
0.0652468577,
-0.0333097503,
0.0493658744,
0.0284265801,
0.0015037538,
0.0126231685,
-0.0526484139,
0.017562123,
-5.980615139e-33,
-0.0208881833,
-0.0101467632,
-0.0194282234,
-0.0158518478,
-0.0149358707,
0.010886156,
0.0341861211,
0.0258071739,
-0.0152108194,
0.0167529322,
0.0074888123,
0.033249367,
0.0183001142,
0.0227159727,
-0.0225914717,
-0.0271216333,
0.0478606038,
-0.051132299,
-0.005087466,
-0.0240828302,
0.0258544106,
0.0363965034,
0.0383463353,
-0.0223133285,
0.0329553634,
-0.0046674958,
-0.0132439267,
0.0092066256,
0.0026988592,
0.0698375851,
-0.0221392903,
0.0816851333,
0.0382037573,
-0.0134556787,
-0.0290208943,
0.0430348404,
-0.0487645417,
-0.0003250771,
0.0286164396,
0.0332860872,
0.0210921839,
0.0242325552,
0.0311008804,
-0.0039876811,
-0.0598422252,
-0.01600888,
0.0225032736,
-0.0558766425,
0.0102443211,
-0.0426810347,
-0.0699328259,
0.033587914,
-0.0052250479,
-0.0255632047,
0.0083415452,
-0.0029863224,
0.0245286319,
0.0241322629,
-0.0522332713,
0.0148709211,
0.0269017927,
-0.0094794501,
0.0183263626,
0.0479582995,
0.0163601749,
0.0845222995,
0.0163772423,
0.0299329888,
-0.0007310262,
0.0364536159,
0.0743702129,
0.029600922,
0.0484224334,
0.0465468913,
0.0453463532,
-0.060491398,
-0.0368412882,
-0.0097299032,
0.0260148589,
-0.0290935505,
0.030272115,
-0.040186774,
-0.0186706427,
-0.0461795516,
-0.0460817628,
-0.0388048328,
-0.0073567638,
-0.023582235,
0.0357309915,
0.0111728618,
-0.0136161139,
-0.0104499254,
0.0340679698,
-0.0225366,
0.003678805,
0.0259822812,
-0.0295151025,
-0.0132090729,
0.0209332965,
0.0380539,
-0.0853685886,
-0.0009357758,
0.0070794863,
-0.0259985253,
0.0186617617,
-0.0165866204,
0.0598903224,
0.0261676647,
-0.0552365184,
0.0198604725,
0.0017666188,
0.0656085834,
0.0135072768,
0.0685036257,
0.0328460336,
0.0522216931,
-0.0215512682,
-0.0163266379,
0.0139675867,
-0.0437040143,
-0.0196897443,
0.0268218815,
-0.0255654082,
0.0369029231,
0.0117064081,
-0.0191992205,
-0.0022572477,
0.007504453,
0.0301174689,
-0.098600328,
0.0192036983,
-0.0333681256,
2.442e-7,
0.0011965586,
0.0356308855,
0.0180600099,
-0.0245172326,
0.0145531213,
0.0385451801,
0.0034171669,
0.0070357001,
-0.0126266694,
-0.001404535,
-0.0298262537,
0.0727115571,
0.0425781161,
0.057541348,
-0.0010763077,
-0.0397359021,
-0.0547925755,
0.0282879174,
-0.069526352,
-0.0144661162,
-0.0263934564,
0.0701922849,
0.0551974811,
-0.0507482775,
-0.0257347655,
-0.0363505334,
0.0229996741,
-0.0593993068,
0.0156741254,
-0.0162386075,
0.0527946576,
0.0407010317,
-0.0188022889,
0.0146064479,
-0.0113361357,
0.0465323292,
-0.0073010814,
0.0333351567,
0.030825831,
-0.0009687942,
0.0018155698,
-0.0071918508,
-0.0247661788,
-0.0340446495,
0.05047708,
-0.0406931452,
-0.0038417862,
-0.063079007,
-0.0112877125,
-0.0167138912,
0.0382103696,
-0.0210432298,
0.0064243693,
-0.0054550553,
0.0431984663,
-0.0144906044,
-0.0235034581,
-0.0970478728,
0.0010130445,
-0.0027929309,
-0.0345173664,
-0.0463999696,
-0.0534854122,
-0.01475811,
0.0285030585,
0.0391915105,
0.0410009623,
1.771914241e-34,
0.0301322397,
-0.0524904877,
0.0053635207,
-0.0347166471,
0.0167496298,
0.0210327525,
0.0132436529,
-0.0133401761,
0.0279527456,
-0.0249221288,
0.0111508034
] |
http://arxiv.org/abs/2402.02135v1 | http://arxiv.org/pdf/2402.02135v1 | Do Moral Judgment and Reasoning Capability of LLMs Change with Language? A Study using the Multilingual Defining Issues Test | 2024-02-03T12:52:36 | 2024-02-03T12:52:36 | Aditi Khandelwal, Utkarsh Agarwal, Kumar Tanmay, Monojit Choudhury | cs.CL | cs.CL, cs.AI | This paper explores the moral judgment and moral reasoning abilities
exhibited by Large Language Models (LLMs) across languages through the Defining
Issues Test. It is a well known fact that moral judgment depends on the
language in which the question is asked. We extend the work of beyond English,
to 5 new languages (Chinese, Hindi, Russian, Spanish and Swahili), and probe
three LLMs -- ChatGPT, GPT-4 and Llama2Chat-70B -- that shows substantial
multilingual text processing and generation abilities. Our study shows that the
moral reasoning ability for all models, as indicated by the post-conventional
score, is substantially inferior for Hindi and Swahili, compared to Spanish,
Russian, Chinese and English, while there is no clear trend for the performance
of the latter four languages. The moral judgments too vary considerably by the
language. | Do Moral Judgment and Reasoning Capability of LLMs Change with Language? A Study using the Multilingual Defining Issues Test: This paper explores the moral judgment and moral reasoning abilities
exhibited by Large Language Models (LLMs) across languages through the Defining
Issues Test. It is a well known fact that moral judgment depends on the
language in which the question is asked. We extend the work of beyond English,
to 5 new languages (Chinese, Hindi, Russian, Spanish and Swahili), and probe
three LLMs -- ChatGPT, GPT-4 and Llama2Chat-70B -- that shows substantial
multilingual text processing and generation abilities. Our study shows that the
moral reasoning ability for all models, as indicated by the post-conventional
score, is substantially inferior for Hindi and Swahili, compared to Spanish,
Russian, Chinese and English, while there is no clear trend for the performance
of the latter four languages. The moral judgments too vary considerably by the
language. | [
0.0390663445,
0.0502869077,
-0.0347866789,
0.0418458208,
-0.0652883202,
0.03475862,
0.0379679613,
0.0181533862,
0.0218266975,
-0.0104947248,
-0.044930283,
-0.0780307353,
0.012526162,
-0.0466412269,
0.0221366808,
0.0100231394,
0.0343855657,
-0.0106531549,
0.0907242224,
-0.0095913727,
-0.0163647458,
0.0575328395,
0.0511826016,
-0.0169504657,
0.025456883,
0.0004803173,
0.0012752939,
-0.0465629324,
0.0193768963,
-0.0328465961,
-0.0137305083,
0.0464865156,
0.0105265817,
-0.0033356787,
0.0000017407,
-0.0549482219,
0.0204127375,
-0.024779126,
0.0458799675,
0.0354995318,
0.0797640905,
0.0102819642,
0.0034100516,
0.0051269052,
-0.0436111614,
0.0707201436,
-0.0098399436,
0.0813215449,
-0.0262491796,
0.0475455038,
0.0012139453,
-0.0413852371,
0.0110962009,
0.0040946039,
0.0086717261,
0.0077237799,
0.0182877481,
0.08808887,
-0.0077435649,
0.01139603,
-0.0099618388,
-0.0079501793,
0.0229863301,
-0.0100163976,
-0.0937445536,
-0.0452337787,
0.0097533679,
-0.0398149863,
0.010182444,
0.0198156238,
0.0407877341,
-0.0094767939,
0.0047023674,
0.0321680978,
0.0142066982,
-0.0125695495,
0.0071107331,
0.0795366466,
-0.0143336616,
0.0189751219,
0.071270816,
-0.0021536385,
-0.0167554393,
0.0171699673,
0.031565059,
0.0181687064,
-0.0313393064,
-0.0049628117,
-0.0022575443,
-0.0145725207,
-0.0794624314,
-0.02144517,
-0.0266426094,
0.0192706902,
0.0011355851,
0.0227859709,
0.0261655822,
0.051487457,
0.0458743572,
-0.0832500234,
-0.0031449422,
0.0379403941,
-0.0111491168,
0.0074926084,
-0.0055597513,
0.0567608289,
0.0064707878,
0.0433019064,
-0.0273210034,
0.0275211576,
0.0279027391,
-0.0304417517,
0.0611897856,
0.03129692,
0.0027144821,
0.0468388163,
-0.0848283097,
-0.0172389038,
-0.0362135507,
-0.0327785462,
-0.0561115704,
-0.0456678011,
0.0018121433,
-0.0205982551,
-0.039460171,
0.0214267485,
-0.0577490367,
0.0090391617,
-0.0202957056,
-0.0410073288,
0.0666291639,
-0.0277941991,
-0.0439756066,
0.0046936572,
0.0243482869,
0.1090321392,
-0.047436703,
-0.0060679838,
-0.0242933203,
0.0028192757,
-0.0156286564,
-0.0899098739,
-0.0115928333,
-0.0328935869,
0.0128182005,
0.0183543116,
-0.0096731428,
-0.0477937609,
0.0447921082,
0.0269675981,
-0.0108754635,
0.0079455432,
0.0705377087,
0.0273869522,
0.0027318024,
0.0334701426,
0.0148358895,
0.0576728955,
0.0033705358,
-0.0041465587,
0.0219056495,
-0.0029707088,
-0.0363545269,
0.0041695633,
-0.0295388754,
0.0035331906,
-0.0473596081,
0.0078757992,
0.0137544004,
-0.0269171968,
0.024369996,
0.0341409892,
-0.0398793556,
0.0298330951,
0.0176551081,
0.1132765859,
0.0149405608,
0.0413686857,
-0.0057375897,
-0.0072959051,
0.0243810192,
0.0412987545,
-0.0505046621,
0.0307848901,
0.0275887288,
-0.0214301907,
-0.0425426997,
0.0532138385,
-0.0147209242,
-0.0356355011,
-0.009027048,
-0.0075303428,
0.0122588659,
-0.0008479439,
-0.0306475256,
-0.0195884816,
0.0059405211,
-0.0663513914,
0.0696373284,
-0.0750804842,
-0.0187816806,
-0.0325323455,
0.0952041671,
0.040263243,
-0.0037809745,
0.0411473438,
0.0090303551,
-0.0094680944,
-0.0147534823,
0.0384415537,
0.0261322577,
-0.0104653519,
-0.0241163857,
-0.0568393916,
0.0120198149,
-0.0341548398,
-0.0325502679,
0.0460602194,
-0.0381985083,
0.0484516136,
0.0357841738,
-0.009297803,
0.0042421934,
-0.006863697,
-0.0234342031,
0.0285757389,
-0.062136054,
0.0133542996,
0.014035847,
-0.0083960071,
-0.0079231132,
-0.0309211668,
-0.0267335158,
0.0005481152,
0.0021766645,
0.0478623323,
-0.0551929139,
0.0096068494,
0.0187958423,
-0.076130189,
-0.0085802246,
-0.0020054895,
0.0209989101,
0.0311524812,
-0.0209731515,
-0.0005832266,
0.0042051268,
-0.0477263406,
0.0286068451,
-0.0850606933,
-0.0138952946,
-0.0039071077,
0.0722942948,
-0.019373212,
0.0388700366,
-0.0114897145,
-0.0023609046,
0.057760708,
0.0089277737,
-0.0322701521,
0.0421858467,
-0.0131173879,
0.0087251691,
0.0258813221,
0.0670251623,
-0.0218456574,
-0.0387503952,
-0.0560349822,
0.0080182487,
0.0239182152,
-0.0098242639,
-0.006548496,
-0.0258738082,
-0.0238796398,
0.0488983393,
-0.0510598682,
-0.0003956495,
0.0512913093,
0.0272479206,
-0.0044521671,
-0.0416313633,
0.0108073577,
0.0070259268,
0.002740877,
-0.0222873371,
-0.0127029764,
0.0364848897,
-0.0275292434,
-0.0136188101,
0.059518747,
-0.0952738151,
-0.0190412812,
-0.0126624247,
-0.0819355547,
0.0054293945,
0.0429373197,
0.0124870893,
0.0595197603,
0.0441319682,
-0.0016608314,
0.0146738887,
0.0288293976,
-0.0250423979,
-0.0286382437,
0.0177416503,
0.0082678804,
-0.0226924121,
-0.0851032659,
-0.0496840738,
0.0737193152,
0.0097985147,
0.0219913535,
-0.0079053948,
0.0129359532,
-0.0784067586,
-0.048519019,
0.0093921935,
-0.0511434898,
-0.0004262177,
0.0659107417,
-0.026585374,
-0.0474699028,
0.0420229994,
0.0406684764,
0.0177222062,
-0.0073252227,
0.035179425,
-0.0267078802,
0.0458167642,
-0.0109329652,
0.0273483023,
-0.0340387411,
-0.0361820646,
-0.0301652122,
-0.0261847321,
0.0325638428,
-0.0295296721,
0.0476296619,
-0.0132351192,
-0.0011213636,
-0.0231040642,
-0.0550382733,
-0.0174227599,
-0.0069822441,
-0.0114679681,
-0.0312172063,
0.003837568,
0.0198400822,
0.0585636683,
-0.005448062,
0.0351283513,
-0.0629406273,
-0.0356736444,
0.0353811644,
0.0045323553,
-0.0682764724,
-0.0024801253,
-0.0101289181,
0.0678988844,
-0.0498101115,
0.0696436688,
0.0354598127,
-0.032157436,
-0.0238532536,
0.0778260306,
0.0509980693,
0.0493271947,
0.0526591949,
-0.0273650251,
0.0272092894,
0.0558016784,
0.0236497503,
0.0357923023,
0.0824335665,
0.0308009014,
-0.0204000138,
-0.0069987467,
-0.0632286146,
-0.0339117534,
0.031927377,
0.0097609153,
-0.0772782415,
0.017917579,
0.0117406296,
-0.0468221568,
-0.089090921,
-0.0369293541,
-0.0051738149,
-0.0617798679,
0.0231593605,
0.0409899466,
-0.0204196498,
-0.021960618,
0.0158699863,
-0.0271915365,
0.0303434618,
0.0657955781,
0.0354831032,
-0.0037891536,
-0.0372079201,
0.0021135786,
0.0258461032,
0.0185539518,
0.0655937567,
0.0130067505,
0.0259543415,
0.0235680006,
0.0411134847,
-0.0352556184,
-0.0400816314,
0.0207442865,
0.0507672578,
-0.0168229714,
0.0095028775,
-0.0331697837,
-0.0053966478,
-0.0089601027,
-0.0787570626,
-0.0006021716,
0.0372670516,
-0.0205025747,
0.0120293843,
0.0488498956,
0.007755958,
-0.0523044318,
0.0289470553,
0.0172631964,
-0.0431282185,
0.0338280089,
0.0072900583,
-0.0305259712,
0.0111334827,
-0.0251079984,
-0.0685988292,
-0.0155538227,
-0.079139933,
0.0131998565,
-0.0239152089,
-0.0405816548,
-0.0368252508,
-0.0734635368,
-0.0340943411,
-0.0314680748,
0.0523861609,
-0.0411335044,
0.0209644306,
-0.0122115742,
-0.0337521248,
-0.0163118616,
-0.0092771966,
-0.0108031267,
-0.0290061273,
0.0707038492,
0.0083900448,
-0.0422314666,
0.0049185441,
-0.0374156162,
-0.0007194958,
-0.0165897366,
0.0244848318,
0.0093299299,
0.0079182675,
-0.0453711562,
0.0158924405,
-0.0015999867,
-0.012774203,
-0.0014947023,
-0.0501507297,
0.0626329929,
0.0115423026,
0.0602515563,
0.0034487839,
0.0192415547,
0.0152380317,
0.0477096774,
0.0302123763,
-0.0304856468,
0.0116553344,
0.0008234457,
0.0394444317,
0.0472673513,
-0.0470687896,
0.0137730138,
0.0019678331,
-0.0679003075,
-0.042692177,
-0.048121199,
-0.043413721,
-0.0305598844,
-0.0236485433,
0.0315129198,
-0.0070064086,
0.0231419019,
-0.0008609956,
-0.0410048366,
-0.0235452875,
0.0101561435,
0.0354278237,
0.0082608061,
0.0486790501,
-0.0155680915,
0.0205150694,
-0.0275402665,
-0.0009997927,
-0.0296984743,
0.0401412956,
-0.0169498585,
0.1082575694,
-0.0190554634,
-0.0133525692,
0.02837562,
0.0055659087,
-0.0319452249,
-0.0415310413,
0.0079188012,
-0.0028866578,
-0.0122263217,
0.0032858336,
0.0158750731,
-0.0162889846,
0.0104353521,
-0.0400730446,
0.0332241133,
-0.0070650256,
-0.0501823723,
0.0076131104,
-0.0115665747,
-0.0260422621,
-0.0221626665,
0.0293358229,
-0.0615891069,
-0.0205319431,
0.0097955139,
-0.0078312494,
-0.0011157943,
-0.0463186838,
0.0209725797,
-0.0575021394,
0.0058039292,
0.0055950293,
-0.0730009228,
-0.0479603149,
-0.0537759177,
-0.0015729087,
-0.0143583231,
-0.0247847158,
-0.0396739654,
0.0849594325,
-0.0175122023,
0.0301723853,
-0.0411907174,
0.0173601061,
-0.0151980929,
0.0218404029,
-0.0156106921,
-5.186774783e-33,
-0.0171130393,
-0.0314105414,
0.0078177536,
-0.0566756129,
0.0064232778,
-0.0279108584,
0.0012489499,
0.0418851711,
-0.0516911075,
0.0345458761,
-0.0067207557,
0.0083983513,
0.0069219684,
-0.0432280153,
0.0112096602,
-0.024465872,
-0.0216023792,
0.014895238,
-0.0166539792,
-0.008077505,
-0.0032744464,
-0.0151382014,
0.075824514,
-0.0394393131,
0.0782065764,
-0.0019040457,
-0.0020861924,
-0.0331740007,
-0.0283694621,
0.0047545959,
-0.0217656028,
0.015179392,
0.0306004528,
-0.0010149291,
0.0012533003,
0.0210801177,
0.0122731533,
-0.0204848573,
0.0852249563,
-0.0159100872,
-0.0152950957,
-0.0678925589,
-0.0071898629,
-0.0148667637,
-0.0364590809,
0.0008271342,
0.0270666201,
-0.0074002682,
0.0260208361,
0.005511262,
-0.0662322044,
-0.0049105687,
-0.0284565724,
0.0169204921,
-0.0666798428,
0.0647683293,
-0.0378870741,
0.0294854678,
0.0209798347,
0.0446222462,
-0.0028134051,
-0.0001489271,
0.0143637341,
-0.0393950902,
0.0078336066,
0.0779636949,
0.0043152547,
0.0245645531,
-0.0077587371,
-0.0118964426,
0.0201701093,
0.0181559995,
0.0156732835,
0.0823590457,
-0.0294235572,
-0.0851091221,
-0.0480242781,
-0.0383973308,
0.0672448799,
-0.0044295765,
0.0156553835,
-0.0188089367,
-0.0245504249,
-0.0161179453,
-0.0085619502,
-0.0230658203,
-0.0213575531,
-0.0360357612,
0.0332072824,
-0.0483822227,
0.0117052346,
-0.0562010743,
0.0446393155,
-0.0093019092,
0.0166908894,
0.0089844055,
-0.0393039174,
0.0050303591,
0.0048341854,
0.0516398512,
-0.0282559395,
-0.0284206439,
-0.0374957249,
-0.0006914765,
0.0376480818,
-0.0065131867,
0.0352611281,
-0.0258737113,
-0.0235119928,
-0.0033832013,
-0.0233688317,
0.0107291527,
0.0419037677,
0.0077487542,
-0.0010886469,
0.0289103333,
0.0050686267,
-0.0377498232,
0.0344816968,
-0.0419228189,
-0.0549746938,
-0.0344179012,
-0.0290814415,
0.0564712919,
-0.0376322977,
0.0073668673,
-0.0339395925,
0.084203437,
0.0193380304,
-0.0516873263,
-0.0276477616,
0.0384364203,
2.456e-7,
-0.0093176868,
-0.0028909992,
-0.0190159529,
0.0135683063,
-0.0190572198,
0.0394467078,
0.0252951067,
0.0034337847,
0.0809052512,
-0.0094927307,
-0.0272815768,
0.0154675227,
0.0573186614,
0.0233628768,
-0.0078434171,
0.0272920169,
-0.0208202023,
-0.0482692756,
-0.0164455269,
0.0411415398,
0.0155011527,
0.0416563787,
-0.0147397164,
-0.00623303,
-0.0198002663,
0.011221027,
0.037638735,
-0.0502918027,
-0.0627131164,
0.0091803437,
-0.0238410644,
-0.0098822638,
0.023301065,
0.0505083352,
0.0322168097,
-0.0252564009,
0.0433284529,
0.0584952421,
0.0195228383,
0.0041710981,
0.0038735117,
-0.0254763141,
-0.0352548361,
-0.0270659737,
0.0945556015,
0.0121271657,
-0.0453839824,
0.0062323655,
0.0425945707,
-0.0275864284,
0.0241693575,
-0.0228824113,
-0.0166446436,
0.0369126573,
0.023870172,
-0.0191261023,
0.0458640382,
-0.0027683617,
-0.0043705604,
-0.0128925154,
0.0230108257,
-0.0852394924,
-0.0225304402,
0.0339309126,
0.0437812321,
-0.0508774035,
0.0042081173,
1.784433141e-34,
0.0318522304,
-0.0278109014,
0.0577481762,
0.0099724317,
0.0656979829,
0.0414264649,
-0.0111965667,
0.0136363097,
0.0069380398,
-0.0089393565,
-0.0027358057
] |
http://arxiv.org/abs/2402.02130v4 | http://arxiv.org/pdf/2402.02130v4 | GITA: Graph to Visual and Textual Integration for Vision-Language Graph Reasoning | 2024-02-03T12:19:47 | 2024-05-24T06:58:05 | Yanbin Wei, Shuai Fu, Weisen Jiang, Zejian Zhang, Zhixiong Zeng, Qi Wu, James T. Kwok, Yu Zhang | cs.CL | cs.CL | Large Language Models (LLMs) are increasingly used for various tasks with
graph structures. Though LLMs can process graph information in a textual
format, they overlook the rich vision modality, which is an intuitive way for
humans to comprehend structural information and conduct general graph
reasoning. The potential benefits and capabilities of representing graph
structures as visual images (i.e., $\textit{visual graph}$) are still
unexplored. To fill the gap, we innovatively propose an end-to-end framework,
called $\textbf{G}$raph to v$\textbf{I}$sual and $\textbf{T}$extual
Integr$\textbf{A}$tion (GITA), which firstly incorporates visual graphs into
general graph reasoning. Besides, we establish $\textbf{G}$raph-based
$\textbf{V}$ision-$\textbf{L}$anguage $\textbf{Q}$uestion $\textbf{A}$nswering
(GVLQA) dataset from existing graph data, which is the first vision-language
dataset for general graph reasoning purposes. Extensive experiments on the
GVLQA dataset and five real-world datasets show that GITA outperforms
mainstream LLMs in terms of general graph reasoning capabilities. Moreover, We
highlight the effectiveness of the layout augmentation on visual graphs and
pretraining on the GVLQA dataset. | GITA: Graph to Visual and Textual Integration for Vision-Language Graph Reasoning: Large Language Models (LLMs) are increasingly used for various tasks with
graph structures. Though LLMs can process graph information in a textual
format, they overlook the rich vision modality, which is an intuitive way for
humans to comprehend structural information and conduct general graph
reasoning. The potential benefits and capabilities of representing graph
structures as visual images (i.e., $\textit{visual graph}$) are still
unexplored. To fill the gap, we innovatively propose an end-to-end framework,
called $\textbf{G}$raph to v$\textbf{I}$sual and $\textbf{T}$extual
Integr$\textbf{A}$tion (GITA), which firstly incorporates visual graphs into
general graph reasoning. Besides, we establish $\textbf{G}$raph-based
$\textbf{V}$ision-$\textbf{L}$anguage $\textbf{Q}$uestion $\textbf{A}$nswering
(GVLQA) dataset from existing graph data, which is the first vision-language
dataset for general graph reasoning purposes. Extensive experiments on the
GVLQA dataset and five real-world datasets show that GITA outperforms
mainstream LLMs in terms of general graph reasoning capabilities. Moreover, We
highlight the effectiveness of the layout augmentation on visual graphs and
pretraining on the GVLQA dataset. | [
0.0170536917,
0.0310738925,
-0.0041769287,
0.0442799851,
-0.0439703502,
0.0184687413,
-0.0241100881,
0.0072575826,
0.0674598515,
-0.0175090786,
-0.0063891741,
-0.0273276195,
0.0255264435,
0.0496503264,
0.0411169082,
-0.0496889912,
0.0149989594,
-0.0425199158,
0.0101539502,
-0.0013080931,
0.0137021979,
0.0233269874,
0.0045460835,
-0.002657179,
-0.0337888561,
-0.0254495814,
-0.0241514798,
0.0273444187,
0.0111828959,
-0.0437299274,
-0.0908493921,
0.0070139379,
0.0649188608,
0.0278125964,
0.0000018556,
0.0098710898,
0.0090406071,
-0.003715124,
-0.0189960152,
-0.0213846099,
0.0147958742,
-0.0721579194,
0.0330735669,
-0.0174349416,
-0.0481584929,
-0.0375544839,
0.0768220574,
0.0249763262,
0.0646497384,
0.0502532981,
-0.0393387638,
-0.0497742556,
0.020943841,
-0.017908046,
0.01810853,
-0.0062101847,
0.021252051,
0.0587274134,
0.0339600369,
-0.0223329812,
-0.0738923922,
0.0408748463,
0.0333559848,
-0.0415349193,
0.0227420069,
0.0511301011,
-0.0290114824,
-0.0253482647,
-0.0073531456,
0.1152718067,
-0.001399711,
-0.0299566202,
-0.0029567287,
0.0021358489,
-0.0153126521,
0.0122399693,
-0.0455555357,
0.0016992601,
-0.0274629407,
0.01845048,
-0.0321820565,
-0.0539802536,
0.0366242565,
-0.0257488899,
-0.0082583278,
0.0630879179,
0.0136008961,
0.0171923935,
0.0199365821,
0.0220579393,
-0.0511104539,
-0.0389689393,
-0.0245416053,
0.0500420108,
-0.0053913295,
0.0164896175,
-0.0342576616,
-0.0128822168,
-0.0410920009,
-0.0084336698,
0.0339411497,
0.0247320123,
-0.0268847793,
0.0466954708,
-0.0327717401,
0.0776630342,
-0.0356243663,
0.1027994379,
-0.0339835659,
-0.0124817891,
-0.0336212628,
-0.0102333101,
-0.057845585,
0.0006092295,
-0.0179915819,
-0.0658983216,
-0.0312758535,
0.0399310887,
-0.0181595702,
0.0146093452,
0.0275482126,
0.0135083562,
0.0013895098,
-0.0330515802,
-0.000945704,
0.0165813435,
0.0224681217,
0.0127211614,
-0.0550474897,
-0.0138246035,
0.0075747273,
0.0446650274,
0.0318241082,
-0.0301599056,
-0.000055204,
0.086586535,
-0.0044849999,
-0.0104725668,
-0.0329768583,
0.001114846,
0.0048681404,
-0.0581032224,
0.0454853363,
-0.0795073286,
0.0348825641,
-0.0055950736,
-0.0083960239,
-0.030939294,
0.0132965976,
-0.0048751878,
-0.0498174094,
0.0826689452,
-0.0281598829,
0.0103044538,
-0.0136672147,
-0.0216459744,
0.0116060358,
0.0418513417,
0.0390240997,
0.0182084758,
-0.0035375028,
0.0042162044,
-0.0159922857,
0.0396090783,
0.0173941553,
-0.0101375738,
-0.0419380479,
0.0026227934,
-0.037064638,
-0.0007815473,
0.0073584523,
0.0765361935,
-0.0438498408,
-0.0442306288,
0.0345662348,
0.0287494361,
0.0240281634,
-0.0300240386,
0.015466759,
-0.0074132793,
0.0121752359,
0.010261395,
-0.0820549652,
0.051013194,
0.0276489314,
0.0365878604,
-0.0508194491,
-0.0198423732,
0.0014618245,
-0.0226490609,
-0.0185196735,
0.0113679525,
0.0558415204,
-0.0094593782,
0.065723516,
0.0285490155,
-0.0349460132,
-0.0320606269,
0.0100598466,
-0.0061982698,
0.0004454334,
0.0083709639,
0.0032235733,
0.0518796109,
0.0390042923,
0.0103911757,
-0.0060874135,
0.0149708046,
0.0054671015,
0.008232031,
0.0209357813,
-0.0103973532,
-0.0180989988,
-0.0513755158,
-0.0297124572,
-0.0003216009,
-0.0279898159,
0.0244852267,
0.0084292274,
-0.0034921884,
-0.0003145941,
-0.0872169659,
-0.0173608102,
0.017985953,
0.0611980334,
0.0356423557,
-0.0580218621,
0.0231425148,
-0.0180869699,
-0.0033444348,
-0.0058650891,
0.0191116426,
-0.0120870294,
-0.0175698902,
0.0125159752,
0.0255976953,
-0.0098619265,
-0.0559077859,
0.0578822345,
-0.035294354,
-0.0488702059,
-0.0066792504,
-0.0428905189,
-0.0351696499,
0.0230964758,
-0.0543294922,
0.0178044271,
-0.017010713,
-0.0200405605,
0.015664082,
-0.0073469859,
0.0031933431,
0.07624688,
0.0279760789,
-0.0323736258,
0.0000507985,
0.0330273658,
0.003586371,
-0.0474705994,
0.0058177658,
0.0429442003,
0.0188217722,
-0.0050465008,
-0.0385145359,
-0.0114441384,
-0.0543079525,
0.0536807328,
-0.0806902573,
-0.0248181559,
0.009017908,
0.0109023601,
-0.0066796085,
-0.0092205908,
0.0036584053,
-0.0343385376,
-0.0464635342,
0.0652530566,
0.0284503177,
0.0199191477,
0.0250621513,
-0.0221303683,
-0.0116736125,
-0.0131822787,
-0.0382496826,
-0.0454772934,
0.0778936297,
0.057893347,
-0.0507829078,
0.01326627,
0.0289223082,
-0.0355531685,
0.0340015739,
-0.0183219481,
-0.0180299859,
0.0187707525,
-0.0234029591,
0.0528241843,
0.0606684685,
-0.0500212498,
-0.0049762749,
0.0155652994,
-0.0160226654,
-0.0038791271,
0.0089004179,
0.0017986869,
-0.0034861783,
-0.0212411154,
-0.0088257762,
-0.0778774992,
-0.0058592698,
-0.0245449301,
0.0124012679,
-0.0543681607,
-0.0112573793,
-0.007676552,
0.0140593247,
-0.0262095854,
-0.0325142555,
0.0159887969,
0.0320370756,
-0.06196668,
-0.0395079106,
0.070805639,
-0.040581435,
-0.0311659966,
0.0564509705,
0.0164460503,
-0.0598228872,
0.0920139328,
0.0757188499,
0.0075067137,
-0.0188759286,
-0.0229564123,
-0.0667631105,
-0.01399926,
0.0189359821,
-0.0008083407,
0.0334896185,
-0.0436545201,
0.0547496453,
-0.032057073,
0.0102248974,
-0.0130043346,
0.0493986458,
-0.0582618937,
0.0120226787,
-0.0329413041,
-0.0375255942,
0.018713722,
0.039976202,
0.0033197033,
-0.0251530558,
-0.0249950029,
0.013984208,
-0.0264816172,
0.0491884872,
0.0443516262,
-0.003243817,
-0.0075853164,
-0.0351471752,
0.0027548987,
0.0616094396,
0.0019448551,
-0.0242999345,
-0.0450685658,
0.0497888699,
0.002228712,
0.0620687194,
-0.0255567916,
-0.0309437811,
0.0327797905,
-0.0482493564,
0.0034264158,
0.0342436358,
0.0036759756,
-0.0351862162,
-0.026313683,
0.022881683,
0.0650771558,
0.0460093655,
0.0374738537,
-0.1129917651,
-0.0237793401,
-0.0003326429,
-0.0260089692,
0.0424972586,
-0.0440813005,
-0.0392024592,
-0.0892201886,
0.0519887097,
0.0442404598,
-0.0011165614,
-0.0512684658,
-0.0059250868,
0.0106394691,
0.0092189442,
-0.0452236235,
-0.0189106613,
-0.0279170778,
0.0195212197,
0.0015211985,
-0.0053718253,
0.0434076525,
-0.0154265221,
-0.0081510851,
-0.0277902745,
-0.0278674979,
0.0168845095,
0.0199516639,
-0.0441850163,
0.0572820492,
0.0333259813,
-0.0096022896,
-0.0197140593,
0.0393820591,
0.0078033693,
-0.042502474,
-0.0202221628,
0.0107347658,
-0.0280365851,
-0.0471475385,
-0.0627118126,
0.086413309,
0.00278479,
-0.0091756713,
-0.006639306,
-0.0045386264,
-0.0468999818,
-0.0072837085,
0.0578534789,
0.0293455198,
-0.0838934705,
-0.0424760468,
0.0240996126,
-0.0562857129,
-0.0623280667,
0.0461634919,
-0.0035949443,
-0.0065566511,
-0.0165807679,
-0.0020847304,
0.0509771295,
0.0667351857,
-0.0064484021,
-0.0311345756,
0.023361383,
-0.0230615735,
-0.0645080358,
0.0300589949,
-0.0448531918,
0.0703083798,
-0.0103845773,
0.0088731293,
-0.0256851483,
-0.014636552,
-0.005002128,
0.015989121,
-0.0516560264,
-0.0609656647,
-0.0414351858,
0.0161233265,
0.0214531533,
0.0479913242,
-0.005300262,
-0.0038080763,
0.0257456005,
0.0158704035,
0.0312879719,
0.008643141,
0.0025056542,
-0.0420976877,
0.0041148942,
0.017962968,
-0.0220744777,
-0.0005010421,
0.0263103042,
0.0377078429,
-0.009934688,
0.0108437398,
0.0013745986,
0.0675820187,
0.0015028856,
0.0279468521,
0.0396277979,
-0.0504108258,
-0.0167063996,
-0.0236543547,
-0.0224260818,
-0.0452095978,
-0.0129728634,
0.0419691466,
0.0247501153,
0.0123031111,
-0.0478558876,
0.0247646291,
0.028355509,
0.008443227,
0.0005402305,
0.0310302936,
-0.0347699113,
0.05368267,
0.0197088588,
0.0146975666,
0.0054003722,
-0.0178270433,
0.0235425048,
0.0012959375,
0.0172684137,
0.0337772518,
-0.0601647384,
-0.038029708,
0.0019762074,
-0.0339453518,
-0.068792738,
0.0205244049,
0.0242718328,
0.049368795,
0.0251594633,
0.0240514614,
-0.0090595977,
0.0026326303,
0.0260996353,
0.0569965243,
0.0025007743,
-0.0620584302,
0.0843053609,
0.0376339629,
0.0043825503,
-0.0163331106,
0.0290949866,
0.0565073118,
0.0010073702,
-0.0063027041,
-0.036738582,
-0.0024478142,
-0.0471289791,
-0.0085108569,
0.0055596111,
0.0065763849,
0.0247479845,
-0.0471473113,
-0.1197076961,
-0.0087184422,
0.0342300534,
-0.0145973563,
0.042824097,
0.0218113717,
-0.0002442208,
0.063741602,
-0.0332867876,
-0.0030809348,
0.0312947147,
-0.0281785615,
-0.0033574197,
0.023242671,
-5.291781326e-33,
-0.027226733,
-0.0679321289,
0.036530707,
0.0204899572,
-0.0659202933,
0.0143644568,
-0.0482417122,
-0.0003180755,
-0.0409876332,
0.034037713,
0.0079727527,
0.023710059,
-0.0002951621,
0.0268986914,
-0.0395532735,
0.0038815232,
0.0055695018,
0.0089537175,
-0.0019590848,
0.0166975558,
0.0064692171,
0.0048429687,
0.047816027,
-0.0574025586,
-0.0156704541,
0.0111957667,
0.0080880728,
0.0489548519,
-0.0053248978,
0.0443116389,
-0.0259105004,
0.0041327495,
0.0289123282,
0.0276636537,
-0.0139865391,
0.0496011935,
-0.0787811205,
-0.0311400462,
0.0590524375,
0.0581798479,
-0.0857023373,
0.0087468522,
0.0406951383,
-0.0380056724,
0.0379714631,
-0.0586985834,
0.0409260094,
-0.0211812146,
-0.0344746634,
0.0639846101,
-0.0276401471,
0.0129002798,
-0.0048336601,
0.0103476057,
0.0627694353,
0.0047196024,
0.020725783,
0.0290265642,
-0.0388637856,
0.0321252234,
-0.0610708073,
-0.0325013958,
0.0602194108,
0.044226639,
0.0238736905,
0.0338013507,
0.0240198616,
-0.0419729799,
-0.0431443229,
0.0138408206,
-0.0064267437,
0.0892585814,
0.0081411907,
0.0023010254,
-0.0374571756,
0.0028922642,
-0.0615600571,
-0.0219703354,
-0.0633037984,
0.0776324421,
-0.0319355316,
0.0312972106,
-0.0111320224,
-0.0499295332,
0.0123520549,
-0.0347413905,
-0.0477694608,
-0.0004541361,
0.0185478888,
-0.0127104986,
-0.0046473681,
0.053332407,
0.0207735468,
-0.045481015,
-0.0126725929,
0.0852849483,
-0.0279844869,
0.0691378042,
-0.0127970176,
0.0496581644,
-0.03018458,
-0.0294640101,
0.0274452586,
0.0057885228,
0.0044916086,
-0.0048329835,
0.0722330064,
0.0193717014,
-0.0288181081,
0.0061135786,
-0.066922456,
-0.037833374,
-0.005579568,
-0.0179457404,
0.0518418215,
0.0403593108,
-0.0283171758,
0.0319089107,
0.0123247895,
0.0051856129,
-0.0312112812,
0.0213477537,
-0.0128203314,
0.036754746,
-0.0095258905,
-0.0207671169,
-0.0278680846,
0.0097849313,
0.0311585926,
-0.0616836622,
-0.0232311655,
0.0152051076,
2.581e-7,
0.0515176877,
0.0233472139,
0.0614845604,
0.030811999,
0.0262442231,
0.0265578311,
-0.0446297079,
-0.005209066,
-0.003008446,
-0.0191277675,
0.030084081,
0.070823893,
-0.0155735416,
0.0208600983,
0.0060996995,
0.0624913387,
-0.0505763851,
-0.0457795598,
-0.0540845767,
0.0166513901,
0.0761373937,
0.0229228325,
0.0105754239,
-0.0284524653,
-0.0093123997,
-0.0785251632,
0.0593667179,
0.0035386553,
0.0121732457,
-0.0278623588,
-0.0304061547,
0.019080719,
-0.029342752,
0.0300960019,
0.0177044068,
-0.000297646,
0.0315504409,
-0.0052893213,
-0.0314038582,
0.0275844578,
-0.0147040905,
-0.0161988754,
-0.0137929423,
-0.0518992506,
0.0761380717,
-0.047644794,
-0.0455746911,
0.0039309282,
-0.0470677391,
-0.0069502168,
-0.0178743228,
0.0407165736,
-0.0033337926,
0.0241017118,
0.0441778414,
-0.0398058966,
-0.016634997,
-0.0436783917,
-0.0159280449,
0.0019860447,
-0.0164483488,
-0.0133025367,
-0.0505485311,
0.0440958105,
0.0161960963,
0.0320595205,
-0.0262223016,
2.039662237e-34,
0.0292599946,
0.008258909,
0.01779221,
0.0013208422,
0.0392122343,
-0.0364828669,
0.0068292879,
0.0191942453,
0.0081961202,
-0.1041310206,
-0.0189297684
] |
http://arxiv.org/abs/2402.02113v1 | http://arxiv.org/pdf/2402.02113v1 | Zero-shot Sentiment Analysis in Low-Resource Languages Using a Multilingual Sentiment Lexicon | 2024-02-03T10:41:05 | 2024-02-03T10:41:05 | Fajri Koto, Tilman Beck, Zeerak Talat, Iryna Gurevych, Timothy Baldwin | cs.CL | cs.CL | Improving multilingual language models capabilities in low-resource languages
is generally difficult due to the scarcity of large-scale data in those
languages. In this paper, we relax the reliance on texts in low-resource
languages by using multilingual lexicons in pretraining to enhance multilingual
capabilities. Specifically, we focus on zero-shot sentiment analysis tasks
across 34 languages, including 6 high/medium-resource languages, 25
low-resource languages, and 3 code-switching datasets. We demonstrate that
pretraining using multilingual lexicons, without using any sentence-level
sentiment data, achieves superior zero-shot performance compared to models
fine-tuned on English sentiment datasets, and large language models like
GPT--3.5, BLOOMZ, and XGLM. These findings are observable for unseen
low-resource languages to code-mixed scenarios involving high-resource
languages. | Zero-shot Sentiment Analysis in Low-Resource Languages Using a Multilingual Sentiment Lexicon: Improving multilingual language models capabilities in low-resource languages
is generally difficult due to the scarcity of large-scale data in those
languages. In this paper, we relax the reliance on texts in low-resource
languages by using multilingual lexicons in pretraining to enhance multilingual
capabilities. Specifically, we focus on zero-shot sentiment analysis tasks
across 34 languages, including 6 high/medium-resource languages, 25
low-resource languages, and 3 code-switching datasets. We demonstrate that
pretraining using multilingual lexicons, without using any sentence-level
sentiment data, achieves superior zero-shot performance compared to models
fine-tuned on English sentiment datasets, and large language models like
GPT--3.5, BLOOMZ, and XGLM. These findings are observable for unseen
low-resource languages to code-mixed scenarios involving high-resource
languages. | [
-0.0260944013,
0.092442967,
0.0112083238,
0.0748434216,
-0.0396339558,
0.0245201234,
-0.0046708947,
0.0076022367,
-0.0193518829,
-0.054275725,
-0.0484619439,
-0.0390352681,
-0.0227915887,
-0.0092845503,
-0.00263215,
-0.0182582568,
0.026754275,
0.019343948,
0.0137166949,
-0.0248744842,
0.0355113409,
0.0329916179,
0.0121443784,
-0.0067762383,
-0.0250636656,
-0.0155829219,
0.02486833,
0.0073882821,
0.0120161949,
-0.0361827686,
0.0140908789,
0.0289497841,
-0.0109580606,
0.0237931348,
0.0000019015,
-0.0110570574,
0.0105263097,
-0.0194003135,
0.0285810512,
-0.0108609311,
0.1037925184,
-0.0815729648,
-0.0112497937,
-0.0088646505,
-0.0287388917,
-0.0363365822,
0.0652566701,
0.0060407408,
0.0299544167,
0.082814157,
-0.04044009,
-0.0276286099,
-0.0374826342,
-0.0121626919,
0.0172760822,
-0.0036876288,
-0.0049310434,
-0.0382876508,
-0.015662536,
-0.0081522139,
-0.0296398811,
0.0173315313,
-0.0068170954,
0.0272983667,
-0.0354926772,
-0.0028718277,
-0.0507366657,
-0.0251885857,
-0.007354693,
0.0525214598,
-0.0535694361,
-0.0164355841,
-0.0403213724,
-0.0287210103,
-0.035516791,
-0.067750372,
-0.0145399338,
0.0049969805,
0.0370217115,
0.0429594927,
0.0193214677,
0.0246370472,
0.0213868469,
0.0106938863,
-0.0322238132,
0.0413894579,
0.004796233,
-0.0207709651,
0.0203806516,
0.071203202,
0.0257828236,
-0.0119219879,
-0.0403715372,
-0.0115616377,
0.0072472049,
0.027739821,
-0.0159917884,
-0.033970885,
0.0220934488,
-0.0636835769,
0.0263215359,
0.0315617472,
-0.0494565405,
0.0223371629,
-0.0217286162,
-0.026196653,
0.0072683878,
-0.0328645147,
-0.0117393266,
-0.0458750688,
-0.0358051732,
0.0287040062,
-0.0009871264,
0.0089685135,
-0.0224646684,
0.0097570959,
-0.0505848527,
-0.0029470457,
0.003573566,
-0.0079693189,
-0.0102612851,
-0.0266820602,
-0.0129770711,
0.0001783715,
0.0128852222,
0.0822764933,
-0.0528700501,
0.0035319061,
-0.0355971716,
-0.082017839,
-0.0138524501,
0.0346501917,
0.0028905452,
-0.0434817597,
0.0392794795,
0.0459074527,
0.0174280833,
-0.0399406329,
-0.0378667079,
-0.0075166151,
-0.006924801,
-0.089842312,
0.0048714336,
-0.0728272945,
0.0324097201,
0.0222854745,
0.0359497927,
0.0125702797,
-0.0000203977,
-0.015722679,
-0.0440513901,
0.0778940171,
-0.0387514941,
-0.0379425101,
-0.0731040686,
-0.0129372943,
0.0431987643,
0.1149516329,
-0.0371881984,
0.0027813553,
0.0481739305,
0.0114012538,
-0.0122621264,
0.0314986445,
0.0066747866,
0.0175734702,
-0.0416795351,
-0.0052494463,
-0.0099354275,
-0.0322728828,
0.0361990258,
-0.0186226182,
-0.028394511,
0.011843754,
0.0287435986,
0.0141140418,
-0.009739873,
0.0306117665,
-0.0051555703,
-0.0106552457,
-0.0259965882,
-0.0087934379,
-0.0074657113,
0.0277610291,
-0.0074431696,
0.0012894009,
-0.0003209719,
0.0382202975,
-0.0371715091,
-0.0213039126,
-0.0123789953,
-0.0553614683,
0.0163491424,
-0.0068813446,
0.0662890971,
0.0114087826,
-0.0401779413,
0.0130902212,
0.0613174587,
-0.0828488767,
-0.0211819075,
-0.0060920627,
-0.0093697449,
0.0542810559,
0.0003898079,
-0.0193373319,
-0.0175296012,
-0.0156686828,
0.0184580572,
0.0451722108,
0.0194426738,
0.0138677601,
0.0707291961,
-0.0528113283,
0.0069652186,
-0.0262501631,
0.0132743316,
0.0548167601,
0.001037007,
0.061208304,
-0.0246472508,
0.0137402127,
-0.0253127161,
-0.0277208705,
0.0064388788,
-0.000948004,
0.0293680187,
-0.005702313,
0.0369223356,
0.0438812934,
-0.0408723988,
-0.0298269056,
-0.001790043,
-0.0505807176,
-0.0223258138,
0.0325498953,
0.0006511962,
-0.0734746382,
0.0238310639,
-0.02704609,
-0.0208974145,
0.0217876602,
-0.0385043956,
-0.0282908846,
0.0589857139,
0.0304325018,
0.0603685491,
-0.0197557136,
-0.0196563322,
-0.0304381624,
0.0408866294,
0.0314368978,
0.024923481,
0.0188235082,
0.0279801972,
-0.0327957124,
-0.018179344,
0.0402647927,
0.0195644572,
-0.0720980316,
0.0553656034,
0.0273744538,
0.0218976494,
0.0428730063,
0.0466914177,
0.0005140491,
-0.0086740917,
-0.1071893275,
-0.0158740524,
-0.0085953232,
0.0154576534,
-0.0092197843,
0.0304392409,
-0.0065532215,
-0.0015727432,
-0.03308459,
0.032534752,
0.0050529693,
0.0418596454,
0.0421569012,
0.0010946901,
0.0476184711,
-0.0316304825,
-0.0032252173,
0.0138912611,
0.0171534661,
-0.0007319665,
-0.0153491721,
0.0301413555,
0.0316038281,
-0.0571485683,
0.0045006885,
0.0059525967,
-0.068482697,
0.0036891703,
-0.0243952069,
0.0058593727,
0.0118227638,
-0.0495784618,
-0.0038067459,
0.0218961108,
-0.0155713847,
-0.0405951813,
0.0256956443,
-0.0104617476,
-0.0189074259,
-0.0086492579,
0.0155221475,
-0.0217033606,
0.0693942755,
-0.0017708672,
-0.0300614368,
-0.0520907864,
-0.0056578116,
-0.0357027613,
-0.0049724299,
-0.0046836445,
0.0450812057,
0.0227408744,
0.0251851361,
0.0216112919,
-0.0188017171,
0.0494970456,
0.0015786256,
-0.0314012207,
-0.0702812374,
0.0043620546,
-0.0401080251,
0.0274646971,
-0.0303989705,
0.0272920132,
-0.0294478517,
0.0510936491,
-0.0039857617,
-0.0288578495,
-0.0230705775,
0.0121272262,
0.0342448466,
0.018933516,
0.0217536539,
-0.014332572,
-0.0590237677,
0.0463322364,
-0.0279133208,
-0.0537231602,
-0.013205844,
0.0254572537,
-0.0377280563,
0.0124636926,
-0.0128509533,
-0.0079252822,
-0.0093335239,
-0.0011054473,
-0.0288816318,
0.0349730812,
0.0052717808,
0.0065983296,
-0.0770329162,
0.0186287239,
-0.0071417484,
-0.0090809604,
0.0373248085,
0.0068308264,
-0.0016853061,
-0.0055114212,
-0.005887039,
0.029499799,
0.0249914937,
-0.0243242923,
-0.0639834628,
0.02479803,
-0.0241932627,
0.0264030844,
0.0947282463,
0.0143113043,
-0.0834519491,
-0.0013712699,
-0.0148328012,
0.0222766474,
0.0450337045,
0.0099469442,
-0.1128134057,
-0.0477244034,
-0.0504181534,
-0.0029482385,
-0.0105015254,
-0.0208611824,
-0.0156739242,
-0.1403095871,
-0.0177076794,
0.0058177174,
-0.0373759381,
-0.0196283087,
-0.034246888,
-0.0158169996,
-0.0332610495,
-0.0213385466,
-0.0782978833,
0.0254554413,
-0.0067609432,
-0.0248198286,
-0.0097331088,
-0.0478916913,
0.0435179174,
-0.0001842463,
0.0348315611,
0.0427134223,
0.0734632015,
-0.0320321657,
0.0159390289,
0.0138886264,
-0.0155531717,
-0.0012571559,
-0.0544536896,
0.0548940897,
-0.0304104201,
-0.0527618863,
0.0425486825,
-0.0273572579,
-0.0686093718,
0.0172642637,
-0.0017217379,
0.0253333393,
0.0107208602,
-0.032091748,
-0.0120483711,
0.0067476649,
0.0426925495,
0.0407047458,
-0.0009535745,
-0.0791398734,
-0.0182649139,
-0.023807589,
-0.0204611626,
-0.0635122433,
-0.029939346,
-0.0224784277,
-0.0016218717,
-0.0477328487,
-0.0355767049,
-0.0677805319,
0.0181125123,
0.043371506,
-0.0286233705,
0.0087834783,
-0.0157440808,
0.0050480049,
-0.0286981519,
0.0135019394,
-0.0097703515,
0.0045943554,
-0.023530677,
-0.025691025,
0.0567636527,
0.0107159615,
0.0097337645,
-0.0269808788,
-0.037807554,
-0.0518101491,
-0.0313326158,
0.0689515397,
0.0090627009,
0.0355869457,
-0.0062022218,
-0.0536397547,
0.025528254,
0.0624334924,
-0.023138525,
0.0117005333,
0.0891322047,
0.061902605,
0.0131503539,
0.0495485738,
0.0026353153,
0.0676617548,
0.0246536825,
-0.002764873,
-0.0276351888,
-0.0151852649,
-0.0196651854,
-0.0078513129,
0.0584792271,
-0.0195593908,
-0.0185864586,
-0.022973517,
-0.0028340667,
-0.0309591461,
-0.0146751916,
-0.0468003452,
-0.0321235321,
0.0258557685,
-0.0211867541,
-0.0582508072,
0.002879729,
-0.0329478122,
-0.0045517306,
-0.0155697213,
0.0554262474,
0.0690102801,
-0.0040657478,
0.0311115496,
0.0345635153,
0.0029790322,
0.0010769252,
-0.0069134878,
-0.0150274569,
-0.0330399908,
0.0236345455,
0.0093630478,
0.0086515304,
0.0230400357,
0.0083443876,
0.0157686006,
0.0231577754,
0.0096427212,
-0.0003421514,
0.0389159769,
-0.0173391216,
0.077465795,
-0.0194836538,
0.0073433472,
-0.0008880419,
0.015801141,
-0.0200197864,
-0.0384101011,
0.0016013174,
-0.0017354738,
-0.0413035676,
0.0283377524,
0.0862960666,
-0.0388919972,
0.0383880958,
0.0425777584,
0.0480383262,
-0.0143854748,
0.016514834,
0.0772570446,
-0.0046731811,
0.0326335691,
0.0409168489,
-0.0691499785,
-0.0696715415,
0.0507316552,
-0.021088602,
-0.0047845556,
0.016335275,
0.0180551428,
0.0601938963,
0.0157538913,
0.0799786448,
-0.0425069071,
-0.004332636,
0.0013330759,
-0.0425949767,
0.0109321903,
-6.276105398e-33,
0.0177421011,
0.0183888916,
-0.0381235927,
-0.0103274891,
0.0256668981,
0.0281309318,
-0.0350141637,
0.0429417975,
-0.0167368203,
0.0335734114,
-0.0098457877,
0.026911037,
-0.0006875182,
0.035759747,
0.0607393198,
0.0083308043,
0.0373496749,
0.0063406448,
-0.0062334901,
0.0282418877,
0.0467606634,
0.0257401094,
0.0848077983,
-0.0479074419,
0.0500474013,
-0.033438772,
-0.0281662736,
-0.0054147565,
0.0775940418,
0.0381308608,
-0.0139760375,
0.0891317949,
0.0216632262,
-0.0003639985,
-0.0096675837,
-0.0000468215,
0.0053424914,
-0.0467924327,
0.0191001985,
0.0128408,
0.0135904728,
-0.0108569385,
-0.0216621831,
-0.000124142,
-0.0229947101,
0.0165879969,
-0.0027177944,
0.0000503914,
0.0168276429,
-0.0550413318,
-0.0821213573,
0.017976081,
-0.0002380531,
-0.0014605994,
0.0538386554,
0.0868489295,
0.0072822543,
0.0522870161,
-0.0618142225,
0.0394677408,
-0.0505150035,
-0.0031026837,
0.0253781322,
-0.0235583652,
-0.0058325138,
0.057438422,
0.0239259023,
0.0239049885,
0.0233186036,
0.0815334395,
0.040614415,
-0.0218341593,
0.0258801058,
0.1125686541,
0.0421754159,
-0.0297437273,
-0.0495502576,
-0.0199775547,
0.0398943275,
-0.0038860578,
0.0507938042,
0.0059243562,
-0.0098610912,
-0.0363586321,
0.0368059091,
-0.0537764207,
-0.0361124352,
-0.0209071673,
0.001036821,
-0.0140715521,
-0.0102805765,
0.0137306955,
0.0453978367,
-0.048820924,
-0.0467534065,
0.0636710599,
-0.0325749479,
0.0270599127,
0.0075700744,
0.0282955356,
-0.0567741618,
-0.0212887842,
-0.0297930744,
0.0281266831,
0.0236796085,
0.0053237262,
0.0760521889,
0.0540323555,
-0.0772246644,
-0.0252110884,
-0.0299498327,
0.0208079033,
-0.0084665958,
0.0725045055,
-0.0123821907,
0.0033274959,
-0.0062875892,
0.0799545273,
0.0354602188,
-0.0100808451,
0.0072053405,
-0.0181017835,
0.0024075557,
0.0069038747,
-0.0447878949,
-0.0514823012,
-0.0181162972,
-0.046597939,
0.0099155502,
-0.0566079803,
-0.0110909054,
0.0031599666,
2.699e-7,
0.0344463624,
0.0493070856,
-0.0447677635,
-0.0029613515,
0.0077931201,
0.0518220849,
-0.0206104293,
0.0215180144,
-0.0296036527,
0.0123396581,
-0.008831555,
0.0276605394,
0.0318195373,
0.0378051586,
-0.0983251557,
0.0083978307,
-0.0196001902,
0.0080543496,
-0.0415617228,
0.023802178,
0.0493151024,
0.131267786,
0.0194269884,
0.0052103498,
-0.0034388441,
-0.0200248938,
0.0480104461,
-0.0166861229,
-0.0329489112,
-0.0042315135,
-0.0410636291,
0.0415872522,
-0.0314842053,
-0.034467347,
-0.0078768972,
0.0160169676,
-0.0063738585,
0.0347018652,
-0.0097688874,
-0.0518393554,
0.0363995284,
-0.0015399449,
-0.0341128111,
-0.0093917958,
0.0541726798,
-0.0353224017,
-0.0399503708,
-0.066734679,
0.0225887969,
0.0302344244,
0.0914848074,
-0.0310021657,
-0.050280001,
0.0195815321,
0.0343713015,
-0.0262119491,
-0.0711181387,
-0.0447694026,
-0.0089581348,
0.0147314807,
0.033656776,
-0.0167217851,
0.0194595456,
0.0250302814,
0.0286462624,
-0.0212621875,
0.0294603948,
2.074143022e-34,
0.0629557297,
-0.0388927348,
0.0077148271,
0.0123009868,
0.0121545373,
-0.0116062351,
-0.0439404398,
-0.0187450852,
0.0069514951,
-0.0098487316,
0.0182456039
] |
http://arxiv.org/abs/2402.02101v1 | http://arxiv.org/pdf/2402.02101v1 | Are Large Language Models Good Prompt Optimizers? | 2024-02-03T09:48:54 | 2024-02-03T09:48:54 | Ruotian Ma, Xiaolei Wang, Xin Zhou, Jian Li, Nan Du, Tao Gui, Qi Zhang, Xuanjing Huang | cs.CL | cs.CL, cs.AI | LLM-based Automatic Prompt Optimization, which typically utilizes LLMs as
Prompt Optimizers to self-reflect and refine prompts, has shown promising
performance in recent studies. Despite the success, the underlying mechanism of
this approach remains unexplored, and the true effectiveness of LLMs as Prompt
Optimizers requires further validation. In this work, we conducted a
comprehensive study to uncover the actual mechanism of LLM-based Prompt
Optimization. Our findings reveal that the LLM optimizers struggle to identify
the true causes of errors during reflection, tending to be biased by their own
prior knowledge rather than genuinely reflecting on the errors. Furthermore,
even when the reflection is semantically valid, the LLM optimizers often fail
to generate appropriate prompts for the target models with a single prompt
refinement step, partly due to the unpredictable behaviors of the target
models. Based on the observations, we introduce a new "Automatic Behavior
Optimization" paradigm, which directly optimizes the target model's behavior in
a more controllable manner. We hope our study can inspire new directions for
automatic prompt optimization development. | Are Large Language Models Good Prompt Optimizers?: LLM-based Automatic Prompt Optimization, which typically utilizes LLMs as
Prompt Optimizers to self-reflect and refine prompts, has shown promising
performance in recent studies. Despite the success, the underlying mechanism of
this approach remains unexplored, and the true effectiveness of LLMs as Prompt
Optimizers requires further validation. In this work, we conducted a
comprehensive study to uncover the actual mechanism of LLM-based Prompt
Optimization. Our findings reveal that the LLM optimizers struggle to identify
the true causes of errors during reflection, tending to be biased by their own
prior knowledge rather than genuinely reflecting on the errors. Furthermore,
even when the reflection is semantically valid, the LLM optimizers often fail
to generate appropriate prompts for the target models with a single prompt
refinement step, partly due to the unpredictable behaviors of the target
models. Based on the observations, we introduce a new "Automatic Behavior
Optimization" paradigm, which directly optimizes the target model's behavior in
a more controllable manner. We hope our study can inspire new directions for
automatic prompt optimization development. | [
0.0316644795,
0.0336356089,
-0.0251319222,
0.0498750843,
-0.0339985676,
-0.0110595562,
0.0028158249,
0.0095763355,
-0.0043908926,
-0.0215556063,
-0.0470330976,
-0.0200155713,
0.0134180626,
0.0053260094,
0.0084063299,
-0.0450169854,
0.0985153988,
-0.0643497333,
-0.0285854265,
-0.0055538747,
-0.0362303257,
-0.0212987848,
0.01032231,
0.0391044356,
0.0029362878,
-0.0321156234,
-0.0213931501,
0.040059533,
-0.0202099793,
-0.0230133161,
0.0041360632,
0.0457779393,
-0.0145679638,
0.0297834259,
0.0000020976,
-0.0978052691,
-0.0189620592,
-0.0047439756,
-0.0466171727,
0.0591727607,
0.0386402793,
0.0126768202,
-0.0368004888,
0.0282172039,
0.0024260895,
0.0147923054,
0.0175205227,
-0.001901648,
0.0563359261,
0.0742262378,
0.0088951541,
0.0220222455,
0.001982325,
-0.0023880301,
0.0022824735,
-0.0189129282,
0.0139878755,
-0.058514744,
0.0168279819,
0.0271919724,
-0.0146023147,
0.0064753769,
-0.0115412343,
-0.0476446562,
-0.0394844115,
0.0038066907,
0.0205452237,
-0.0565639734,
-0.0509265214,
0.0281965081,
-0.0010614805,
-0.0098077934,
0.0124647636,
-0.0438206308,
0.0420163125,
0.0115492865,
-0.0073543009,
0.0467850454,
-0.0233529489,
-0.0132249761,
0.0063085509,
-0.0112124542,
0.0156755429,
-0.0114714047,
-0.0235615987,
0.0964075178,
0.0241373479,
-0.0177136511,
0.0097097177,
-0.0290785618,
-0.0030767396,
-0.0760427639,
-0.026127601,
0.045056995,
-0.1095879674,
0.0236813612,
-0.0300219581,
0.0061295251,
0.0382683501,
0.0430320837,
0.0561810695,
0.0716410503,
-0.0230164118,
0.0032126827,
-0.0185186435,
0.0813533962,
-0.0060958723,
-0.013015938,
-0.0275980588,
0.0488342009,
-0.0045754109,
-0.009314022,
-0.0685001239,
0.026436368,
-0.0161208324,
0.0279286224,
-0.0132439546,
0.0883040056,
0.0112874517,
0.0635440275,
-0.0169266518,
-0.0252963118,
-0.0259060897,
0.00177235,
-0.0216971468,
0.0019650755,
-0.0208590161,
-0.0125007909,
0.0178236552,
-0.0684247166,
-0.0020285358,
0.0066546449,
-0.0347677022,
-0.0067821671,
-0.0083467877,
0.0827109143,
-0.0052496875,
-0.0060891393,
-0.0117475037,
-0.0615673773,
-0.0140254358,
-0.0234459415,
-0.0367759056,
-0.0603115,
-0.0166656617,
0.0164936744,
-0.016334936,
-0.0171814542,
-0.00781387,
0.0007122688,
0.0044290833,
0.05712796,
-0.0412496552,
0.0311869849,
0.0171443708,
0.0090673333,
-0.0297229569,
0.0656816885,
0.0005184201,
0.0318589061,
0.0387704484,
-0.0172879826,
0.0072284583,
0.06929297,
-0.0024329775,
0.0239374954,
-0.0132668996,
-0.0068146107,
0.0527186207,
0.0265338533,
0.0398308635,
0.0389219485,
-0.0001062016,
0.0131425792,
0.0274020582,
0.0273306668,
-0.0125095304,
0.0007213056,
0.0545905121,
0.0002426932,
-0.0003605465,
0.0400013253,
-0.0607879311,
-0.0191713199,
-0.04971141,
0.0331258141,
-0.0175639018,
0.0245410316,
-0.0030996983,
-0.048023928,
-0.0064313216,
0.0265551656,
-0.0463639908,
0.0113989441,
-0.0131208049,
0.0243118871,
-0.0597472638,
0.0513266437,
0.0164661147,
-0.0312939733,
-0.0357552432,
-0.0757280067,
0.0416885354,
0.0529066399,
0.0365041234,
-0.0422632731,
-0.0527588539,
-0.0069328425,
0.0224775374,
0.0200680047,
0.0850806758,
-0.0278313588,
-0.0181177929,
-0.012727445,
-0.0322148502,
-0.0248054117,
0.0222218819,
0.0610995293,
-0.0039586704,
-0.0218535792,
0.007471926,
0.0368223228,
-0.0118415197,
0.0329150893,
-0.0166002139,
-0.0144655565,
0.0133705214,
0.0141011793,
0.0223483015,
-0.000760468,
-0.0155234393,
0.018363744,
0.0216012914,
0.0120054288,
0.0008664585,
-0.0136990389,
-0.0513926484,
-0.0512563847,
0.0386367403,
0.0297323745,
-0.0172841791,
-0.007601554,
-0.0313356258,
-0.044350177,
-0.0251589436,
-0.0390394591,
0.0199319515,
-0.0177819263,
0.0904678702,
-0.0424120836,
-0.0141012333,
-0.0628035665,
0.0364779755,
0.0071006515,
0.0157807097,
0.0866872594,
0.055260241,
0.0354656205,
-0.0570308119,
-0.021662863,
0.0028851735,
-0.0252171438,
-0.0014801067,
0.0447573699,
0.0333010741,
-0.0284427498,
-0.0487604849,
-0.0731810704,
0.0231985394,
-0.0215555616,
-0.0006332524,
0.0087978505,
0.0449946001,
-0.0476888046,
-0.0051972694,
0.0014272335,
-0.0010006083,
0.0236946288,
0.045464769,
0.059022069,
0.006110338,
-0.061015401,
-0.0403880775,
-0.0015983923,
-0.0321820304,
0.0077042398,
-0.0193703491,
-0.0177085865,
-0.0931802914,
0.084860526,
-0.0011572206,
0.0233704969,
-0.0032323443,
-0.0049217581,
0.0185724925,
-0.0049139461,
0.0472365543,
0.0505405553,
0.0055214078,
0.0034591581,
0.0400881022,
-0.0057490598,
-0.0887952447,
0.0166752897,
0.0061034579,
-0.0468784347,
0.0009349962,
-0.0309709739,
0.0545455515,
0.077819176,
0.0063014659,
-0.012310707,
-0.0227526575,
0.0242168587,
0.0010593943,
-0.0338485427,
-0.0324351043,
-0.0540503487,
-0.0446672998,
0.0830025449,
0.0003365361,
-0.0666510835,
0.077123791,
-0.0560516343,
-0.0080167586,
-0.0016413064,
0.0195223391,
-0.0409008376,
0.0182825606,
-0.0139419809,
-0.0062264842,
0.0426864587,
-0.0624758452,
-0.0415557437,
0.0009513321,
0.0538990982,
0.0043038321,
0.0515188836,
-0.0230730362,
-0.0014734467,
-0.0263758469,
-0.0546182655,
-0.0585244149,
-0.0741905048,
-0.0059716976,
-0.0183097273,
0.0024227542,
0.0182163548,
0.0145348227,
-0.002015145,
0.0064880527,
0.0329097211,
-0.0332698375,
-0.0235083457,
-0.0390131585,
-0.0245081224,
0.0214078501,
-0.0057605784,
0.0194962621,
-0.0251486041,
0.0722315684,
0.0053602471,
0.0365662053,
-0.0022436299,
-0.010756488,
0.0127984127,
0.0684695169,
-0.0594641976,
-0.0656081215,
0.0579782538,
0.0696221218,
-0.0574231111,
-0.0202706046,
0.1106890142,
0.0532519221,
-0.0387453139,
0.0214877836,
-0.0724909082,
0.0492931679,
0.0477872938,
0.0256549492,
-0.0610120036,
-0.018829016,
0.0396696329,
0.0283356681,
-0.0221357159,
-0.0036678489,
-0.0100155929,
-0.1204925627,
0.0524894148,
-0.0156115117,
-0.0068791546,
0.0155703956,
-0.0331320353,
-0.0028839416,
0.0289702397,
0.048695758,
0.0533543415,
-0.0416126512,
-0.0124871163,
0.0149987703,
-0.0299164541,
-0.0103422161,
0.0008292114,
-0.0264882334,
0.0406386442,
0.0934400409,
-0.0217769668,
0.0126834502,
-0.0234708022,
0.0678598508,
-0.0039178212,
0.0424926803,
-0.0311244596,
-0.0023803383,
0.025355015,
0.0029118098,
0.0104439035,
0.0066634458,
-0.0060285502,
-0.0348282307,
0.0320970118,
0.0498088337,
-0.0395146199,
0.0061054043,
0.0005155105,
-0.0015561662,
0.016889276,
0.0538442284,
0.0484817848,
-0.0283090398,
-0.0400402509,
-0.0486819446,
-0.0347290859,
-0.0191024709,
-0.0115111284,
0.0061035911,
-0.0020284902,
-0.0128610143,
-0.037391834,
-0.0327449217,
-0.003488183,
0.0407017358,
0.0162025541,
-0.017377153,
-0.0594640225,
0.0009834114,
-0.004592157,
0.031762097,
-0.026710229,
0.0253791474,
-0.087926656,
0.0433970727,
0.0191020742,
-0.0352551118,
0.0000479858,
-0.0132660428,
-0.0254364312,
-0.0823674574,
-0.0095236795,
0.0714760646,
0.0259113479,
0.057630226,
-0.023814803,
-0.0502087176,
0.0420201384,
0.0326319784,
-0.0034983582,
0.017008489,
0.0340420641,
0.0315841697,
-0.0244746152,
0.0226299465,
-0.044040937,
-0.0098435534,
0.0184807517,
0.0167453587,
0.0087609105,
-0.0928438604,
0.0258405935,
0.079892002,
0.0366837531,
-0.0317454934,
-0.0149618592,
-0.0227661915,
0.0205024462,
-0.0367731974,
-0.0300605539,
-0.0178587008,
0.0481363609,
-0.0019897358,
-0.0310363509,
0.0227702018,
-0.0164752044,
-0.0591620319,
0.0397157557,
-0.0313061848,
0.042495124,
-0.0159894619,
0.0138299474,
0.0587970428,
0.0159984846,
-0.0002536991,
0.0518192686,
0.0401428379,
0.0246828683,
-0.0116887307,
0.0677894279,
0.0313200951,
0.0350296311,
-0.0349446423,
0.0303571355,
-0.0392075479,
-0.0742456242,
0.0124206785,
0.0469332114,
0.0568659678,
0.0102861812,
0.0155681046,
-0.0163378324,
-0.016773684,
-0.0058090994,
0.0033571469,
0.0564595945,
-0.0701763108,
0.0293420944,
0.0209377762,
-0.0246223174,
-0.0292195156,
0.0533579513,
-0.0373875611,
0.0454878919,
0.0098788366,
-0.0068793087,
-0.0428654328,
-0.0001188499,
0.0215806607,
-0.019088788,
-0.0345643535,
-0.0463341549,
-0.0691932812,
-0.0354473032,
0.0154172564,
0.0307744388,
-0.0114870723,
-0.0192891806,
0.0228749588,
0.0501874723,
0.0223418921,
-0.0191246439,
-0.0324398801,
0.00534043,
-0.0027255313,
-0.0254901182,
-0.0187177062,
-5.396566362e-33,
-0.0506945401,
-0.0152405733,
0.0122990692,
0.0005042752,
-0.0215036627,
-0.0145797459,
-0.0386565886,
0.0207963623,
0.0130441254,
-0.0172642972,
-0.0649882108,
-0.0322674438,
0.0206482597,
-0.037149366,
0.0211108569,
0.0018804361,
0.0418191329,
0.0034963109,
-0.101635471,
-0.0275296513,
-0.0168352053,
0.020417992,
-0.0122053083,
0.0061825826,
0.0214752201,
0.0117992656,
-0.0112294136,
0.0476290546,
-0.0102271866,
0.0411510244,
0.007195449,
0.025046099,
-0.0021043455,
-0.0609860122,
-0.0476129204,
0.0500231273,
-0.0446647368,
-0.0754821002,
0.0419660397,
0.0226874556,
-0.0375282913,
-0.0587621331,
0.0022340689,
-0.0072889286,
-0.0464534834,
0.0417625308,
-0.0063647944,
-0.0085349642,
-0.0113150785,
-0.0097024869,
-0.061713513,
-0.0079412358,
-0.0088628363,
-0.0196615811,
0.0278652236,
0.0318013169,
0.021889234,
-0.0002304944,
0.0064086583,
0.0165063459,
0.0316828452,
0.0393922329,
0.0824617594,
0.0423725434,
-0.0156216631,
0.017618496,
0.0365385674,
0.0473265536,
0.0110393958,
0.0065723844,
0.0162930805,
0.0180990528,
0.0252887923,
0.0267305654,
-0.0344675332,
-0.0266227126,
-0.0091485316,
0.0533002354,
0.05333516,
-0.0001737697,
0.0109855114,
0.0138818389,
-0.0035056884,
-0.0360056646,
-0.0680053309,
0.0288731586,
-0.0395895056,
-0.0135987056,
0.0402281471,
-0.0276160501,
-0.0033427246,
-0.0489771031,
0.0297012851,
-0.0181355849,
-0.0071211248,
-0.0040231533,
-0.0388445966,
0.0167300496,
-0.0134480754,
0.0398687571,
0.0088065406,
-0.034909524,
0.0060016569,
-0.0211390611,
0.0440371707,
0.037794672,
-0.0147137046,
0.0244051237,
-0.0716157481,
0.0242787022,
-0.0367961563,
0.0320461839,
0.0010587982,
-0.0146881305,
0.0208242908,
-0.0343085974,
-0.0105150426,
0.046282392,
0.0165280681,
-0.0354990587,
-0.0528080575,
0.0295531154,
-0.0082741044,
0.0238373261,
-0.0160471648,
0.0073662954,
-0.0061581377,
0.062273141,
0.0256235767,
-0.0055406555,
-0.0360708609,
-0.0002528737,
2.713e-7,
0.0270863697,
0.04042162,
0.0249855947,
0.0762401596,
0.0063139936,
0.0396715701,
-0.029463293,
0.0100387195,
0.0310481712,
-0.0156944375,
0.0171100907,
-0.0184154734,
0.0218501333,
0.0001675724,
-0.0879516974,
0.0064742337,
-0.0355115235,
0.0225380566,
-0.0672148541,
0.0103106564,
0.0058916453,
0.0477277003,
-0.0219810326,
-0.0208710767,
-0.0090347985,
-0.0020622744,
0.0120567828,
-0.0512196757,
-0.0147850756,
0.071661748,
0.0575379767,
-0.0417549461,
0.0072572054,
0.026093604,
0.0090576205,
0.0090955421,
-0.033203315,
0.0553575158,
0.0230620429,
-0.0076826615,
0.0013517538,
-0.0131112337,
0.006025597,
-0.0544135086,
0.063172318,
-0.0119344695,
0.0013351677,
-0.0451026037,
0.0039565763,
-0.0519678406,
-0.0073392522,
-0.0371283256,
0.0200966876,
0.0470169969,
-0.0083252853,
0.0211542137,
-0.0234025437,
-0.0291901045,
-0.0334544405,
-0.0106717991,
-0.0304599199,
-0.0667504072,
-0.0244907327,
0.0215955488,
0.0016659548,
-0.1016538665,
-0.0206462611,
3.071034782e-34,
0.0318435319,
-0.0322927311,
0.0456703641,
0.0083777495,
0.0291405283,
0.0171798225,
-0.0163718183,
0.0522463918,
0.0450684726,
-0.0495586246,
-0.0495950207
] |
http://arxiv.org/abs/2402.02099v1 | http://arxiv.org/pdf/2402.02099v1 | Analyzing the Evaluation of Cross-Lingual Knowledge Transfer in Multilingual Language Models | 2024-02-03T09:41:52 | 2024-02-03T09:41:52 | Sara Rajaee, Christof Monz | cs.CL | cs.CL, cs.AI, cs.LG | Recent advances in training multilingual language models on large datasets
seem to have shown promising results in knowledge transfer across languages and
achieve high performance on downstream tasks. However, we question to what
extent the current evaluation benchmarks and setups accurately measure
zero-shot cross-lingual knowledge transfer. In this work, we challenge the
assumption that high zero-shot performance on target tasks reflects high
cross-lingual ability by introducing more challenging setups involving
instances with multiple languages. Through extensive experiments and analysis,
we show that the observed high performance of multilingual models can be
largely attributed to factors not requiring the transfer of actual linguistic
knowledge, such as task- and surface-level knowledge. More specifically, we
observe what has been transferred across languages is mostly data artifacts and
biases, especially for low-resource languages. Our findings highlight the
overlooked drawbacks of existing cross-lingual test data and evaluation setups,
calling for a more nuanced understanding of the cross-lingual capabilities of
multilingual models. | Analyzing the Evaluation of Cross-Lingual Knowledge Transfer in Multilingual Language Models: Recent advances in training multilingual language models on large datasets
seem to have shown promising results in knowledge transfer across languages and
achieve high performance on downstream tasks. However, we question to what
extent the current evaluation benchmarks and setups accurately measure
zero-shot cross-lingual knowledge transfer. In this work, we challenge the
assumption that high zero-shot performance on target tasks reflects high
cross-lingual ability by introducing more challenging setups involving
instances with multiple languages. Through extensive experiments and analysis,
we show that the observed high performance of multilingual models can be
largely attributed to factors not requiring the transfer of actual linguistic
knowledge, such as task- and surface-level knowledge. More specifically, we
observe what has been transferred across languages is mostly data artifacts and
biases, especially for low-resource languages. Our findings highlight the
overlooked drawbacks of existing cross-lingual test data and evaluation setups,
calling for a more nuanced understanding of the cross-lingual capabilities of
multilingual models. | [
-0.0276803076,
0.028702572,
-0.023117302,
0.0649614111,
-0.0394324623,
0.0177457966,
0.0436730571,
-0.0211909208,
-0.0303257704,
-0.0381261669,
-0.0775854737,
-0.0305743814,
0.0033965432,
-0.014257336,
0.0310654566,
-0.0394643731,
0.0539209582,
-0.049616605,
-0.0312474705,
-0.0112312259,
-0.0210804418,
0.0075536626,
-0.011025642,
-0.0019240759,
0.0014804426,
-0.0071306443,
0.0312091019,
-0.0151567301,
0.0104284901,
0.006509332,
-0.0200980306,
0.0306601729,
-0.0061096987,
0.0494077466,
0.0000018719,
-0.0483218879,
-0.0095802182,
0.0177042,
-0.0127839185,
0.0059906715,
0.0476939529,
-0.0743990541,
-0.0120384665,
-0.0165115818,
-0.0128576253,
-0.0368904024,
0.0546580218,
0.0274756327,
0.010092238,
0.0678221583,
-0.0213632956,
0.0129542286,
0.0275276285,
-0.0397565179,
0.0175457541,
-0.0584712625,
0.023094425,
0.0274961144,
-0.0391922183,
-0.0110377222,
-0.0455264077,
0.0393019877,
0.0121420091,
0.036014054,
-0.0401720405,
0.0378209054,
-0.0333835706,
0.0131849833,
-0.0129912803,
0.0498118438,
-0.0621196963,
-0.0194499269,
-0.032090839,
-0.0041095768,
-0.0120504834,
0.0037450048,
-0.0776313022,
0.0227813125,
0.0215552486,
-0.0043687448,
0.0368960947,
0.0006433013,
0.0523517393,
-0.0166672729,
-0.0145387072,
0.0573725365,
0.0372506902,
-0.0337178372,
0.0444189422,
0.0472602174,
0.033837866,
-0.0372223668,
-0.0549170822,
-0.004567056,
0.0134136379,
0.0304718316,
-0.0384445824,
-0.0356965363,
0.0129235853,
-0.0180321317,
0.0653256029,
0.0233275928,
-0.0254724342,
0.0565145202,
-0.0605523475,
0.044829376,
-0.0025903326,
0.039039813,
-0.0254591443,
-0.0733804628,
-0.0652050674,
0.0178027544,
-0.0044153873,
0.0152716376,
0.0074456739,
0.0006866147,
-0.0024792992,
-0.0219689179,
-0.0082289539,
-0.0285542533,
0.0029397171,
-0.0543118604,
-0.038758561,
-0.0087836096,
0.0103351492,
0.0737015083,
-0.0160435848,
0.032764975,
-0.0085854139,
-0.0507538058,
-0.0145446258,
0.0222921763,
0.0026931672,
-0.0755647719,
0.0142731583,
0.0768055022,
-0.0542528592,
-0.0159137845,
-0.0367344692,
0.0178898014,
-0.0086618345,
-0.0869415253,
0.0438019931,
-0.0624051578,
0.0605263598,
0.0111739701,
0.0255320985,
-0.0199797377,
-0.021879578,
-0.007956787,
-0.0567888394,
0.0514053628,
0.0182534158,
0.0007840228,
-0.0037666734,
-0.012396385,
-0.0425111726,
0.0916306078,
0.011598764,
-0.0164738167,
0.0388262756,
-0.0084673343,
-0.0481646731,
0.0482023731,
0.0426249728,
0.0495524444,
-0.0139828501,
-0.0036053525,
0.0360354707,
0.002115279,
0.0474516116,
0.0014355684,
-0.0316955037,
0.0287459083,
0.053444583,
0.0236331709,
0.0285847038,
0.0117442561,
0.0551347062,
-0.0304953083,
-0.0265137665,
-0.0066621806,
-0.0244114734,
0.0211709216,
0.0027225472,
-0.0071706437,
-0.0115719251,
0.0486160852,
-0.0406085476,
0.0233811419,
0.0158035401,
-0.0305126235,
0.0154086361,
-0.005593447,
0.0706744641,
0.037469957,
-0.0179452728,
0.0628357902,
0.0678900555,
-0.043456547,
0.0210873559,
-0.0144006712,
0.0009828971,
0.0162688568,
-0.0107920943,
0.0211853571,
-0.0321247987,
0.0018612322,
0.0373305455,
0.0148894489,
0.0828892216,
-0.0146930963,
0.0572406054,
-0.0419919156,
-0.010705133,
-0.0404496416,
0.0109106619,
0.0621094108,
-0.0127477478,
0.0720260814,
-0.0000965787,
-0.0496222377,
-0.0162448399,
0.0177753288,
-0.0878161937,
-0.0041510933,
-0.0046825008,
0.0209923293,
0.0078898231,
-0.0016394409,
-0.0042078113,
-0.0184463933,
-0.0114055453,
-0.0427561961,
-0.0158954263,
0.0315170176,
-0.0145589598,
-0.047367122,
0.0127149392,
-0.0272313841,
-0.0064849579,
0.0264937989,
-0.0176070873,
-0.0359579436,
0.0438400917,
0.0392253734,
0.0346505456,
-0.0265097357,
-0.0014341223,
-0.0191186201,
0.0496928096,
0.0218332019,
0.0396951362,
0.0253907293,
0.0191081129,
-0.0067135887,
-0.032880079,
0.0333644934,
-0.0052427268,
-0.0214722808,
0.0383380093,
-0.0301757865,
0.0186100043,
0.0075571234,
0.0169673022,
-0.0093073202,
0.0072130314,
-0.0883805752,
0.0278649796,
-0.0215648934,
0.0313083194,
0.0204833988,
0.0122174826,
-0.0093208505,
-0.0004518976,
-0.044400312,
0.0081429016,
0.0132314637,
0.0404440202,
0.0034578333,
0.0067301453,
0.0382385738,
-0.0340680517,
-0.0115686459,
0.0072171274,
0.067339763,
-0.0296375416,
-0.0041900212,
-0.0049031898,
0.0476886593,
-0.038028203,
-0.0030996632,
-0.0109433345,
-0.0184088256,
-0.0114461537,
0.0110182343,
0.0165024009,
0.0508909076,
-0.0631604865,
-0.0248839743,
0.0468700416,
-0.0413964279,
-0.0588364713,
0.0150303897,
0.0237773824,
-0.0104469955,
-0.0530650355,
-0.0221887548,
-0.0169473402,
0.0741343573,
-0.0038747727,
-0.0301670842,
-0.058840245,
-0.0057178224,
-0.0330028981,
-0.027110735,
-0.039951127,
0.0438597053,
0.0146017931,
0.0377435721,
-0.0450633466,
-0.0054959692,
0.0093188565,
-0.0363389105,
-0.0035624118,
-0.0172223318,
0.0072729546,
-0.034897469,
0.0451057144,
-0.0232230201,
0.0309126116,
-0.0142947249,
0.0138819618,
-0.0564301461,
-0.0006004347,
0.01947942,
0.0021351315,
0.0332267731,
-0.0449389853,
-0.0045357752,
-0.0181218516,
-0.0427111126,
0.0050136554,
0.0062675015,
0.0034249958,
0.0465763845,
-0.0046820021,
-0.006147583,
-0.0386812575,
-0.0088256979,
0.0417257622,
0.0314520784,
-0.0024372772,
0.0083143227,
0.0030040415,
0.0165075492,
0.0211169142,
-0.0077410857,
0.0018255174,
-0.0476146154,
0.0149373682,
0.0074220081,
0.0506137796,
0.0155117568,
-0.0108774034,
0.0094202105,
0.0238550305,
-0.003161134,
-0.0604810156,
-0.021902103,
0.0576768294,
-0.0247080214,
0.0010741022,
0.045630198,
-0.0280109588,
-0.0161179192,
-0.0017683858,
-0.0303391442,
-0.004611013,
0.0195822697,
0.0270231795,
-0.0866144225,
-0.0367607772,
0.005259065,
-0.024493698,
0.0343766101,
-0.0388992429,
-0.0207031015,
-0.1506478339,
-0.0085726678,
-0.0126269516,
-0.1027217582,
-0.0293796211,
-0.0385966487,
-0.001811169,
-0.0255591255,
0.0258867871,
-0.0527788326,
-0.0020851723,
-0.012177065,
0.0173090547,
0.00485744,
-0.0843584538,
0.072693862,
-0.0485130847,
0.0556231029,
0.0335900858,
0.041812975,
-0.0317783579,
-0.0057844422,
0.0127494317,
-0.0085037313,
-0.0292730853,
0.0234628059,
0.0558475703,
-0.0150775844,
-0.0161943864,
0.0745171085,
-0.0176868588,
-0.0390837193,
0.0119857108,
0.0250374097,
0.102582261,
0.0586958267,
-0.041839283,
0.0308425762,
-0.0443030894,
0.0227506738,
0.0681845695,
0.0065850802,
-0.0180199239,
-0.0091839619,
0.0052311788,
-0.0407897271,
-0.0554223694,
-0.0330051146,
0.0451290384,
-0.0351038389,
-0.0436603874,
-0.0699554235,
-0.0806003809,
0.0324594639,
0.0224931613,
0.023477314,
0.0134007111,
-0.0439109728,
-0.0203290656,
-0.0135038421,
0.0036761737,
-0.0446774513,
0.042826511,
-0.0262513701,
-0.0010119379,
-0.0076133152,
-0.0282933414,
-0.0042903735,
-0.0213108938,
-0.059326604,
-0.0358025692,
-0.0413618311,
0.0793532655,
0.0111377239,
0.0518686101,
-0.015645178,
-0.03650916,
0.0169471297,
-0.0039504748,
-0.0363416895,
-0.0023231537,
0.0596391633,
-0.0250296406,
0.0157088339,
0.0599528477,
-0.0475764759,
0.0227874219,
0.0350521542,
0.0316116996,
-0.0189947337,
-0.0041156807,
-0.0089335917,
0.0038544897,
0.0235142894,
-0.0097869067,
-0.035439942,
-0.031425748,
0.0127431918,
-0.0298902057,
-0.0105718691,
-0.0419807881,
-0.035829626,
0.0537986346,
0.0274412408,
0.0161900241,
0.0204818137,
0.0042769923,
-0.0035247745,
-0.0437635258,
-0.0057743848,
0.0264573004,
0.0337034091,
0.0229411554,
0.0441732705,
-0.0057681552,
-0.0296129044,
0.0111452146,
0.0243432764,
-0.0540944822,
0.0061831931,
0.0048514479,
0.0068691056,
0.0266421344,
-0.0135300178,
-0.0394175611,
-0.0239690375,
0.0222713221,
-0.0180199444,
0.060866829,
-0.0443395935,
0.1012198478,
-0.0107807377,
-0.0071832519,
-0.0327736139,
0.0252748691,
-0.0201740917,
-0.0084401779,
0.0235545225,
0.0115486048,
0.0089458656,
0.0214185957,
0.1012320518,
0.0159890316,
0.05475609,
-0.0323121771,
0.0139239989,
-0.0102064563,
-0.0147580951,
0.0580375381,
-0.0188652389,
-0.0198304709,
0.023196429,
-0.0548430607,
-0.0877285898,
-0.0041113291,
-0.0277886558,
0.0088210674,
0.010864011,
-0.0106094927,
0.0928872824,
-0.0248568561,
0.0113880234,
-0.0574841872,
-0.0266897399,
0.0115313418,
0.0233665481,
0.0101830214,
-5.778926022e-33,
-0.007659859,
0.0423553474,
0.0485994294,
-0.002244255,
0.0161892194,
0.0148718599,
-0.0301137306,
0.0019224618,
-0.0411422178,
0.0253856406,
-0.0039426717,
-0.0238693226,
0.0152173387,
0.0319801308,
0.0445980765,
0.0091970656,
0.0255432259,
0.002417471,
-0.0125563955,
0.0113911554,
0.0470671989,
0.0420863442,
0.1049128547,
-0.0127685936,
0.0356542841,
0.0164312236,
0.0284148939,
-0.0114312377,
0.0164160039,
0.0494105294,
-0.008181233,
0.0461027846,
0.0380426161,
-0.0685343444,
-0.0169030782,
-0.0206627902,
-0.0658188462,
-0.0443137735,
0.0387705974,
0.030462604,
0.001971506,
-0.0322830677,
-0.0095915748,
-0.0123534007,
0.0215148013,
-0.0484867766,
0.0193968359,
-0.0318089239,
-0.0322837159,
0.003059231,
-0.1173120216,
0.0104120476,
-0.0016494398,
-0.0263478104,
0.0636748597,
0.0626882464,
0.0119767245,
-0.0033081127,
-0.0687208399,
0.0144919874,
-0.0492302328,
0.0241036396,
0.0925962552,
0.0008828067,
-0.0376595818,
0.0371665508,
0.0445304215,
0.0266179927,
0.0359543711,
0.0243154708,
0.0661454424,
0.036204163,
0.0211647339,
0.0761187747,
0.0318628848,
-0.004787745,
-0.0829143822,
-0.0006362413,
-0.0014381157,
0.0205702689,
0.0106676724,
-0.0039515248,
0.0036436985,
-0.0321774259,
-0.0411908887,
-0.0292872339,
-0.0234922525,
-0.0301916134,
0.0299650896,
-0.0093280915,
0.0097689591,
0.0081276307,
0.07093779,
-0.0382945649,
-0.0487037562,
0.0204017758,
-0.0229065791,
0.0156639293,
-0.0079228031,
0.0425258726,
-0.1028269827,
-0.0447805487,
-0.0112670083,
0.0255576987,
0.029990891,
0.0106192064,
0.0970165879,
0.0226350296,
-0.0525948964,
-0.0190158933,
-0.0654958859,
0.0064882548,
0.0217047259,
0.0546571687,
0.009402141,
0.0166726224,
-0.0134890927,
0.0324714147,
0.0336430594,
-0.0447377712,
0.0031926876,
-0.0100376839,
-0.0140422434,
-0.0083311023,
-0.0303072985,
-0.0193933006,
-0.0616129376,
-0.0287558232,
0.0350523777,
-0.0231635291,
-0.0245101489,
0.0158826429,
2.506e-7,
0.051949393,
0.008822294,
-0.0357034765,
0.0063436888,
0.0042137778,
0.0314116962,
-0.068079941,
0.0398862176,
-0.0059455563,
0.0105683692,
-0.0344851427,
0.0555250794,
0.0860369205,
0.0358181484,
-0.0612129383,
-0.0326327085,
-0.0104318187,
0.0124621606,
-0.0286586992,
-0.0047320877,
-0.0112534733,
0.0520436838,
-0.0019585451,
-0.0204217788,
-0.0116347829,
-0.0219213795,
0.0880350471,
-0.0417354293,
-0.0227648783,
0.0059462534,
-0.0071793818,
0.0203092191,
-0.0197510775,
0.0670917481,
-0.0363242701,
0.0204415601,
0.0456053726,
0.0437108278,
0.010675366,
-0.0071170568,
0.0089016436,
-0.0178369712,
-0.039791666,
-0.0432763435,
0.0887816176,
-0.0519935861,
-0.0268815141,
-0.0251848679,
-0.0138914073,
0.0530885756,
0.0622009486,
-0.0262228493,
-0.0416877009,
0.0187475979,
0.055682499,
0.0117535936,
-0.0628768429,
-0.0162774399,
-0.0040660715,
0.044714842,
0.0008421899,
-0.0441209711,
-0.0106876995,
-0.0550830811,
0.0045164619,
-0.0559975952,
0.0257678255,
1.815157165e-34,
0.0144381849,
0.0037164015,
0.0275744125,
-0.0207734481,
0.0170029979,
-0.0553707965,
-0.0259661805,
0.0081922207,
-0.0260133222,
-0.0129969707,
0.0191127174
] |
http://arxiv.org/abs/2402.02084v1 | http://arxiv.org/pdf/2402.02084v1 | Revisiting the Markov Property for Machine Translation | 2024-02-03T08:50:50 | 2024-02-03T08:50:50 | Cunxiao Du, Hao Zhou, Zhaopeng Tu, Jing Jiang | cs.CL | cs.CL | In this paper, we re-examine the Markov property in the context of neural
machine translation. We design a Markov Autoregressive Transformer~(MAT) and
undertake a comprehensive assessment of its performance across four WMT
benchmarks. Our findings indicate that MAT with an order larger than 4 can
generate translations with quality on par with that of conventional
autoregressive transformers. In addition, counter-intuitively, we also find
that the advantages of utilizing a higher-order MAT do not specifically
contribute to the translation of longer sentences. | Revisiting the Markov Property for Machine Translation: In this paper, we re-examine the Markov property in the context of neural
machine translation. We design a Markov Autoregressive Transformer~(MAT) and
undertake a comprehensive assessment of its performance across four WMT
benchmarks. Our findings indicate that MAT with an order larger than 4 can
generate translations with quality on par with that of conventional
autoregressive transformers. In addition, counter-intuitively, we also find
that the advantages of utilizing a higher-order MAT do not specifically
contribute to the translation of longer sentences. | [
0.0068786154,
0.0397073291,
0.0129461819,
-0.0116474768,
-0.0288536176,
0.0042321738,
0.0407947935,
0.0160665307,
-0.0608430021,
-0.0017271506,
0.0064677238,
-0.0268428661,
0.0495837294,
0.0069716354,
0.0284515899,
-0.0563636348,
0.0121160839,
-0.0039071501,
-0.0667471439,
0.0135169784,
-0.0256385896,
-0.0211606678,
-0.0322759114,
0.0528889969,
-0.0607672371,
0.014092138,
0.0030737363,
0.0082740802,
0.0541673638,
-0.0233729742,
-0.0365718529,
0.0266155768,
-0.0113844201,
0.0462526008,
0.0000019105,
0.0047029173,
0.0121368431,
-0.0005388986,
0.050823193,
-0.0409076884,
-0.0146206534,
-0.0580086596,
-0.0342396386,
0.0280233379,
-0.0404612385,
-0.0386916213,
0.0513667129,
0.0316328816,
0.0582666844,
0.1221283749,
-0.037598405,
-0.0658286586,
0.0480560139,
-0.082406126,
0.0060913782,
-0.0521410741,
-0.0242618583,
0.0132650398,
-0.0396786742,
-0.0167349391,
-0.030775601,
0.0841908678,
0.0211130884,
0.0044293306,
0.0224058554,
0.0282627847,
0.0237768833,
-0.0120412633,
-0.0364232771,
0.0605475381,
-0.0269606318,
0.0194343384,
-0.0103104105,
0.004356158,
-0.0174251273,
-0.0112643456,
-0.0264217407,
0.0459451564,
0.0178350825,
0.0412780009,
-0.0143557452,
0.0841353238,
0.0144256502,
-0.0315051824,
-0.0739654154,
0.0543962233,
0.0297742859,
-0.0166203007,
-0.0313912705,
0.0234675128,
0.0465475246,
0.0032929005,
0.0421435125,
0.0201010089,
0.0306338072,
0.0103821196,
-0.0419104695,
-0.0086284811,
-0.0231605172,
-0.0312116053,
0.0712680444,
0.0549734235,
0.0185034368,
0.0287638251,
-0.0296294503,
0.0804720446,
-0.0760370344,
-0.0183041226,
-0.0591463894,
0.0027745331,
-0.0485436544,
0.0413970873,
-0.0571440831,
0.0006670172,
0.0047395825,
-0.0163603537,
-0.0456122123,
0.0272056535,
0.031583596,
-0.0293425117,
0.0245456975,
-0.0460235253,
-0.1068259403,
-0.0425662622,
-0.0274471175,
0.0727881715,
0.0127860941,
-0.0467014685,
0.0405401289,
-0.0567937046,
-0.0238146354,
0.0201504063,
-0.0099402992,
-0.0486430191,
-0.0158366095,
0.092841506,
-0.0155677823,
-0.0255752746,
0.0334159434,
0.0095380591,
-0.0268497597,
-0.0347708762,
0.0270041004,
-0.0260420237,
0.0232002065,
0.0094030211,
0.0250982903,
0.0019637148,
-0.000814587,
0.0175725278,
-0.0146174869,
0.0322208442,
0.0534345321,
-0.0304973945,
0.0098076817,
-0.0254074764,
0.014113633,
-0.0017437332,
0.0568899773,
0.0379301719,
-0.0063264226,
-0.0022768315,
0.0226166677,
0.0276673418,
-0.0134910457,
0.0116026476,
-0.0137205096,
-0.0265230834,
-0.0192493778,
0.0332577303,
0.0403457545,
0.0776977837,
-0.0166201852,
0.0167356301,
0.0643052757,
0.0358135365,
0.0074519007,
0.0677357391,
0.0493161641,
-0.0251976419,
-0.022029331,
0.0117469514,
-0.014839231,
0.060276188,
0.0105676288,
-0.009974204,
-0.0032620819,
0.0347299092,
-0.0214770194,
-0.0330574214,
0.0059914007,
-0.0145652359,
0.0272478797,
-0.0618116744,
0.0182305574,
0.0282014981,
-0.0489693061,
0.0596789792,
-0.0499601476,
-0.0599342845,
0.0084511833,
-0.0802801028,
-0.0647009015,
0.0196694843,
0.0346768945,
0.013692311,
0.0034764893,
-0.0470556654,
0.0092816306,
0.0516673625,
0.0670208111,
0.0663110092,
0.0254287552,
-0.0394380502,
-0.0243469439,
0.0042712921,
-0.0019308334,
0.0363934301,
0.0052830772,
0.0706108063,
0.0452918932,
0.012922788,
-0.023279639,
0.0323926806,
0.0198195055,
-0.0297918655,
0.0580728836,
0.0221208297,
0.0376973674,
-0.011459955,
-0.0116408523,
0.0202773735,
-0.0050319633,
-0.060328193,
-0.0213223249,
0.0731457844,
-0.0977940261,
-0.026134409,
0.0214015115,
-0.0781186298,
-0.0057848357,
0.0356385298,
-0.0280195102,
-0.0607408285,
-0.0034500468,
-0.0539900698,
0.0150743639,
-0.033663433,
0.0209414586,
0.0459323302,
0.0462418012,
-0.033777222,
-0.0100354338,
-0.0183109306,
-0.0423124097,
-0.0223991144,
-0.0458425619,
0.0265822336,
-0.0241454542,
-0.020499438,
0.0295650084,
0.0146218752,
0.0306764077,
0.048368834,
0.0064520757,
-0.0068442384,
-0.056936428,
0.0098159881,
0.0050812769,
-0.0370456837,
0.0235943776,
0.0416438915,
0.0597866364,
0.0357256681,
-0.0034254391,
-0.0544399954,
0.0296170991,
0.0041599171,
-0.0080985427,
0.0355650224,
0.0084208436,
0.0228334069,
-0.0025729858,
-0.0385511294,
0.0152579714,
0.0521884933,
-0.0094667925,
-0.0127311954,
-0.0839852169,
-0.0069515333,
0.0118771642,
0.0177407637,
-0.0171966776,
-0.0783831999,
0.0076704356,
-0.0567954108,
0.0743166879,
0.0364289656,
0.0499394909,
0.040890038,
0.0275770631,
-0.0198899321,
-0.0123792905,
0.0255386792,
-0.0471941829,
-0.0119887628,
-0.0288164318,
-0.0474116318,
0.0091887796,
0.0609864853,
0.0101245493,
0.0141829578,
-0.061513897,
-0.0616339557,
0.0150465732,
-0.0023526195,
0.0492662974,
0.0291565191,
-0.0189171769,
0.0044436129,
-0.0084950114,
-0.0361701027,
0.0607706048,
-0.031824477,
0.0110808909,
-0.0269674193,
-0.0160006844,
-0.0027849306,
0.0079924101,
0.0067086164,
0.0112987226,
-0.0342427827,
-0.0261575561,
0.0011176347,
0.0038794028,
0.0048684441,
-0.0089521389,
0.0372874141,
-0.0215460137,
0.0056799185,
-0.0227346644,
0.0088757221,
-0.0252466016,
-0.0075365785,
-0.0038762297,
-0.0142869921,
0.0369814001,
-0.0308176018,
-0.0688570589,
0.0436972193,
0.0213421416,
0.0693335235,
-0.0193081144,
-0.021886386,
-0.0181757454,
0.0525118001,
0.0013524607,
-0.0003268158,
0.0423168615,
-0.0260136537,
0.0195081569,
-0.0048175193,
0.018330181,
-0.019016603,
-0.0826745182,
-0.0294595305,
0.0078679072,
-0.0079839891,
-0.0061179027,
0.0002822465,
0.0241746046,
-0.038406983,
-0.0116599994,
0.0667874143,
-0.0104437126,
-0.0527750812,
-0.0202200972,
0.0477003604,
0.038624987,
-0.0044636284,
0.0246260054,
-0.0544365868,
-0.0141024534,
0.0031043019,
0.0049873856,
0.0402541533,
-0.0046150931,
-0.0134251192,
-0.0493411496,
-0.0015151778,
0.0183296092,
-0.0497140326,
-0.0692618415,
-0.0033016407,
-0.0435902402,
-0.0487326235,
0.0361634158,
0.0035447909,
-0.0156077873,
0.0030339968,
0.0093499636,
0.0098292483,
-0.0360682458,
-0.0221465658,
-0.0297952984,
0.0026485515,
0.0980514884,
0.0291582197,
-0.0170488972,
-0.0676366463,
0.0297646802,
0.0135550639,
0.0501731932,
0.0178858247,
-0.0138662346,
0.0197128262,
-0.0477583855,
-0.0048347111,
-0.0198935159,
-0.0094075613,
0.0345162377,
-0.0097025381,
0.0673297793,
-0.0441441014,
-0.0249238089,
-0.0019852854,
-0.0134891802,
-0.0136337951,
0.0348820798,
0.0152377561,
-0.0223207045,
-0.0513363108,
-0.0204039719,
0.0362444036,
-0.0701905414,
-0.0224633496,
-0.0072761178,
-0.0132715832,
-0.0586405955,
-0.0890968516,
-0.0568656251,
-0.0177563485,
0.0776822791,
-0.0074208458,
-0.0312877074,
0.0402626991,
0.0076409089,
-0.0279290378,
0.0113186305,
-0.0449289605,
0.0067297854,
-0.0784059092,
0.0381876342,
0.0310246702,
-0.0006604959,
0.0169117805,
-0.0034993882,
-0.1137382612,
-0.0733101517,
-0.0140081951,
0.0777836069,
0.0213135928,
0.032528013,
-0.0051701167,
-0.0024671196,
0.0382708572,
-0.0169200841,
0.0057674628,
0.0081531312,
0.0248520225,
-0.0169247892,
-0.0169312991,
0.0214532893,
-0.0036084501,
0.0241387859,
-0.0371470675,
0.0614045188,
-0.0141570913,
-0.0386974216,
0.0137962792,
-0.0013139209,
0.0555796847,
-0.037101116,
0.0175613128,
0.0053453962,
0.0196763054,
-0.0097598843,
0.0327886567,
-0.0465477332,
0.0254334901,
0.0577966198,
0.0000062758,
-0.0104687633,
0.0134397792,
-0.0157570392,
-0.041212976,
0.0197351985,
0.0210614465,
-0.0543442853,
-0.0139971478,
0.0169618074,
0.0344293788,
-0.0134502454,
0.0193486437,
0.0391988158,
-0.0079837283,
-0.016320968,
0.033835534,
0.0069378628,
0.0117813163,
-0.0191668198,
0.0176402256,
-0.0728592053,
-0.0037891422,
0.0000493016,
-0.0083449362,
0.0660027117,
0.0098435711,
0.1097710729,
-0.0204579476,
0.0143064354,
0.019763235,
0.0552815944,
0.0020360625,
0.0199088827,
-0.0064721452,
-0.0116625857,
0.0158422925,
0.0186490063,
0.0306675546,
-0.0330878608,
0.0051164986,
0.0364925005,
0.0137415612,
-0.0319001637,
0.0061248466,
-0.048763644,
-0.061467547,
0.0234753396,
-0.0282496717,
0.0100852316,
-0.0267519578,
0.0179310031,
-0.0072229821,
0.0227154475,
-0.0292515326,
0.0180821177,
0.0460150987,
-0.0101843039,
0.0155130476,
-0.0305888485,
0.005122235,
0.04189061,
-0.0197580587,
-0.048768688,
-6.10138102e-33,
-0.018823456,
0.0003569016,
0.0250669774,
-0.004501462,
-0.0268806554,
-0.0195198264,
-0.0144232288,
-0.0191182606,
-0.026891157,
0.0194696486,
-0.0182258617,
0.0260457397,
0.024198221,
0.0019945758,
0.020758817,
0.0344187319,
0.0748445019,
-0.0316891,
-0.0153425895,
0.0239710379,
0.0097159231,
0.0316925943,
0.0789341182,
0.0026271546,
-0.04056032,
0.0052217604,
-0.0229099579,
-0.0004260514,
-0.0005024697,
-0.0171086937,
-0.0077944477,
0.0113588376,
0.0330863409,
-0.0278592464,
-0.0121508269,
-0.0332778394,
-0.0967913792,
0.0036015471,
-0.0249046441,
0.0547471382,
-0.0090916976,
-0.0354927294,
0.0497389585,
-0.039815709,
-0.0206134599,
-0.0348099619,
0.0118440688,
-0.0262217522,
-0.0367340371,
-0.0466290042,
-0.0256276038,
0.0136167435,
0.0009938867,
-0.0309115369,
0.0390634611,
0.0612948984,
-0.0437079221,
-0.0133089321,
-0.0583264083,
0.0196390599,
-0.0345300622,
0.0526750684,
0.0332728997,
0.0222087651,
0.0018246969,
-0.0124423606,
0.0310570169,
0.0158903226,
0.0212095436,
0.0455565639,
0.015557684,
0.0076272748,
0.0637320951,
-0.018163709,
0.0664543062,
-0.0152300615,
-0.0405307412,
0.0319481567,
-0.0021858101,
0.0550689995,
0.013456394,
-0.0152314203,
-0.0219486635,
0.0258653164,
-0.0132495733,
-0.0372069217,
-0.0152681284,
-0.0176308937,
0.0047078202,
0.0469049215,
0.0081613595,
0.013322087,
0.0869488269,
-0.0127852652,
0.0005873151,
0.0133683262,
-0.0446414016,
-0.003183997,
-0.0456326269,
0.0178467501,
0.0020980595,
-0.0393198319,
0.0027622045,
0.0415175892,
0.0017291154,
-0.0204478577,
0.0090802601,
0.0502213761,
-0.0625755265,
-0.0279892348,
-0.027192466,
0.049224589,
0.0274168029,
0.004636988,
0.0141400564,
-0.0405627042,
-0.0009189434,
0.0547390543,
-0.0182164833,
-0.0232564323,
0.0106449369,
0.0299887825,
0.0394949652,
0.0463127643,
-0.0467515551,
0.0366970599,
-0.0349769518,
0.0752860606,
0.04684655,
-0.0578574538,
-0.0494788401,
0.026675038,
2.634e-7,
0.0340350904,
0.0662192926,
0.0346458815,
-0.0014807223,
0.0044746599,
-0.0024170543,
-0.0466113687,
0.0126642715,
-0.0095894141,
0.0722200498,
0.0144028747,
-0.0040728007,
0.0926988795,
0.0193775818,
-0.0402204394,
-0.0329982527,
-0.0668835714,
0.0183598641,
-0.03484907,
-0.0257424619,
0.0423803516,
0.0612455271,
-0.0294360053,
0.0215051044,
0.0227703694,
-0.0143950023,
0.0588494539,
-0.0043519791,
0.0234718788,
-0.0019265572,
0.0570920035,
0.0072043273,
-0.0430183597,
-0.0457816683,
-0.0276864562,
0.0103861392,
0.010022114,
0.0588981248,
-0.0188853033,
-0.0474239402,
0.0374437533,
-0.0024672735,
-0.0388211869,
0.0014841883,
0.0256366543,
-0.0940468311,
-0.004167757,
-0.0085562244,
-0.0692017078,
0.0176779013,
0.0475554802,
-0.0174000859,
-0.0205014162,
-0.005250724,
0.0443347506,
-0.011369844,
-0.0049837246,
-0.0705523267,
0.0004990683,
-0.0086441822,
-0.0257795006,
-0.0150472466,
-0.0437449887,
-0.0422459915,
0.0324053541,
0.0333908647,
-0.0407880172,
1.863143967e-34,
0.044367075,
-0.0238922443,
0.0261150654,
-0.0163401607,
0.0043945801,
-0.0142306518,
-0.0126475906,
-0.0021683103,
-0.014786981,
-0.087955378,
-0.0673424602
] |
http://arxiv.org/abs/2402.02082v1 | http://arxiv.org/pdf/2402.02082v1 | GliDe with a CaPE: A Low-Hassle Method to Accelerate Speculative Decoding | 2024-02-03T08:44:11 | 2024-02-03T08:44:11 | Cunxiao Du, Jing Jiang, Xu Yuanchen, Jiawei Wu, Sicheng Yu, Yongqi Li, Shenggui Li, Kai Xu, Liqiang Nie, Zhaopeng Tu, Yang You | cs.CL | cs.CL | Speculative decoding is a relatively new decoding framework that leverages
small and efficient draft models to reduce the latency of LLMs. In this study,
we introduce GliDe and CaPE, two low-hassle modifications to vanilla
speculative decoding to further improve the decoding speed of a frozen LLM.
Specifically, GliDe is a modified draft model architecture that reuses the
cached keys and values from the target LLM, while CaPE is a proposal expansion
method that uses the draft model's confidence scores to help select additional
candidate tokens for verification. Extensive experiments on different
benchmarks demonstrate that our proposed GliDe draft model significantly
reduces the expected decoding latency. Additional evaluation using walltime
reveals that GliDe can accelerate Vicuna models up to 2.17x and further extend
the improvement to 2.61x with CaPE. We will release our code, data, and the
trained draft models. | GliDe with a CaPE: A Low-Hassle Method to Accelerate Speculative Decoding: Speculative decoding is a relatively new decoding framework that leverages
small and efficient draft models to reduce the latency of LLMs. In this study,
we introduce GliDe and CaPE, two low-hassle modifications to vanilla
speculative decoding to further improve the decoding speed of a frozen LLM.
Specifically, GliDe is a modified draft model architecture that reuses the
cached keys and values from the target LLM, while CaPE is a proposal expansion
method that uses the draft model's confidence scores to help select additional
candidate tokens for verification. Extensive experiments on different
benchmarks demonstrate that our proposed GliDe draft model significantly
reduces the expected decoding latency. Additional evaluation using walltime
reveals that GliDe can accelerate Vicuna models up to 2.17x and further extend
the improvement to 2.61x with CaPE. We will release our code, data, and the
trained draft models. | [
0.0250107329,
0.0718344226,
0.0336526036,
0.0159126054,
-0.0255210809,
0.0087615708,
-0.0499460399,
0.0282281768,
-0.0602520332,
-0.0048232959,
0.0080530215,
-0.031787999,
-0.0013815279,
0.0173686296,
0.0088339662,
-0.0107137728,
0.0232007615,
-0.034923967,
0.0290500913,
-0.0107335774,
-0.0527487807,
0.0277571231,
-0.0034686679,
0.04560294,
0.0325783342,
-0.006561317,
-0.0603688322,
0.0381684043,
0.0033585401,
-0.07869827,
-0.0219664276,
0.0698733106,
-0.0179698095,
0.0170368478,
0.0000021182,
-0.0040339483,
-0.0415261872,
0.0261074584,
-0.0707658455,
0.0459846854,
0.0557078198,
0.006071229,
-0.0435777046,
0.0215666257,
-0.0118611399,
-0.045964364,
0.0257697068,
0.0059877983,
0.003862089,
0.0769726783,
-0.0024676127,
-0.0248132069,
0.0202637054,
0.0202577412,
-0.0147018787,
-0.0962076262,
0.0314207673,
-0.0254285261,
-0.0333835706,
0.0072545656,
-0.0011575196,
-0.0038148121,
0.0376353562,
0.0207257401,
0.0073806844,
0.0230170805,
-0.0189521834,
0.0015674796,
-0.0471613295,
0.0283460896,
0.0012817585,
-0.0693466812,
-0.0024808131,
0.018957084,
-0.0202112254,
-0.0407528579,
-0.0200160965,
-0.0027638329,
-0.0444713421,
0.0129146893,
0.040591903,
0.0195166189,
0.0548167266,
-0.0057222513,
-0.0236463714,
0.0649181753,
0.03696955,
-0.0034878021,
-0.0239814073,
0.0120880753,
0.0229884274,
-0.0692146271,
0.0183554683,
-0.0014540438,
-0.0302584339,
-0.028655231,
-0.0515212677,
-0.0228612293,
0.0060647316,
-0.0222987738,
0.0572983213,
0.074750334,
-0.0246532783,
-0.0104494179,
-0.0442143157,
-0.0186599791,
-0.0384329781,
0.0055418378,
-0.0208523199,
-0.0068520615,
-0.0059682429,
-0.0057702432,
-0.0361124016,
0.043915946,
-0.0277333949,
-0.0178025663,
-0.0359604657,
0.009928396,
0.0570722446,
0.0222945437,
0.0552209392,
-0.0239821877,
-0.0084736208,
-0.017637223,
0.0161512773,
0.0044691213,
-0.0112503013,
0.0148054743,
-0.0123349782,
-0.0729574412,
-0.0174576994,
0.0252886377,
0.0016788319,
-0.0112792589,
0.0029889564,
0.0519007854,
-0.0209001135,
-0.0375369415,
-0.0144487042,
-0.0235483348,
0.03328887,
-0.0286946539,
0.0051433868,
-0.0545466505,
-0.0005118837,
0.0174196605,
0.0051888344,
-0.0003903981,
-0.0381742008,
0.0046089157,
0.0146019785,
0.052906964,
-0.0170250498,
-0.0059539685,
-0.0178162977,
-0.0000450794,
-0.0259139817,
0.0888433382,
-0.0068217842,
0.0577801988,
-0.0043727853,
-0.0248661544,
-0.022633113,
0.0627987608,
0.0462819375,
-0.0145290084,
0.0036954703,
0.0078208325,
0.0326305628,
-0.0227582864,
0.0318854302,
0.024492342,
-0.0579750538,
-0.0142148007,
0.037043184,
0.0739633515,
0.0087136533,
0.0112177674,
0.0091794543,
0.0302041546,
0.024949912,
-0.0331406556,
-0.031317763,
0.0286604166,
-0.0431255996,
0.013119326,
0.0375192016,
0.063954778,
-0.0156504177,
0.005444041,
-0.0296961982,
-0.0253994279,
-0.0084165512,
-0.0060093701,
0.0411165394,
-0.013047847,
-0.0940303877,
0.0819982663,
-0.0147762662,
-0.0557569116,
-0.062963143,
-0.0228613745,
0.0115063218,
0.0769622847,
0.0730613172,
0.047877267,
-0.003679652,
0.0162826367,
-0.0258879811,
0.0210754108,
0.0276864544,
0.0017054774,
0.02271243,
-0.0427717008,
-0.0351192579,
0.0742207244,
0.0305189081,
0.0509343222,
0.0705678612,
0.0075502237,
-0.0081603583,
-0.0279597435,
0.0117963562,
0.017130211,
0.0605060011,
0.0167770721,
0.0276564583,
0.0010640848,
0.0584696718,
0.0234197173,
0.0113795195,
0.0063062678,
0.0539873615,
-0.0423870645,
-0.0085056815,
0.0198674016,
-0.0281259753,
-0.0273990221,
0.0406025648,
0.0185074564,
-0.0234034266,
0.0357081592,
-0.0117422882,
-0.0171495024,
0.0065270988,
-0.0114520015,
0.0485030748,
-0.0420360416,
0.0414175428,
-0.006837727,
0.0215888135,
0.0065710405,
0.0312075187,
-0.0688840598,
-0.0101511413,
-0.0193583053,
0.0120334364,
-0.0013462778,
-0.0393953212,
0.0194768701,
0.0407638438,
-0.0353360437,
0.0062324489,
-0.0272912048,
0.0203639288,
-0.0030144488,
-0.0307464767,
-0.0907470286,
-0.0077011148,
-0.0450830534,
0.0579835624,
-0.0103521887,
-0.0090140095,
-0.0180284176,
-0.0180285051,
-0.00324614,
-0.0122333337,
-0.0257219318,
-0.0192296468,
-0.020242041,
-0.0226002913,
-0.0750365779,
0.0211983994,
-0.0080815786,
0.0221326258,
0.0494201221,
-0.0643324628,
-0.0404505618,
-0.0438161567,
0.0119509092,
0.052588284,
-0.0420522913,
0.0196828227,
-0.0406826884,
0.034764003,
-0.0244900025,
-0.0122168465,
-0.0068051922,
-0.0016026751,
-0.0056618415,
-0.0058548837,
0.0092146778,
-0.0411332808,
-0.004753307,
0.0127882268,
0.0002059635,
-0.0734716877,
-0.0297519602,
0.0191081446,
0.0326043218,
-0.0078229168,
0.0145703349,
-0.0646661893,
-0.0191699583,
0.0486820824,
-0.0505071357,
0.0073763444,
-0.0073038158,
-0.0131491506,
0.0332732201,
-0.0185690671,
-0.046795398,
0.0307658948,
-0.0801088735,
-0.0150420498,
-0.0141144656,
0.0176070798,
-0.0244501121,
0.0188618004,
-0.0444216207,
0.0041289791,
-0.0518311448,
-0.061417941,
0.0460580289,
-0.0149333831,
-0.0046478934,
-0.0689458996,
0.0277254041,
-0.0096843261,
0.0375643,
-0.0305001289,
-0.0281503089,
-0.029811576,
0.0122871436,
-0.0624258034,
-0.025413068,
0.0039073019,
-0.0627371669,
-0.0068283593,
0.0259286687,
-0.0822103992,
0.0450167619,
-0.0167417601,
0.0100423144,
-0.0275031012,
-0.0051852474,
-0.0117812939,
-0.0066889008,
0.0131433494,
-0.0148944939,
0.0437889174,
0.0852770433,
0.0555648655,
0.0079943351,
-0.0848044232,
0.0263153147,
0.0431888327,
-0.006444653,
0.0277330447,
0.0855493024,
0.0396092907,
-0.0280539263,
0.0159804579,
0.0878212303,
-0.0023709724,
-0.058211863,
-0.0153370583,
0.0087075783,
0.0444725417,
-0.0050152335,
0.0097683286,
-0.0703588799,
-0.0435511097,
0.0690436289,
0.0472781248,
-0.0065690302,
0.0070534036,
-0.0212758388,
-0.070789516,
0.0745455623,
0.0262553878,
-0.0223406032,
0.0286030322,
-0.0625176728,
-0.0654149875,
0.0345521569,
0.0388299972,
0.0365597159,
0.0561903752,
-0.0143494094,
0.0086872699,
-0.0259695258,
-0.0674134046,
0.0141610578,
-0.0141831376,
0.0639461502,
0.0478145853,
-0.0219479371,
0.0389019661,
-0.0084533682,
0.0017851674,
0.0069635566,
-0.0460805073,
-0.0369891189,
-0.0133207096,
-0.0149995703,
0.0157173388,
0.0126693752,
0.003678865,
-0.0462773517,
-0.0256937109,
-0.0338203423,
0.031986177,
-0.01503886,
0.0009504328,
-0.0116355401,
-0.0166491475,
0.0004913091,
0.0612579323,
0.0130604692,
0.0182379168,
-0.0181433465,
-0.0442046598,
-0.031935975,
-0.1037220731,
-0.0286623165,
0.0130380606,
0.0442975909,
-0.0222238358,
-0.0153411571,
0.0296149123,
0.0038343277,
0.0547245331,
-0.0306824092,
0.069913879,
-0.025061978,
-0.0102641163,
-0.1071662158,
-0.0080970461,
-0.0309844892,
0.0222885236,
-0.0764719173,
0.0169602763,
0.0171673242,
-0.0248591769,
-0.0160038825,
0.0414660089,
-0.0897833779,
-0.038924668,
-0.0008420176,
0.0367898643,
0.034891732,
0.0091142608,
-0.0510380119,
-0.0418067947,
0.0442457683,
0.0379602611,
-0.0404399373,
0.035347309,
0.0085939029,
-0.0356356464,
0.0144580994,
0.0067700921,
-0.0288506914,
-0.0211553518,
-0.0104666715,
0.032255318,
0.0183387622,
-0.0161782242,
0.0272677131,
0.0354864039,
-0.0071383654,
-0.0399535522,
-0.013164104,
0.0452775396,
0.0818033218,
0.0006870007,
0.0276084337,
0.0127193313,
0.0161368046,
0.0069876378,
-0.0525671206,
0.0100739179,
-0.0088488255,
0.0159058142,
0.0271823686,
0.0277671423,
0.0040219249,
-0.0215832535,
0.0361385271,
0.0553249829,
-0.048282966,
-0.0026631306,
-0.0180280358,
0.0111921476,
0.0080619948,
-0.0353605822,
-0.0371369384,
0.0123455385,
0.0303384084,
-0.0291225091,
0.0335069112,
-0.0361324847,
-0.0370789804,
-0.031022124,
0.0297331233,
0.07284569,
0.0435122289,
0.0317423753,
-0.0307804905,
0.0128554907,
-0.0408591926,
0.0046284855,
-0.0101904292,
-0.019295672,
0.0453118347,
0.0054226299,
-0.0127852997,
0.0064596254,
0.0166129582,
0.0079343803,
0.0735331923,
-0.0191609822,
0.0153946215,
-0.0406967774,
-0.0547689646,
-0.0092398142,
0.0289169271,
-0.0598850735,
-0.0143765593,
-0.0159372073,
-0.0173603017,
0.0522684827,
-0.010046307,
0.024837682,
0.0146991387,
0.0090202,
0.0567756258,
-0.0038727159,
-0.0482239947,
-0.0062421658,
-0.0461403392,
0.0372913927,
-0.0444876477,
0.0229609143,
-6.841969277e-33,
-0.0062651462,
-0.0022827939,
0.0146103892,
-0.0417894572,
-0.0123611419,
-0.0014253858,
-0.0261014998,
-0.0271753855,
0.035914693,
-0.0163375903,
-0.0189330541,
0.0313586518,
0.0563388541,
0.0150530674,
-0.0078743724,
-0.0010332416,
0.0843534991,
-0.0204595756,
-0.036629077,
-0.0255115703,
0.0249277912,
0.0254815761,
0.0154712396,
-0.0050365198,
-0.0203046575,
-0.0058817174,
-0.0419032015,
-0.0371710435,
0.0055078045,
0.0609981157,
-0.0113810226,
-0.0255093146,
0.0058217426,
-0.0031426519,
-0.0290999599,
-0.0207478795,
-0.0953616798,
0.0516161881,
0.0313918591,
0.0541087203,
0.0335100368,
-0.0898479149,
-0.0317227282,
0.0343469977,
-0.0534298569,
0.0081955176,
-0.0585653409,
-0.0237644725,
-0.0427021012,
-0.0515547059,
0.0136006214,
0.0341470279,
0.0152707724,
0.0391187407,
0.0925706029,
-0.0238723308,
-0.0213885326,
0.0092291906,
-0.0516898297,
0.0173589587,
-0.0309129227,
0.041966524,
0.045741234,
-0.0244288389,
-0.0379674956,
0.0465319157,
0.0464337766,
-0.0534059368,
-0.0438235439,
0.0320790894,
0.0385752693,
0.0214393027,
0.0115295742,
0.0129324542,
0.0683165118,
-0.0309738293,
-0.0329299755,
0.0412888341,
0.0251975246,
0.0287276804,
0.0653786436,
0.0231634807,
-0.0330896899,
-0.0095875291,
-0.0438711196,
-0.0386123322,
-0.0318764783,
-0.0361344181,
0.0525091738,
0.0462259427,
-0.0253031682,
-0.016627105,
0.0665283948,
-0.0261789337,
0.0127633708,
0.0118329395,
0.0138397207,
0.0511409193,
-0.0219352897,
0.008314305,
0.0244744308,
-0.0348032117,
0.0550339818,
-0.0396344773,
-0.0040394329,
-0.0287426468,
-0.0229629688,
0.0215386171,
-0.0661131889,
0.0074483394,
-0.0049141613,
0.0235143509,
0.0365939923,
0.0181158558,
0.0484607071,
-0.0321344174,
0.0179351624,
0.0286642928,
-0.056643378,
0.0365822986,
-0.0076433062,
0.0164553858,
0.0351556204,
0.080615975,
0.0199199002,
0.0435541719,
-0.0394299813,
-0.0234653745,
0.0145729277,
-0.0122317215,
-0.0208185483,
0.0498136729,
2.818e-7,
0.0214952026,
0.0489943735,
0.0853714645,
0.0058459765,
0.005109366,
0.0041566142,
-0.0433772355,
0.0019716665,
-0.023396261,
-0.0227856152,
-0.0215455908,
0.0364768505,
0.0833320692,
0.0110998265,
-0.0739779323,
-0.0334866084,
0.0139813107,
-0.002554463,
-0.0451944023,
-0.0194530897,
-0.0143349422,
0.0698524266,
-0.0214339402,
-0.0357172042,
0.0561823808,
0.0288802478,
0.0073312786,
-0.0092198625,
-0.0252951458,
0.0020508191,
-0.0030427687,
-0.0076530045,
0.0171458405,
0.0026856677,
-0.0016158823,
0.0001818478,
-0.0337224975,
0.0547330379,
0.0143103078,
-0.0586837828,
0.0122250151,
-0.0213684998,
-0.0136059709,
0.0338093825,
0.0350437351,
0.0486306958,
-0.0207846723,
-0.0181086566,
-0.0399118103,
-0.0106670829,
-0.0436161123,
0.0072234585,
0.0436759107,
0.0637395009,
0.0264393762,
0.0113379294,
0.0027247733,
-0.0542385466,
-0.0382501408,
-0.0401807465,
-0.048555091,
-0.0264986325,
-0.0515277162,
0.0570759773,
0.0284140166,
-0.0399026871,
-0.0085676452,
3.02521943e-34,
0.0710059702,
-0.0154263424,
0.0056357104,
0.1167433634,
-0.0036491463,
-0.0312902667,
0.0355994292,
0.0348392464,
-0.0149327191,
-0.0598581098,
-0.0144357327
] |
http://arxiv.org/abs/2402.02080v1 | http://arxiv.org/pdf/2402.02080v1 | Translation Errors Significantly Impact Low-Resource Languages in Cross-Lingual Learning | 2024-02-03T08:22:51 | 2024-02-03T08:22:51 | Ashish Sunil Agrawal, Barah Fazili, Preethi Jyothi | cs.CL | cs.CL | Popular benchmarks (e.g., XNLI) used to evaluate cross-lingual language
understanding consist of parallel versions of English evaluation sets in
multiple target languages created with the help of professional translators.
When creating such parallel data, it is critical to ensure high-quality
translations for all target languages for an accurate characterization of
cross-lingual transfer. In this work, we find that translation inconsistencies
do exist and interestingly they disproportionally impact low-resource languages
in XNLI. To identify such inconsistencies, we propose measuring the gap in
performance between zero-shot evaluations on the human-translated and
machine-translated target text across multiple target languages; relatively
large gaps are indicative of translation errors. We also corroborate that
translation errors exist for two target languages, namely Hindi and Urdu, by
doing a manual reannotation of human-translated test instances in these two
languages and finding poor agreement with the original English labels these
instances were supposed to inherit. | Translation Errors Significantly Impact Low-Resource Languages in Cross-Lingual Learning: Popular benchmarks (e.g., XNLI) used to evaluate cross-lingual language
understanding consist of parallel versions of English evaluation sets in
multiple target languages created with the help of professional translators.
When creating such parallel data, it is critical to ensure high-quality
translations for all target languages for an accurate characterization of
cross-lingual transfer. In this work, we find that translation inconsistencies
do exist and interestingly they disproportionally impact low-resource languages
in XNLI. To identify such inconsistencies, we propose measuring the gap in
performance between zero-shot evaluations on the human-translated and
machine-translated target text across multiple target languages; relatively
large gaps are indicative of translation errors. We also corroborate that
translation errors exist for two target languages, namely Hindi and Urdu, by
doing a manual reannotation of human-translated test instances in these two
languages and finding poor agreement with the original English labels these
instances were supposed to inherit. | [
-0.0110266032,
0.0126880929,
-0.0377419703,
0.0563389473,
-0.0540374741,
0.0266330112,
0.0536226518,
0.0037963192,
-0.0395310521,
-0.0237532184,
-0.0823026896,
-0.0392402448,
0.0625345632,
-0.0630377606,
0.0047475109,
-0.0339850672,
0.0887486115,
-0.0056085014,
-0.0194014218,
-0.005261506,
-0.0219873488,
0.0165808182,
-0.0338315889,
0.0110130627,
-0.0134075386,
-0.0088514769,
0.0186147168,
-0.0117482357,
0.0512573905,
-0.0313978009,
-0.0422505997,
0.0235931817,
-0.0050103208,
0.0783513039,
0.0000018505,
-0.0471042879,
-0.0285791103,
0.0098952223,
-0.0073623094,
0.0205928348,
0.0653091446,
-0.0776282027,
-0.0070993695,
-0.0062168567,
0.0098211048,
-0.056368202,
0.0597229041,
0.0150764659,
0.0016455802,
0.0579002164,
-0.0072172349,
0.0167996157,
-0.0091624847,
-0.019598268,
0.0500609875,
-0.0595107526,
-0.0031774393,
0.0178362075,
-0.0145223867,
0.0062597911,
-0.0622364394,
0.0364405178,
0.002073382,
0.0246759821,
-0.0352171548,
0.0068901135,
-0.0435780771,
-0.0216245875,
0.0153074404,
0.0729314163,
-0.0324504524,
-0.0163541194,
-0.0471718274,
-0.0001677677,
-0.0255191829,
0.0191854127,
-0.0356369391,
0.038557034,
-0.0025369483,
0.0006805147,
0.034817379,
0.0349350385,
0.027343994,
-0.0211949199,
-0.0098127229,
0.0121086286,
0.0215591174,
-0.0460143685,
0.010008269,
0.0415407792,
0.0066739614,
-0.0120619955,
-0.0069735069,
-0.0205876715,
-0.025236642,
0.035149578,
-0.034708485,
-0.0686657503,
0.0183464326,
-0.0094775073,
0.0419107862,
0.0543078296,
-0.0352224074,
0.03185606,
-0.0376408435,
0.1057278737,
0.0292797498,
0.0237253066,
-0.0545254387,
-0.0287721939,
-0.0403692164,
0.0264180154,
-0.0019081439,
-0.0100547755,
-0.0380284972,
0.014687147,
-0.0610610023,
-0.0250452943,
-0.0103598526,
0.0295618977,
-0.007856152,
-0.0712411106,
-0.0730243176,
0.0007286809,
0.0050045066,
0.0779409036,
-0.0310825277,
0.0015577112,
-0.0206275564,
-0.0829482824,
-0.0188384932,
0.029817218,
-0.0464349687,
-0.034517426,
-0.0234301426,
0.101517871,
-0.0408643223,
-0.0160999466,
-0.0266190916,
0.0069115278,
-0.0142211998,
-0.0688420981,
0.0156645607,
-0.0562502071,
0.0564198084,
0.0295966491,
0.0316671096,
-0.0026374664,
-0.0057361382,
0.0038948872,
-0.0418520495,
0.0690987334,
0.0073945657,
-0.0064957133,
0.0015099655,
0.0131966993,
-0.0212886017,
0.086376451,
0.0181023665,
-0.0148797361,
0.0374091379,
0.0006628314,
-0.0529691689,
0.0171617139,
0.0227581821,
0.0226080026,
-0.0361161754,
-0.0137048839,
0.0496253744,
0.0410099886,
0.0198788326,
0.009184544,
-0.0341636725,
0.0314830579,
0.0259906203,
0.0556613468,
0.0106416931,
0.0222394727,
0.0288611352,
0.0085983891,
-0.0163304731,
0.0384281613,
-0.0240894929,
0.0214355178,
0.0254124664,
-0.0029102236,
0.0120086838,
0.019810576,
-0.020263955,
-0.0084622642,
-0.0207833014,
-0.0174241588,
0.0161775164,
-0.0453604162,
0.0622621141,
0.0193459876,
-0.0451011769,
0.0351641923,
0.0479472615,
-0.0390251614,
0.0409048796,
-0.0056278617,
0.0014035885,
0.046799425,
0.0001876677,
0.0141517119,
0.0073925476,
0.0332702622,
0.0219536591,
0.0523278415,
0.0964123979,
-0.0019320328,
0.0174892563,
-0.0390356183,
-0.0131619945,
-0.0108161252,
0.022814624,
0.0521437079,
-0.0271884967,
0.0852405503,
0.018261591,
-0.0658547133,
-0.060821902,
0.0247350372,
-0.0594434328,
-0.0156114176,
0.0010962366,
0.0085864197,
-0.0006800556,
-0.0060582403,
-0.0293449834,
-0.014723951,
-0.0273656547,
-0.0072129099,
-0.0028902302,
0.0335506275,
0.0034656399,
-0.0446175598,
-0.0023515979,
-0.0918218493,
-0.008638937,
0.0261614379,
-0.0317157432,
-0.0564306825,
0.0409751348,
0.0084571717,
0.0030275022,
-0.0204628762,
0.0356512628,
-0.0226390492,
0.0141104246,
-0.0228017569,
0.0300771054,
0.045181945,
0.0398032553,
-0.0124056609,
-0.0003774444,
0.0561968423,
-0.0069034938,
-0.0123968795,
0.0660569668,
0.0037938359,
0.014663008,
0.0206998289,
0.0214999057,
-0.0563251823,
0.0106756212,
-0.0642446503,
0.0146999573,
0.0000026605,
0.0448430292,
0.0282755904,
0.015203028,
0.0213825665,
0.0038653971,
0.001856355,
0.0297038909,
0.0185100008,
0.0346817039,
0.0121432366,
-0.0008493409,
0.0361475125,
-0.0066543436,
-0.0129626244,
0.006128713,
0.0435575508,
0.0102971699,
-0.0126859816,
-0.0411748141,
0.0439437032,
-0.0447258502,
-0.0323172435,
0.0070878766,
-0.0457482785,
0.0116615389,
0.0260805283,
-0.0183781032,
0.0269933268,
-0.0457224622,
-0.0151399085,
0.0386561677,
-0.040961884,
-0.0616832078,
-0.0155995814,
0.022531217,
-0.0592805371,
-0.0134489285,
-0.0371610112,
-0.0260100663,
0.0518227108,
-0.0000346409,
-0.0184949897,
-0.0799095482,
-0.0241625477,
-0.0869877711,
-0.01991136,
-0.038411092,
0.0455384254,
0.0310372617,
0.0599036068,
-0.0320830569,
-0.0118534174,
0.0376404114,
-0.0186542533,
0.0059875534,
-0.0740898475,
0.0400270708,
-0.0292497091,
0.0032125481,
-0.0483906902,
0.013090644,
-0.0570927858,
-0.0169952419,
-0.0601750761,
0.0118298,
0.0337160714,
-0.0108303083,
0.0062062591,
-0.0655604899,
0.0051918207,
-0.0484496951,
-0.0013397556,
-0.0134077966,
-0.0052105081,
0.0213512946,
-0.0019603439,
-0.0060769981,
-0.0150259957,
-0.0441918708,
0.0102319233,
0.0109706763,
0.0052558845,
-0.0172361024,
-0.0008950806,
0.0078364871,
-0.0106567647,
-0.0154817235,
-0.0402218327,
0.0228791535,
-0.069442071,
0.0339501351,
0.0099019371,
0.0311771929,
0.0145754255,
0.0025502674,
0.0030689025,
0.0259032194,
-0.0073539675,
-0.0507403463,
-0.0089662364,
0.0595811941,
-0.0136224767,
-0.0258319415,
0.0552925766,
0.0127157588,
-0.0344044492,
-0.0012543442,
-0.0476589166,
-0.0220834371,
-0.0085343365,
0.0035882667,
-0.1035454497,
-0.0139738591,
0.0038245106,
0.0291035399,
0.0173520688,
0.002610137,
-0.0141053991,
-0.1502272934,
-0.0091710771,
0.0406285487,
-0.1037082523,
-0.0046601412,
-0.0153116183,
0.006656941,
-0.0261000507,
0.0907434449,
-0.0024495223,
-0.0095246788,
0.0060053659,
0.0179865249,
0.0493254736,
-0.0649506599,
0.0159622934,
-0.0206663404,
0.0294667464,
0.0101967938,
0.0127387755,
-0.0288300291,
-0.052703701,
0.0140147712,
-0.0475037582,
-0.0225095954,
-0.0007581685,
0.0389636159,
0.0131759867,
-0.0200881362,
0.0402153097,
-0.0311235674,
-0.0060333977,
-0.0068203034,
-0.0149061484,
0.0681507066,
0.0216729883,
-0.021839533,
0.0463994406,
0.0083599938,
0.000928537,
0.0637977049,
0.0093547711,
-0.0242145378,
-0.017222926,
0.0213477015,
-0.0152395247,
-0.0673279464,
-0.0250939932,
0.0757736713,
-0.0823773369,
-0.0679678693,
-0.0423550457,
-0.0975208655,
0.0192704313,
0.0255972557,
0.0360683203,
0.0045923404,
-0.0533510484,
-0.0080261752,
0.0048077242,
0.0069695474,
-0.0269457158,
0.0287416913,
-0.0032148061,
0.0301407035,
-0.0107020065,
-0.0064178854,
0.0020384898,
0.0012582978,
-0.0500727817,
-0.0394380465,
0.0061750221,
0.0385259427,
0.01915011,
0.0169613399,
0.0004295198,
-0.0122183524,
0.0223611258,
-0.0024788962,
0.0042253905,
0.0257956311,
0.0233350471,
0.0105162719,
-0.0045209546,
0.0480300374,
-0.0319561176,
-0.0033369027,
0.0111013055,
0.0308805183,
-0.0157462787,
-0.0252940245,
0.0139201097,
0.0188701153,
0.0227319933,
-0.0410477221,
-0.0147475824,
-0.0112099955,
0.0243593026,
-0.0357001871,
-0.0120068323,
-0.0162369926,
-0.0702724978,
0.0633844063,
0.0097477771,
0.0239943136,
0.0264812987,
-0.0143106012,
0.002342809,
-0.0234116782,
0.0153026721,
0.0158308223,
0.0884434134,
-0.0221801475,
0.0450903066,
-0.0074613215,
-0.0438617282,
0.0026260405,
0.053469412,
-0.0351530723,
0.0570272058,
0.0149482908,
0.0301767997,
0.033423055,
-0.0050336812,
0.0161892828,
-0.0302783158,
0.0268934947,
-0.0266204588,
0.049484957,
-0.0030544514,
0.1055591777,
-0.0037913974,
-0.0083066812,
-0.01869254,
0.040305607,
-0.0311808903,
0.0111969793,
0.0009806765,
-0.0202120282,
0.0153394584,
0.0064769736,
0.070050858,
-0.0038838093,
0.0173872095,
-0.0224841367,
0.0191313904,
-0.0037245748,
-0.0068454025,
0.0198627822,
-0.0349588431,
0.0046635945,
-0.0294622201,
-0.0754951164,
-0.0702881515,
0.017354425,
0.0226869322,
0.0335588157,
0.0221808888,
-0.01699269,
0.0773975626,
-0.0208197981,
0.0154096717,
-0.0490245931,
-0.0136822267,
0.0116696609,
-0.0081930859,
-0.0036098561,
-5.676586749e-33,
-0.0399630591,
0.0053833555,
-0.0038936993,
-0.0410852022,
-0.0175336357,
0.0276888274,
-0.0212404337,
-0.0013797918,
0.0059773689,
0.0116571868,
0.0219389349,
0.0013547336,
0.0248400848,
-0.0071279109,
0.0459499434,
-0.0139322123,
0.0314328261,
0.0124564273,
-0.0351307057,
-0.0066196746,
0.0620930418,
0.0174000207,
0.1240399256,
-0.0566727966,
0.0828397796,
-0.0115014631,
0.0006832996,
-0.0128923031,
0.0244099051,
0.0309405532,
0.0109413136,
0.0480705425,
0.0252656043,
-0.0588703938,
-0.0109999785,
-0.0356709696,
-0.0882382244,
-0.0478317924,
0.0558692366,
0.0636334345,
-0.0255991537,
-0.0208798684,
0.0063610198,
-0.0408194959,
-0.0437424146,
-0.0242183339,
0.0085433442,
-0.0214196164,
-0.0201436505,
0.040059302,
-0.0939542577,
0.0123179322,
-0.0259860978,
0.0312719233,
0.055960957,
0.0777830631,
-0.0228071939,
-0.0063132299,
-0.0788569003,
-0.0048032515,
-0.0467296839,
0.0339119099,
0.0856399089,
0.0110221459,
-0.0502311774,
0.0141525557,
0.0627599806,
0.000980226,
0.0825330466,
0.0014629813,
0.0475422703,
-0.0164509118,
0.0249026716,
0.0973504707,
0.0470059589,
-0.0334987752,
-0.0618639365,
0.0334299132,
-0.035390906,
0.043222744,
-0.0144121237,
0.0139485952,
-0.0390574113,
-0.0233393982,
-0.0344575755,
-0.0494609661,
-0.0379928835,
-0.0467499681,
0.0151568707,
-0.0230479352,
-0.0080678361,
0.0075662122,
0.0794316158,
-0.0073707127,
-0.0115875676,
-0.0258571859,
-0.0310310218,
0.0072833654,
-0.0277607292,
0.037514206,
-0.0512018614,
-0.0717427433,
-0.0217971765,
0.0587946847,
0.0485416278,
-0.0056775808,
0.0783128589,
-0.0102361515,
-0.050981652,
0.0075941957,
-0.037864998,
-0.0089777503,
0.0001397523,
0.0252811592,
0.0316181779,
0.0132728769,
-0.0211533736,
0.0080690766,
0.0423247218,
-0.0201932415,
-0.0009611133,
0.0192104373,
-0.0068739895,
-0.0255998652,
-0.0411938876,
0.0193097554,
-0.02758497,
-0.0056697289,
0.0415771157,
-0.0462436043,
-0.0209189393,
0.023165442,
2.621e-7,
0.0202393625,
0.0373113193,
-0.0034701556,
0.0013100606,
-0.0157239996,
0.0115807317,
-0.0791739598,
0.0317082778,
-0.0091911191,
0.0247314572,
0.0107072201,
0.0332165472,
0.0972094014,
0.0463836379,
-0.0177399796,
-0.0355462842,
-0.0099461684,
0.0328033678,
-0.042663157,
0.0306574497,
0.0235968139,
0.030191198,
-0.0104784537,
0.010629666,
-0.0195401534,
-0.0221815668,
0.0800976902,
-0.0145658385,
-0.0269718748,
-0.0046443809,
-0.0073632384,
0.0368956663,
-0.0229550526,
0.0384581722,
-0.0228436124,
-0.0130856549,
0.0450745113,
0.0277242512,
-0.0006579135,
-0.011124583,
-0.0002118937,
-0.0080245612,
-0.0116281575,
-0.0188474674,
0.0470916443,
-0.0148775978,
-0.0197837148,
0.0089413207,
-0.0261132456,
0.0464477278,
0.0637456402,
-0.0042952625,
-0.0076142475,
0.0291316789,
0.0455976576,
0.0265107341,
-0.0439072698,
0.0102204802,
0.0033908617,
0.042200584,
-0.0049477937,
-0.0596415326,
0.0074247494,
0.0124458494,
-0.0036700477,
-0.0313753225,
-0.0084933974,
1.903450681e-34,
0.0507134162,
-0.0272636507,
0.0300901085,
-0.0550634488,
0.0224308744,
-0.0317616239,
-0.0134598436,
0.0018641163,
-0.0084225703,
-0.0217751618,
0.0082217278
] |
http://arxiv.org/abs/2402.02078v1 | http://arxiv.org/pdf/2402.02078v1 | Exploring the Robustness of Task-oriented Dialogue Systems for Colloquial German Varieties | 2024-02-03T08:16:43 | 2024-02-03T08:16:43 | Ekaterina Artemova, Verena Blaschke, Barbara Plank | cs.CL | cs.CL | Mainstream cross-lingual task-oriented dialogue (ToD) systems leverage the
transfer learning paradigm by training a joint model for intent recognition and
slot-filling in English and applying it, zero-shot, to other languages. We
address a gap in prior research, which often overlooked the transfer to
lower-resource colloquial varieties due to limited test data. Inspired by prior
work on English varieties, we craft and manually evaluate perturbation rules
that transform German sentences into colloquial forms and use them to
synthesize test sets in four ToD datasets. Our perturbation rules cover 18
distinct language phenomena, enabling us to explore the impact of each
perturbation on slot and intent performance. Using these new datasets, we
conduct an experimental evaluation across six different transformers. Here, we
demonstrate that when applied to colloquial varieties, ToD systems maintain
their intent recognition performance, losing 6% (4.62 percentage points) in
accuracy on average. However, they exhibit a significant drop in slot
detection, with a decrease of 31% (21 percentage points) in slot F1 score. Our
findings are further supported by a transfer experiment from Standard American
English to synthetic Urban African American Vernacular English. | Exploring the Robustness of Task-oriented Dialogue Systems for Colloquial German Varieties: Mainstream cross-lingual task-oriented dialogue (ToD) systems leverage the
transfer learning paradigm by training a joint model for intent recognition and
slot-filling in English and applying it, zero-shot, to other languages. We
address a gap in prior research, which often overlooked the transfer to
lower-resource colloquial varieties due to limited test data. Inspired by prior
work on English varieties, we craft and manually evaluate perturbation rules
that transform German sentences into colloquial forms and use them to
synthesize test sets in four ToD datasets. Our perturbation rules cover 18
distinct language phenomena, enabling us to explore the impact of each
perturbation on slot and intent performance. Using these new datasets, we
conduct an experimental evaluation across six different transformers. Here, we
demonstrate that when applied to colloquial varieties, ToD systems maintain
their intent recognition performance, losing 6% (4.62 percentage points) in
accuracy on average. However, they exhibit a significant drop in slot
detection, with a decrease of 31% (21 percentage points) in slot F1 score. Our
findings are further supported by a transfer experiment from Standard American
English to synthetic Urban African American Vernacular English. | [
0.0580638014,
0.0307533275,
-0.0044555925,
0.0321676582,
-0.0294547249,
-0.0008769131,
0.0083993897,
-0.0210110471,
0.003188167,
-0.0415591486,
-0.0763820559,
-0.0535401255,
0.0155915162,
0.0349180885,
0.0108073354,
-0.0284203049,
0.0368928686,
-0.0155229568,
0.0090898965,
0.0228060819,
-0.0071288357,
0.0000204376,
-0.0426018313,
0.055035539,
-0.0108168172,
0.0149745932,
-0.0374855734,
0.0226751957,
-0.0125106042,
0.0184190329,
0.0340884477,
0.068884559,
-0.0150250103,
0.0245781578,
0.0000019032,
-0.0730130896,
0.0022368252,
-0.045695018,
-0.0651434734,
-0.025662493,
0.0003404215,
-0.0031602818,
-0.0293791853,
0.0128459735,
-0.048397392,
-0.0248266645,
0.0572039858,
0.0500916168,
0.0326141901,
0.0798261613,
-0.0124307405,
0.0256321002,
0.0274600908,
-0.0126187475,
0.0124348523,
-0.0739405304,
0.0319193453,
0.0430015139,
-0.0025156657,
-0.0068206247,
-0.0248951372,
0.0130053908,
0.0091801211,
-0.0431974903,
0.011118955,
-0.0007246824,
-0.0131066935,
-0.0081987428,
-0.0603435636,
0.0873929411,
0.0186630636,
-0.0150594404,
-0.0282931011,
0.015058985,
-0.0040598176,
-0.0186165869,
-0.054573983,
-0.0180521179,
0.0034835325,
-0.0035115418,
0.0142670209,
-0.0098682512,
-0.0275649503,
0.0040146839,
-0.0257870108,
0.0828060433,
0.0104636448,
-0.0390571393,
0.0214949977,
0.0110699935,
0.0504135825,
-0.0813059881,
0.0228912197,
-0.0069409064,
-0.0716709644,
0.0072249994,
-0.0154011091,
-0.0122577408,
0.0019933817,
-0.0093850503,
0.0405640006,
0.043893896,
-0.057264924,
0.0266993605,
0.0317122526,
0.0570222922,
0.0078138942,
0.0242897421,
-0.0453863479,
-0.0180048961,
-0.0804388002,
0.0259655025,
-0.0410308838,
0.0108355731,
-0.0520922989,
0.0069207679,
-0.0292067062,
0.0207062848,
0.0076716668,
-0.022307843,
0.0136501091,
-0.0063504092,
-0.0406880118,
0.0029183074,
-0.0270022918,
0.0824560151,
0.0127152074,
0.043086987,
0.003265854,
-0.0584470779,
-0.0406270996,
0.0140773803,
0.0150282318,
-0.062986955,
-0.0039027822,
0.0220340174,
-0.054070875,
-0.025292797,
-0.024557218,
-0.0002925678,
-0.06410072,
-0.0876260921,
-0.0405215658,
-0.047993917,
0.0087605687,
0.0116955545,
0.0344143286,
-0.0406105705,
-0.0047868481,
0.0043588234,
-0.0456293002,
0.070106782,
-0.0091140429,
0.0165809747,
-0.0008146151,
-0.0166694727,
-0.0560320467,
0.0391311273,
0.012756452,
0.0083175935,
0.0329335928,
-0.0137546482,
-0.0181939099,
0.0163910631,
0.0097760279,
0.0060403808,
0.0557130873,
0.0106292544,
0.0164767839,
-0.0064504924,
0.0526232757,
0.071017459,
-0.0247725379,
-0.0008161129,
0.0610365756,
0.0399644896,
0.019582767,
0.0056315428,
0.0563359968,
0.072988525,
0.0079725645,
0.024300497,
-0.0542168617,
-0.0289687868,
-0.0325259119,
-0.0286464039,
0.0032181037,
0.025886938,
-0.0291641839,
-0.0044170334,
0.0013101803,
-0.0036103991,
0.0505739264,
-0.0172428992,
0.046857819,
0.0273028836,
-0.0252050757,
0.0583187304,
0.0263291784,
-0.0244439952,
-0.0315435864,
-0.0357561633,
-0.0150541943,
0.0398204997,
-0.0016508836,
0.0057868767,
-0.0708194301,
-0.0392620862,
0.029959783,
0.0124297868,
0.0733863935,
-0.0144922035,
0.0715501979,
-0.051695291,
0.0045123529,
-0.0009175855,
-0.0006507034,
0.0075144866,
-0.020054508,
0.0097424379,
0.0040800385,
-0.0292098392,
0.0170883015,
-0.0052448167,
0.0040763048,
0.0027785187,
0.0617581047,
0.015532543,
0.0120330006,
0.0377771892,
0.0125485482,
-0.0312257372,
-0.0041250493,
-0.0206968151,
0.0253967475,
0.0775892511,
-0.0645594224,
-0.0428924784,
0.0141222076,
-0.0026207708,
-0.039920494,
0.01011083,
0.0016704343,
-0.0817747265,
0.0221543517,
-0.0318730734,
0.0284895804,
-0.0687278956,
0.0067808712,
0.025690956,
0.0398763232,
-0.0171412416,
0.0004665019,
0.0405658446,
-0.0058220644,
0.060567379,
-0.0332783647,
0.0397391506,
-0.029443441,
0.0114785451,
0.0029573874,
-0.0559289493,
-0.0203326941,
-0.0026665058,
0.0357956253,
-0.0537909754,
0.0064531411,
-0.1021791324,
0.0100883776,
-0.0350271463,
0.0233680606,
0.0125414189,
0.0179871637,
-0.0097111771,
-0.0202384423,
-0.0259001441,
-0.0133130085,
0.0179860536,
-0.0059550391,
0.0196037851,
-0.000615351,
-0.011112608,
-0.0198429264,
0.0326043181,
-0.0043079313,
0.0177914854,
-0.05719034,
-0.0344053246,
-0.0415647328,
0.0424555838,
0.0257639084,
0.0047195801,
0.0050072391,
-0.0206626765,
-0.0018345303,
0.0200661384,
0.0427663811,
0.0506072864,
-0.0119051309,
-0.0059625907,
-0.0061151688,
-0.0637044609,
-0.0425709449,
0.0318322219,
-0.0008568691,
-0.0059784288,
-0.0726239607,
-0.0495649986,
0.0081766145,
0.0402806588,
0.0132107744,
0.0122962072,
-0.0616307221,
-0.0210613888,
-0.0148919746,
-0.0481027775,
0.000352013,
0.0228192359,
-0.0428964458,
0.0031152584,
-0.0297302622,
-0.0088767176,
0.0537639521,
-0.0092974035,
-0.059359163,
-0.044717107,
-0.0040538581,
-0.0307439677,
0.0322415419,
-0.0318689384,
0.0124581428,
0.0822848529,
-0.0713882297,
-0.0126644811,
0.0019033458,
-0.0057875598,
0.0021740405,
0.0593687557,
-0.0495096929,
-0.0055536712,
-0.0145323323,
-0.0774107277,
-0.0496447496,
-0.050621815,
-0.0143255806,
0.036514774,
0.021751266,
-0.0222444367,
0.018312864,
0.0034456993,
0.0028655049,
0.025790995,
-0.0176310129,
-0.0019528549,
-0.0000905326,
0.0465457253,
0.0400134549,
0.0812087879,
-0.0318055078,
0.0439732373,
0.0350757129,
0.028352309,
0.0100300247,
0.0053968532,
0.0034745114,
0.0269430336,
0.0524157695,
-0.0000356689,
-0.0280747879,
0.0152796432,
0.0967581645,
0.0073472122,
-0.0233169068,
0.0885764211,
0.023344012,
-0.011372433,
-0.0124172261,
-0.0328817181,
0.0270414744,
0.0072222487,
0.0393862799,
-0.1049725935,
-0.0508791246,
0.0258107558,
-0.004019158,
-0.0244031288,
-0.0202125106,
-0.0289829019,
-0.1258757859,
0.0408925861,
-0.0247545782,
-0.0436486751,
0.0123638064,
-0.0115164192,
0.0300775729,
-0.0332484581,
0.0330991186,
-0.0165390391,
-0.0060229106,
-0.0387828387,
0.0403617583,
-0.0099817403,
-0.0524139404,
0.0187391732,
-0.013221167,
0.086595431,
0.0003943362,
0.0319734029,
0.0123509523,
-0.0148129798,
-0.0213632099,
0.0112819504,
-0.0178663302,
-0.0207739398,
0.0299416967,
0.0097894864,
-0.0116099063,
0.0004233858,
0.0358498991,
-0.0233031269,
-0.0346080475,
0.017543193,
0.0598869175,
-0.0382857323,
0.018182002,
0.0237805899,
0.005371036,
0.0755555406,
0.0973536819,
0.049215842,
-0.0427518524,
-0.0133403791,
-0.0241414104,
-0.0235085189,
-0.0567698367,
-0.0064754309,
0.0175422393,
-0.022753315,
-0.0202580616,
-0.0263071675,
-0.0330410637,
0.0103995921,
0.0470941588,
0.0588020943,
0.0020634034,
0.0191881284,
-0.0124334265,
-0.0231439304,
0.0048801405,
-0.0217164513,
0.0593893826,
-0.0499029569,
-0.035375841,
0.0178663321,
-0.0315528028,
0.0009322803,
-0.0109656556,
-0.0804271623,
-0.0581122078,
0.0102629457,
0.0478508696,
0.0222764313,
0.0274798237,
-0.0098292595,
-0.0212654993,
-0.0287624635,
-0.0271295309,
-0.0174843557,
0.0108966166,
0.0852533653,
-0.0217969026,
0.0003233881,
0.0321210511,
-0.0002517819,
0.0344473161,
0.0169043839,
0.0350232571,
-0.0563769713,
-0.0135827092,
0.0149918795,
-0.0083688227,
0.0576183386,
0.0111806672,
-0.0583179444,
-0.0287327636,
0.0547776259,
-0.0033939856,
0.0044567082,
-0.0542783551,
-0.0116505167,
0.0112727676,
0.0024556217,
0.0043636891,
0.0058367983,
-0.0607236587,
0.0489998944,
-0.031215515,
-0.0036938409,
0.0467395373,
-0.0099090915,
-0.0084747784,
-0.002192548,
-0.0275045168,
0.0114598004,
0.0760179833,
0.0604275838,
0.0029156141,
0.0792931467,
0.0075129187,
-0.0512632281,
0.0349977538,
0.0257895924,
-0.036675483,
0.0151100596,
0.0148350941,
0.014234419,
0.0348901078,
-0.0140000461,
0.0787553862,
-0.0216971003,
0.0271079689,
-0.0152167901,
0.0325899608,
-0.0077495603,
0.0068079904,
0.0394482501,
-0.0131371636,
-0.0151124503,
0.0228634458,
0.0228219852,
-0.0231221337,
-0.0093263919,
0.0223344788,
-0.0009995381,
-0.0429112725,
-0.0296133105,
0.0349410065,
0.0107246917,
-0.0056762928,
0.0031532899,
-0.0562702864,
-0.0523662865,
0.021015713,
-0.0325230248,
0.0342114121,
-0.0151829012,
0.0139687406,
0.0429956354,
-0.0267013591,
-0.0129299341,
-0.02708284,
-0.0146110123,
0.0058688428,
-0.0184513386,
-0.037989147,
-6.270363843e-33,
-0.0011014304,
-0.0059923776,
0.0000109723,
-0.0061933016,
-0.0329963155,
-0.0102445949,
-0.0439498499,
0.0107626244,
-0.0158851109,
0.0181607064,
-0.0186736751,
-0.0245741289,
0.0351953618,
0.0427097119,
0.0106697604,
-0.0038481555,
0.0209822953,
0.0271142181,
-0.0215892084,
-0.0014901559,
0.033243943,
0.0412708111,
0.078795135,
-0.0484829955,
0.0251576137,
0.0526784882,
-0.0139570218,
0.0237001982,
0.0388608873,
0.0577164441,
-0.0001195971,
0.0705616176,
0.0266853981,
-0.0555766709,
0.0135634718,
0.0655870661,
-0.0926963538,
-0.0275089331,
0.0757329091,
0.0292846877,
0.0320757516,
-0.0201786812,
-0.0235472079,
-0.0243340805,
-0.0278001633,
0.0329650827,
-0.0028930418,
-0.0182628371,
0.0043612272,
0.0296522286,
-0.1039847061,
0.0110083288,
-0.0126378546,
-0.0753539875,
0.0012154354,
0.0761570334,
-0.0025081055,
-0.0965869948,
-0.0147891575,
0.0336926468,
-0.0361070335,
0.0079902215,
0.0375423618,
0.0263637453,
-0.0304874089,
0.0503598265,
0.0300808679,
0.0081792772,
0.0232733246,
-0.034688212,
0.0136517882,
0.048833359,
-0.0126544125,
0.0500225499,
0.0165450089,
0.0186828356,
-0.0340413824,
0.016155351,
0.0458746217,
0.0402986221,
-0.0187175237,
0.0060063913,
-0.0684901699,
-0.0467900708,
0.0003916574,
0.0061904597,
-0.017128408,
-0.045250088,
0.0560556762,
-0.0203798376,
0.0000838778,
0.0300930981,
0.0727369487,
-0.0401759073,
-0.0152695943,
0.0140248528,
-0.0417535156,
0.0061384891,
-0.0219396781,
0.0051201065,
-0.1007401645,
-0.0441647358,
0.0279680174,
0.0521985479,
0.0311052985,
0.0153174233,
0.0523907244,
0.0284346268,
-0.0424156338,
0.0006114014,
-0.0425163545,
0.0586948805,
-0.0538754575,
0.0654424652,
0.0344275087,
0.0458273031,
-0.033409752,
0.0535700954,
0.0192301124,
-0.0256692991,
-0.0489012524,
-0.0349006131,
0.0261204913,
0.0356946476,
-0.0153227104,
0.0024639778,
0.0203771628,
-0.0342838541,
-0.0367722139,
-0.0239044167,
-0.0118379006,
-0.0162452236,
2.778e-7,
0.0095412582,
0.0286022983,
0.0085464111,
0.0158441197,
-0.0169389769,
0.0371732116,
-0.028857803,
0.0362497754,
0.0373895057,
-0.000143216,
0.0178062357,
-0.0064432397,
0.0801407993,
0.0308779683,
-0.129758656,
-0.0407497659,
0.0023968397,
-0.0263076443,
-0.0388340987,
0.0190583523,
0.0536581762,
0.0766738355,
0.0102686752,
-0.0279808827,
0.0020110598,
-0.024733521,
0.0115321316,
-0.0480360501,
-0.0059638671,
0.033805076,
0.0269448236,
-0.035004098,
0.0008267342,
0.0600462593,
-0.020420406,
0.0359191485,
0.0007201664,
0.044239644,
-0.0002387357,
-0.0801402628,
-0.0614261776,
-0.0041055693,
0.0115606161,
-0.0541044585,
0.0578946508,
-0.0487361997,
-0.0345188677,
-0.0231912229,
-0.005667388,
0.0112820771,
0.0215945896,
-0.01667629,
-0.0032329294,
-0.0134172756,
0.0469471365,
-0.012947401,
-0.0065110214,
-0.0441343226,
-0.0054654591,
0.0262691956,
-0.0067232843,
-0.0439952761,
-0.0012250753,
-0.0491303019,
-0.0280596726,
-0.0313074961,
0.0133824553,
2.312278433e-34,
0.0282463338,
0.0052581187,
0.0742037892,
0.0389091112,
0.0114041176,
0.0009317466,
-0.0133862365,
-0.0040307557,
0.02468732,
-0.0525864623,
-0.0055253115
] |
http://arxiv.org/abs/2402.02077v1 | http://arxiv.org/pdf/2402.02077v1 | Investigating Content Planning for Navigating Trade-offs in Knowledge-Grounded Dialogue | 2024-02-03T08:16:39 | 2024-02-03T08:16:39 | Kushal Chawla, Hannah Rashkin, Gaurav Singh Tomar, David Reitter | cs.CL | cs.CL | Knowledge-grounded dialogue generation is a challenging task because it
requires satisfying two fundamental yet often competing constraints: being
responsive in a manner that is specific to what the conversation partner has
said while also being attributable to an underlying source document. In this
work, we bring this trade-off between these two objectives (specificity and
attribution) to light and ask the question: Can explicit content planning
before the response generation help the model to address this challenge? To
answer this question, we design a framework called PLEDGE, which allows us to
experiment with various plan variables explored in prior work, supporting both
metric-agnostic and metric-aware approaches. While content planning shows
promise, our results on whether it can actually help to navigate this trade-off
are mixed -- planning mechanisms that are metric-aware (use automatic metrics
during training) are better at automatic evaluations but underperform in human
judgment compared to metric-agnostic mechanisms. We discuss how this may be
caused by over-fitting to automatic metrics and the need for future work to
better calibrate these metrics towards human judgment. We hope the observations
from our analysis will inform future work that aims to apply content planning
in this context. | Investigating Content Planning for Navigating Trade-offs in Knowledge-Grounded Dialogue: Knowledge-grounded dialogue generation is a challenging task because it
requires satisfying two fundamental yet often competing constraints: being
responsive in a manner that is specific to what the conversation partner has
said while also being attributable to an underlying source document. In this
work, we bring this trade-off between these two objectives (specificity and
attribution) to light and ask the question: Can explicit content planning
before the response generation help the model to address this challenge? To
answer this question, we design a framework called PLEDGE, which allows us to
experiment with various plan variables explored in prior work, supporting both
metric-agnostic and metric-aware approaches. While content planning shows
promise, our results on whether it can actually help to navigate this trade-off
are mixed -- planning mechanisms that are metric-aware (use automatic metrics
during training) are better at automatic evaluations but underperform in human
judgment compared to metric-agnostic mechanisms. We discuss how this may be
caused by over-fitting to automatic metrics and the need for future work to
better calibrate these metrics towards human judgment. We hope the observations
from our analysis will inform future work that aims to apply content planning
in this context. | [
0.0823096335,
0.048282221,
-0.0034924443,
-0.0080865147,
-0.0278463513,
-0.0107298587,
-0.003869029,
-0.0141764609,
0.000257518,
-0.0430747904,
-0.0115361651,
0.0151306922,
-0.0096548824,
0.0367336906,
0.0093080839,
-0.0879712626,
0.0171715468,
-0.0065778131,
0.0187671017,
0.001483249,
-0.0034383265,
0.0133036328,
-0.0177660715,
0.0367496088,
-0.0719164014,
-0.0026125549,
-0.0461251065,
0.0138218049,
0.0209866576,
-0.0168678276,
0.015153382,
0.0574734174,
-0.0163706895,
0.0062798001,
0.0000019614,
-0.0354832262,
0.020903049,
-0.0362453349,
-0.0191401988,
-0.024876792,
-0.0228723753,
-0.0055967602,
0.0274008419,
0.0085311076,
-0.072036922,
0.045110438,
0.0015180224,
0.0651378855,
0.042656064,
0.0803395808,
-0.0197889674,
-0.0103729088,
-0.022627933,
-0.0183398668,
0.0003768924,
-0.054086823,
-0.0266759694,
0.0544010811,
0.0089041274,
0.0384433493,
-0.0119170016,
0.002601333,
0.0258690026,
-0.0567916371,
0.039620515,
0.0157828331,
-0.0159101374,
-0.0445408449,
-0.0287077576,
0.07594046,
0.0661124885,
0.0103349853,
0.0418496206,
-0.041573029,
0.0281053912,
-0.022728933,
-0.0334171467,
0.0020401408,
-0.0015847785,
-0.020475192,
-0.0654087216,
0.0169319678,
-0.0004601035,
0.0248179361,
0.0059711756,
0.0519194715,
-0.0325893983,
-0.0195380468,
0.0163439773,
0.0262285192,
0.0089490823,
-0.1112875789,
-0.0018900814,
0.0609019101,
-0.0409058891,
0.0107336845,
-0.0477223434,
0.0318162404,
0.0283571538,
0.0006134737,
0.0046857158,
0.0383634456,
0.014415672,
0.0392495766,
0.0065781218,
0.0824766904,
-0.0018031403,
-0.0050623892,
-0.0238928292,
0.0483992733,
-0.032573957,
-0.0157830883,
-0.0393459573,
0.0223993324,
-0.0798561051,
-0.0629801527,
-0.0027465587,
0.0497514196,
-0.0088716056,
0.0305791702,
0.023475213,
0.0463396162,
-0.038670741,
0.0017004285,
-0.0492935814,
0.0095854439,
-0.0878614038,
0.0194779597,
0.0468857065,
-0.0238642767,
-0.0022381844,
0.0624916479,
0.0041731447,
0.0112338644,
-0.0129942307,
0.0471325368,
-0.0398371033,
0.0022573047,
0.0150340637,
-0.0875284672,
-0.0587391183,
-0.0554555692,
0.019336354,
-0.0224914961,
0.0310802571,
0.0228573177,
-0.0285943486,
-0.0258502942,
0.0009451532,
-0.004024813,
-0.0030072001,
0.0087337112,
0.0383332148,
0.0258564595,
0.021832874,
0.0274684653,
-0.0517488644,
0.0771862119,
0.0321120247,
0.0437427312,
0.0237955786,
0.0257517416,
-0.0420567431,
-0.0193051212,
0.0293957386,
0.0082080821,
0.0301340502,
0.0189042948,
-0.0202184562,
0.0308728702,
0.0227716547,
0.0359045081,
-0.0376363173,
-0.0019789687,
0.0473631248,
0.037948329,
-0.0084849494,
0.0268670507,
0.0425306372,
0.0365821309,
-0.017274674,
0.0484757647,
-0.004720768,
0.0634560734,
0.0377136059,
-0.0019946818,
0.070748806,
0.0037611632,
-0.0397826098,
-0.0587702468,
-0.0079084504,
0.009191568,
0.0056974776,
-0.0133070042,
-0.0227640625,
0.0358244553,
-0.041403003,
0.0344849974,
-0.0483083948,
-0.0350127108,
0.0237071849,
-0.0276212581,
0.0565056279,
0.0343058594,
0.0650637448,
-0.0134119894,
-0.040185228,
0.0011577324,
0.0194592513,
0.0593369454,
0.0820827633,
-0.0409090072,
-0.0073880106,
-0.0594041981,
-0.0338974632,
-0.063449271,
0.0025153628,
0.0168529544,
0.0504937693,
0.0249735042,
0.0222652853,
0.0113665015,
0.017919058,
0.0241597798,
0.0046064435,
0.0038735399,
0.0187674314,
0.0036347541,
0.0377021059,
0.0383733734,
0.0166167468,
0.0115846265,
0.0240900163,
0.00790866,
-0.0299906041,
0.0474190302,
-0.070497714,
-0.0125717176,
0.0160965454,
-0.0401250869,
-0.0097399224,
-0.0559727997,
-0.0163331274,
-0.0654966459,
0.0400045514,
-0.0244210549,
0.003626511,
-0.0452504829,
-0.0107799647,
0.0447458327,
-0.0192179661,
-0.0133958729,
0.039001707,
0.0035796801,
-0.0078306729,
0.0587746613,
-0.0079017831,
0.0345487073,
-0.0319896191,
0.0404571071,
0.0152511988,
-0.0265340507,
-0.0074441875,
0.0528425798,
0.0187446084,
-0.0647371635,
-0.063143827,
-0.1117607579,
-0.0132917957,
-0.0378706604,
-0.0006859986,
0.0208492465,
0.0279889442,
0.0008537156,
-0.0192999654,
-0.0639673397,
-0.0204528365,
0.0322835222,
0.0039089336,
0.0313506946,
0.0356021188,
0.0752903521,
-0.0161296204,
0.0313380696,
0.0299700759,
-0.0022993919,
0.0410245173,
-0.0045808223,
-0.0423027463,
0.055429168,
-0.0658626631,
-0.006122699,
-0.0063275318,
-0.0232851543,
0.0376924016,
0.010844944,
0.0429034531,
0.0343291983,
0.0302255247,
0.0153428679,
0.0135588534,
-0.0185578857,
-0.0768435672,
0.0563557111,
-0.0492522046,
0.0061934977,
-0.0377477407,
0.0011485224,
0.0266562346,
0.0435835607,
-0.000197595,
0.0215390883,
-0.082198061,
0.0136584053,
-0.0034651852,
-0.0636376143,
0.0026742332,
-0.0269322898,
-0.0152747976,
0.0749100149,
-0.0099204564,
-0.0260786712,
0.0662607774,
0.0070435186,
-0.0089497091,
-0.0255814828,
0.0474283285,
-0.0269702002,
0.0306187924,
-0.0311490204,
0.0005778699,
-0.0544998944,
-0.0127233388,
-0.0523935594,
-0.0147423819,
0.0252644792,
0.002788428,
0.0067499271,
-0.0398576707,
-0.0065705911,
-0.0042646639,
-0.0384405553,
-0.0456687249,
-0.0064389142,
-0.0172410533,
-0.0032276218,
-0.0084504215,
-0.0294506103,
-0.0200404432,
-0.0091403825,
0.0213034395,
0.0492439978,
-0.0246528648,
-0.000715966,
-0.0101171788,
0.0389573574,
0.0401449874,
0.0530302934,
0.0341330171,
-0.0254223142,
0.0376603156,
0.0457512476,
0.0322683938,
0.01137996,
0.0268652961,
0.0010454883,
0.0022890619,
0.0162770711,
-0.0028363578,
0.035942588,
0.0782783478,
0.0017317197,
-0.0434207097,
0.0960919783,
0.0360348076,
-0.009812423,
-0.006598114,
-0.0581335835,
0.0768157691,
0.0120607866,
0.023671044,
-0.0764046907,
-0.0100513631,
0.0043168943,
0.0318776332,
-0.0234701186,
-0.0578157455,
-0.0209745187,
-0.0623494349,
0.0399022698,
-0.0186871476,
-0.0064825742,
-0.000508143,
-0.0153976372,
0.0009427509,
0.0197707079,
-0.0024554918,
-0.0112041011,
0.0277202521,
0.0161892287,
0.0262369961,
-0.0232398584,
-0.0164112225,
-0.0251809862,
0.0237696972,
0.0209740587,
0.0488348603,
0.0593361929,
0.0581175163,
-0.0097293872,
-0.0182549022,
0.0155625883,
0.0369050205,
-0.0143844504,
0.0243631341,
-0.0174938422,
0.0195582006,
-0.012338995,
0.0585881732,
-0.0074117463,
-0.0383919738,
0.0305052269,
0.0363288336,
-0.0618265644,
-0.0135888187,
-0.0251887999,
0.0582984388,
0.0425392576,
0.0886190608,
0.0555346608,
-0.0352736041,
-0.0324667506,
0.0180789605,
0.0460612774,
-0.1011517644,
0.0260611959,
-0.0121794492,
0.0091407532,
0.0394269973,
-0.1095237285,
-0.071511887,
0.0091490103,
0.0300498307,
-0.0029603059,
-0.0640856847,
-0.0300425123,
0.0026613479,
-0.0058885831,
0.0063150763,
0.0036139549,
-0.0127756782,
-0.068653293,
-0.0348391533,
-0.001837821,
-0.0581198819,
0.0061221169,
0.0239981934,
-0.0685260966,
-0.0615455247,
-0.0248847101,
0.0509866551,
0.0277331546,
0.0565713942,
0.0005908103,
-0.0230603851,
0.0451043397,
0.0410736613,
-0.098021917,
0.0379442498,
0.0318641327,
0.0410234891,
-0.0218172129,
0.0436958522,
-0.039685145,
0.0168225672,
-0.0056669414,
0.0409093648,
-0.0170877632,
-0.0795542672,
-0.0137763973,
-0.0104019642,
-0.0274688527,
-0.0183724128,
-0.0129410829,
-0.0159108918,
0.0295468681,
-0.0078699393,
-0.0325965025,
-0.0450128466,
-0.0044485861,
-0.021983033,
0.016242899,
-0.0052832025,
0.0267141014,
-0.0197057128,
0.0695254579,
0.0008683373,
-0.0056274757,
0.0315859728,
-0.0106722815,
0.0016831473,
-0.0010154118,
-0.0211967342,
0.0086978283,
-0.0194332562,
0.0346264243,
0.0215977821,
0.0868696794,
0.0385043807,
-0.0228631031,
0.0380297564,
-0.0085059367,
-0.0721845925,
-0.0446614623,
0.0048589827,
0.0258379206,
-0.0241255071,
0.0009415316,
0.047265783,
-0.0590488873,
-0.0015875051,
0.0095135411,
0.0231337827,
0.0054532108,
-0.0402957126,
-0.003264179,
0.0264471602,
-0.0147114396,
-0.0036935818,
0.0457810462,
-0.0443646982,
0.0371526368,
-0.0387304798,
-0.0152726285,
-0.0660904571,
-0.0666290447,
-0.0049457871,
0.0230805408,
0.0125121977,
0.0071455822,
-0.0858168378,
-0.0475914888,
0.0011948878,
0.018012898,
-0.0207369495,
-0.0161141362,
0.0400690213,
0.0496026799,
-0.0097849211,
0.0025605424,
-0.0328894109,
0.0033546127,
-0.0397375524,
-0.0295956545,
-0.0348752923,
-5.715969483e-33,
-0.0547411814,
-0.0540679879,
0.0243179891,
-0.0142862108,
-0.0362391733,
0.0106792403,
-0.0488631912,
-0.0008619417,
-0.0009998062,
0.0153697552,
-0.038095843,
-0.0081622405,
0.0077073984,
0.016312873,
-0.0040540462,
-0.0147954505,
0.0302260239,
-0.0335728675,
-0.0520057455,
0.066982165,
0.0482507199,
-0.0058762846,
-0.0137448469,
-0.0441633351,
0.0062556523,
0.0430489592,
-0.0128444592,
0.0054069581,
0.0254583005,
0.0600797236,
-0.0301149301,
-0.0093452074,
0.0205371585,
-0.0084167635,
-0.0205961876,
0.0510479994,
-0.0498970002,
-0.057079874,
0.0363135338,
0.0043748622,
-0.0460199043,
-0.0305493139,
-0.009672734,
-0.0121604893,
-0.0124509381,
-0.0137970606,
0.0138466181,
-0.0001562411,
-0.0205160566,
-0.0333593078,
-0.0538987443,
0.0082736937,
-0.0442406051,
-0.0242162868,
0.0216891021,
0.0776825249,
0.0222635195,
-0.0270934794,
0.002169878,
0.0398980975,
-0.0218266621,
0.0218341853,
0.0394485369,
0.0133022815,
0.0087387562,
0.0696831569,
-0.0197821651,
0.0139273796,
-0.0306195002,
-0.0273104291,
0.011042594,
0.0403411835,
0.0226717368,
0.0548589677,
-0.0322849937,
-0.0041593481,
-0.0231022071,
0.0248730481,
0.0071621719,
0.0225936733,
0.0168581028,
-0.0098649757,
-0.0250953436,
-0.0087672267,
-0.0281384457,
-0.0273123793,
-0.014661734,
-0.0249330569,
0.0339721851,
0.0613240786,
-0.0288088564,
0.0055883322,
0.0342893302,
-0.0233916752,
0.0424402282,
0.0710899532,
-0.0550202094,
0.0239756051,
0.0042907526,
-0.0031121974,
-0.0590054728,
-0.0026258975,
0.0542779826,
0.018956136,
0.0069146547,
0.023339469,
0.0023929239,
0.0201659985,
-0.0273777526,
-0.0001082894,
-0.0351607502,
0.0824435875,
-0.037957713,
-0.0299400464,
-0.0007094942,
0.0064525353,
-0.0497412272,
0.0309824999,
0.0075896946,
-0.0208872799,
-0.0257652476,
-0.0485703908,
0.0362123884,
0.0094341999,
-0.0638237,
-0.0097988723,
0.0264250301,
-0.0258039106,
0.0137705533,
-0.0272628274,
-0.0163591169,
0.0368345007,
2.658e-7,
-0.0213510673,
0.0130108753,
0.0616692863,
-0.0190171208,
0.0171435382,
0.005458931,
0.0220448188,
0.0319591202,
0.003091098,
-0.0530783124,
0.05309885,
0.0119117685,
0.0650539547,
0.0170068406,
-0.1087147295,
0.0136254327,
-0.0720646605,
-0.0450502299,
-0.0253546126,
-0.0570520908,
0.0296583343,
0.0009263973,
-0.0050076507,
0.0024023431,
-0.0156233162,
0.0137970056,
0.0402179621,
-0.0307361558,
0.0295713432,
0.0140035981,
0.0770375356,
-0.032414522,
-0.0061136037,
0.0336577259,
-0.0206533838,
0.0388553739,
0.0101479469,
0.0603935979,
0.0375766605,
-0.0848282278,
-0.0284261107,
-0.0157098807,
0.0329227746,
-0.0631024465,
0.0320795327,
-0.0255411603,
-0.0022834174,
-0.0132334484,
-0.0312893614,
-0.0288245827,
-0.0200123955,
-0.0220263712,
0.0115263388,
0.0406099074,
0.0016673447,
0.0063705551,
0.0628253967,
-0.0702588111,
-0.0067566326,
-0.0280512664,
-0.0631239638,
-0.0466903076,
-0.0621305965,
0.0061173383,
-0.0200461484,
-0.0478232801,
-0.0275802054,
2.129999896e-34,
0.0094483793,
-0.0036791218,
0.0021049017,
-0.0118972892,
0.0318760984,
0.0174329169,
0.0155213,
-0.0037642792,
0.0509286113,
-0.0711480528,
-0.0378429517
] |
http://arxiv.org/abs/2402.02057v1 | http://arxiv.org/pdf/2402.02057v1 | Break the Sequential Dependency of LLM Inference Using Lookahead Decoding | 2024-02-03T06:37:50 | 2024-02-03T06:37:50 | Yichao Fu, Peter Bailis, Ion Stoica, Hao Zhang | cs.LG | cs.LG, cs.CL | Autoregressive decoding of large language models (LLMs) is memory bandwidth
bounded, resulting in high latency and significant wastes of the parallel
processing power of modern accelerators. Existing methods for accelerating LLM
decoding often require a draft model (e.g., speculative decoding), which is
nontrivial to obtain and unable to generalize. In this paper, we introduce
Lookahead decoding, an exact, parallel decoding algorithm that accelerates LLM
decoding without needing auxiliary models or data stores. It allows trading
per-step log(FLOPs) to reduce the number of total decoding steps, is more
parallelizable on single or multiple modern accelerators, and is compatible
with concurrent memory-efficient attention (e.g., FlashAttention). Our
implementation of Lookahead decoding can speed up autoregressive decoding by up
to 1.8x on MT-bench and 4x with strong scaling on multiple GPUs in code
completion tasks. Our code is avialable at
https://github.com/hao-ai-lab/LookaheadDecoding | Break the Sequential Dependency of LLM Inference Using Lookahead Decoding: Autoregressive decoding of large language models (LLMs) is memory bandwidth
bounded, resulting in high latency and significant wastes of the parallel
processing power of modern accelerators. Existing methods for accelerating LLM
decoding often require a draft model (e.g., speculative decoding), which is
nontrivial to obtain and unable to generalize. In this paper, we introduce
Lookahead decoding, an exact, parallel decoding algorithm that accelerates LLM
decoding without needing auxiliary models or data stores. It allows trading
per-step log(FLOPs) to reduce the number of total decoding steps, is more
parallelizable on single or multiple modern accelerators, and is compatible
with concurrent memory-efficient attention (e.g., FlashAttention). Our
implementation of Lookahead decoding can speed up autoregressive decoding by up
to 1.8x on MT-bench and 4x with strong scaling on multiple GPUs in code
completion tasks. Our code is avialable at
https://github.com/hao-ai-lab/LookaheadDecoding | [
-0.0004216499,
0.0981527865,
0.0012011249,
0.0653081462,
-0.0120112691,
0.0342320129,
-0.0214729998,
0.0071956823,
0.0174019448,
-0.0116869435,
-0.0122309728,
-0.0201049689,
0.0088994801,
0.0470333956,
0.0095350901,
0.0081849676,
0.0483171791,
-0.0589301623,
0.011724012,
-0.0117177619,
-0.0116804074,
0.0265684854,
0.0008524363,
0.0474718846,
0.0370019116,
-0.0378806181,
-0.0117107527,
0.0036307804,
-0.0103589483,
-0.0010477625,
-0.0056619295,
0.009027943,
-0.022999227,
0.01767141,
0.0000022422,
-0.0065890136,
-0.0154841393,
-0.0205797385,
-0.0367416181,
0.0356971063,
0.031327907,
-0.051323235,
-0.0367783606,
0.0153668365,
-0.0072889477,
-0.0612996221,
0.0530885234,
-0.0185509324,
0.0122503368,
0.039596092,
0.006348236,
-0.0472728722,
0.0477710664,
-0.0467382073,
0.0102958847,
-0.0597315878,
0.0007477375,
0.0130459396,
0.0019387706,
-0.007908971,
-0.0352902077,
0.0488454811,
0.0196225159,
0.0095732193,
-0.0101197418,
0.027051419,
0.013255612,
-0.0263264496,
-0.0166875292,
0.0526948348,
-0.0480849408,
-0.0397914648,
-0.0101071391,
-0.022418702,
-0.025817398,
-0.015781574,
-0.0257631708,
0.0618028007,
-0.0382231101,
0.0151066612,
0.0884148851,
-0.0092171868,
0.0447698981,
-0.0031259886,
-0.0518091395,
0.0792398155,
0.0175538566,
0.0192534402,
-0.0208564717,
0.0250208005,
0.0218584463,
-0.0360170454,
0.0466861054,
0.0138130076,
-0.0109163141,
0.0159342326,
-0.0488587432,
-0.0311236456,
0.0023908142,
-0.0277326684,
0.0569511168,
0.1003236696,
-0.0215251297,
0.0249698479,
-0.0397194624,
0.0379338674,
-0.0440328047,
-0.0379360802,
-0.0543866158,
0.0049469173,
0.0328039117,
-0.0077295452,
-0.0387019217,
0.0308803506,
-0.0478071086,
0.001026138,
-0.0116862059,
0.0561501049,
0.0118550388,
0.006546481,
-0.006739839,
-0.0060212505,
-0.0431257002,
-0.0059962082,
0.0426974632,
0.0512971804,
-0.0340252221,
-0.0298989695,
-0.0390330851,
-0.0930205882,
-0.0374043547,
0.0239163283,
-0.0064147254,
-0.0150982495,
0.0034396946,
0.0563375503,
-0.0134386383,
-0.0053449147,
-0.0592645518,
-0.0023830731,
0.0482933559,
-0.0144348228,
0.0001977777,
-0.0573440306,
0.0309708416,
0.0300017949,
0.01091584,
-0.0581403188,
-0.0033145137,
0.0308561008,
-0.0336501524,
0.0317970812,
-0.0269144978,
-0.0281361751,
0.0018360727,
0.0083119124,
-0.0144500593,
0.0860232264,
0.0238433722,
0.0355769694,
-0.011624625,
-0.0104201809,
0.0182813946,
0.0791090205,
0.0205210913,
-0.022504149,
-0.0206002183,
-0.0010587954,
0.0001519515,
-0.0035549973,
0.1076938733,
0.0344702899,
-0.0430814065,
-0.0387826264,
-0.0187480878,
0.0534157492,
0.0227800235,
0.0328358784,
0.0268376339,
0.0159284472,
0.0012042952,
-0.0085595548,
-0.0130499061,
0.0497578457,
0.0018965503,
0.0304466411,
0.0375361927,
0.0299469233,
0.0042079156,
-0.0173288751,
-0.0042521292,
0.0013334431,
-0.024204148,
0.0289202891,
0.0349479131,
-0.0074881688,
-0.0604300983,
0.0374854766,
0.0272007771,
-0.0275130142,
-0.053176336,
-0.0439515114,
0.0171814598,
0.0503827333,
0.0496940166,
0.021363914,
-0.0270549823,
0.0063707205,
-0.0319187157,
0.0152562074,
-0.003982476,
-0.0005999419,
0.0085985763,
-0.0586065724,
0.0074636661,
0.0241995398,
0.0021856078,
0.0566345192,
0.0395895019,
-0.0040399046,
-0.0004669084,
0.0232055187,
-0.0371958949,
-0.0083778184,
-0.0244363286,
-0.0184209775,
0.001433099,
-0.0113176117,
0.0336413346,
-0.0113260867,
-0.0122922752,
0.0256032776,
0.0403165035,
-0.0613697655,
0.0347379632,
0.0031145995,
-0.0351056419,
-0.0450603254,
0.0222175755,
-0.0165052824,
-0.0224552006,
0.0250983816,
-0.0289220624,
-0.019216666,
0.0431344174,
-0.0328022987,
0.073862955,
-0.0227433182,
0.0270850174,
0.0251798257,
0.0323971659,
-0.009274736,
0.0413097329,
-0.0521263592,
-0.0078436527,
0.0226253401,
0.0532466918,
-0.0134112332,
-0.0031499732,
-0.0072195423,
0.0537380688,
0.0026481042,
-0.0186598245,
-0.0377316773,
-0.0220412239,
-0.0025846462,
-0.0125393104,
-0.0574108548,
0.0114891846,
-0.0260134824,
0.0456435643,
0.0180534255,
0.0304317717,
0.0058308733,
-0.0210240819,
0.0185471121,
0.0300053209,
-0.0229470376,
0.0282534659,
-0.0143187186,
-0.0196655393,
-0.0364269242,
-0.0150667876,
-0.0444939472,
-0.0056580971,
0.058519505,
-0.0156467948,
-0.0550344512,
-0.0738676339,
0.0343359374,
0.0248632953,
0.0148586947,
-0.0299626198,
-0.0465295762,
0.0090067163,
-0.0531502441,
0.0435712449,
0.0596548542,
-0.0191399064,
0.0128781805,
0.0215189997,
-0.0195752867,
-0.0597432032,
0.0219349042,
0.0410810709,
-0.0282316934,
-0.0139743295,
-0.0680878684,
-0.0164317433,
0.1134197041,
0.0051085805,
-0.0051539871,
-0.0367964581,
-0.0473534912,
0.022763066,
-0.0276753269,
0.0042351615,
-0.0066833538,
-0.0155652845,
0.013828679,
-0.023270892,
-0.0656487271,
0.0532785393,
-0.0591239743,
-0.034216322,
-0.005027818,
0.0013721335,
-0.0338655263,
0.0552367195,
-0.0009561144,
0.0015516393,
-0.0735644102,
-0.0642042011,
0.0542465784,
0.0555387065,
-0.0303893257,
-0.0485113822,
0.0550341234,
-0.0356751792,
0.0312762745,
-0.0405499488,
-0.0070303567,
-0.0362722911,
0.0213239063,
-0.0742061213,
-0.0149176214,
0.0463207513,
-0.03256841,
-0.0131366868,
0.0113254283,
-0.0058900481,
0.0627052784,
-0.0018559434,
0.0004887535,
-0.0145289563,
-0.013916254,
0.0261780582,
-0.0418320149,
-0.0091408463,
-0.0197217464,
0.0057892464,
0.0648611486,
0.0443997979,
0.0088783931,
-0.0730685368,
0.0243911371,
0.0815534368,
0.0111256028,
-0.0187571831,
0.0513283275,
-0.0159638394,
-0.0742936656,
0.0039338619,
0.077170752,
0.0094741359,
-0.0774701908,
-0.0104858214,
-0.006433357,
0.0569475703,
0.0153723136,
0.0370272323,
-0.0904139802,
-0.053837806,
0.030692704,
0.0184905976,
0.0188528579,
-0.0210900307,
-0.0240385272,
-0.1009699404,
0.040153794,
0.0243723169,
-0.0054836916,
-0.0366345569,
-0.0244333372,
-0.0328247063,
0.0109811537,
-0.0155333541,
0.0619114861,
-0.0006324574,
0.0195214041,
-0.0123976059,
0.0080389483,
0.0506631657,
0.0054718321,
-0.0068075759,
0.0684873983,
0.0723070055,
-0.0034691582,
0.0286728498,
-0.0142233521,
0.0254871957,
-0.0180631001,
-0.003362308,
-0.0521133095,
0.0010792007,
-0.0435242392,
-0.0096859364,
0.0115715452,
0.0192546584,
-0.0534169339,
0.0142255435,
-0.0092920512,
0.0730179176,
-0.022285033,
-0.0266522598,
-0.0165239871,
-0.0170631409,
-0.0228661057,
0.0564443506,
0.0179158375,
0.0085630147,
-0.0199376121,
-0.0979115665,
-0.022298893,
-0.0706407204,
-0.0403387882,
0.0071694744,
0.0563991331,
-0.0517829098,
0.0133430278,
-0.0077201184,
0.03005784,
0.0027657414,
-0.0098430188,
0.0155296111,
0.030518854,
-0.0392154716,
-0.091375038,
-0.0006935545,
-0.0425424725,
0.0067348946,
-0.0763799474,
-0.0106834453,
0.039013613,
-0.0114603192,
-0.0058041117,
0.051478073,
-0.1204266399,
-0.0475176349,
-0.0004066558,
0.0322587416,
0.019721847,
0.0410688333,
-0.0306709912,
-0.0522294864,
-0.0361535028,
0.0095534343,
-0.0536367968,
0.0471683294,
0.0238914173,
0.002047756,
0.0158798918,
0.0259055831,
-0.0210779905,
-0.0108907996,
-0.0150751034,
0.0200116448,
-0.0117377657,
-0.0368631557,
-0.0082991468,
0.0210161209,
-0.0051827924,
-0.0038917977,
-0.0059036277,
0.0085999081,
0.0288913399,
0.0177447628,
0.0082424106,
-0.0535725802,
0.042890694,
0.0401011594,
-0.0552315451,
-0.008984142,
-0.0196478367,
0.0156415701,
-0.0226022676,
-0.0075366134,
0.0371838436,
-0.0023986204,
0.0050805882,
0.0617390126,
-0.0316939056,
-0.0212910995,
0.0082113352,
0.0161115844,
0.0427147895,
-0.0248976368,
0.0158788189,
0.0055596065,
0.001394049,
-0.0111910114,
0.0019874638,
-0.0411711037,
-0.0488920771,
-0.0242185611,
0.0079902122,
0.0560207404,
0.0145915216,
0.051447399,
-0.0028806941,
0.0217304304,
-0.0143402293,
0.0169270411,
-0.0051695537,
-0.0403001457,
0.0236585941,
0.0344367325,
-0.0137310373,
0.0038106074,
0.0614973493,
0.0278153624,
0.0403988697,
0.0071864268,
0.0041055703,
-0.0335181765,
-0.0663533807,
-0.0249705371,
-0.0017034836,
-0.0444833934,
-0.011107171,
-0.0060663517,
-0.0448350571,
0.0250003636,
0.0160084758,
0.0003122943,
-0.0276584812,
-0.0188141391,
0.0472725481,
0.0169473998,
-0.0412458405,
-0.0295603592,
0.062156748,
0.0164239313,
-0.0387049355,
-0.0164727923,
-6.513069631e-33,
-0.0151533401,
-0.019267492,
0.0103998668,
-0.045894485,
-0.009600359,
0.0119832819,
-0.0402268618,
-0.0235098377,
-0.0025069322,
-0.0229863301,
-0.0124225141,
0.0607125834,
0.0184529182,
-0.0053557991,
0.0258507337,
0.0156076895,
0.0401034951,
-0.0028537798,
-0.0258884318,
0.0178455636,
-0.0003048973,
0.0727944896,
0.0191344712,
0.0113951853,
-0.0026825157,
-0.0415340438,
0.0347145684,
-0.0122910291,
-0.0050994987,
0.0213031899,
-0.0431087129,
0.0107654696,
0.002174665,
-0.0426044501,
-0.0346262269,
-0.002621525,
-0.1017623842,
-0.0037310147,
0.0650293827,
0.040134076,
-0.0114299534,
-0.0445090681,
-0.0014788565,
0.0471758544,
0.0071703088,
-0.0279769432,
-0.0270849559,
-0.0168180615,
-0.0389715061,
-0.006222656,
-0.0185797308,
0.0398835614,
0.0014120274,
0.0377028398,
0.1096350625,
0.0058686095,
0.0082490379,
0.0585716404,
-0.0409381278,
0.0007631401,
-0.0031857493,
0.0946178064,
0.0987352952,
0.0107446024,
-0.0042019212,
0.024765674,
0.0332244746,
-0.0341226272,
-0.0564138852,
0.0367948003,
0.0124365781,
-0.0000658231,
-0.0548025966,
0.0096983099,
0.0763455778,
0.0067010843,
-0.0306747463,
-0.0031121355,
0.0559065901,
0.0408712402,
0.0303822439,
0.0332523137,
-0.0504909419,
-0.0316563882,
-0.0097220782,
-0.0114201,
-0.0570224561,
-0.0008814371,
0.0276930705,
0.0088617019,
-0.0380820222,
0.006723315,
0.0602737814,
-0.0383057818,
0.0125207556,
0.0486539863,
-0.0044263056,
0.0297796149,
-0.0068910108,
0.0457978658,
-0.0014228504,
-0.0234889146,
-0.0100409891,
-0.0228781998,
0.0679190531,
0.0077977497,
-0.0298437532,
0.0102772871,
-0.0556826219,
-0.0358148292,
-0.0598001108,
-0.008597984,
0.0018032215,
0.0184832923,
0.0552120432,
-0.0052329716,
0.0238143709,
0.052471891,
-0.0263228118,
0.0240233094,
-0.0242265891,
0.0140095884,
0.004706149,
0.0221921969,
-0.0472461767,
-0.0238482058,
-0.0589617155,
0.047590889,
0.0169056132,
-0.0794959739,
-0.038038332,
0.060687013,
2.857e-7,
0.0047592944,
0.0351923779,
0.0625254437,
0.0578864701,
0.0096102292,
0.0058003538,
0.0100233909,
-0.0100184083,
0.0439607725,
0.0004244261,
-0.0024478782,
0.0288256612,
0.0702114552,
0.0000372448,
-0.0657728538,
-0.0097315004,
-0.007034508,
0.0100717042,
-0.0335083306,
-0.0585556738,
0.0090170512,
0.085091792,
-0.0416646525,
0.0069129365,
0.0606134422,
0.014237036,
-0.0028282639,
-0.0001157773,
0.001151672,
0.005744061,
-0.0104690902,
0.0116377948,
-0.032808084,
-0.0101209981,
0.0001405686,
-0.009461673,
-0.0348369032,
0.0710807741,
0.019034652,
0.005427483,
0.0272533298,
-0.0436147191,
-0.0077029127,
-0.0235932823,
0.0588839799,
0.0190895032,
0.0102549233,
-0.0197084248,
-0.0306196027,
0.0062556942,
-0.049630709,
0.0004639038,
0.0584944263,
0.0272026751,
0.0388853922,
-0.0614260286,
-0.0330630876,
-0.04986085,
-0.0802661628,
-0.0192275885,
-0.0420628861,
-0.0190553181,
-0.0689801872,
0.015409044,
0.0329733752,
-0.05674107,
0.0058837901,
2.82895871e-34,
0.0601035543,
-0.0108933793,
0.0118472017,
0.0949267969,
0.0393391624,
-0.037539769,
-0.0005358922,
0.0444456004,
-0.0303461365,
-0.085137099,
-0.0656146258
] |
http://arxiv.org/abs/2402.02056v1 | http://arxiv.org/pdf/2402.02056v1 | AnthroScore: A Computational Linguistic Measure of Anthropomorphism | 2024-02-03T06:36:11 | 2024-02-03T06:36:11 | Myra Cheng, Kristina Gligoric, Tiziano Piccardi, Dan Jurafsky | cs.CL | cs.CL, cs.AI, cs.CY | Anthropomorphism, or the attribution of human-like characteristics to
non-human entities, has shaped conversations about the impacts and
possibilities of technology. We present AnthroScore, an automatic metric of
implicit anthropomorphism in language. We use a masked language model to
quantify how non-human entities are implicitly framed as human by the
surrounding context. We show that AnthroScore corresponds with human judgments
of anthropomorphism and dimensions of anthropomorphism described in social
science literature. Motivated by concerns of misleading anthropomorphism in
computer science discourse, we use AnthroScore to analyze 15 years of research
papers and downstream news articles. In research papers, we find that
anthropomorphism has steadily increased over time, and that papers related to
language models have the most anthropomorphism. Within ACL papers, temporal
increases in anthropomorphism are correlated with key neural advancements.
Building upon concerns of scientific misinformation in mass media, we identify
higher levels of anthropomorphism in news headlines compared to the research
papers they cite. Since AnthroScore is lexicon-free, it can be directly applied
to a wide range of text sources. | AnthroScore: A Computational Linguistic Measure of Anthropomorphism: Anthropomorphism, or the attribution of human-like characteristics to
non-human entities, has shaped conversations about the impacts and
possibilities of technology. We present AnthroScore, an automatic metric of
implicit anthropomorphism in language. We use a masked language model to
quantify how non-human entities are implicitly framed as human by the
surrounding context. We show that AnthroScore corresponds with human judgments
of anthropomorphism and dimensions of anthropomorphism described in social
science literature. Motivated by concerns of misleading anthropomorphism in
computer science discourse, we use AnthroScore to analyze 15 years of research
papers and downstream news articles. In research papers, we find that
anthropomorphism has steadily increased over time, and that papers related to
language models have the most anthropomorphism. Within ACL papers, temporal
increases in anthropomorphism are correlated with key neural advancements.
Building upon concerns of scientific misinformation in mass media, we identify
higher levels of anthropomorphism in news headlines compared to the research
papers they cite. Since AnthroScore is lexicon-free, it can be directly applied
to a wide range of text sources. | [
0.0242570322,
0.0971388891,
-0.0389235467,
-0.0070898384,
-0.0033685379,
0.0145995338,
0.0455105528,
-0.0345407762,
-0.0008168534,
-0.1037104353,
-0.0262494497,
0.0211992767,
0.0152268829,
-0.0073637734,
0.0463212579,
-0.0228240974,
0.0374548733,
0.0431649573,
0.0159062911,
0.0063054184,
-0.0374299064,
0.0328578688,
0.0186889805,
0.0352233425,
0.010560764,
-0.0368237011,
0.0610294864,
0.0071860016,
0.0235954132,
-0.0470664948,
0.0076021492,
0.0149343675,
0.0291263182,
0.0176897049,
0.0000019546,
-0.0711007342,
0.0443772189,
0.0295137595,
0.0142480899,
-0.0082460064,
0.0546484552,
0.0188120343,
-0.0239837281,
-0.0164147373,
-0.0149388518,
-0.0383693427,
0.0493342094,
0.0226469785,
-0.0149788717,
0.0293949209,
-0.0130035505,
0.0014624001,
-0.0376827046,
0.0302393641,
-0.031355001,
0.0230162125,
0.0211784039,
-0.0437447987,
-0.0299781486,
0.0223960597,
0.0140444236,
0.0190075189,
-0.0039701918,
0.0046722307,
0.0511311293,
0.0045392648,
-0.018953288,
-0.0359325893,
0.021082744,
0.0353237167,
0.0580070205,
-0.0404148698,
0.0141513878,
0.05305323,
0.0232258756,
0.0402098671,
0.0028952074,
0.0386386476,
0.0158208869,
-0.0495215468,
-0.0464014709,
0.0006219381,
0.0239290893,
0.0775211379,
0.0021783144,
0.1081139147,
0.0003344149,
-0.0185229219,
0.0172358714,
-0.0053445189,
0.0126886666,
-0.0260477029,
0.0061222217,
0.0151425628,
0.0070016282,
-0.0030575912,
-0.0319134034,
-0.013311537,
0.0191094559,
-0.0899834782,
-0.0156714488,
0.0013192515,
0.0115257902,
0.0255339202,
0.0086680287,
0.030005144,
-0.0137587367,
-0.0234522913,
-0.0040830253,
0.0259614438,
-0.050473541,
-0.02979015,
0.0122246277,
0.0064547108,
0.0303628165,
0.0030968685,
0.0029006954,
-0.0002941867,
-0.0726007074,
-0.008978094,
-0.0999235809,
0.0273252334,
-0.0097025195,
-0.0193293877,
-0.0012691397,
-0.0291857608,
-0.0384572558,
0.0393977724,
-0.0461845435,
-0.0560822301,
0.0223356877,
-0.0012889681,
-0.0194345079,
-0.0188342463,
-0.00253998,
0.0801714212,
0.0170176402,
-0.0149797639,
0.0419707745,
0.0108211692,
0.0151312463,
-0.0002716695,
0.0217847247,
-0.0160288811,
-0.0092256581,
-0.0016635843,
0.0063387537,
-0.0528608114,
-0.0143432328,
0.0215162057,
-0.0364933088,
0.0641632676,
-0.0317023285,
-0.0059504253,
0.0550567433,
0.0227835737,
0.0086636422,
0.0391456522,
0.0331218764,
0.0247764848,
0.0371423922,
-0.0075824591,
-0.0410632901,
-0.0000592085,
-0.0006994164,
-0.0086849341,
0.0185228046,
0.0159583986,
-0.0091924872,
0.0248324946,
0.0198283549,
0.0181877874,
0.0059059737,
0.0025099847,
0.0273623653,
0.0682024807,
-0.0747371688,
0.0633601248,
0.0348247327,
-0.0057499507,
-0.0135304928,
0.0074712532,
-0.0707208142,
0.0050211051,
-0.0251901783,
-0.0572607741,
-0.0245359205,
0.0466124825,
-0.03723168,
-0.0186892208,
0.0106183346,
0.0211353209,
0.073843807,
-0.0128816431,
0.0300725941,
0.0114118336,
-0.0004875879,
-0.0094416263,
-0.0088085802,
-0.023227267,
-0.0086069973,
-0.0494960956,
-0.025201112,
0.0130549641,
0.0191634689,
-0.0137106376,
0.0444291532,
-0.0345314443,
0.0073534288,
0.0602151342,
0.0264732912,
-0.005825995,
-0.0524644554,
0.0295128245,
0.036779739,
-0.010415961,
0.0258894637,
0.0553732403,
-0.0128100552,
0.040356487,
0.0270899162,
0.0223272014,
-0.0472603329,
0.0360009372,
-0.008397771,
-0.0134768905,
-0.0347542912,
-0.0045338087,
0.019931145,
0.0186999757,
-0.0153294923,
0.0298961066,
0.0149485543,
-0.0154790292,
-0.035559997,
0.0308719315,
0.0223366544,
0.0634560362,
0.0255693477,
-0.0990778878,
-0.0303098764,
0.071132265,
0.0215362683,
0.0129666273,
0.0847300291,
-0.0488302559,
0.0374852158,
-0.0355425924,
0.0159002114,
-0.0598009937,
0.0210018251,
-0.0146597978,
0.0079069547,
0.0145132681,
0.011403949,
-0.0245121792,
-0.0358968712,
0.0236785468,
0.0158668905,
-0.0423384905,
0.0110044284,
-0.0165192764,
0.015443855,
0.0471575595,
-0.0665143132,
-0.0477624089,
0.0180532038,
-0.047968816,
-0.0298718996,
0.0360662788,
-0.0080123171,
0.0117654735,
-0.0455289409,
0.0011654079,
0.0094369985,
0.0134121953,
-0.0554020926,
-0.041890163,
0.0408947431,
0.0178267322,
-0.0435909629,
-0.0048444145,
-0.031239789,
-0.0093304701,
-0.0637136176,
0.0654740632,
0.0453615151,
-0.0097805476,
-0.0257655568,
0.0022014035,
-0.067765303,
-0.010039269,
-0.0013893086,
-0.0195045006,
0.0335539691,
0.0575062074,
-0.0155636985,
0.0695740804,
0.0437452085,
0.0485369191,
0.0455429889,
0.0465962291,
-0.0209098738,
-0.0459060334,
0.021698406,
-0.0202787519,
0.047284089,
-0.0905002877,
0.0506020039,
0.0303416718,
0.0000061172,
-0.032180395,
-0.039960973,
0.1085661128,
-0.0700701252,
-0.0684994459,
-0.0210697949,
-0.0359592736,
-0.0046472237,
0.0348031893,
-0.0185814667,
-0.0130331898,
0.0412187018,
-0.0086360788,
0.030931022,
0.0223433487,
0.0140622202,
-0.0031747145,
0.0096914694,
0.0309514198,
0.0519484952,
0.0585653707,
-0.0119119752,
-0.0723156109,
-0.0017074738,
-0.0099965846,
-0.0211933665,
0.0116296308,
-0.0251340885,
0.0194043927,
0.0386894681,
0.0009026934,
0.011352716,
-0.0443358943,
-0.0094421832,
0.0177178625,
-0.0106894653,
-0.0325221345,
-0.0448157042,
-0.033012148,
0.0322406404,
-0.0234078877,
0.018024018,
-0.0065776394,
0.0166747235,
-0.0359101668,
0.0288455971,
0.0349253528,
0.0363018736,
0.0367776901,
-0.0063966624,
0.0266264956,
0.0030032038,
0.0086170072,
-0.0492929332,
0.0188074112,
-0.0170308575,
-0.035043858,
0.0145104546,
-0.0065901517,
0.0140331145,
-0.0235104207,
0.0410382673,
0.0701266676,
0.0082317991,
-0.0313719399,
0.0398185477,
0.0288798884,
-0.0578206331,
0.0509851947,
0.007367637,
-0.0547331274,
0.0457359441,
-0.0225403272,
-0.0135208881,
-0.0707024783,
-0.0598811992,
-0.0144895324,
-0.061072506,
-0.0010493655,
-0.0055722734,
-0.015326988,
0.0057648877,
-0.0402098894,
0.0431523956,
0.0454275645,
-0.0305098034,
-0.006092947,
-0.0065929275,
-0.012599173,
-0.0251431745,
-0.0437271558,
-0.0145881129,
0.0380114913,
0.0057659042,
0.0206209831,
0.0650773272,
0.0967161879,
-0.0371839441,
0.0026637011,
0.0222144313,
-0.000042137,
-0.0277101956,
-0.0125509063,
0.0068986099,
-0.0504658371,
-0.0056058327,
-0.0177601352,
-0.0064245313,
0.0248190966,
0.0077101057,
-0.008599258,
0.0332538635,
0.0102272714,
-0.065404065,
-0.0628497377,
-0.0226413738,
-0.0268277824,
0.0576113462,
0.0276314691,
-0.021577606,
0.0068721538,
-0.0191631354,
0.029969709,
0.0475700609,
-0.0986207128,
0.0027264766,
-0.0217170753,
-0.0752553567,
-0.072346054,
-0.060595233,
-0.0140373083,
-0.0523160286,
0.0455523171,
-0.0627910122,
0.0535296202,
-0.0015499722,
-0.098474212,
0.0411852822,
-0.0022591413,
0.0182665102,
0.0419485383,
-0.0024981999,
0.0461088121,
0.0473862253,
-0.0077703171,
0.0564094111,
0.0686760545,
-0.056178797,
0.0234784912,
-0.0126765911,
0.0241785273,
-0.0033731265,
-0.0022963644,
-0.0488581732,
0.0216543935,
0.0363694206,
-0.0238875151,
0.041600734,
-0.0086700292,
0.0211150832,
-0.0245250892,
0.011808156,
-0.0235189591,
0.040430259,
-0.0226203427,
-0.0488047935,
-0.0460994802,
-0.0170592777,
0.0044621462,
0.0171154812,
0.0200297348,
-0.0531136021,
0.0454968251,
0.0226595309,
-0.0650426522,
-0.0191742964,
-0.0160594452,
0.0048649143,
-0.0623911098,
0.0005627665,
0.047023192,
-0.0024279975,
0.0026216868,
-0.0407736748,
0.0601836182,
0.0000766771,
0.0144676846,
0.0067618052,
0.0680999383,
-0.0366532505,
0.0087735681,
-0.0060351701,
-0.0134949973,
-0.0532274656,
0.0361329392,
0.0103566665,
0.0083244061,
0.031247573,
0.0382797644,
-0.0405012332,
0.0571372174,
-0.0491169654,
-0.0288172625,
-0.0128680877,
-0.0233642124,
-0.0267829709,
0.0202479847,
0.0393523872,
0.0172861163,
-0.0249091424,
0.0096422667,
0.0142835332,
0.0031988637,
0.0046411282,
0.0185467079,
0.0343681052,
-0.0106702466,
0.0207756273,
0.0192249417,
0.0187507886,
0.0093422811,
0.0071731186,
-0.0350113139,
-0.0099636847,
-0.0277201161,
-0.0200102143,
-0.0334006958,
0.011629086,
-0.0034662746,
-0.0635283142,
-0.0800455362,
-0.0190298203,
0.0066563431,
0.0274217222,
-0.0231023114,
-0.0176668242,
0.0397232659,
-0.0044824374,
-0.0498420037,
-0.0266633853,
0.0398501791,
-0.0351585634,
-0.0135766277,
-0.0100834593,
-6.10218256e-33,
-0.0384289809,
-0.0473591462,
-0.0186190214,
-0.0071836021,
-0.066050902,
0.0106891142,
-0.0515837409,
0.0141154006,
-0.0359091572,
0.0196571369,
0.0174179841,
-0.0038021221,
-0.0083213421,
-0.0317664258,
0.046881564,
0.0321000069,
0.0520488471,
-0.0189658944,
-0.0053124935,
-0.0210379045,
-0.0151863098,
-0.0399117656,
0.0131985573,
-0.0911109522,
0.112419121,
0.0224410929,
-0.0305964965,
-0.0216087326,
0.0541081093,
-0.0286991093,
-0.0160196349,
0.0370471403,
0.0373633578,
-0.0002178526,
0.0178592838,
0.0981695652,
-0.0627582669,
-0.0119403275,
0.0248891022,
-0.0149267074,
0.021359697,
-0.0356884152,
0.0270779897,
-0.0159426983,
-0.0567157604,
0.0142156612,
0.0783325508,
-0.0376063175,
0.0340968892,
-0.0023654196,
-0.0688338578,
-0.025919972,
-0.0548244379,
0.0255199727,
-0.0120368758,
0.0669079646,
-0.0420799032,
0.0081847506,
-0.0879729092,
0.0221893266,
-0.0318066888,
-0.0435022824,
0.0279666148,
0.0218881536,
0.0740319192,
0.0648224801,
0.0713502094,
0.0710816234,
-0.0561907515,
0.0234535914,
0.0691141412,
-0.0087403152,
0.0032962062,
0.0286563747,
0.0080439271,
-0.0321090408,
0.0117579885,
-0.01836765,
-0.0000841019,
0.0218880493,
0.0282314494,
-0.0079160007,
-0.0512123369,
-0.0382070839,
0.0072812592,
-0.0748906657,
-0.0191157348,
-0.0618605576,
-0.0099251932,
-0.0618174635,
-0.0412818715,
0.0501361787,
-0.002344023,
-0.0478163399,
-0.0026998527,
-0.0452973768,
0.0062784725,
0.004530346,
0.0259661134,
-0.0110558961,
-0.0413241982,
0.0027583356,
-0.0006414538,
0.0649929047,
0.0818616748,
-0.0056185164,
-0.0412877984,
-0.0315139703,
-0.0292655565,
-0.0255835038,
0.0036344761,
-0.0590637214,
0.0069733188,
0.0011208983,
0.0045994199,
0.01564564,
-0.0022763063,
-0.0430469401,
0.0389854535,
0.0114110829,
-0.0359339081,
0.0960154533,
-0.0472424328,
0.016633423,
-0.0299660619,
-0.0472517051,
-0.0324194878,
0.0390465446,
0.0726227313,
-0.086534664,
-0.0290890913,
-0.0424078554,
2.821e-7,
0.0063546645,
0.071887359,
-0.0191594008,
-0.0146327382,
0.0274720471,
0.0561812818,
0.0316633917,
0.0515595675,
0.0479341,
-0.040622998,
-0.0285635907,
-0.0023708029,
0.0054877512,
0.0135095697,
-0.0229008459,
-0.022388462,
-0.0374447219,
-0.0081670769,
-0.0284053739,
0.0302285627,
0.0013032035,
0.0018643138,
0.0164425969,
-0.0068237297,
-0.0570430569,
-0.0206264388,
0.0158168934,
-0.0237314925,
-0.009709226,
0.0033840067,
-0.0373776704,
0.0765739903,
-0.0100366743,
0.0332051292,
-0.004146439,
0.0096207615,
0.0655339137,
-0.0168086533,
-0.0260221679,
0.0065844143,
0.0093932794,
0.0114875911,
-0.0218226183,
-0.0305764526,
0.0319590829,
-0.0304279793,
-0.0465075262,
-0.0450957566,
-0.0452136211,
0.0053801518,
0.0040886318,
-0.0064402502,
-0.0222901925,
0.0333914831,
-0.0046866308,
-0.0220259372,
0.0234160591,
-0.0607504621,
-0.0095756361,
0.0139952917,
-0.0307320803,
-0.0188295357,
0.0089538386,
0.0283125322,
0.0338629447,
-0.0117985792,
0.0486800149,
2.044100417e-34,
0.0256362464,
0.0199310351,
0.0109143313,
0.0047723902,
0.0494214222,
0.0190592464,
0.0253030434,
-0.0176533591,
0.0673922449,
-0.0543169789,
0.0075871921
] |
http://arxiv.org/abs/2402.05120v1 | http://arxiv.org/pdf/2402.05120v1 | More Agents Is All You Need | 2024-02-03T05:55:24 | 2024-02-03T05:55:24 | Junyou Li, Qin Zhang, Yangbin Yu, Qiang Fu, Deheng Ye | cs.CL | cs.CL, cs.AI, cs.LG | We find that, simply via a sampling-and-voting method, the performance of
large language models (LLMs) scales with the number of agents instantiated.
Also, this method is orthogonal to existing complicated methods to further
enhance LLMs, while the degree of enhancement is correlated to the task
difficulty. We conduct comprehensive experiments on a wide range of LLM
benchmarks to verify the presence of our finding, and to study the properties
that can facilitate its occurrence. Our code is publicly available at:
\url{https://anonymous.4open.science/r/more_agent_is_all_you_need}. | More Agents Is All You Need: We find that, simply via a sampling-and-voting method, the performance of
large language models (LLMs) scales with the number of agents instantiated.
Also, this method is orthogonal to existing complicated methods to further
enhance LLMs, while the degree of enhancement is correlated to the task
difficulty. We conduct comprehensive experiments on a wide range of LLM
benchmarks to verify the presence of our finding, and to study the properties
that can facilitate its occurrence. Our code is publicly available at:
\url{https://anonymous.4open.science/r/more_agent_is_all_you_need}. | [
0.0220290963,
0.0531233512,
-0.0389430709,
0.03530268,
-0.0262363553,
-0.0124644144,
0.0236723404,
0.0099874558,
0.0083003798,
-0.0230127312,
-0.0305421837,
-0.0498488732,
-0.0151407607,
-0.0094334483,
0.0290211178,
-0.0422029682,
0.0169768184,
-0.078549698,
-0.0034942471,
-0.0494856536,
-0.0354578458,
-0.0160517897,
0.0223098733,
0.0478397161,
0.0413409024,
0.0078312578,
-0.0035639128,
-0.0219532996,
-0.00700142,
0.0010583781,
-0.0478293262,
0.0321076624,
0.029676944,
0.0251561683,
0.0000021013,
-0.0422781743,
0.0128806904,
-0.0062912572,
-0.0463308357,
0.0616863258,
0.0624507703,
-0.0102715781,
-0.0033804048,
0.0306771342,
-0.0255025234,
-0.0095114494,
0.027855603,
-0.0104234144,
0.0225913823,
0.0557103902,
0.0059993654,
0.0046743387,
0.0028792655,
-0.0335467421,
-0.0193703193,
-0.0473766141,
0.0169701651,
0.0161898136,
0.0509198047,
0.0175458156,
-0.0392837673,
0.0424229428,
0.0214202069,
-0.0165575016,
0.0033271159,
0.0702626705,
-0.0304939728,
-0.0048477822,
-0.0064788489,
0.0453958362,
-0.0161169861,
0.0078606252,
-0.015389719,
-0.0124989422,
-0.0014873982,
0.0278456993,
-0.058851704,
0.0395019725,
0.0106387613,
0.0164744165,
0.0338143483,
-0.0036815049,
0.042211324,
0.0073538818,
-0.0302279387,
0.0794179738,
0.0284413788,
-0.0081266109,
-0.0193589702,
0.0014890493,
-0.0156938992,
-0.0418796614,
-0.0200812779,
0.0160999931,
-0.0455646515,
0.0210223775,
-0.0438846797,
-0.0063635353,
0.0170192197,
-0.0364254154,
0.0737727806,
0.0121479295,
0.0184888113,
0.078541778,
-0.0214752425,
0.0423868559,
-0.025276294,
0.0419607088,
-0.0351082608,
-0.0038852578,
-0.047390338,
0.0071917484,
-0.0563014485,
0.0273599941,
-0.0106217219,
-0.0103695039,
-0.0202230141,
0.0567922853,
0.035246253,
-0.0211556647,
-0.0221277624,
-0.0607674271,
-0.0196549892,
-0.0193252116,
-0.0074339313,
0.0102629866,
-0.0290116072,
0.025864454,
-0.0146074481,
-0.0719437301,
-0.0030694031,
-0.0065946132,
0.0182591509,
-0.018384777,
-0.003990809,
0.0614128746,
0.0090343999,
-0.0436324887,
-0.024832096,
0.0253780596,
0.006856997,
-0.0529849045,
-0.0750542805,
-0.0802412257,
0.0242302064,
0.0092342384,
-0.0026976408,
-0.0253897924,
-0.0017020931,
-0.0172496177,
-0.0640399307,
0.0476838276,
-0.0110775679,
0.0590010136,
0.0140879517,
-0.0115826651,
-0.0219310932,
0.080013819,
0.0507843159,
0.0159220453,
0.0295423102,
-0.0086196223,
0.0480048545,
0.0645319521,
-0.0217249393,
0.065417178,
0.0465330258,
-0.0151329264,
0.0006769841,
0.0318248309,
0.0136663914,
0.0761518404,
-0.0435343646,
-0.0163787473,
0.0779586881,
0.0634361506,
-0.0010484826,
-0.0154356975,
0.0156087233,
0.0067316005,
0.0021977052,
0.0010191054,
-0.0631071106,
0.0388283022,
0.0143959848,
0.0144977709,
0.026352128,
0.044133611,
-0.05282069,
-0.0198619273,
-0.0189786144,
0.0098140752,
0.0322603993,
0.0074911606,
0.0716008618,
-0.0163377561,
-0.0248817578,
0.0112593817,
0.0460689701,
-0.0601823144,
0.0113518871,
-0.0885182247,
0.0320931785,
0.0573784299,
0.0523332059,
-0.0144324871,
-0.0065148873,
-0.005081147,
0.0358615965,
0.0742479563,
0.0593817458,
-0.0036439914,
0.0274953283,
-0.0343886018,
-0.0137916021,
0.0063532433,
-0.0041533038,
0.0497340597,
0.0192563925,
-0.0084437486,
-0.0072844611,
-0.0268531386,
-0.0466011241,
0.0155715877,
-0.0515436418,
-0.0056542028,
-0.0105882278,
-0.0052975006,
0.0538377687,
-0.0137052676,
-0.0263967365,
-0.002419665,
-0.0011038452,
-0.0576676764,
-0.0284145195,
-0.0158530325,
0.0134143652,
-0.033007849,
0.0318796635,
0.0120779993,
0.0111285206,
0.0358055122,
-0.0222690199,
-0.0286485627,
0.0433164462,
-0.0408226587,
0.0615768656,
-0.0171039514,
0.0508659258,
-0.0406533517,
0.0195266996,
-0.037028268,
0.0463158414,
-0.0137916794,
-0.0014588149,
0.0119331973,
0.0445397682,
-0.0066219568,
-0.0312232971,
-0.0243026,
-0.010148529,
0.0223271307,
0.0446662083,
0.0498203039,
-0.0216897316,
-0.0017641499,
-0.0607215874,
-0.0628270879,
-0.0034797986,
-0.0164232217,
-0.0149077922,
-0.0130663812,
0.0075222589,
-0.0038546515,
0.0105137024,
-0.0384975076,
0.0402500816,
0.0007982057,
0.0563010946,
0.0106455795,
0.01256561,
-0.1089154556,
-0.0580469407,
-0.0183676947,
0.0094266841,
0.0169936437,
0.0222079027,
-0.0175307505,
-0.0531679317,
0.054692544,
-0.0046179718,
0.0091310488,
-0.0345860086,
-0.0542245284,
0.0052606687,
-0.0522801541,
0.0219202843,
0.0696176365,
-0.0145135801,
0.0260234252,
0.0558593161,
0.0022814439,
-0.0545131303,
0.0148526616,
-0.0209970903,
-0.023442097,
-0.0677734464,
-0.0576702729,
0.0154976863,
0.08107014,
-0.0043277992,
-0.0358921289,
-0.0337259844,
0.0076311952,
-0.0311634112,
-0.0352873616,
0.0051099048,
-0.0016969854,
-0.0069145714,
0.0364075378,
-0.0372687727,
-0.0611032099,
0.0373578072,
-0.0456703156,
-0.0265968926,
-0.0296136718,
-0.021178009,
-0.0404574387,
0.0759259835,
0.0470923856,
0.0187799074,
-0.0227942877,
-0.0179284308,
-0.0357798599,
-0.0140610337,
0.0215026494,
-0.012609628,
0.0698605329,
-0.0526839495,
0.037207704,
-0.0242778175,
-0.0533666909,
-0.0214677658,
-0.0517163165,
-0.0324631222,
0.0241812728,
0.0312181339,
-0.023532562,
-0.0480544679,
0.0313489698,
0.0200013164,
0.0386908874,
-0.0118481405,
-0.009187663,
-0.0201882757,
0.0035477735,
0.0419663191,
0.0212044679,
0.0264550094,
-0.0991135016,
0.0538960211,
0.0213295631,
0.0634829253,
0.0123408632,
-0.0273401532,
-0.0329397544,
0.0260799024,
-0.0426422954,
-0.0184765216,
0.0044090175,
0.0305772834,
-0.0337918885,
0.0135378726,
0.1091903523,
0.0050347024,
-0.0308012031,
0.0197142009,
-0.0130067002,
0.0490124822,
0.0303786695,
0.033469595,
-0.027280353,
-0.0115632331,
0.0050447676,
0.0045465869,
-0.0065558581,
-0.0306319539,
0.0071572703,
-0.1135667041,
0.0375667065,
0.015788151,
-0.0322813764,
0.0040240292,
-0.0388212092,
0.0127555458,
-0.0256921332,
0.04005678,
0.0377495997,
0.0026893923,
-0.009819462,
0.0201371815,
-0.0299006104,
-0.0196851119,
0.0134764109,
-0.0621111952,
0.0943719149,
0.0601683818,
0.0132326661,
-0.030665718,
-0.0110099874,
0.0197223146,
0.0418072008,
0.0000485695,
0.007309135,
0.0349343196,
0.0196931548,
-0.0285014845,
0.0013145307,
0.0116844978,
-0.0217311233,
0.0214446541,
0.0203231573,
0.0415211916,
0.0165399518,
-0.0357945189,
0.0061643454,
-0.0411810353,
-0.0215752255,
0.0706062093,
0.0358237922,
-0.0609586723,
-0.0203450415,
-0.0483119711,
-0.0532101803,
-0.0366175734,
-0.0320347697,
0.0186688751,
-0.0344251394,
-0.0090276673,
-0.0532066785,
-0.0359724611,
0.0532379523,
0.023870891,
0.0213973857,
0.0593177304,
0.0586641245,
-0.0019931893,
0.0108556896,
-0.0330299735,
-0.066660814,
0.0229741372,
-0.0420810357,
-0.0544189513,
0.0423875041,
-0.0093366485,
0.0099092098,
-0.0295099225,
-0.0986199081,
-0.0839334503,
0.0027771192,
0.0446150005,
0.0190213285,
0.0233829878,
-0.0222507287,
-0.0640504509,
0.0464180335,
-0.0048543015,
-0.0260744039,
0.0579273477,
-0.0053192033,
0.0017269041,
0.01200171,
0.0610396154,
-0.0475754142,
0.025595421,
0.028777767,
0.0530900508,
-0.0726058334,
-0.0200685747,
-0.0074120518,
0.0247164927,
-0.0155661553,
-0.0098483777,
0.019055292,
-0.009200288,
0.0097473292,
-0.0322822705,
0.0100013297,
-0.0496173836,
-0.0035612104,
0.0365636498,
0.0060772588,
0.0081969593,
0.0072241509,
0.0096685067,
0.0411753431,
0.0219537411,
0.0243131388,
-0.0268907659,
-0.0079304893,
0.0674481764,
0.0238073878,
-0.0272523239,
0.0274982434,
-0.0078479573,
0.0257239938,
0.0218386687,
0.0153365536,
-0.008876876,
-0.0159426313,
-0.0214383993,
-0.0001226004,
-0.0936938673,
-0.0799774304,
0.0254286435,
0.0708164498,
0.0591756962,
0.0148440935,
0.0809230357,
-0.0219100099,
0.0098828003,
-0.0363795273,
-0.0002316045,
0.0122625129,
-0.0375873484,
0.0148292966,
0.0325739235,
-0.0087880911,
0.0069325757,
0.1004098356,
0.0226070695,
0.0398430265,
0.0107825007,
0.0090990411,
-0.019599691,
-0.0206412785,
-0.0130938981,
-0.0048658461,
-0.0314633101,
0.0178675149,
-0.0380054079,
-0.0325795002,
-0.0228906479,
0.0209889356,
0.0073325005,
-0.0434910022,
-0.0016895341,
0.053699147,
0.0071606548,
-0.0070288409,
-0.0695270598,
-0.0058574523,
-0.0171801355,
-0.0099691497,
0.0184466224,
-6.127650747e-33,
-0.0465174615,
-0.0042358371,
0.0342818275,
0.0125062456,
0.0082712946,
-0.0078516137,
-0.0530562587,
0.0227988046,
-0.0107138716,
0.0061882366,
-0.0207242575,
0.003733478,
0.0102752047,
0.0200487971,
0.0225450844,
0.0115490248,
0.0543382093,
0.0102854641,
-0.0217842031,
0.0030458379,
-0.0165828001,
0.0057329931,
-0.0142276976,
-0.0242157504,
-0.035090033,
0.0001622824,
0.0138347577,
0.0049978583,
0.003674279,
0.0650649592,
-0.0095259072,
0.0285490863,
0.0178043302,
-0.0567459129,
-0.0451578461,
-0.0174739901,
-0.0163160209,
0.0038966693,
0.0306719039,
0.0161986314,
-0.007307074,
-0.042784337,
0.0168734435,
-0.0148399044,
-0.008214538,
0.0116569893,
0.025549436,
-0.046046067,
-0.0701999366,
0.0194399599,
-0.0568916015,
0.0381745584,
0.020546278,
0.0051008151,
0.0085168388,
-0.0349538624,
-0.024532726,
0.0087209102,
-0.0196829289,
0.0224158634,
0.0415645465,
0.0482870787,
0.0332339294,
0.0549110323,
0.0146476636,
0.0474205986,
0.0439531393,
0.0229186248,
-0.0378375016,
-0.0201357566,
0.027017504,
0.0297347363,
0.0408724546,
0.0740253031,
-0.0211966131,
-0.0041504749,
-0.0515884943,
0.002049244,
0.0768669993,
-0.0104739713,
0.0755064115,
0.0257824529,
-0.0419083349,
-0.0483396463,
-0.0411944166,
0.0143347951,
-0.0234562922,
-0.0294614621,
0.0252749547,
0.0242785495,
-0.0188019536,
-0.0799233615,
0.052622512,
-0.0562197231,
-0.0428850427,
-0.0053397939,
-0.0111657195,
0.0377221368,
-0.0288909506,
0.0054361653,
-0.0602313317,
-0.0379726812,
-0.0471208058,
0.0042379904,
-0.0026319488,
0.0176103171,
0.0066399081,
0.0248955395,
-0.067722477,
-0.0099212192,
-0.0228744466,
-0.0111931702,
0.029897457,
0.0288365129,
0.0168782547,
-0.0324206911,
0.0231245495,
0.0290385373,
-0.0003563955,
-0.0380820148,
-0.0102294795,
0.0063442509,
-0.0161015317,
0.0039699422,
-0.0369988941,
0.0061592143,
-0.0772087798,
0.0218165088,
0.0463928208,
-0.0280447192,
-0.0313294046,
-0.0182342958,
2.622e-7,
-0.0000186689,
0.0402756594,
-0.0098045431,
0.0152073577,
0.0289354678,
0.0312027242,
-0.0432856455,
0.0106731709,
-0.0140996324,
0.0117111653,
-0.0133718019,
0.0652781576,
0.0421918444,
0.0216355994,
-0.0928632617,
-0.0018153539,
-0.029795995,
0.0392924733,
-0.0265736766,
-0.0391370989,
-0.0066077854,
0.1019868106,
0.0459749699,
-0.0080290586,
-0.0197689123,
-0.095156081,
0.0563763082,
-0.0391250849,
-0.0310664922,
0.0694669262,
-0.0160807557,
0.0158690959,
-0.0170353893,
-0.0018735868,
0.0144260107,
0.0223743767,
-0.0156227984,
0.0812202469,
-0.0023044341,
0.0195468143,
0.0485741384,
-0.0360756181,
-0.0104144886,
-0.0180395208,
0.0487382673,
-0.043846976,
-0.0266164634,
-0.0153465541,
0.0101853954,
0.0024681191,
-0.0124113951,
-0.0209934842,
-0.0266663972,
0.0151540302,
0.0204316564,
-0.0533611551,
0.0099468259,
-0.055298347,
-0.0068212277,
-0.0346150324,
-0.0340428576,
-0.117933251,
-0.0268045515,
0.0088842176,
0.0245981254,
-0.0290345121,
0.0032429299,
2.45730413e-34,
0.0723876953,
0.0133303031,
0.0204699822,
0.0580613576,
0.0303465258,
-0.0051152236,
-0.0474318452,
-0.0078760963,
0.0424089655,
-0.1180407852,
-0.0061988416
] |
http://arxiv.org/abs/2402.02037v3 | http://arxiv.org/pdf/2402.02037v3 | EffiBench: Benchmarking the Efficiency of Automatically Generated Code | 2024-02-03T05:24:39 | 2024-06-07T09:21:21 | Dong Huang, Yuhao Qing, Weiyi Shang, Heming Cui, Jie M. Zhang | cs.SE | cs.SE, cs.CL | Code generation models have increasingly become integral to aiding software
development. Although current research has thoroughly examined the correctness
of the code produced by code generation models, a vital aspect that plays a
pivotal role in green computing and sustainability efforts has often been
neglected. This paper presents EffiBench, a benchmark with 1,000
efficiency-critical coding problems to assess the efficiency of code generated
by code generation models. EffiBench contains a diverse set of LeetCode coding
problems. Each problem is paired with an executable human-written canonical
solution, which obtains the SOTA efficiency on the LeetCode solution
leaderboard. With EffiBench, we empirically examine the ability of 42 large
language models (35 open-source and 7 closed-source) to generate efficient
code. Our evaluation results demonstrate that the efficiency of the code
generated by LLMs is generally worse than the efficiency of human-written
canonical solutions. For example, GPT-4 generated code has an average
\textbf{3.12} times execution time that of the human-written canonical
solutions. In the most extreme cases, the execution time and total memory usage
of GPT-4 generated code are \textbf{13.89} and \textbf{43.92} times that of the
canonical solutions. The source code of EffiBench is released on
https://github.com/huangd1999/EffiBench. We also provide the LeaderBoard at
https://huggingface.co/spaces/EffiBench/effibench-leaderboard. | EffiBench: Benchmarking the Efficiency of Automatically Generated Code: Code generation models have increasingly become integral to aiding software
development. Although current research has thoroughly examined the correctness
of the code produced by code generation models, a vital aspect that plays a
pivotal role in green computing and sustainability efforts has often been
neglected. This paper presents EffiBench, a benchmark with 1,000
efficiency-critical coding problems to assess the efficiency of code generated
by code generation models. EffiBench contains a diverse set of LeetCode coding
problems. Each problem is paired with an executable human-written canonical
solution, which obtains the SOTA efficiency on the LeetCode solution
leaderboard. With EffiBench, we empirically examine the ability of 42 large
language models (35 open-source and 7 closed-source) to generate efficient
code. Our evaluation results demonstrate that the efficiency of the code
generated by LLMs is generally worse than the efficiency of human-written
canonical solutions. For example, GPT-4 generated code has an average
\textbf{3.12} times execution time that of the human-written canonical
solutions. In the most extreme cases, the execution time and total memory usage
of GPT-4 generated code are \textbf{13.89} and \textbf{43.92} times that of the
canonical solutions. The source code of EffiBench is released on
https://github.com/huangd1999/EffiBench. We also provide the LeaderBoard at
https://huggingface.co/spaces/EffiBench/effibench-leaderboard. | [
0.0149240419,
0.0608986989,
-0.0081677521,
0.0421119556,
-0.0294512082,
-0.0201584063,
-0.0516531207,
0.0353120863,
-0.0035117525,
-0.0246687289,
-0.0103295725,
0.0819613561,
-0.0037013027,
0.0459842123,
0.0124022113,
-0.0025580281,
0.0609187596,
-0.0024903652,
-0.0446607023,
-0.005828131,
-0.0255087633,
-0.0082637323,
0.0313560478,
0.0084147276,
-0.0032789435,
-0.0040874137,
-0.0155468481,
0.0812898129,
-0.0055092326,
-0.0086293211,
0.0330939852,
0.0535603538,
-0.0231758747,
0.0011471761,
0.0000025731,
-0.0224255193,
-0.0086147888,
0.0194378663,
-0.0022476346,
0.0848074183,
0.0458442681,
-0.0119491657,
-0.018025903,
0.0230377447,
-0.0034989899,
-0.0186238252,
0.0352526121,
-0.0832368806,
0.0382429995,
0.0120577291,
0.0077003068,
0.0052431403,
0.0178533942,
0.0138304541,
0.0294783674,
0.0218403656,
-0.0160538964,
0.013220937,
0.0001250213,
-0.0149308853,
0.0074126129,
0.039091792,
0.0296525341,
-0.0148979519,
0.0179131627,
0.0035840473,
-0.0455926806,
-0.0134818992,
-0.0186719913,
0.0235442128,
-0.022485802,
-0.0227795057,
-0.001005044,
-0.019747287,
-0.0580518171,
-0.020904487,
-0.0069667865,
0.035541296,
0.0061034034,
0.0007588,
0.0051719304,
0.03461878,
-0.0121206716,
0.0033581958,
-0.0020567919,
0.0509513579,
-0.0061257286,
-0.0171210226,
0.0047668521,
-0.0006498831,
-0.0283047091,
-0.0274415724,
-0.0001633451,
-0.0149822365,
0.0366470441,
-0.0196305905,
-0.0200112388,
0.0108878752,
0.0306392703,
-0.0340621062,
0.0538901128,
0.0609607287,
0.0808634907,
-0.035558451,
-0.0327125452,
0.0565583222,
-0.0577212796,
-0.0507000796,
-0.0149610154,
0.020028444,
-0.0130112721,
-0.0063283807,
-0.0470483676,
0.0170563906,
-0.0044992235,
0.0382532366,
-0.0171628203,
0.0288874228,
-0.0096583152,
0.0367086045,
0.0271396451,
-0.0005055732,
0.0380830728,
-0.0083963331,
0.0438103527,
-0.0251350086,
-0.0076695853,
-0.0360640027,
-0.0294927489,
-0.0960729644,
0.0355330631,
0.0159726739,
-0.0234998353,
0.0065321014,
0.0181661285,
0.0645409599,
0.0284437593,
0.0126068518,
-0.0173746087,
-0.0535624586,
0.0449271426,
-0.0230360366,
-0.0380147062,
-0.064686209,
-0.0213955007,
-0.0151343895,
0.0136347637,
-0.0277198367,
0.0205180477,
-0.0392423831,
-0.0109791458,
0.0183560252,
-0.0858791098,
0.0663468391,
0.011800888,
0.0544118769,
-0.0051583522,
0.0585574768,
0.0551959984,
-0.0029517363,
0.0123928543,
-0.0749824345,
0.0283409152,
0.0327110663,
-0.0080619687,
0.0095368633,
0.0098911831,
0.0213837717,
-0.0123785893,
0.0046309996,
0.0006316948,
0.0405509062,
-0.0005410317,
0.0389445163,
0.0014919546,
0.0188976806,
-0.004455572,
0.0141195282,
0.00718715,
0.0398801453,
0.0227447916,
-0.0169051327,
-0.0152739799,
0.0565506741,
-0.0361054577,
-0.0175833441,
-0.0030966052,
0.0158197451,
0.0075873332,
-0.0635063946,
-0.0300480332,
-0.0178338625,
-0.0323248357,
-0.0100768171,
-0.0267634522,
-0.0300314911,
-0.0550244823,
0.0587213524,
-0.0309185907,
-0.1255665272,
-0.0576286092,
-0.0477351137,
0.0465363637,
0.0266419444,
0.0285720788,
-0.0003550375,
-0.0100999782,
0.0146775935,
-0.0447075814,
-0.0484877713,
0.0582483932,
-0.0149146691,
0.0006521069,
-0.0501201227,
-0.0109356046,
-0.0315833837,
-0.011592662,
0.0199616551,
0.0400986709,
-0.0303301457,
0.0524543896,
-0.0187836792,
-0.0222238414,
0.0108809043,
0.0249190386,
0.013938549,
0.0223506279,
-0.0159947257,
0.0402787812,
-0.0292121787,
-0.0369505137,
0.0019754535,
0.056979131,
-0.0150054786,
-0.0292465426,
-0.0145737557,
-0.0307115596,
0.0209575854,
-0.0002597824,
-0.0601394475,
-0.0271630716,
0.0784067884,
-0.0155199999,
-0.0076090856,
-0.0161623508,
-0.0233614706,
-0.0009345251,
0.0108934753,
0.0423213802,
-0.0444887169,
0.0122577092,
0.0016096667,
0.0506411456,
-0.0097030252,
0.0525908731,
-0.0063308333,
0.0702446327,
0.0603830516,
0.0198347606,
0.0200553257,
-0.0033123859,
-0.0352228507,
-0.0147546967,
0.0388519093,
0.0611229166,
-0.0144003266,
-0.0323060118,
-0.106731832,
0.0584593527,
0.0056428663,
0.0355837122,
0.0484572984,
-0.002237519,
0.0054164645,
0.0623902678,
0.0090836138,
0.0305352528,
0.0021799272,
0.0206475779,
0.0044547399,
-0.0623966344,
-0.0480810329,
0.0093362527,
0.0010348115,
-0.0367455035,
0.0376403593,
-0.0191493891,
-0.0154367536,
-0.0817965716,
0.1044923887,
0.031875968,
-0.0121317832,
-0.0014774001,
-0.0567969233,
-0.0295134168,
-0.0219510198,
0.0798969939,
0.0315279141,
0.0098240217,
0.0010031144,
0.0125230635,
-0.0007088843,
-0.0647839904,
0.0260903072,
0.0761368871,
-0.0299809054,
0.0761177614,
-0.0692162961,
0.0498359166,
0.0456770286,
-0.0227414649,
-0.032452926,
-0.0083084926,
0.0511142127,
0.0056801066,
-0.0219327491,
-0.0052803657,
-0.019950483,
0.0134592308,
0.062266387,
-0.0049576531,
-0.1123654246,
-0.0088610556,
-0.05265994,
-0.0421918705,
0.0109380279,
-0.0245728362,
-0.036076881,
0.0381649211,
0.0178209879,
0.0430764519,
-0.022596648,
-0.0097900908,
0.0084487591,
0.0381577685,
0.0008288312,
-0.0282190386,
0.0352464579,
-0.0433829017,
0.0098997392,
-0.0356595665,
-0.055043634,
-0.0126757212,
0.0025590546,
0.0262329839,
-0.0189608764,
0.0165696759,
-0.0236713607,
0.0392372049,
0.0267816596,
0.0404211991,
0.0272238199,
0.0228298567,
-0.0389044136,
0.0292992331,
-0.0045062886,
-0.0474418215,
-0.0490382388,
-0.0114895813,
0.034789212,
0.0062126839,
0.0615585893,
0.0128960004,
0.0631515831,
-0.0250572935,
0.0122866146,
0.0577937886,
0.0296928249,
-0.0668478534,
0.0628085583,
0.0439554974,
-0.0354344919,
0.0074924952,
0.0292100459,
0.0163986254,
-0.0067514325,
0.0320826843,
-0.021230204,
0.03690942,
0.0490711443,
0.0331948958,
-0.0743293017,
-0.0184062086,
0.0306825992,
0.0236827433,
-0.0384494737,
0.0027323137,
-0.0491204113,
-0.0456682481,
0.0562838539,
0.0119107608,
-0.0223296173,
-0.0164225623,
-0.0326168127,
-0.045883067,
0.0027544855,
-0.0685135424,
0.0364024267,
0.0096484218,
0.0321863256,
-0.0065534515,
-0.0372592919,
0.0109963035,
-0.0010326684,
0.0197955891,
0.0056872489,
0.0859885812,
0.0252057016,
0.005069145,
-0.0291602351,
0.037847627,
-0.0248055905,
0.0024685673,
-0.0625851229,
-0.0241362099,
0.0458078496,
-0.0253136903,
0.0297482815,
0.0175981466,
-0.0702912211,
0.012814139,
0.0518875644,
0.0612425506,
0.0290753767,
-0.028660886,
-0.0138686951,
-0.0284170639,
-0.0056877998,
0.0455504209,
0.0203273259,
0.0238627512,
-0.0791379213,
-0.0978532881,
-0.009026737,
-0.058912497,
-0.0245396569,
-0.0485526659,
-0.0166219715,
-0.050737109,
0.0213835128,
0.038407173,
-0.0269214865,
0.0234356355,
-0.0108344741,
-0.05613143,
0.0052700462,
-0.021640284,
-0.0696849227,
-0.0306799151,
-0.0601532646,
0.0250822883,
-0.0567700379,
-0.0337849483,
0.0602627471,
0.0506894737,
0.0131710069,
0.0298004039,
-0.0554000065,
-0.0618736893,
-0.0249354076,
0.0843564719,
0.0252544265,
-0.0426407717,
0.0268864594,
0.0308815073,
-0.0458308235,
0.0743829757,
-0.0172211137,
0.0645216256,
0.0046835635,
0.0098053906,
0.0210221689,
-0.025286952,
-0.0105387243,
0.0133870896,
-0.008817059,
0.0421691649,
-0.0012919947,
-0.013453152,
0.0282216147,
0.0937271565,
-0.0190421604,
-0.0227058493,
-0.0390367843,
0.0608575866,
-0.0166215524,
0.0025091073,
-0.0340935811,
-0.0723100379,
0.0358970501,
-0.0091061313,
0.002035189,
0.0125048049,
0.0497126579,
-0.0478284098,
0.0183944814,
0.0128657594,
0.0482612438,
-0.0368410721,
0.0398445167,
0.0679060966,
-0.0017200656,
-0.0371682383,
-0.0155537594,
0.0248809699,
-0.0079310704,
-0.0344246589,
0.0089841271,
0.0076026563,
0.0622908883,
0.0240992904,
0.0364494845,
0.003530758,
-0.053992264,
0.0004069973,
0.009507359,
0.0304888841,
0.0764537975,
-0.0057170955,
-0.0390651003,
-0.0170743987,
-0.0232455339,
0.0267688707,
-0.0245221369,
-0.0111045642,
0.0137143582,
0.0210204944,
-0.0511189811,
-0.0817931145,
0.1073409766,
0.045700781,
0.0056087486,
-0.0426937677,
-0.0021348421,
-0.0314717703,
-0.0104997987,
-0.0122420592,
-0.0478768349,
0.0186484903,
-0.0185045768,
-0.0348499753,
-0.0335667655,
0.0032809682,
0.0359547362,
0.0700412244,
0.0118699735,
0.0324318483,
0.037470974,
-0.0103148259,
-0.0464600027,
0.0051680347,
0.0008364021,
-0.0462711304,
-0.0249001477,
0.0337460637,
-6.977741079e-33,
-0.0244511496,
-0.0541533493,
-0.0146565214,
0.038130112,
0.0031775909,
-0.0005991458,
-0.0314411484,
-0.0037701714,
-0.003386762,
-0.0244173817,
0.0096009662,
-0.0240766238,
-0.0060904659,
-0.0243231189,
0.0226381347,
0.0450270213,
0.0065539018,
0.0325479358,
-0.0600165837,
-0.0310375094,
-0.0265400335,
0.013126702,
-0.0274837725,
-0.0115849338,
-0.0660883933,
-0.0228562318,
0.0053851921,
-0.0088669984,
-0.0160369184,
-0.0232636444,
-0.0035630574,
0.0159291867,
-0.0070599499,
0.0047178119,
-0.0318786912,
0.0109210294,
-0.0326331779,
-0.0524829961,
0.0186947454,
-0.0193187799,
-0.0235090517,
-0.0696560517,
0.0377339683,
0.0373543054,
-0.0717535168,
0.0223567206,
-0.0342092812,
0.0030947991,
-0.0268671345,
-0.0247608032,
-0.0424452871,
0.0061040618,
-0.0168478973,
0.0430514887,
-0.0320949405,
0.0173050109,
0.0019163617,
0.0404781997,
0.0054594255,
0.016382644,
0.0228920393,
0.049835816,
0.065330416,
-0.068974264,
0.0178783014,
-0.0221817158,
0.0337740183,
-0.0056787259,
-0.0723320246,
0.0079150479,
0.022765249,
0.000383244,
0.0197262894,
0.0088836495,
0.0465679094,
-0.0368189327,
-0.0239385944,
0.0460771732,
-0.0035631966,
-0.0341819897,
-0.0322814584,
0.0135598108,
0.0049444754,
-0.046865467,
-0.0435743071,
0.0362860747,
-0.0186816733,
-0.0149023877,
0.0417653471,
-0.0281697623,
-0.0339374729,
-0.0297215059,
0.0034316818,
-0.0153480237,
0.0338632464,
0.0696634203,
-0.0250485167,
0.0015962623,
0.0151681602,
0.0500714965,
0.0022038177,
-0.0423421822,
-0.0446861722,
0.0322511457,
0.0880962238,
0.0114622824,
-0.0099836104,
0.0312264618,
-0.0735247731,
0.0237923265,
-0.0652406663,
-0.0088848919,
0.0032311794,
-0.0563605726,
-0.0432454981,
-0.0307300035,
0.0350339636,
0.0333577283,
0.0024625824,
-0.028868651,
-0.0434544832,
-0.0160579681,
0.0142644495,
0.0494436957,
0.0054399963,
-0.0219979677,
-0.0601441972,
0.0210102312,
0.0159068815,
0.0132952053,
-0.0731081963,
0.056968689,
3.281e-7,
-0.0030038559,
-0.0080669383,
-0.0066685025,
0.019777514,
0.0079328986,
0.0354638435,
0.0080722971,
-0.0125649096,
-0.0415805057,
0.0075952108,
-0.0294167716,
-0.0244283006,
0.0578779168,
0.004187,
0.0251348168,
0.0082707833,
-0.004192967,
-0.0232431442,
-0.0347890444,
-0.0013640472,
0.0060170591,
-0.0401126258,
0.0130198672,
-0.0187139735,
0.0110847512,
0.0008045264,
0.025219271,
-0.058170341,
0.0548891686,
0.0240455847,
-0.0734742507,
0.0051277648,
-0.0148810241,
-0.0256949589,
-0.0013634426,
-0.0488801748,
0.0172051452,
0.0848123804,
0.0278916508,
0.0471247137,
0.0726402029,
-0.053625863,
-0.026604414,
-0.0219519604,
0.0331244543,
0.0493570603,
-0.0172689948,
0.0225807279,
-0.0194412619,
-0.0053876908,
-0.0289420299,
-0.0424091555,
0.0273469985,
0.0008257062,
-0.014878137,
-0.0070532793,
-0.0238527954,
0.0247095358,
-0.0101785082,
-0.0520817414,
0.0000312113,
-0.0653025582,
-0.0290609822,
0.0818495303,
0.0664518327,
0.0146558583,
-0.0187978167,
3.238403366e-34,
0.009088818,
-0.0448235385,
0.0363351963,
0.0078924224,
0.0561677888,
0.0262140315,
-0.0266469512,
-0.0303971674,
-0.0426853672,
0.0324237049,
-0.0336463936
] |
http://arxiv.org/abs/2402.02030v2 | http://arxiv.org/pdf/2402.02030v2 | Panacea: Pareto Alignment via Preference Adaptation for LLMs | 2024-02-03T05:01:04 | 2024-05-23T13:49:25 | Yifan Zhong, Chengdong Ma, Xiaoyuan Zhang, Ziran Yang, Haojun Chen, Qingfu Zhang, Siyuan Qi, Yaodong Yang | cs.CL | cs.CL | Current methods for large language model alignment typically use scalar human
preference labels. However, this convention tends to oversimplify the
multi-dimensional and heterogeneous nature of human preferences, leading to
reduced expressivity and even misalignment. This paper presents Panacea, an
innovative approach that reframes alignment as a multi-dimensional preference
optimization problem. Panacea trains a single model capable of adapting online
and Pareto-optimally to diverse sets of preferences without the need for
further tuning. A major challenge here is using a low-dimensional preference
vector to guide the model's behavior, despite it being governed by an
overwhelmingly large number of parameters. To address this, Panacea is designed
to use singular value decomposition (SVD)-based low-rank adaptation, which
allows the preference vector to be simply injected online as singular values.
Theoretically, we prove that Panacea recovers the entire Pareto front with
common loss aggregation methods under mild conditions. Moreover, our
experiments demonstrate, for the first time, the feasibility of aligning a
single LLM to represent an exponentially vast spectrum of human preferences
through various optimization methods. Our work marks a step forward in
effectively and efficiently aligning models to diverse and intricate human
preferences in a controllable and Pareto-optimal manner. | Panacea: Pareto Alignment via Preference Adaptation for LLMs: Current methods for large language model alignment typically use scalar human
preference labels. However, this convention tends to oversimplify the
multi-dimensional and heterogeneous nature of human preferences, leading to
reduced expressivity and even misalignment. This paper presents Panacea, an
innovative approach that reframes alignment as a multi-dimensional preference
optimization problem. Panacea trains a single model capable of adapting online
and Pareto-optimally to diverse sets of preferences without the need for
further tuning. A major challenge here is using a low-dimensional preference
vector to guide the model's behavior, despite it being governed by an
overwhelmingly large number of parameters. To address this, Panacea is designed
to use singular value decomposition (SVD)-based low-rank adaptation, which
allows the preference vector to be simply injected online as singular values.
Theoretically, we prove that Panacea recovers the entire Pareto front with
common loss aggregation methods under mild conditions. Moreover, our
experiments demonstrate, for the first time, the feasibility of aligning a
single LLM to represent an exponentially vast spectrum of human preferences
through various optimization methods. Our work marks a step forward in
effectively and efficiently aligning models to diverse and intricate human
preferences in a controllable and Pareto-optimal manner. | [
-0.0106387231,
0.0722791627,
0.0053348909,
0.0375350751,
0.0045902873,
0.0246731658,
-0.0261711478,
-0.0147003345,
-0.0283156577,
-0.0257932283,
-0.0703292266,
-0.0299143195,
-0.0017618636,
-0.0199655257,
0.0328592733,
-0.0552081428,
0.0540379696,
-0.0052294657,
-0.019338185,
-0.0270776469,
0.0430062748,
0.0207548719,
0.0040066415,
0.0040828502,
0.0341159217,
-0.0229768492,
0.0142318774,
0.0118374769,
0.0538594201,
0.0043727332,
0.0158190038,
0.0419504121,
0.0348311514,
0.0164362341,
0.0000017866,
-0.0561492331,
-0.0407366008,
-0.0216578711,
-0.0286653638,
0.0060011633,
0.0345867276,
0.019625511,
-0.0374206379,
-0.0020190317,
-0.0133148544,
0.014284947,
0.0660879239,
-0.0106998039,
0.043503046,
0.0858706161,
-0.005267614,
-0.0531214327,
-0.0190451015,
-0.018904591,
-0.0228143316,
-0.0409601144,
0.0166428294,
-0.0646417439,
-0.0366734341,
-0.0204054266,
-0.0132321874,
0.0263712704,
-0.0324824713,
0.0302315559,
0.0310571734,
0.0124028781,
-0.0141210807,
-0.0134819569,
-0.0243670363,
0.0866465494,
-0.03162333,
-0.0190713741,
0.0180470683,
-0.0059487629,
0.0100841122,
-0.0407527238,
-0.0410355926,
-0.0230599046,
0.0273926388,
0.0409361161,
0.009374422,
-0.0350474119,
0.0075258161,
0.0121906986,
-0.0283639915,
0.0827477574,
0.0418016762,
-0.0448556393,
0.0249146204,
-0.0133075323,
-0.0305978209,
-0.023050813,
0.0183250681,
-0.0061276504,
-0.0778822303,
0.0342441052,
-0.0664663836,
-0.0452912934,
0.0005161151,
-0.0766762197,
0.0667559206,
0.0517075285,
-0.008281054,
0.0283604953,
0.0284036919,
0.0713330731,
-0.0674540326,
0.0148622282,
-0.0712422058,
0.0087098442,
0.0042360937,
0.0039069573,
0.0204988196,
0.0286406819,
-0.0150599293,
0.0080974195,
-0.048871696,
0.0567564964,
0.0102395704,
0.0494649895,
0.0126803545,
0.0175890196,
-0.0325464495,
-0.0445621759,
-0.0180101525,
0.1010406166,
-0.0148946512,
0.0168375466,
-0.0054346994,
0.0236331169,
0.0138246315,
0.0686795637,
-0.003106466,
-0.0284378733,
0.0006085854,
0.0717371926,
-0.0485291258,
0.0146222627,
-0.0007608708,
0.0007765855,
-0.0363765061,
-0.0447810739,
-0.0179907735,
-0.0500606932,
0.0515887588,
0.0059429519,
-0.0126574375,
-0.0636145324,
-0.020209888,
-0.0092006978,
-0.0296766628,
0.0678484589,
0.0416090786,
0.0134029165,
0.0102094542,
-0.0373267755,
0.0034961523,
0.0601054728,
0.0301154573,
0.0369526036,
0.0161140356,
-0.032569129,
-0.0501269549,
0.0776782408,
-0.014050208,
0.0143883508,
-0.0622792505,
0.0278962348,
0.0302559603,
0.0514716543,
0.0698600262,
0.0334650613,
0.0186320525,
-0.0334278047,
0.0728606358,
0.052526243,
0.0289589819,
0.0176801588,
-0.0273379534,
0.0390129946,
-0.041683767,
0.0119971307,
-0.0404602401,
0.0353652872,
-0.0437007584,
0.0126559166,
0.0269768834,
0.0517676137,
-0.0136144143,
-0.0206565373,
-0.018947497,
-0.0054150485,
-0.0364837274,
0.0058461321,
0.0322432108,
-0.065686442,
-0.054491207,
0.0634724796,
0.0266777053,
-0.0786836296,
0.0180859007,
0.0016794095,
0.040225435,
0.077191256,
0.0132959168,
-0.0139395343,
-0.0543838963,
0.0093458015,
-0.0182584636,
0.0550115071,
0.0371810012,
-0.0212819036,
-0.0136042964,
-0.0558876358,
-0.0324853845,
0.0517730899,
-0.0078218756,
0.042905096,
-0.0112520894,
-0.0393791534,
-0.0154765164,
-0.0054053958,
-0.0015105753,
0.0192190707,
-0.0242597144,
-0.0166315995,
0.003170308,
0.0291803293,
0.036949005,
0.0068505942,
0.0138083352,
-0.0223195106,
0.0200062394,
-0.0121830646,
0.0023055649,
0.0051109977,
-0.0190664437,
-0.0003201965,
0.0549925491,
-0.0372418575,
-0.0434906259,
0.0625725389,
-0.0349298231,
-0.0393034853,
-0.0295710452,
-0.029465029,
0.0084269587,
-0.0262782518,
0.0431315079,
-0.0160912424,
0.0037200705,
-0.0495676883,
0.0627785847,
-0.0110749295,
0.000967437,
0.0036168888,
-0.0154567417,
0.0015058219,
-0.0822618306,
-0.0232924838,
0.0189538542,
0.0030303737,
0.04164676,
0.0155613851,
0.0113095297,
-0.0379096195,
-0.0266868714,
-0.0572382733,
0.0352399871,
-0.012500532,
0.0347587615,
0.0342807248,
-0.016274929,
-0.0036447896,
-0.0290814172,
-0.0629824474,
0.0203533471,
-0.0197552722,
0.0115951402,
0.0462651886,
-0.0105923945,
-0.0262075569,
-0.0659594312,
0.0169704184,
-0.0096905315,
-0.011271216,
-0.051548522,
0.0137169622,
-0.0658542514,
0.0646376833,
-0.0423025899,
0.0066061025,
0.0007890525,
-0.0264060814,
-0.0188712627,
0.006867019,
0.0358468778,
0.074009344,
0.0210027825,
0.0003846302,
0.0666025952,
0.0130524123,
-0.0548621416,
-0.0054974784,
-0.0418093503,
-0.0163024571,
-0.0642287359,
0.0113989916,
0.0214657113,
0.0451805554,
-0.0072312397,
0.0191466063,
-0.070877865,
0.0111827236,
0.016622806,
-0.0143885966,
-0.033341296,
0.0486692451,
-0.0570798256,
0.0352299586,
-0.0160257947,
-0.0330515653,
0.0691563338,
-0.0154015077,
-0.0197853763,
-0.0380758867,
-0.0128995953,
-0.0409150682,
-0.0263684113,
-0.0148944445,
0.0233623944,
-0.0327171758,
-0.0042788521,
-0.1084045991,
-0.0195477791,
0.0111020515,
-0.0256743841,
0.071295701,
-0.0328887142,
0.0082057575,
-0.0260536559,
-0.0897661895,
-0.0359320864,
-0.0462327637,
-0.0058085634,
-0.0281063057,
0.0097451005,
-0.0513340048,
-0.0468979776,
-0.0253242608,
0.0127281733,
0.00625447,
-0.0038840536,
0.0239459239,
-0.0352121927,
0.0212487374,
0.0857939422,
0.0652236864,
-0.0094226953,
-0.0034441848,
-0.0103335781,
0.0384535305,
0.0513708107,
-0.0113174561,
-0.0015872627,
0.01070162,
0.07024353,
-0.0366110802,
-0.0051534879,
0.0109961405,
0.0176067315,
0.0105813658,
0.0290517733,
0.1052512676,
0.014941019,
-0.0761767477,
-0.0001174519,
-0.0130387442,
0.0695373341,
0.024402773,
0.0374301001,
-0.044263307,
-0.0067438604,
-0.0102678742,
0.0425131992,
0.0327725299,
0.013525106,
-0.0182793234,
-0.1052372307,
0.0019824982,
0.0388992801,
-0.0313354358,
0.0293786023,
-0.016562501,
0.0260272846,
-0.0550016351,
0.077035524,
0.0191808082,
-0.0123452572,
-0.0253954176,
0.0091454955,
-0.0117708789,
-0.0197986793,
0.0013165581,
-0.0206915103,
0.0554571673,
0.058220081,
0.0367390066,
-0.0005928149,
-0.0743262321,
0.0143304151,
0.0254481807,
-0.0425652191,
-0.0015999894,
-0.0036433716,
-0.0185422488,
-0.0055043828,
-0.0153397564,
0.063862212,
-0.0460912324,
0.011016164,
-0.0078000454,
0.0830176994,
0.0076110112,
0.0354106016,
0.0188452024,
-0.0213885047,
-0.0283351224,
0.0622933842,
-0.0003753199,
-0.0761889294,
-0.0124280807,
-0.0380249843,
-0.0464651212,
-0.0363219827,
-0.0366340764,
0.0034030227,
-0.0412754901,
-0.014280105,
-0.0548718534,
-0.0040915883,
-0.0062248134,
0.0346720144,
0.0073125577,
-0.0086279763,
0.0082096905,
-0.0223645326,
-0.0125240879,
0.0339666009,
-0.0329501256,
-0.0101062581,
-0.0740131736,
-0.0086561926,
0.0326108038,
-0.0198394973,
-0.0030609029,
-0.0414607078,
-0.0423619449,
-0.0886514187,
-0.0019717363,
0.0190549139,
0.0172857996,
0.0556135364,
0.0152456257,
-0.0747445673,
0.0304934792,
0.036821261,
0.021321075,
0.0528669953,
0.0278476123,
0.0356512628,
-0.0115947984,
0.0513849705,
-0.0569641888,
-0.0089267828,
0.0051599997,
0.0179396458,
-0.032093022,
-0.048302833,
0.0182385836,
0.0239723902,
0.0229512788,
-0.0419376083,
-0.0408681706,
-0.019745145,
0.0222956147,
-0.01325492,
-0.0225969851,
0.0040480415,
-0.0437322818,
0.0385165028,
-0.0164340306,
0.0027189455,
0.0179961547,
-0.0057613943,
0.0284742881,
-0.0522844084,
-0.0085676899,
-0.0006621869,
-0.0105634807,
0.0399421304,
0.0514362678,
-0.0061442927,
0.0312908255,
0.0125549641,
0.0525411181,
-0.0088595413,
0.0165881775,
0.0317138396,
0.0120330183,
-0.0274150651,
0.0404688567,
-0.0032772345,
-0.0569877587,
-0.0380580686,
-0.0290738177,
0.0512276813,
-0.0077833734,
0.0871244892,
0.0125401635,
-0.0183395091,
-0.0319255926,
0.0421441197,
-0.0455629937,
0.012871013,
0.03393225,
0.0008013935,
-0.0187699217,
0.0074061309,
0.045327194,
0.0059789773,
0.0091714077,
-0.0504485369,
0.0157321934,
-0.0521403924,
-0.0269705188,
-0.013931579,
0.0124419332,
-0.0363573395,
-0.0516804196,
-0.0408556797,
-0.0552558415,
0.0116650751,
0.0033519182,
-0.0109862294,
-0.0053310716,
0.0362981223,
0.0379454754,
-0.0588302054,
0.0141488994,
0.0070028305,
-0.021509815,
-0.0279860236,
-0.0306797065,
-0.0174975544,
-6.148264878e-33,
-0.0443208255,
-0.0138566503,
0.0231833886,
0.007025104,
-0.0259917565,
-0.0713473111,
0.02138393,
0.0419978723,
0.0123622296,
-0.0361307152,
0.0050716866,
-0.0382417738,
0.0207609478,
0.0067626485,
0.0408590436,
0.0011673127,
0.0318846554,
-0.0104125934,
-0.0438389853,
-0.0275881849,
0.0528376922,
0.0186398365,
0.0485853739,
-0.0541836433,
-0.0013432604,
0.0214615166,
-0.0231298525,
0.038595181,
0.0094546722,
0.0780520439,
-0.0004355985,
0.0456365012,
0.0045388774,
0.0058771707,
-0.0202447735,
0.0457364395,
-0.0109335296,
-0.005106885,
0.0480512679,
0.0421637483,
0.0458324812,
-0.0503183007,
0.0384685099,
0.0014639553,
-0.0514437482,
-0.0012790167,
0.0427573845,
-0.0335611589,
-0.0227517113,
0.0333658978,
-0.0470548645,
-0.0003523734,
-0.0045019924,
-0.0033207485,
0.0406702571,
0.0450593047,
-0.0013906538,
0.0389130078,
-0.0532680079,
0.0440646596,
-0.0379095152,
0.0300154369,
0.1169059277,
0.028796507,
-0.0312707163,
0.0429549813,
0.035671182,
-0.0455593579,
0.0324852467,
-0.0384678617,
0.0305554755,
0.0144537985,
0.0173908882,
0.04413753,
-0.0103974761,
0.0012769017,
-0.0159482993,
0.0465640612,
0.0094038583,
-0.0151175428,
0.0147705777,
0.0788559839,
-0.02134162,
-0.0199508127,
-0.0252766516,
-0.0183843412,
-0.04461601,
-0.017956974,
0.0218655095,
-0.013881824,
-0.0190413818,
-0.0075751869,
0.0440592542,
-0.0360883623,
-0.0082763871,
-0.0074940161,
-0.0351957083,
0.0273043681,
0.0126196807,
0.0057916017,
-0.0481235385,
-0.0321748555,
-0.0107426932,
-0.0182391852,
-0.0006759099,
0.0183845889,
0.0249782093,
0.010182308,
-0.0647021383,
0.0232413076,
-0.0053115347,
-0.063290976,
0.0021999306,
0.0417952426,
0.0398532227,
-0.0246362053,
0.0119064478,
0.0825408325,
0.0212403852,
0.0103696436,
-0.0417540297,
-0.0342834815,
0.0011931058,
0.0109068314,
-0.0175997224,
0.0355674215,
0.0119366935,
0.005108661,
0.0257342905,
-0.0163836479,
0.0011390835,
-0.0273062382,
2.509e-7,
0.0361086428,
0.0490089133,
0.000661038,
-0.031393636,
-0.0072236937,
0.0550123155,
-0.0298481602,
0.0311891157,
-0.0094199516,
-0.0086639114,
0.0299811158,
0.0022031618,
0.050247062,
0.0018308368,
-0.0783130676,
-0.0000936426,
-0.0416968651,
0.0391767696,
-0.048255451,
0.0042111767,
0.0399858169,
0.0971147195,
0.0354125053,
-0.0095230145,
-0.0192024186,
0.047597073,
0.0692423359,
-0.0417155325,
-0.0160556119,
-0.0023965493,
-0.0024036746,
-0.0315837339,
-0.0495356694,
-0.0330752581,
0.0024984533,
0.053360682,
-0.0192465056,
0.1085270494,
-0.0145472279,
-0.0676410943,
-0.0064131413,
0.0265428368,
-0.0082203047,
-0.0382717326,
0.0229394846,
-0.0348585024,
-0.0184865408,
-0.0374712721,
-0.0188399721,
0.0243365113,
0.0074974461,
-0.0290641245,
0.0200071242,
-0.0040547131,
0.0230585523,
-0.0140667558,
-0.0191488322,
-0.0558090359,
-0.0205390695,
0.0436191,
-0.0661386997,
-0.045319207,
-0.0065215621,
0.0315139592,
0.038246274,
-0.0363551565,
-0.0026175543,
2.036376822e-34,
0.0018824298,
0.011615566,
0.0102233673,
0.0463341288,
-0.0111686597,
-0.015239899,
0.000203964,
0.0322505571,
0.0384088717,
-0.0846773833,
0.0020049717
] |
http://arxiv.org/abs/2402.05119v4 | http://arxiv.org/pdf/2402.05119v4 | A Closer Look at the Limitations of Instruction Tuning | 2024-02-03T04:45:25 | 2024-05-28T01:55:38 | Sreyan Ghosh, Chandra Kiran Reddy Evuru, Sonal Kumar, Ramaneswaran S, Deepali Aneja, Zeyu Jin, Ramani Duraiswami, Dinesh Manocha | cs.CL | cs.CL, cs.AI | Instruction Tuning (IT), the process of training large language models (LLMs)
using instruction-response pairs, has emerged as the predominant method for
transforming base pre-trained LLMs into open-domain conversational agents.
While IT has achieved notable success and widespread adoption, its limitations
and shortcomings remain underexplored. In this paper, through rigorous
experiments and an in-depth analysis of the changes LLMs undergo through IT, we
reveal various limitations of IT. In particular, we show that (1) IT fails to
enhance knowledge or skills in LLMs. LoRA fine-tuning is limited to learning
response initiation and style tokens, and full-parameter fine-tuning leads to
knowledge degradation. (2) Copying response patterns from IT datasets derived
from knowledgeable sources leads to a decline in response quality. (3)
Full-parameter fine-tuning increases hallucination by inaccurately borrowing
tokens from conceptually similar instances in the IT dataset for generating
responses. (4) Popular methods to improve IT do not lead to performance
improvements over a simple LoRA fine-tuned model. Our findings reveal that
responses generated solely from pre-trained knowledge consistently outperform
responses by models that learn any form of new knowledge from IT on open-source
datasets. We hope the insights and challenges revealed in this paper inspire
future work in related directions. | A Closer Look at the Limitations of Instruction Tuning: Instruction Tuning (IT), the process of training large language models (LLMs)
using instruction-response pairs, has emerged as the predominant method for
transforming base pre-trained LLMs into open-domain conversational agents.
While IT has achieved notable success and widespread adoption, its limitations
and shortcomings remain underexplored. In this paper, through rigorous
experiments and an in-depth analysis of the changes LLMs undergo through IT, we
reveal various limitations of IT. In particular, we show that (1) IT fails to
enhance knowledge or skills in LLMs. LoRA fine-tuning is limited to learning
response initiation and style tokens, and full-parameter fine-tuning leads to
knowledge degradation. (2) Copying response patterns from IT datasets derived
from knowledgeable sources leads to a decline in response quality. (3)
Full-parameter fine-tuning increases hallucination by inaccurately borrowing
tokens from conceptually similar instances in the IT dataset for generating
responses. (4) Popular methods to improve IT do not lead to performance
improvements over a simple LoRA fine-tuned model. Our findings reveal that
responses generated solely from pre-trained knowledge consistently outperform
responses by models that learn any form of new knowledge from IT on open-source
datasets. We hope the insights and challenges revealed in this paper inspire
future work in related directions. | [
0.0597260334,
0.055529099,
-0.0308817811,
0.0169179309,
0.0025956705,
0.0150064388,
-0.0132384188,
-0.0167115778,
-0.0139996856,
-0.0637025163,
-0.0888582468,
-0.0209867507,
-0.0204544216,
0.0109586744,
0.0014657163,
-0.0457698032,
0.0430019498,
-0.0729748309,
-0.0019676657,
-0.0178224687,
-0.0029959991,
0.0144621944,
-0.0223127119,
0.054387901,
0.0175149217,
-0.0256379955,
-0.0414847881,
0.0106463498,
-0.0149425957,
-0.0138148516,
0.0445533916,
0.0610999241,
0.0383173004,
-0.0045405682,
0.0000020789,
-0.0575195402,
0.0112577137,
-0.0512314215,
-0.026677575,
0.019260047,
0.0377106369,
0.0184223074,
0.0175186135,
0.0319293477,
-0.0575423315,
-0.0384715647,
0.0154441996,
0.0441851877,
0.0507891141,
0.0707406551,
0.0070650731,
-0.042609565,
0.0369063169,
-0.0045845541,
-0.005972404,
-0.0310124699,
-0.001448157,
0.014090416,
-0.0011550526,
-0.00142515,
-0.0400324762,
0.0155175673,
0.0253275651,
0.0027651822,
0.0188974403,
0.0018086752,
-0.0690541714,
-0.0419153422,
-0.052791737,
0.0945385918,
-0.0037771668,
-0.0042392178,
-0.0193877146,
-0.0062597645,
0.0198387001,
-0.0040496481,
-0.0254664179,
0.0182288419,
0.0213226117,
0.0199221652,
-0.0187768694,
-0.0230148062,
-0.0145088555,
-0.0072618369,
-0.0223989282,
0.0475764833,
0.0000323366,
-0.0354889669,
0.0229635183,
-0.0244874712,
-0.0167470165,
-0.0662065521,
-0.0241566058,
0.0617515817,
-0.0437088571,
0.0201367512,
-0.0583654121,
0.0372243077,
0.0379390903,
0.0027782819,
0.0743372962,
0.0306375343,
-0.0228203963,
0.0034953398,
-0.0249689408,
0.0449838825,
0.0150383255,
0.0565491244,
-0.0276290551,
0.0105285067,
-0.0712208822,
-0.0174927544,
-0.0349194482,
0.0220289752,
-0.045117069,
-0.0236891173,
0.0034616559,
0.0470222831,
0.0081790639,
-0.0101480661,
-0.0215284117,
-0.0096464166,
-0.0058614369,
0.0032332405,
-0.075717099,
0.027094502,
-0.0206674431,
0.0592497736,
0.0262548272,
-0.1077099815,
-0.0110655362,
0.0121431137,
0.024409987,
0.0136410771,
0.0284482855,
0.0430878252,
-0.0153027615,
-0.0445379838,
-0.0170948375,
-0.024403207,
-0.0080851335,
-0.1067600474,
0.0161115304,
-0.036060255,
0.0074208174,
0.0244632363,
-0.0265632737,
-0.0170584265,
0.0169569124,
0.0001402016,
-0.0538245775,
0.0356464647,
-0.0076378975,
0.0345215425,
0.007921773,
0.0137138329,
0.0181069877,
0.0553414002,
0.0313256197,
0.0001986156,
-0.000907388,
-0.0415006317,
0.0283301976,
0.0336023271,
0.013086603,
0.0150013249,
0.025809288,
0.0330531336,
0.0324841477,
0.0106641585,
0.0614069849,
0.0179175213,
-0.0342895128,
0.0204876307,
0.0838859454,
0.0108666476,
0.006802062,
-0.0010898723,
0.0381257832,
0.0461948961,
0.0266195629,
0.0107537508,
-0.0677742362,
0.017059654,
-0.0059215953,
0.0146551766,
0.0315870568,
0.0114085926,
-0.0629837736,
-0.0504535027,
0.0166962985,
0.0022912992,
0.0154227931,
0.0440465957,
-0.0070888405,
0.0558675155,
-0.0910211354,
0.0423011295,
0.0533523895,
-0.0314473696,
-0.003697658,
-0.0781812817,
0.0358967185,
0.0624376573,
0.0197628774,
-0.0049619959,
-0.041799061,
-0.0173078328,
0.0608771443,
0.0559260435,
0.0440027677,
-0.0040215808,
0.037175741,
-0.0745135248,
-0.0225385074,
-0.0526970252,
0.0026852028,
0.0304271933,
0.0100903893,
0.0022399237,
-0.0184806939,
0.0097870613,
-0.0334835984,
0.0332053564,
-0.0121678533,
-0.0349972546,
-0.0294908378,
0.0017274505,
0.0178044438,
-0.0128446762,
0.0057804151,
0.0082984464,
0.0283690467,
-0.0020653615,
0.0138248121,
0.0040109647,
-0.0626922697,
-0.0192429256,
0.0071694981,
0.0232309382,
-0.0177070219,
0.0225338694,
-0.0200910084,
-0.0416005068,
0.0368879214,
-0.0137298107,
0.0707237273,
-0.0150527498,
0.0326867066,
0.0013142999,
0.035662096,
-0.0260148924,
0.0209765676,
-0.0003823335,
-0.0071563679,
0.0328563489,
0.0415190905,
-0.0221815016,
-0.0311154649,
0.0145729445,
-0.0112082493,
-0.0049581449,
0.0228733197,
-0.002765201,
0.0195506271,
-0.0723742172,
-0.041859597,
-0.0734175593,
0.0087126344,
-0.0138323894,
0.0201928858,
0.0215533469,
-0.0133606931,
-0.0328666903,
-0.0247753803,
-0.0350918733,
0.0290803015,
0.00461127,
0.0261146724,
0.0025529384,
0.0217141695,
-0.0509681217,
-0.048451487,
0.0278190766,
0.0145976767,
-0.0037120243,
-0.0150606474,
0.0511253886,
-0.0544864312,
0.0632286072,
-0.0243944693,
0.0422866791,
-0.0197842233,
-0.0694964305,
0.0369804762,
-0.0033459435,
0.0370608345,
0.0673382133,
0.0317531452,
0.0179192517,
0.031953603,
-0.0251506176,
-0.0701318905,
0.0333667807,
-0.0361899026,
-0.0660151988,
-0.0682362393,
-0.0068625212,
-0.0043545798,
0.0553304851,
0.0265396852,
0.0441470593,
-0.0713311657,
-0.0015081831,
0.0172991008,
-0.0804271102,
0.0168879181,
0.0042225444,
-0.0214903057,
0.0020977587,
-0.0246058945,
-0.046622701,
0.0456404798,
-0.0252953693,
-0.0356771648,
-0.0334830545,
0.0059035085,
-0.0331680365,
0.0177719351,
-0.0022970997,
-0.0008912144,
-0.0197595917,
-0.0248494707,
-0.0551443286,
-0.0167521127,
0.002242279,
-0.0142878173,
0.0373363048,
-0.0541430637,
-0.0225272346,
0.0135135856,
-0.0397075526,
-0.0521687306,
-0.0305228774,
-0.0001909193,
0.0095041608,
-0.0006487034,
0.0128196897,
-0.0287793186,
-0.0026670266,
0.0185882989,
0.0230147056,
-0.0187353529,
-0.0063178991,
-0.0436122306,
0.003211519,
0.0727202818,
0.0335105658,
0.0093529709,
-0.0399845578,
0.0586549267,
0.0214161165,
0.0381382331,
0.0022181359,
-0.0062723849,
0.0371532738,
0.0289061088,
-0.0041119889,
-0.0013666844,
0.0348921865,
0.0937602967,
-0.037107978,
-0.0295938551,
0.121540308,
0.014834119,
-0.0263195653,
-0.0030143203,
-0.0246915519,
0.06605611,
0.0327917039,
0.0073201582,
-0.0636355206,
-0.0397298671,
-0.0220914893,
0.0269796308,
-0.0169151835,
-0.0038380816,
0.0007321699,
-0.1245645061,
0.0374491289,
-0.0034348643,
-0.0700514838,
-0.0070026848,
-0.0307433177,
0.0152392657,
0.0228716657,
0.0558426045,
0.0115521587,
-0.0482433252,
-0.0068131364,
0.038754452,
-0.0059874221,
-0.036858391,
0.0370061025,
-0.0354739763,
0.0808778629,
0.0758950114,
0.0176657178,
0.0270242598,
-0.0416316129,
0.0305923764,
-0.0151501857,
0.038234219,
0.0021946724,
0.0402436405,
-0.0053006662,
0.0067498023,
-0.0148309683,
0.0419739261,
-0.0337603912,
-0.0071349484,
0.0070708287,
0.022884775,
-0.0273970757,
-0.0116903679,
0.0197222065,
0.0182686374,
0.0334946699,
0.0582464486,
0.0576953702,
-0.0550974272,
-0.047465466,
0.0067239213,
0.013214102,
-0.0290855225,
-0.0307911709,
0.0407086276,
-0.012221369,
0.0230759382,
-0.0352588445,
-0.0780245364,
0.0539885648,
-0.0169338156,
0.001190468,
-0.0001095772,
-0.0203081425,
0.0066982429,
0.0040331972,
0.0631054565,
-0.0493213311,
-0.0196428783,
-0.092717886,
-0.075326845,
0.0153861661,
-0.0515822992,
0.0052786176,
0.0010480818,
-0.054458037,
-0.0703113973,
0.0181657299,
0.0904975235,
0.0226079151,
0.0621718168,
0.0259660948,
-0.0449194126,
0.0353992693,
0.0154119264,
-0.0406797864,
0.0366100408,
0.0398742482,
0.0132454718,
-0.0065951352,
0.0411190838,
-0.0350643322,
0.0432802401,
0.0440105423,
0.038661994,
-0.0074606068,
-0.0271264203,
0.0357238352,
0.0027818023,
0.0252148304,
0.0174362455,
-0.0119020641,
-0.0144021977,
0.0388973393,
-0.0103001548,
0.0211342797,
-0.0522787198,
-0.0554767326,
0.0107486825,
0.007850226,
0.0126807122,
-0.0148856584,
-0.0448315628,
0.0369532406,
-0.032150574,
0.0307187084,
0.0397926793,
0.0280689094,
0.0500577018,
0.0156736486,
-0.0465929322,
0.0369059965,
0.0139793269,
0.0326248817,
0.0225294754,
0.0807069018,
-0.011495987,
-0.0107909814,
-0.0122246211,
0.0369725302,
-0.0472826846,
-0.0848161951,
-0.0048011593,
0.0187492613,
0.006461869,
-0.0012768714,
0.0553536378,
-0.0077496758,
0.0217278264,
0.0025314221,
0.0151899084,
-0.0096793398,
-0.031811282,
0.036550153,
0.0248622503,
-0.0189622659,
-0.0029826919,
0.0803232044,
0.000737466,
0.0887839198,
-0.0198736694,
0.0485544577,
-0.0377682634,
-0.0294328835,
0.0266814176,
-0.00355181,
-0.0059613516,
0.0264494792,
-0.0222074296,
-0.0472011752,
-0.0380142257,
-0.0023966746,
-0.0113445669,
-0.0434043594,
0.0503638685,
0.0180108566,
-0.0240368545,
-0.0375991724,
-0.0481723472,
-0.0067140092,
-0.0146946693,
-0.0149848461,
-0.0061084474,
-6.623231084e-33,
-0.0334331207,
0.0073781908,
0.0390441529,
0.0127508482,
-0.0144327767,
-0.0386204422,
-0.0437388383,
-0.002658668,
-0.0417824313,
-0.0232856683,
-0.008953589,
0.0023919588,
0.0136321429,
0.0201358441,
0.0069043837,
-0.0312447604,
0.0417439826,
-0.002815122,
-0.026207922,
0.0259269048,
0.0134459725,
0.0038155208,
0.0381260142,
-0.0437995866,
-0.0082209213,
-0.0280862171,
-0.0026144043,
0.000984836,
-0.0252124704,
0.062571153,
-0.0190889277,
0.0390439481,
0.0061351755,
-0.0362751856,
-0.0162566025,
0.0233400725,
-0.0470311753,
-0.034519244,
0.0290321056,
0.0232321732,
-0.0256734844,
-0.0489220284,
0.0351259559,
-0.0143430112,
-0.0326603167,
0.0045594112,
0.0155619802,
-0.0035230471,
-0.0091220969,
0.0397450365,
-0.0988640115,
-0.0129325837,
-0.0015485029,
-0.070202969,
-0.0171331912,
0.0578422844,
-0.0124495495,
-0.0151054068,
-0.0674254075,
0.0379814133,
0.0335132442,
0.0023753152,
0.0401811115,
0.0424396284,
0.0128174406,
0.0446173921,
0.0363448821,
0.0399080031,
0.0333012715,
-0.0170367155,
0.0299345367,
0.0619148761,
-0.0381322317,
0.0105495397,
0.0057679252,
0.005447566,
-0.0355023593,
-0.0130161857,
-0.0113181816,
0.0557744838,
0.0204905793,
0.0080930535,
-0.0320542678,
-0.0138171893,
-0.0479086228,
0.0134798512,
-0.0251966994,
-0.0316477939,
0.066876255,
0.0321246833,
0.0109310942,
-0.0418645777,
0.0432446077,
-0.0173263997,
-0.009730015,
-0.0189590976,
-0.0392079465,
0.0238809865,
-0.0302929357,
0.0092244018,
-0.095733434,
-0.0620795041,
0.0482109562,
-0.0374333039,
0.0212844703,
0.0385221057,
0.0059055286,
0.0357703678,
-0.0564254671,
-0.0174853913,
-0.0255728979,
0.0656019375,
0.0212598685,
-0.0044549648,
0.0192449261,
0.0195588674,
-0.0356338471,
0.0522394031,
-0.008687106,
-0.0287708733,
-0.0266696233,
-0.0452123284,
-0.0008124721,
0.0151624037,
-0.042320285,
-0.0090238266,
-0.0462901779,
-0.0430908315,
-0.0177581348,
-0.0480081663,
-0.0130702518,
0.0146172121,
2.842e-7,
-0.0104948897,
0.0300165042,
-0.0004925689,
0.0389297605,
0.0160687584,
0.0209024213,
-0.0129568307,
0.0269368943,
0.0564656183,
-0.011379783,
-0.0197419599,
0.0437952764,
0.0512687378,
-0.0044339644,
-0.0953982323,
-0.0405283421,
-0.023349775,
0.0073789661,
-0.0520220138,
-0.0148341581,
0.0204102453,
0.0699097514,
-0.0147090219,
-0.0095686289,
-0.0242853928,
-0.0032686389,
0.0721194968,
-0.0301015265,
-0.015575272,
0.0535625666,
0.0068675559,
0.0039210087,
-0.0085011479,
0.0592359714,
0.0142507115,
0.0291128512,
-0.0465146415,
0.0441064648,
-0.0023613742,
-0.0767741054,
0.0129520129,
-0.0000066309,
-0.000958169,
-0.0408915505,
0.0784766972,
-0.0625868291,
-0.0400881581,
0.0100060618,
-0.0128577584,
-0.0275499858,
0.0033247243,
-0.0161350388,
-0.0169663671,
0.0170907509,
0.0116534047,
-0.0035304944,
0.0060234512,
-0.0668524876,
0.0247762166,
0.0158214737,
-0.0305742398,
-0.0739444047,
-0.0328752436,
-0.0048485841,
0.0125751123,
-0.0374920331,
0.0250541586,
2.359972968e-34,
0.0076894704,
0.0090689249,
0.0620404035,
0.0613737777,
0.0276064929,
0.0056425491,
0.0071830195,
0.0334681123,
0.0363275222,
-0.1273566186,
0.0018369677
] |
http://arxiv.org/abs/2402.02008v1 | http://arxiv.org/pdf/2402.02008v1 | How well do LLMs cite relevant medical references? An evaluation framework and analyses | 2024-02-03T03:44:57 | 2024-02-03T03:44:57 | Kevin Wu, Eric Wu, Ally Cassasola, Angela Zhang, Kevin Wei, Teresa Nguyen, Sith Riantawan, Patricia Shi Riantawan, Daniel E. Ho, James Zou | cs.CL | cs.CL, cs.AI | Large language models (LLMs) are currently being used to answer medical
questions across a variety of clinical domains. Recent top-performing
commercial LLMs, in particular, are also capable of citing sources to support
their responses. In this paper, we ask: do the sources that LLMs generate
actually support the claims that they make? To answer this, we propose three
contributions. First, as expert medical annotations are an expensive and
time-consuming bottleneck for scalable evaluation, we demonstrate that GPT-4 is
highly accurate in validating source relevance, agreeing 88% of the time with a
panel of medical doctors. Second, we develop an end-to-end, automated pipeline
called \textit{SourceCheckup} and use it to evaluate five top-performing LLMs
on a dataset of 1200 generated questions, totaling over 40K pairs of statements
and sources. Interestingly, we find that between ~50% to 90% of LLM responses
are not fully supported by the sources they provide. We also evaluate GPT-4
with retrieval augmented generation (RAG) and find that, even still, around
30\% of individual statements are unsupported, while nearly half of its
responses are not fully supported. Third, we open-source our curated dataset of
medical questions and expert annotations for future evaluations. Given the
rapid pace of LLM development and the potential harms of incorrect or outdated
medical information, it is crucial to also understand and quantify their
capability to produce relevant, trustworthy medical references. | How well do LLMs cite relevant medical references? An evaluation framework and analyses: Large language models (LLMs) are currently being used to answer medical
questions across a variety of clinical domains. Recent top-performing
commercial LLMs, in particular, are also capable of citing sources to support
their responses. In this paper, we ask: do the sources that LLMs generate
actually support the claims that they make? To answer this, we propose three
contributions. First, as expert medical annotations are an expensive and
time-consuming bottleneck for scalable evaluation, we demonstrate that GPT-4 is
highly accurate in validating source relevance, agreeing 88% of the time with a
panel of medical doctors. Second, we develop an end-to-end, automated pipeline
called \textit{SourceCheckup} and use it to evaluate five top-performing LLMs
on a dataset of 1200 generated questions, totaling over 40K pairs of statements
and sources. Interestingly, we find that between ~50% to 90% of LLM responses
are not fully supported by the sources they provide. We also evaluate GPT-4
with retrieval augmented generation (RAG) and find that, even still, around
30\% of individual statements are unsupported, while nearly half of its
responses are not fully supported. Third, we open-source our curated dataset of
medical questions and expert annotations for future evaluations. Given the
rapid pace of LLM development and the potential harms of incorrect or outdated
medical information, it is crucial to also understand and quantify their
capability to produce relevant, trustworthy medical references. | [
0.0670182183,
0.0468538702,
0.0090476489,
-0.037247207,
-0.0228715185,
0.0062638088,
0.0242982022,
0.0368811563,
0.0265168045,
-0.0436295383,
-0.0112527637,
-0.0262200739,
-0.0197240692,
-0.005537299,
0.0132230399,
0.004591574,
0.0163907129,
0.0455179065,
0.0628444999,
-0.0039506857,
-0.0150192017,
0.0176079459,
-0.0421317145,
0.0496927053,
0.0240509566,
0.0005323266,
0.0569939762,
-0.0116189886,
0.0242696106,
-0.0753296614,
-0.0021262749,
0.0257347319,
-0.0348221734,
-0.0041756281,
0.0000018947,
-0.051894661,
-0.0122020803,
0.0314225368,
-0.0292724576,
0.016200887,
0.0147230998,
-0.0213993397,
0.0095881764,
0.0239585638,
-0.0540733337,
-0.0493914895,
0.0298677087,
0.0187426358,
0.0682663769,
0.0533573143,
-0.0108822268,
0.0212336034,
0.0456465669,
-0.0292808935,
0.0423263311,
-0.0464899242,
0.0430800021,
0.0347579904,
0.037674997,
0.0313987695,
0.0488571785,
0.0554069169,
-0.0280472804,
0.0241514184,
0.023094438,
0.0348260924,
-0.0481636971,
-0.0103029115,
0.0282988809,
0.0405782275,
0.0552454814,
-0.0179084539,
0.0261420347,
0.0040316405,
-0.013034042,
0.0619481541,
-0.0447581857,
0.0100121535,
-0.0240279678,
-0.0063137449,
-0.0002921776,
-0.0319593996,
0.0482841581,
0.0201717038,
0.011492677,
0.0505558178,
0.0019454972,
-0.0252260584,
0.0310673639,
0.0078910738,
0.0003212335,
-0.0610252693,
0.0311098751,
0.019664675,
-0.0167894494,
0.0131763602,
0.0162580386,
-0.0077756238,
0.1019973233,
-0.0335882045,
-0.0139424941,
-0.0141276913,
-0.0318483412,
0.0299381223,
-0.019491693,
-0.0098791355,
-0.0497938134,
-0.0011466574,
-0.0395730287,
0.0127358986,
-0.0476171635,
-0.0308297798,
0.0016541645,
0.0057746111,
-0.0080938563,
0.0242623296,
-0.0770919025,
-0.0142944902,
-0.0317459889,
0.0208331887,
-0.0053980346,
0.0192083642,
0.0034970355,
0.013251056,
-0.0463171303,
0.0366129652,
-0.0015952563,
0.0394381396,
-0.0137700308,
-0.0718633458,
0.023913702,
0.0084892921,
0.0053209243,
-0.0121751102,
0.0360867828,
0.0499485694,
0.0256511904,
0.0041221557,
-0.0026038345,
-0.0449143499,
-0.0505731031,
-0.0355877168,
0.0223841835,
-0.0726325065,
0.0296456479,
0.0410735458,
0.0112095773,
-0.0663021132,
0.023465503,
0.0591631196,
-0.0694758296,
0.0350752287,
-0.0568228029,
0.030779222,
0.0351743177,
0.0238421224,
0.0151438154,
-0.0048269262,
0.0299119875,
-0.0677396879,
0.0078712301,
-0.0016377458,
-0.0022926051,
-0.0022204362,
0.033189889,
-0.001650298,
0.0306197703,
0.0460545421,
-0.0009957042,
0.04629777,
-0.0423085354,
0.0904170275,
-0.0222594813,
-0.0080525409,
0.0383191705,
0.0543646775,
-0.0119447121,
-0.0054096142,
0.0282256324,
-0.0223734397,
0.0312235411,
-0.0071577062,
-0.0226148423,
0.0083857672,
0.0143953227,
-0.0141279902,
0.0288468655,
-0.0082475329,
-0.0101584867,
-0.0243653283,
-0.0506417304,
0.0074517652,
-0.0545283929,
0.013501768,
0.0147670554,
-0.0289744288,
-0.0188833661,
0.0984959677,
0.0209012534,
-0.0455111936,
-0.0099720228,
-0.0298241004,
0.002080627,
0.020609241,
0.061713025,
-0.0179680884,
-0.0378192775,
-0.0251456387,
0.0131309573,
0.0153775532,
0.025327865,
-0.006396235,
-0.0116878226,
-0.0321488716,
0.0003664356,
0.0294720437,
0.0196462218,
-0.0492535792,
0.0261701774,
0.0115800826,
-0.0019084014,
-0.0461943932,
-0.0352619439,
0.0430011414,
-0.0683122203,
-0.002111061,
-0.0023196789,
-0.0345374048,
0.0402313545,
0.0306440536,
-0.0196634866,
0.0606013685,
0.0391772799,
-0.0209949762,
-0.0101938099,
0.0106488941,
-0.0072169746,
0.0300727971,
0.0465339683,
-0.0432277434,
-0.0047795624,
-0.032887511,
-0.0342746489,
-0.0297917351,
-0.0418095998,
-0.053883262,
0.0099193985,
-0.0334871076,
-0.0149387103,
0.0349475369,
-0.0051739095,
-0.0499093011,
0.0357184671,
-0.0152303716,
0.0380012132,
0.0323162451,
-0.0327175744,
0.0251084659,
-0.0379423499,
-0.0381613225,
0.0114807542,
0.0016759651,
0.0169020444,
0.0264831707,
0.0035774827,
-0.0471421853,
-0.0352710299,
-0.1058040932,
0.0032766911,
0.0166426785,
0.0132835945,
0.0236075148,
0.0000862484,
0.0051418855,
-0.004160827,
-0.0444687754,
-0.0311948564,
0.0374810174,
0.0314868763,
0.0337827876,
0.0398198254,
0.044483047,
-0.0484678335,
-0.0165065005,
-0.0422583036,
0.0079701757,
0.0517749786,
0.0026346385,
-0.0276106298,
-0.0072242711,
-0.07601621,
-0.0170993246,
0.0056467433,
0.010763376,
0.0278878529,
-0.0033027388,
-0.0165384151,
0.0084639732,
-0.000578639,
-0.042539835,
0.0576527789,
0.0063839462,
-0.0616970994,
0.0148196435,
-0.010026345,
-0.0662866682,
0.0024767881,
-0.0691945329,
0.0155285029,
0.0227838494,
0.012431805,
-0.0217998121,
-0.0306418259,
0.0071934722,
-0.0332070403,
0.0249811504,
-0.0435544215,
-0.0020678288,
-0.0329876132,
0.0708431154,
0.0131948385,
-0.0200408977,
0.0679109395,
-0.0374054834,
0.0238992106,
-0.0584466644,
0.0130134923,
-0.0471603721,
-0.0185886715,
-0.0069186101,
0.0050386596,
0.0156795643,
0.0137468288,
-0.02065924,
0.0341587216,
0.012890894,
-0.0854493678,
0.069532305,
-0.0250591971,
0.0502008535,
0.0119376546,
-0.0467817374,
-0.0382809117,
-0.0298557337,
-0.0584849231,
-0.0059378403,
-0.0134187089,
0.0088312626,
-0.062609449,
-0.0015549168,
0.041199293,
0.042506326,
0.0141767282,
-0.0181183256,
-0.0001494736,
0.0181338377,
0.0504725948,
0.0103425235,
0.0125198252,
0.0115991673,
-0.0221416056,
0.0744430795,
0.019558968,
0.0371681638,
-0.0117445914,
-0.0306798927,
0.0136392545,
-0.0438436307,
-0.0100176772,
0.0118184583,
0.018497834,
-0.0468531251,
-0.0021078526,
0.0963558108,
0.0303799547,
-0.0530442223,
-0.0174801853,
-0.0281825606,
-0.0252039433,
0.0106462222,
0.0225894116,
-0.0994495749,
-0.0099908141,
-0.0282924715,
0.0480345227,
-0.0181472749,
-0.0318073295,
0.0139943203,
-0.0818227828,
0.0337408334,
-0.0055363891,
-0.0235786643,
-0.0132959159,
-0.0504691526,
0.0201354623,
-0.0010902084,
-0.0099716708,
0.0345950052,
-0.008783089,
0.0127384812,
0.0162617303,
-0.0596140921,
-0.0756393671,
0.045999404,
-0.0249992609,
0.0664855912,
0.0749901086,
0.0322110578,
-0.0201771967,
-0.0366748497,
0.0134509532,
-0.0302948989,
-0.0213913191,
-0.0116119226,
0.0812802985,
0.0218819026,
0.0100692073,
-0.0470052063,
0.0202565063,
-0.0210020449,
-0.0133437207,
0.0146570625,
0.0779802576,
-0.0236453712,
0.0000340389,
-0.0518140793,
0.0539454892,
-0.0019865595,
0.0679019243,
-0.0015379433,
0.0065726647,
-0.0007181194,
0.0119105875,
0.0548109822,
-0.0058924779,
-0.0397365987,
0.0481118225,
-0.0202002358,
0.0050423639,
-0.071793288,
-0.0091416175,
0.0229599196,
-0.0419001952,
0.0105546145,
0.0367785506,
0.0206640624,
-0.0005364218,
-0.0490833521,
0.0562806726,
0.0013734896,
0.0051285238,
-0.0353727639,
-0.0198805369,
0.0935797319,
-0.0015087047,
-0.0173327923,
0.0416715071,
0.0047465214,
-0.0112941684,
-0.0335737839,
0.0285267252,
0.023287354,
-0.0228874628,
-0.0149395606,
-0.0333871022,
-0.0002861581,
0.056047,
-0.0489670783,
-0.0045377966,
-0.0156061519,
-0.0022536055,
-0.0078622205,
-0.0086555835,
-0.0572661683,
0.0225883964,
-0.019708965,
0.0391869172,
-0.0073303725,
-0.0287053622,
0.0440950654,
0.0456141569,
0.0584378839,
-0.0227998141,
0.0237257667,
0.0106702279,
0.0078900419,
-0.0110047795,
0.0013030497,
-0.0220640805,
0.0157877486,
0.0072203665,
-0.0221391413,
-0.0168034583,
-0.0007445068,
-0.0553393811,
0.052898936,
-0.037582621,
0.0283320118,
0.0074653598,
0.007522244,
0.0118182693,
-0.0456119478,
-0.0200896859,
-0.0118177235,
-0.0137608452,
0.030919008,
-0.0102136219,
0.0448400229,
-0.0196159147,
0.013278353,
0.0513323955,
-0.0002869378,
-0.0845801234,
-0.0418718606,
-0.0188311357,
0.0038961468,
0.0322644636,
0.014154681,
0.0784254074,
-0.0341869481,
-0.0168723501,
-0.048850365,
-0.012376057,
-0.0020107355,
0.0208754316,
0.0543173999,
0.0128104398,
-0.0063180476,
-0.0117215328,
0.0707507804,
0.0233065039,
0.0440852605,
-0.0228090696,
-0.0641758069,
-0.0441949777,
-0.0561644137,
-0.068613328,
0.0249408912,
-0.0497852787,
0.0082825311,
-0.074487634,
-0.037803147,
0.0196591094,
-0.0026889909,
-0.0123799657,
-0.0221987069,
0.0141142653,
0.0575327389,
0.0173118487,
-0.0308911745,
-0.049463179,
0.0742463917,
-0.0487954468,
0.0004653473,
-0.0069998265,
-6.03374932e-33,
-0.0019280643,
-0.0503560528,
0.0247540884,
-0.0402978361,
-0.0003429901,
0.0369022377,
-0.0434558764,
0.0004164269,
-0.0404535942,
0.0322397612,
-0.0316814482,
0.0086179348,
0.0125605455,
0.0237918142,
0.0272154324,
-0.0204168968,
0.0194634683,
0.0077836965,
-0.0245753638,
0.0273334868,
-0.0010691257,
0.0105539057,
0.0333856605,
-0.118995443,
0.059005022,
-0.0356612839,
0.0200853199,
-0.0090171192,
-0.0278676357,
0.0853374749,
-0.017764762,
0.03829588,
0.0405473039,
-0.0294812787,
0.0158800762,
-0.0344718285,
-0.0734445378,
0.0024293831,
0.0276507959,
-0.0486797877,
-0.0468299873,
-0.0163514409,
0.063704066,
-0.0187956579,
-0.0418732204,
-0.0163211375,
0.0383600779,
-0.0501652136,
-0.0082620485,
-0.0500508808,
0.0022759433,
0.0470796488,
0.0311880838,
0.0452776887,
-0.0339611322,
0.0627311319,
-0.0296506397,
0.0574485399,
-0.0239252169,
0.0161628854,
0.006180116,
0.039945893,
0.0123660108,
-0.0731444284,
-0.0153482594,
0.0380814038,
-0.0053634355,
-0.0317007042,
-0.0369135477,
-0.0378469527,
-0.0520452373,
0.0334243253,
0.0502474383,
0.0692441463,
0.0027467269,
-0.0083255293,
-0.0918400511,
0.033673957,
-0.0196530242,
0.0653914139,
0.0415131375,
-0.0471416153,
-0.025209086,
0.0136108724,
-0.0331414118,
0.0061030043,
-0.0249891263,
-0.0357503146,
0.0282743778,
0.0056443862,
0.0169384163,
0.0479237773,
0.0179106053,
-0.0165453851,
0.0030254403,
0.0280458443,
-0.0098774564,
-0.0005641379,
-0.0195402242,
0.0193661731,
-0.0523691475,
0.0305624772,
0.0014258984,
-0.0115220025,
0.044974789,
0.023159897,
0.0053559383,
-0.0463488623,
-0.0646457896,
-0.0161978621,
0.0113484049,
0.0265342277,
-0.0426059924,
-0.000987606,
0.0120367091,
0.048553884,
0.0282114763,
0.0391152762,
0.019028198,
0.0187315065,
-0.0373946242,
-0.049816303,
0.0038893288,
-0.0078626499,
-0.0423910469,
0.020624714,
0.0027691338,
0.074210003,
0.0022344852,
-0.0259785615,
0.0050822911,
-0.0067443121,
2.7e-7,
0.1158426926,
0.0492209941,
-0.015156121,
-0.0254176483,
0.0363559276,
-0.0068986337,
-0.0130811492,
-0.0150083425,
0.0376937054,
-0.0299194325,
0.0101453084,
-0.0106334593,
0.0322923176,
0.0567951985,
-0.0593169592,
-0.0544292927,
-0.1075165048,
-0.0275223982,
-0.0115004051,
-0.0279221218,
0.0251305457,
0.0475197844,
-0.0415778533,
-0.0233099386,
-0.0243885349,
-0.035579592,
0.0586449318,
-0.0334749222,
-0.0094594639,
0.0616371892,
0.0710356757,
0.0150735099,
0.0320074297,
0.0456927158,
-0.0274153147,
-0.0111954967,
0.0164512526,
0.0903554261,
-0.0203869771,
-0.0031825707,
0.0081750704,
-0.0162487067,
0.0505284034,
-0.0564983822,
0.0535873547,
-0.0271160658,
-0.0272439234,
-0.0430994444,
-0.0584618412,
-0.0248750318,
0.0690851212,
-0.0026101377,
-0.0299646165,
0.0253095273,
-0.0064035044,
0.0338546708,
-0.0531463139,
-0.0543342419,
-0.0356993452,
0.0394167267,
-0.0500877947,
-0.0666268319,
0.028832972,
-0.0264984742,
-0.0351743549,
-0.0800557509,
0.0348226987,
2.122257705e-34,
0.0149026923,
0.0238577444,
0.0127635915,
0.0169250071,
0.0325100347,
0.0186125115,
-0.011471739,
0.0282300841,
-0.0066748075,
-0.0585905649,
0.0138882389
] |
http://arxiv.org/abs/2402.01981v1 | http://arxiv.org/pdf/2402.01981v1 | Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypes | 2024-02-03T01:40:11 | 2024-02-03T01:40:11 | Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Tong Yu, Hanieh Deilamsalehy, Ruiyi Zhang, Sungchul Kim, Franck Dernoncourt | cs.CL | cs.CL, cs.AI, cs.CY, cs.LG | Large language models (LLMs) have shown remarkable advances in language
generation and understanding but are also prone to exhibiting harmful social
biases. While recognition of these behaviors has generated an abundance of bias
mitigation techniques, most require modifications to the training data, model
parameters, or decoding strategy, which may be infeasible without access to a
trainable model. In this work, we leverage the zero-shot capabilities of LLMs
to reduce stereotyping in a technique we introduce as zero-shot self-debiasing.
With two approaches, self-debiasing via explanation and self-debiasing via
reprompting, we show that self-debiasing can significantly reduce the degree of
stereotyping across nine different social groups while relying only on the LLM
itself and a simple prompt, with explanations correctly identifying invalid
assumptions and reprompting delivering the greatest reductions in bias. We hope
this work opens inquiry into other zero-shot techniques for bias mitigation. | Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypes: Large language models (LLMs) have shown remarkable advances in language
generation and understanding but are also prone to exhibiting harmful social
biases. While recognition of these behaviors has generated an abundance of bias
mitigation techniques, most require modifications to the training data, model
parameters, or decoding strategy, which may be infeasible without access to a
trainable model. In this work, we leverage the zero-shot capabilities of LLMs
to reduce stereotyping in a technique we introduce as zero-shot self-debiasing.
With two approaches, self-debiasing via explanation and self-debiasing via
reprompting, we show that self-debiasing can significantly reduce the degree of
stereotyping across nine different social groups while relying only on the LLM
itself and a simple prompt, with explanations correctly identifying invalid
assumptions and reprompting delivering the greatest reductions in bias. We hope
this work opens inquiry into other zero-shot techniques for bias mitigation. | [
-0.0072708759,
0.080675222,
-0.0118047064,
0.0569082536,
0.02014401,
0.0372226909,
0.0544594936,
-0.0290009379,
0.0007778384,
-0.0736001134,
-0.0395842828,
-0.0123357335,
0.0282812044,
0.0145918019,
0.0361474007,
-0.0393229574,
0.0424003787,
-0.0032904581,
0.0322118364,
-0.0590374731,
0.0185105745,
0.0358354896,
0.0082332836,
0.029559724,
-0.0094084879,
-0.0301684942,
0.0383326039,
-0.0126133533,
-0.0119761303,
-0.0262722299,
0.0295465607,
0.0691399053,
0.0152260391,
0.0504429899,
0.0000021217,
-0.0556669347,
0.0065837009,
-0.0147822611,
-0.0253545698,
0.0156124141,
0.0494927838,
-0.0356561467,
0.0114551606,
0.0316556357,
-0.0266820304,
-0.0411858447,
0.0565510392,
0.0092597827,
0.0522650778,
0.0219459049,
0.0012841302,
-0.043579556,
-0.0730559155,
-0.0193033107,
0.0014102436,
0.0131371394,
0.0486866124,
-0.0377677195,
-0.0753669217,
-0.0346562192,
-0.039872624,
-0.0037507857,
-0.0077232923,
0.0125168907,
0.0364428423,
0.0308506135,
-0.04897606,
-0.0383938961,
0.0053079841,
0.0746220723,
-0.0067945872,
-0.0605443902,
0.0031075103,
-0.0015480526,
0.0137065705,
0.0100318566,
-0.0275996812,
0.0552954711,
0.0296832565,
0.0295801796,
-0.0338949561,
0.0206835922,
0.0419933237,
0.0220169239,
-0.029186815,
0.0479609892,
0.0230024662,
0.0081339115,
-0.0497626923,
0.0353129171,
0.0256857611,
-0.0614871234,
-0.0375525095,
0.0302751381,
0.0074560232,
0.0208633207,
-0.0382758491,
0.0005557883,
0.0079029156,
-0.0245896932,
0.0073283459,
0.0375637524,
-0.0077165724,
0.008130542,
-0.0153093953,
0.0114470012,
0.0247298256,
-0.0220146514,
-0.0015985575,
-0.0027678749,
0.0027401228,
-0.0213935003,
-0.0227545332,
0.0047892914,
0.0194469355,
0.0444815159,
-0.031662751,
-0.0038105294,
-0.0371926278,
0.0045698611,
-0.0109299095,
0.0335728042,
-0.0486708879,
-0.0371216387,
-0.0110361194,
0.0474235229,
-0.0289646648,
-0.0213535149,
-0.0354282446,
-0.1020495221,
-0.0077170795,
0.00341252,
0.0074719666,
-0.0264894646,
-0.0233523622,
0.0613467172,
0.001372042,
-0.024236979,
-0.0291637294,
-0.0468194894,
-0.0419988334,
-0.0617253855,
0.0024367876,
-0.0558670796,
0.0110322153,
0.0046286541,
0.0129486816,
-0.0693411604,
0.0020928206,
0.0056151333,
-0.0536282994,
0.0724948943,
-0.0015693036,
0.018819226,
0.0535321049,
0.0211301707,
-0.0095063895,
0.1120910123,
0.0061697341,
-0.0172990859,
0.0231005475,
-0.0288446937,
0.0231322907,
0.0387957655,
0.056297224,
-0.003126218,
-0.043337781,
-0.0232319515,
0.0064510461,
-0.0092474474,
0.0798180252,
0.0321825109,
-0.0213112347,
-0.0319345072,
0.0676763803,
0.0494636483,
-0.0378119014,
0.0257565267,
0.0314403884,
0.0209519789,
-0.0170141421,
0.0325185098,
-0.0706974268,
0.0203582775,
-0.0077762874,
-0.023140382,
0.0046287607,
0.0558979325,
-0.0383539423,
0.0001594998,
0.0168613847,
-0.0183148049,
0.0396606401,
0.0065802983,
0.0292993169,
0.042491585,
0.0033995265,
0.0746476278,
0.0534748696,
-0.0364437811,
0.0128326416,
-0.0621526539,
0.0466167666,
0.0296520647,
0.0382546559,
0.0131301936,
-0.0009280712,
-0.0282664504,
-0.0169817656,
0.0581159331,
0.0569721088,
-0.0126356902,
0.023884993,
-0.0411046408,
0.006535159,
0.0022130641,
0.0363775343,
0.02589068,
0.041420348,
0.0273208003,
0.0318761468,
-0.0146809518,
-0.0118835103,
0.0140672391,
-0.0332517065,
-0.0605471842,
0.0176580623,
0.0215563457,
0.0057365717,
0.0205631498,
-0.0256821439,
-0.0092367958,
-0.0270178542,
0.007185285,
0.0055692866,
0.0594814643,
-0.1056096703,
-0.0505113862,
0.0240175724,
-0.0442696847,
-0.0415882953,
0.0187012795,
-0.0075847362,
-0.0149180545,
0.0802721083,
0.0054641343,
0.0307528023,
-0.0209769346,
0.0065385019,
-0.0320723504,
0.042989511,
-0.0187629871,
0.0363970064,
0.0239076968,
-0.0457049757,
0.0238571633,
-0.0519097075,
0.0089826724,
-0.0384076945,
-0.035239473,
0.0172433257,
-0.0086169466,
-0.0162312631,
0.0161732323,
-0.0316523723,
-0.0156460833,
0.0273468923,
-0.0597101487,
-0.0240344219,
-0.0138820661,
-0.006892975,
0.0319443345,
0.0101230321,
-0.0102706868,
-0.0258973632,
-0.0460679382,
-0.0154163241,
0.0196433831,
0.0155639453,
0.0276436731,
-0.0306190122,
-0.0075022005,
-0.0143124862,
-0.0007089337,
-0.0283243749,
0.0709007904,
0.0172970183,
-0.0005389731,
-0.0476102717,
-0.0005841317,
-0.0526714548,
0.0249088034,
0.0273770634,
-0.0357733332,
0.0099771414,
0.0143576423,
0.0055260193,
0.0692367181,
0.0081657637,
0.0057168934,
0.0721358135,
-0.0137463855,
-0.0454930104,
-0.0415637903,
-0.0038820952,
0.0281017888,
0.0254072975,
-0.0691915005,
-0.0436082445,
0.0894862264,
0.042265743,
-0.0211931076,
-0.0661351755,
0.0581708588,
-0.0285963435,
-0.0643319562,
-0.005259343,
-0.032735195,
-0.0076131318,
0.0193334334,
-0.047698915,
-0.0542789847,
0.0636417866,
-0.0318060629,
-0.0018293306,
0.0266578533,
0.0237193443,
-0.0244288892,
0.0428677611,
0.0000889754,
0.0470108278,
0.0411095619,
0.0079108682,
-0.0391435362,
-0.0506509356,
0.0379478298,
-0.0116471387,
0.0134477224,
-0.0563707538,
0.0134607079,
-0.0084372852,
-0.0452374555,
-0.0522790067,
-0.063090682,
-0.0455121621,
-0.0178586524,
-0.0186619926,
-0.0339924842,
0.0090433965,
0.0082840975,
0.0094612045,
0.0082721869,
-0.0234678239,
0.010039065,
0.0046129641,
0.0210584495,
0.0450770743,
-0.0265892036,
0.0156705603,
-0.0178289302,
0.0471838303,
0.0298265964,
-0.0106119765,
0.056586355,
-0.0318933353,
-0.0103887394,
0.01288509,
-0.0164369252,
0.0021424636,
0.0033283683,
0.0621828549,
-0.0318565033,
0.0016613663,
0.1346308291,
0.0059442539,
-0.0579013936,
0.0184726957,
0.0033681232,
0.0069558141,
0.0308648907,
0.0013829517,
-0.1067616269,
-0.0240983497,
0.0051733213,
0.0194040593,
-0.0108005265,
-0.0620825104,
-0.0303107686,
-0.1504678577,
0.0226746444,
0.0488978289,
-0.0259119589,
-0.0202674791,
-0.0167005714,
0.0072714556,
-0.0233358145,
0.03114748,
-0.0131405788,
0.0005533443,
0.0123608708,
0.00781589,
-0.0522398353,
-0.0509624258,
0.0121147949,
-0.0217930116,
0.1009123996,
0.0254122596,
0.0311125759,
0.0082496684,
0.0386649556,
0.0062451251,
-0.0159323309,
-0.0015582104,
-0.0180152878,
0.0686366111,
-0.0394972935,
-0.0347512178,
0.000792949,
0.035637062,
0.0076670102,
0.0255233441,
-0.0157022458,
0.0325235128,
0.0189237352,
-0.0444758572,
-0.0130714215,
-0.0008544473,
0.0016133913,
0.1029764861,
0.0550911091,
-0.0108919684,
-0.0069612591,
-0.0697842389,
-0.0048700357,
-0.006898961,
-0.046622891,
0.0267480481,
0.0119745256,
-0.0529654026,
-0.0494112037,
-0.0447311848,
0.0429117791,
0.0046338597,
0.0071774763,
-0.0380617492,
0.0117032165,
-0.0015359746,
-0.0530170761,
-0.0004836549,
-0.081682086,
0.0182074662,
0.0155681539,
0.0028383057,
0.0114975739,
-0.0415086448,
0.008510069,
0.0086128721,
-0.026517231,
-0.0875847042,
0.0275805742,
0.0391460508,
0.0226467457,
0.0827333555,
0.0020366057,
-0.1004060656,
0.0781714469,
0.0494023003,
0.0062447237,
0.0320525542,
0.0082437852,
0.0272223819,
0.0097148437,
0.0466700904,
-0.019463148,
0.0536903366,
0.0004735191,
-0.0317682177,
-0.0489641204,
-0.0179655552,
0.0015378707,
0.0172575545,
0.0283836573,
-0.0558626652,
0.0634324178,
-0.0316902027,
-0.0520521365,
-0.0777239799,
-0.0274912305,
-0.0109721962,
-0.0012667283,
0.0256925281,
-0.0088733854,
-0.011816225,
-0.0404242724,
-0.0637123063,
0.0389012769,
-0.0210371148,
0.0193727519,
0.0192524102,
0.0094219297,
0.016404992,
0.0116528533,
-0.0109250955,
0.0410130844,
-0.0209877789,
0.0171783753,
0.0119584883,
0.0283994637,
0.0149825951,
0.0286326502,
-0.0293744858,
-0.0108567635,
-0.0016633313,
-0.0278845187,
0.0003550646,
0.029203061,
0.0430922136,
-0.0064158272,
0.0758243278,
0.024738906,
-0.013047819,
0.0105563514,
0.0120725948,
0.0052368566,
-0.0266124196,
0.0284610763,
0.0047095353,
-0.0293774195,
0.0145332348,
0.0902488828,
0.027009761,
0.0599496029,
0.001559,
-0.0145894848,
-0.0324060209,
-0.0062714131,
0.0305431187,
-0.0252502877,
-0.0017127586,
-0.0093964189,
-0.0633412451,
-0.0934702158,
-0.0478541218,
0.0017870709,
0.0201342013,
-0.0144855408,
0.0159374289,
0.0197023544,
-0.0243473053,
-0.0266384203,
-0.0187956244,
0.0271956865,
-0.0223548859,
-0.0066321357,
-0.0022213098,
-6.928549579e-33,
-0.0159607045,
-0.044702366,
-0.0221513212,
-0.0518084243,
0.0019110066,
-0.0133609138,
-0.0468063653,
0.011760409,
-0.0228438284,
0.0257579107,
0.0024090125,
0.0055589406,
0.0082078381,
-0.0105270324,
-0.0020540978,
0.015423811,
0.0996844023,
-0.038408082,
-0.0497765206,
0.0144656124,
0.0207178518,
0.0149803571,
0.0377718024,
-0.0290936995,
0.0305420384,
-0.0208871905,
-0.0374904647,
0.0238386039,
0.0348845497,
0.01818645,
-0.0462180637,
0.0409668647,
0.0423122086,
-0.0065359501,
-0.0198570099,
0.0496527664,
-0.0608470254,
-0.0241014529,
-0.0048498977,
0.0253314897,
0.0102622788,
-0.0296438485,
0.0264935363,
-0.0122075854,
-0.0566892847,
-0.0088478765,
0.0210535973,
-0.015250558,
0.0300393812,
0.0393522978,
-0.0536786132,
0.0196326561,
-0.0283829253,
-0.015464833,
0.0364153534,
0.0096043311,
0.0118043469,
0.0165703166,
-0.0247558579,
0.0390365086,
0.0030336003,
-0.0135630965,
0.0851138607,
0.0306222085,
0.0086760418,
0.0427827537,
0.0585270971,
0.0351145454,
-0.0099856807,
-0.0345788747,
0.0374095254,
-0.0327459201,
0.0012610808,
0.089261651,
0.0535717718,
-0.0306442417,
-0.0178482123,
-0.0102246115,
0.050719142,
-0.0107318684,
0.0277394336,
-0.0123727378,
-0.0643213242,
-0.0525076091,
-0.0417850651,
-0.0859088227,
-0.0221227687,
-0.0342362747,
0.0175931584,
0.0140087921,
-0.0646341294,
0.017044479,
0.0356484354,
-0.0634312928,
-0.0200064071,
0.0320206955,
-0.0204604324,
-0.028861491,
0.0081845066,
-0.0061950744,
-0.061412733,
0.0133849299,
0.0114155915,
0.0512732379,
0.0206692424,
0.0262204893,
0.0174680389,
0.0218838528,
-0.0504944585,
0.0188595355,
-0.0276754051,
-0.0042492952,
0.0180036277,
0.0173917338,
0.0231058523,
0.0228919294,
-0.0102894977,
-0.0044380995,
0.0345351845,
-0.0109109543,
-0.0213027783,
0.0347416699,
-0.0009995974,
0.0309450328,
-0.0450884774,
-0.028665524,
-0.0029460806,
-0.0242085289,
0.0742576942,
-0.0421579368,
-0.0195358638,
-0.0535255447,
2.657e-7,
-0.0079267621,
0.0630378723,
-0.0057009012,
0.0285822246,
0.0581777878,
0.0516551584,
0.0502619743,
0.0402197726,
-0.0008622266,
-0.0076164813,
-0.0088066803,
0.0728002116,
0.0225835014,
0.0415901169,
-0.0607356988,
0.0118431225,
-0.0719591677,
0.013842958,
-0.0202564821,
-0.0318936184,
0.0201703738,
0.0657501668,
0.0121741695,
-0.0098384619,
0.0059482204,
-0.0309933964,
0.0467882492,
0.0008239477,
0.0130494246,
0.0124524506,
0.0107555119,
0.0319961198,
-0.0074977493,
-0.0066876668,
0.0094989119,
0.0106599145,
-0.0455835238,
0.0223341491,
-0.0271302462,
0.0134241125,
-0.0041395766,
0.029122049,
-0.0225239396,
-0.0815047175,
0.0648467541,
-0.0498990268,
-0.0485389233,
-0.0649989545,
-0.0252372585,
-0.011362358,
0.0137855401,
-0.0223614834,
-0.0263553392,
0.0236354023,
0.0274116676,
-0.0100188218,
-0.0229899157,
-0.0719593093,
-0.0499836504,
-0.036868982,
-0.0260388348,
-0.0324001089,
0.0223525167,
-0.0180915631,
0.0137517769,
-0.0400545299,
0.0541454256,
2.209229134e-34,
0.0240336843,
0.0043557091,
0.0106625967,
-0.0271750074,
0.0258276872,
-0.0149437105,
-0.038417995,
-0.0066911587,
0.0586783849,
-0.032786753,
-0.0143383667
] |
http://arxiv.org/abs/2402.01980v2 | http://arxiv.org/pdf/2402.01980v2 | SOCIALITE-LLAMA: An Instruction-Tuned Model for Social Scientific Tasks | 2024-02-03T01:33:16 | 2024-03-14T15:30:41 | Gourab Dey, Adithya V Ganesan, Yash Kumar Lal, Manal Shah, Shreyashee Sinha, Matthew Matero, Salvatore Giorgi, Vivek Kulkarni, H. Andrew Schwartz | cs.CL | cs.CL | Social science NLP tasks, such as emotion or humor detection, are required to
capture the semantics along with the implicit pragmatics from text, often with
limited amounts of training data. Instruction tuning has been shown to improve
the many capabilities of large language models (LLMs) such as commonsense
reasoning, reading comprehension, and computer programming. However, little is
known about the effectiveness of instruction tuning on the social domain where
implicit pragmatic cues are often needed to be captured. We explore the use of
instruction tuning for social science NLP tasks and introduce Socialite-Llama
-- an open-source, instruction-tuned Llama. On a suite of 20 social science
tasks, Socialite-Llama improves upon the performance of Llama as well as
matches or improves upon the performance of a state-of-the-art, multi-task
finetuned model on a majority of them. Further, Socialite-Llama also leads to
improvement on 5 out of 6 related social tasks as compared to Llama, suggesting
instruction tuning can lead to generalized social understanding. All resources
including our code, model and dataset can be found through
bit.ly/socialitellama. | SOCIALITE-LLAMA: An Instruction-Tuned Model for Social Scientific Tasks: Social science NLP tasks, such as emotion or humor detection, are required to
capture the semantics along with the implicit pragmatics from text, often with
limited amounts of training data. Instruction tuning has been shown to improve
the many capabilities of large language models (LLMs) such as commonsense
reasoning, reading comprehension, and computer programming. However, little is
known about the effectiveness of instruction tuning on the social domain where
implicit pragmatic cues are often needed to be captured. We explore the use of
instruction tuning for social science NLP tasks and introduce Socialite-Llama
-- an open-source, instruction-tuned Llama. On a suite of 20 social science
tasks, Socialite-Llama improves upon the performance of Llama as well as
matches or improves upon the performance of a state-of-the-art, multi-task
finetuned model on a majority of them. Further, Socialite-Llama also leads to
improvement on 5 out of 6 related social tasks as compared to Llama, suggesting
instruction tuning can lead to generalized social understanding. All resources
including our code, model and dataset can be found through
bit.ly/socialitellama. | [
0.0627434179,
0.0637576282,
-0.017463794,
0.0215341039,
0.0242220517,
0.0353801139,
-0.0296127498,
-0.0034134083,
0.0357882939,
-0.0548087098,
-0.046082072,
0.0199323911,
-0.0288301297,
0.0004824208,
0.0373896174,
-0.0650947839,
0.0413096286,
0.0088260518,
0.0465569682,
-0.0110055115,
0.033061035,
0.0340233371,
-0.0046625617,
0.0244528577,
0.0271902792,
-0.0173882563,
-0.0033660973,
0.0387669243,
0.0143568488,
-0.0265675653,
0.0558222048,
0.0542990416,
0.0301405452,
-0.0434084721,
0.0000020055,
-0.0394004099,
0.019280728,
-0.0438672043,
0.000419428,
0.0461421311,
0.0722456053,
-0.0355266631,
-0.0045753894,
0.0110455295,
-0.0537987277,
-0.00159053,
0.0246908013,
0.0255761743,
0.0436699614,
0.0580152981,
0.0053243423,
-0.0045872829,
-0.0016195626,
-0.0256981105,
0.0012920541,
-0.0572302677,
0.0410784371,
0.0029012521,
0.0088057239,
0.0472184755,
-0.0331009328,
0.0351038985,
-0.0007071852,
-0.0200897828,
0.0183844045,
0.0223910362,
-0.0332973935,
-0.0023928762,
-0.0516220033,
0.0799816623,
0.0346545763,
-0.0267652106,
0.0170768891,
0.0048958804,
0.0266670212,
0.0384484753,
-0.0520476028,
-0.0020672251,
0.010145463,
0.0232275464,
0.0514600202,
-0.0023301574,
0.000929678,
0.0659493059,
0.0688970834,
0.0586122088,
-0.0363326371,
0.0057187448,
0.0184435118,
0.0312042832,
-0.0213792287,
-0.0308131427,
0.0605292507,
0.0411102399,
0.0084268469,
0.014948722,
-0.0417806394,
0.0187574234,
0.025888782,
-0.0271708686,
0.0317862146,
0.0448808745,
-0.0101813022,
0.0499466285,
0.0213165805,
0.0030802872,
-0.0183575582,
0.037828207,
0.0220386926,
-0.0076924697,
-0.0624263212,
0.0130817927,
-0.0272043124,
0.0171705093,
-0.0411652587,
0.0057234182,
-0.0259373803,
0.0154228713,
-0.0219911318,
-0.0273569115,
-0.0017293238,
0.0157819018,
-0.052400995,
0.0058511253,
-0.0295062698,
0.028800834,
-0.0669848472,
0.0490326285,
-0.0154468445,
-0.0663481951,
-0.0443878919,
-0.0170478802,
0.0451696552,
-0.0208779071,
0.0064766374,
0.0018058037,
0.0001547333,
-0.0503159165,
-0.0441970564,
-0.0298440848,
-0.0065917526,
-0.0611763932,
0.0224160384,
-0.0441643149,
0.0569860153,
0.01367738,
0.0132546313,
-0.0417503715,
-0.0033201298,
0.0123993885,
-0.0409407541,
0.040014416,
0.0004230241,
0.0006701415,
0.0052281655,
0.0505834259,
0.0195828509,
0.10170497,
-0.0094318623,
-0.0077143097,
0.0328642726,
-0.0074174814,
-0.0143380435,
0.0319969356,
0.0249512568,
0.0037874207,
0.0091382442,
0.0144748762,
-0.0285316464,
-0.0048065241,
0.0798639581,
0.0314859226,
-0.0245602988,
-0.0158525407,
0.0363292471,
0.0453918539,
-0.0235140063,
0.0631942376,
0.0280205477,
0.0518225022,
0.0249842107,
0.0029057951,
-0.0536933206,
0.0591949187,
-0.0646596774,
-0.0474314429,
-0.0135519635,
0.0228976589,
-0.0440493263,
-0.0227158014,
-0.0046710931,
-0.0043672463,
0.0436739437,
0.021355547,
0.0217646081,
0.0140905008,
-0.0457873717,
0.0257834978,
0.0553655177,
-0.0342079587,
-0.0187449604,
-0.0219815969,
0.0221792161,
0.0547855794,
0.0286926255,
-0.0359308496,
-0.0426569432,
-0.0130936475,
0.0542384908,
0.0353736617,
0.0635276213,
-0.0442927033,
0.0090618692,
-0.0409851633,
0.0026965681,
-0.0340164676,
0.0272302777,
-0.009940302,
0.0844107196,
0.0164124817,
-0.0033014645,
0.0106616281,
0.0062737106,
0.030496031,
-0.0658406615,
-0.0182009079,
-0.0593523532,
-0.0168990903,
0.0241248328,
-0.006179553,
0.0276983231,
-0.0174681004,
0.0063802893,
-0.0246591568,
-0.0054142233,
0.022031812,
-0.0246925838,
-0.0582771413,
0.0424410738,
-0.0072718076,
-0.069017671,
0.0536570363,
-0.0297406968,
0.0156648085,
0.0589641146,
-0.0426339507,
0.0226163492,
-0.0299430639,
-0.0205478109,
-0.0385002643,
0.0418550558,
-0.0069467276,
0.0604150891,
-0.0130094085,
-0.0385531969,
0.0203336179,
-0.0308835488,
-0.0454519615,
-0.0364573486,
0.0032182094,
-0.039127335,
0.0253829025,
-0.0397124477,
-0.0218426734,
0.0067753978,
-0.0787980407,
-0.0610300042,
-0.0665079281,
0.0253856704,
0.0007573118,
0.0111357309,
0.0122365449,
-0.0138550568,
-0.0264212936,
0.004459376,
0.0151904933,
0.0314689577,
-0.0362064727,
0.0109735681,
0.0416022204,
-0.0119441813,
-0.0075223311,
-0.074288629,
0.0070855585,
-0.0491711199,
-0.0043537733,
0.0182068646,
0.0114140119,
-0.058062125,
0.0367519371,
-0.0378880873,
0.0201945696,
-0.0130311977,
-0.074156642,
0.008797992,
0.0120860562,
0.0427692756,
0.1136626303,
-0.0204360653,
0.0510454401,
0.0147291441,
-0.0106754294,
-0.0740034506,
0.0241830405,
-0.0235232711,
0.0448663756,
-0.0230072197,
-0.0322687849,
-0.0470119864,
0.0645189434,
-0.0158420689,
0.0135180298,
-0.0693584085,
0.0222109705,
-0.0053307293,
-0.0358593613,
0.010603983,
-0.0099098077,
0.0107595064,
-0.0161422696,
-0.0492772385,
-0.0219603945,
0.0718278363,
0.0121932225,
-0.0313619971,
-0.0375715196,
-0.0215469263,
-0.024777811,
0.0638851672,
-0.0497898012,
0.0062239384,
-0.0349132568,
-0.0084294397,
-0.0284978133,
-0.0440010056,
-0.0164631959,
-0.0252099168,
0.0622218512,
-0.051260449,
-0.0149945123,
0.0241634473,
-0.038709376,
-0.0096780332,
-0.0620342717,
-0.0615764707,
-0.0054827393,
-0.0130842328,
-0.0236556809,
-0.000029102,
-0.0489308983,
0.0084637618,
-0.0260265991,
-0.0257259365,
-0.0325904228,
-0.0142528806,
-0.0264329072,
0.0677425936,
0.0313232392,
0.0208154693,
-0.030186208,
-0.0218051784,
0.0492160954,
0.0398032218,
-0.0270676948,
-0.0263128076,
0.007994079,
0.037558198,
0.0066710841,
-0.006380727,
-0.0031613228,
0.1222516671,
-0.035679359,
-0.0135326954,
0.1067489386,
0.0037745703,
-0.0024050158,
0.0217879377,
-0.0166734401,
0.0174124818,
0.041449856,
0.0455615297,
-0.0070793298,
-0.0558622293,
-0.0119700888,
0.0044438392,
0.0148861902,
-0.0495578051,
-0.0198357087,
-0.1009838805,
0.013874609,
0.0085944934,
-0.0124220522,
-0.0030883695,
-0.0506415293,
0.0313187987,
-0.0016646767,
-0.0255797058,
-0.0196502097,
-0.0340459533,
0.0089184139,
0.0241116732,
0.0054446277,
-0.0374490544,
0.0697075874,
-0.0319210589,
0.0980732813,
-0.0139353732,
0.0769207105,
0.0078041549,
-0.0411252454,
0.0268625095,
-0.0077804644,
0.0083955629,
-0.0347384252,
0.0241575912,
-0.0180120915,
0.0043212981,
-0.0255842581,
0.0685542822,
-0.051448334,
0.0043129595,
-0.01044802,
0.0612308756,
0.0060408376,
-0.0288822763,
-0.0285656489,
0.038020391,
0.010299773,
0.075687781,
0.0513400845,
-0.0113541679,
-0.0295296423,
-0.0344983339,
-0.0146080833,
-0.0546919294,
-0.0461508408,
0.0031262911,
-0.006533741,
-0.0371057764,
-0.0460030623,
-0.041845087,
0.0471031629,
0.0337582789,
-0.0041267527,
-0.0397898555,
0.0419597514,
0.0119834589,
-0.0436546244,
0.0642983988,
-0.0568990819,
0.0107039018,
0.0095644621,
-0.0491475649,
0.0335317515,
-0.0139151178,
-0.0163379461,
0.0184539985,
-0.0596323349,
-0.0716321319,
0.0209529921,
0.0430127606,
0.0179736018,
0.0173901562,
-0.017785253,
-0.0756663978,
-0.0103365518,
0.0477507561,
-0.0432436019,
0.0720117912,
0.011146537,
0.0162228756,
0.0261743236,
0.0607517175,
-0.0015707585,
0.025165651,
0.0113053834,
0.0320824534,
-0.049595762,
-0.009587124,
-0.0222640894,
-0.0045111845,
0.0305006895,
0.0154031012,
-0.0796294659,
-0.0320958458,
-0.0559799075,
-0.0015300382,
-0.0036587454,
-0.0323824137,
-0.0153560368,
0.0045312336,
-0.02248675,
-0.0485269502,
-0.0363931656,
-0.0463836752,
0.071978651,
-0.0472781286,
0.0317220874,
0.0343065299,
0.0107349195,
-0.0079237763,
0.0098471455,
-0.0324570499,
-0.0039067958,
-0.0135002527,
0.0342471004,
0.0018394713,
0.0502003431,
0.0127795264,
-0.00217062,
0.0005881843,
0.0124970563,
-0.0081164455,
0.0180898607,
0.0279602967,
0.0536501035,
0.0146024516,
-0.0209143627,
0.0608935282,
-0.0093869139,
-0.0012990918,
0.0179588795,
0.030671956,
-0.0387099311,
-0.0623930432,
0.0856299177,
0.0118822167,
0.0015947466,
-0.0189693514,
0.0566427223,
-0.0326121524,
0.0665776432,
0.0346283168,
0.0036342919,
-0.0113894567,
-0.0318800732,
0.0287936702,
0.0128430128,
0.0202040449,
0.023753766,
-0.0550370887,
-0.0488477871,
-0.0016924889,
-0.0137851378,
-0.0367008001,
-0.0382317081,
0.0248994827,
0.0642383322,
-0.0349672027,
-0.042399779,
-0.0367983133,
0.0435603298,
-0.0474569537,
-0.0110438196,
-0.021628486,
-6.734418156e-33,
-0.0532435328,
-0.0424511582,
0.0099747125,
0.0489212796,
-0.0361914784,
0.0254462827,
-0.0866474733,
0.0040366068,
-0.0406048819,
0.038773004,
-0.0188100897,
0.0522740372,
-0.0041451594,
-0.0296648666,
0.0094397347,
-0.0341750495,
0.0199261308,
0.0085380357,
-0.0155611774,
0.0213402938,
0.0121166268,
0.0101248706,
0.0614225753,
-0.091246821,
0.0106290281,
0.012736964,
0.0066492404,
0.0124705434,
0.0326141641,
0.0262541845,
-0.0159684159,
0.0513218604,
0.0205794647,
-0.0010283693,
0.0234753247,
0.0607727915,
-0.0678127855,
-0.0700804964,
0.0429021232,
-0.0102918744,
-0.0528481714,
-0.0414955877,
0.0497217998,
0.0059147771,
-0.0172467213,
-0.0163284987,
0.0350505523,
-0.0139834713,
0.0286818389,
0.0043662456,
-0.0599508025,
0.0031790836,
-0.023713544,
-0.0426234864,
-0.0149437776,
0.0011710165,
0.0181691851,
0.0421854407,
-0.0403102003,
0.0320492312,
-0.0346200168,
-0.0034229765,
0.0059116748,
-0.0499923155,
0.0191960111,
0.0688074529,
-0.0301957577,
0.0269295387,
-0.016440019,
0.0078299977,
0.0389766097,
0.0275688749,
0.0055729747,
0.0261979997,
0.0015698763,
-0.0032060973,
-0.0534965582,
0.0169644374,
0.0929993913,
0.0085132411,
0.0004056166,
0.0002690984,
-0.0534592681,
-0.0004781448,
0.0007630239,
0.0184087381,
-0.0021150303,
-0.0152613539,
0.0155144623,
0.0175425299,
-0.0574841537,
0.0399444327,
-0.0181736443,
-0.0295630675,
-0.0157319661,
0.0140423309,
-0.0164891202,
0.013203972,
0.0046202238,
-0.0120539265,
-0.0800457895,
-0.0416254699,
0.005049157,
0.0061991122,
0.0452303402,
0.0262078121,
0.027170537,
0.0257929917,
-0.0402739979,
-0.0093094036,
-0.0340941884,
0.0266037025,
-0.0211984124,
-0.0073607238,
0.0074880659,
0.0565131009,
-0.0160393827,
0.0143077234,
0.0456642732,
0.0220018607,
-0.0097522736,
0.0162560791,
-0.0108723147,
0.0121276807,
-0.0413438529,
-0.0418564193,
-0.0122194029,
-0.0539426208,
-0.0361960642,
-0.0555803888,
-0.0165670235,
0.0114948312,
2.695e-7,
0.0180782322,
0.0975803807,
0.0098107513,
0.0116983084,
0.0288659669,
0.023561582,
-0.0031204256,
0.0501502752,
0.0251154937,
-0.0457499325,
-0.0367894508,
0.0719147548,
0.0143171763,
0.0228020921,
-0.0872231349,
0.0051751551,
-0.0156179797,
0.0214580726,
-0.0231825132,
-0.0303764585,
-0.0036770795,
0.051856488,
0.0583905242,
0.0120985145,
0.0001203825,
-0.0323538706,
0.0608447082,
0.0052628969,
-0.0093842708,
0.0389794782,
-0.0187200811,
0.0231671259,
0.0025697467,
0.0140321404,
-0.0055922405,
-0.006814843,
-0.0181495454,
0.1021117941,
-0.0146714561,
-0.0172019508,
0.0264630131,
-0.0223875549,
-0.0197643433,
-0.0336788148,
0.0572312213,
-0.0703763142,
-0.0449153073,
0.0202922467,
-0.0105242645,
-0.0029314156,
-0.001848205,
0.0226269588,
0.0055657011,
0.0050393767,
0.0073332698,
-0.0411853828,
-0.0425621755,
-0.0761752576,
-0.0126012508,
-0.0303726755,
-0.0189754143,
-0.0405393504,
-0.0107536642,
-0.0186586324,
0.0033048254,
-0.0671179593,
0.0243333802,
2.376532056e-34,
0.0019952862,
0.0101488531,
0.0235813428,
-0.0123488484,
0.0435594656,
0.0081415251,
-0.0490942709,
0.0107908864,
0.0710921511,
-0.0519826822,
0.022609761
] |
http://arxiv.org/abs/2402.01976v1 | http://arxiv.org/pdf/2402.01976v1 | MasonPerplexity at ClimateActivism 2024: Integrating Advanced Ensemble Techniques and Data Augmentation for Climate Activism Stance and Hate Event Identification | 2024-02-03T01:06:33 | 2024-02-03T01:06:33 | Al Nahian Bin Emran, Amrita Ganguly, Sadiya Sayara Chowdhury Puspo, Dhiman Goswami, Md Nishat Raihan | cs.CL | cs.CL | The task of identifying public opinions on social media, particularly
regarding climate activism and the detection of hate events, has emerged as a
critical area of research in our rapidly changing world. With a growing number
of people voicing either to support or oppose to climate-related issues -
understanding these diverse viewpoints has become increasingly vital. Our team,
MasonPerplexity, participates in a significant research initiative focused on
this subject. We extensively test various models and methods, discovering that
our most effective results are achieved through ensemble modeling, enhanced by
data augmentation techniques like back-translation. In the specific components
of this research task, our team achieved notable positions, ranking 5th, 1st,
and 6th in the respective sub-tasks, thereby illustrating the effectiveness of
our approach in this important field of study. | MasonPerplexity at ClimateActivism 2024: Integrating Advanced Ensemble Techniques and Data Augmentation for Climate Activism Stance and Hate Event Identification: The task of identifying public opinions on social media, particularly
regarding climate activism and the detection of hate events, has emerged as a
critical area of research in our rapidly changing world. With a growing number
of people voicing either to support or oppose to climate-related issues -
understanding these diverse viewpoints has become increasingly vital. Our team,
MasonPerplexity, participates in a significant research initiative focused on
this subject. We extensively test various models and methods, discovering that
our most effective results are achieved through ensemble modeling, enhanced by
data augmentation techniques like back-translation. In the specific components
of this research task, our team achieved notable positions, ranking 5th, 1st,
and 6th in the respective sub-tasks, thereby illustrating the effectiveness of
our approach in this important field of study. | [
0.0162380841,
0.1211113036,
0.0037854209,
0.0272722021,
-0.0195522923,
0.0361205675,
0.0085202316,
-0.0251065902,
0.0105482908,
-0.0493106022,
0.0225192364,
-0.0311894957,
-0.0102105308,
-0.012886445,
-0.0162264723,
-0.042442739,
0.0134168379,
0.0344898589,
0.0030825185,
-0.0310300644,
0.0365207121,
-0.0141287399,
-0.012071725,
0.0301677007,
-0.0129358256,
0.0180315431,
0.0028586637,
0.010206921,
0.07428395,
-0.0013124002,
0.0203097723,
0.0468338169,
0.0093698464,
0.0905430317,
0.0000019678,
-0.0330117457,
-0.043789804,
-0.0189752504,
0.018951118,
-0.0285344627,
0.0240555778,
0.0105725434,
0.0243231002,
0.0253539719,
-0.051469624,
0.0101151494,
0.0662950277,
0.0273506008,
0.0465734899,
0.0602717809,
-0.009705306,
-0.0114964619,
0.0088221738,
0.0049165171,
0.0047669946,
0.0493726358,
0.0378713682,
-0.0409355424,
-0.0779608712,
-0.0013843394,
-0.0052615432,
0.03060266,
-0.0138190826,
0.0189493503,
0.0094763543,
-0.0100408569,
-0.0211193953,
-0.0074870628,
-0.0295260046,
0.0498511381,
0.0165697001,
-0.0039000628,
-0.0067231506,
0.0112750819,
-0.0143196294,
0.0127486959,
-0.0501361154,
-0.0162368398,
-0.0279862583,
0.0325270183,
-0.0131786894,
0.0769176111,
0.0331567079,
0.0590465628,
-0.0249254089,
0.06335067,
-0.037482582,
-0.0080233011,
-0.0084356684,
0.031623628,
-0.0057100165,
0.021706719,
0.0438616946,
-0.0188965052,
0.0055321692,
0.0235490371,
0.0035332753,
-0.0540659428,
-0.0051157083,
-0.0712463856,
0.0266242959,
0.0339056812,
0.0267830025,
0.0493809879,
0.0267514773,
0.0128219752,
0.0087982714,
-0.0008833563,
-0.0047321827,
0.0132015254,
-0.089200668,
0.0280467086,
0.0131073073,
-0.000983543,
-0.0088695027,
0.0301995352,
-0.0338385887,
-0.0262808129,
-0.0477987193,
-0.0150349382,
0.0189113114,
0.0154728014,
0.0333260037,
-0.0129665928,
0.0073768618,
0.0742831081,
-0.0251805652,
0.0246089473,
0.0043624355,
-0.0487756282,
0.0055732452,
0.0225740373,
0.0088852001,
-0.0101001989,
-0.0088699544,
0.0049497462,
0.0459275693,
0.0068476913,
0.0309350677,
0.0022719093,
-0.0286278464,
-0.0512319691,
0.0229686741,
-0.0628285185,
0.0499607399,
0.0046863374,
0.0132002421,
-0.0602001622,
0.0038468339,
-0.0070205415,
-0.0232900418,
0.0853584185,
0.000788566,
-0.0086788656,
0.0343941934,
0.0288052727,
0.0076248543,
0.0825774893,
0.0160649624,
0.0537982807,
0.0355970822,
-0.0190583151,
0.0068175164,
0.0262635071,
0.019663563,
0.0208128802,
-0.0660827532,
0.00739541,
-0.0484566614,
-0.0137891937,
-0.0504422598,
0.0442746878,
-0.0026342925,
-0.0452448316,
0.0613236427,
0.0231702253,
-0.0477079973,
0.0053511169,
-0.0099333012,
0.0035883293,
-0.0176549871,
-0.0380921587,
-0.0085370401,
0.0131308837,
-0.0476630554,
-0.0505571924,
0.0306440461,
0.0695405304,
-0.0252756961,
-0.0611711703,
-0.0302714612,
0.0230765659,
0.0260853432,
-0.0179070327,
0.0055030137,
-0.046425771,
0.0016625786,
-0.0451493077,
0.0114153884,
-0.0740811303,
0.004701498,
-0.0314170234,
-0.0132224634,
0.0667209998,
-0.0102440557,
-0.0061952169,
-0.0517373942,
-0.0803451836,
0.0185942277,
0.0148545876,
0.0347130857,
0.0029978168,
-0.004164075,
-0.0183438454,
-0.050340347,
0.0413941517,
0.0254379101,
0.0073764361,
0.0358917117,
0.0134163424,
-0.0194335226,
0.0095792934,
-0.0228473749,
0.0367680751,
0.0158613846,
-0.014158302,
0.0119609274,
0.0287667178,
0.0649011731,
0.0415334329,
0.0000913769,
-0.0294187404,
0.0517136641,
-0.034308102,
-0.1160669848,
0.0459228493,
0.0021787174,
-0.0560897477,
0.029680565,
-0.046379704,
-0.0234583076,
0.0219562296,
-0.0319633409,
-0.003984313,
0.0681381896,
-0.020924883,
-0.0260739308,
-0.0122596715,
-0.0225162655,
-0.0577643886,
0.0395541266,
-0.0354144573,
0.0635181069,
-0.0476088151,
-0.0147611843,
-0.0289430898,
-0.1103875935,
0.0516174175,
0.0078559844,
-0.0221805386,
-0.0193017144,
0.0198023692,
0.0459571108,
0.0142283598,
-0.0672879294,
-0.0429299437,
0.0049065128,
-0.0618426092,
0.0092233205,
-0.0307940971,
0.025432542,
0.0034895181,
-0.0052875667,
-0.0212975089,
0.0322919935,
-0.0026158798,
-0.0282266717,
0.0041563907,
0.0277154259,
-0.0132795945,
-0.0267689787,
0.006205183,
-0.0410012193,
-0.0096255764,
-0.0101218019,
0.0553006157,
0.0365394503,
-0.0906204432,
0.0209317561,
-0.0363827161,
-0.061206203,
-0.0295010265,
-0.0285488795,
-0.0237352103,
0.0002271944,
-0.0130458623,
-0.0197966509,
0.0088799326,
0.0503992923,
0.0744299665,
0.0095818862,
0.0215708409,
-0.0584259108,
-0.0084883906,
-0.0547684804,
0.0303555969,
0.0069702393,
-0.0303445123,
0.004971398,
0.0483037867,
-0.0075210962,
-0.0036476427,
-0.0257635973,
0.065272972,
-0.023394892,
-0.0456471369,
-0.0309293866,
-0.0460798442,
0.0242832098,
-0.0159692243,
-0.014735139,
-0.054527726,
0.0674428567,
0.007601392,
0.0409849733,
-0.0625753254,
-0.0376621261,
-0.0113212224,
0.0391713679,
0.0113460775,
0.022959942,
-0.0330964774,
0.0451353043,
0.0327115618,
-0.0455207974,
-0.0073985266,
-0.0390656628,
0.0014599747,
-0.0073621566,
0.0238058567,
0.022731008,
-0.0097240238,
-0.0423537865,
-0.061622221,
-0.0444139093,
-0.0199014358,
0.0119781829,
-0.037064109,
-0.0560731553,
0.0430897884,
0.0214375369,
-0.0467164926,
-0.0025919604,
-0.0157438163,
0.0425377265,
-0.0132601885,
-0.0119505199,
0.0051805116,
0.0135379322,
-0.0397430994,
0.0066381744,
0.03977165,
0.0318373367,
-0.0128984014,
-0.0803789124,
-0.0382477045,
0.0182375256,
0.0126800202,
0.0118356999,
-0.0805339739,
0.0542491786,
-0.0234363563,
0.0303396117,
0.061175663,
0.0126687447,
-0.0379901044,
-0.017744014,
0.0172402337,
0.0422689095,
0.0100123491,
-0.0039416994,
-0.0202515405,
-0.0102425897,
-0.0069808955,
0.0250987019,
-0.0359551199,
-0.0562804975,
0.0036885284,
-0.0428698286,
0.0456536189,
0.010349174,
-0.0606190786,
-0.0183295906,
-0.0527003892,
0.0486117937,
-0.0406542718,
0.0030853064,
-0.0586264245,
0.0490371697,
0.0310135968,
0.0352527574,
-0.0555522032,
-0.030067917,
-0.0196461696,
-0.0292907711,
0.0112012336,
0.0567108467,
0.0380951725,
-0.0042545632,
-0.0209524445,
-0.0052474542,
-0.024647668,
-0.0042936588,
-0.036071673,
0.0700237304,
-0.0026726963,
-0.003849919,
-0.0135851139,
0.0536251701,
-0.0527604446,
0.0081953341,
0.0064805453,
0.0835075751,
-0.0194474738,
-0.0048758234,
-0.0455195457,
0.0258176029,
0.05690483,
0.0578252599,
0.0214846898,
-0.0135512454,
0.0087201074,
-0.0870189667,
0.0022253611,
-0.0645981953,
-0.0034007223,
-0.0207388159,
0.0123497387,
-0.0356581584,
-0.0343705453,
-0.0200280752,
0.0534645431,
0.0409562103,
-0.0216816962,
0.022126928,
-0.0122576514,
0.004481195,
-0.0667356551,
0.0458523706,
0.0159343444,
0.0723155215,
0.0346339084,
-0.0555403791,
0.0104453992,
0.0270678587,
-0.0273791701,
0.0178681426,
-0.0531949773,
-0.050951492,
0.0262041315,
0.0527686402,
0.0267084278,
0.0251607448,
0.0070046685,
-0.0341347195,
0.0266274679,
0.1322359145,
0.0511813946,
-0.0102579622,
0.0128300227,
0.0167805813,
0.0031657221,
0.0143150054,
0.0351409279,
0.0455062501,
0.0353578702,
-0.012123107,
-0.0897681639,
-0.0120451516,
0.0003685994,
-0.0170915369,
0.0598101802,
0.0082289912,
-0.0733537152,
-0.0292279702,
-0.0162271373,
0.0100902924,
-0.0279570501,
-0.0222133771,
-0.0246891622,
0.0468323268,
-0.0243289415,
-0.0379970446,
0.0510041006,
-0.0093519473,
0.0036524988,
-0.0781842694,
0.0736602247,
0.0247505475,
0.0412981398,
0.0086617293,
0.0170048699,
-0.0150372284,
0.0020157504,
0.0338718183,
-0.0262758657,
0.0119726332,
-0.0056732008,
0.0016474093,
0.0363442972,
0.0088317096,
-0.0037250079,
-0.0407567434,
-0.0246523358,
-0.0510620438,
0.0107133379,
-0.0297344662,
0.0129006784,
0.0377159491,
0.0269618332,
-0.0328663252,
-0.0386910848,
0.0380932614,
-0.0081504183,
-0.0272218678,
0.0320831276,
0.0225737542,
-0.0313110501,
0.0091999425,
0.0252690036,
-0.0622495376,
0.0291561745,
0.0488718115,
-0.0472018197,
-0.0350894332,
-0.0006386258,
0.0041452912,
-0.0199346524,
0.0145715177,
0.0144788604,
-0.0495329946,
-0.0200123526,
0.0292607732,
-0.0360796228,
0.0289986692,
0.0064562238,
-0.0070906389,
0.0360478424,
-0.0289454162,
0.0458124541,
-0.0250493716,
0.0185608119,
-0.0772153065,
-0.0323546305,
-0.004984315,
-5.854196599e-33,
-0.0101859914,
-0.0382648483,
-0.0219191238,
-0.0125295259,
-0.0077280845,
-0.0326468609,
0.0029514513,
0.0446830168,
-0.003103835,
0.0325898975,
0.0167762525,
-0.0111537911,
0.013426722,
0.0126724998,
-0.0029559294,
-0.009603438,
0.0198335238,
-0.0161478762,
-0.0159439575,
0.0222029854,
0.0713875368,
0.0179124046,
0.0405432135,
-0.0906600505,
0.0205653459,
0.0131933121,
-0.023837667,
-0.0193552766,
0.0081495969,
0.0495100915,
-0.0004135236,
0.0567733683,
0.0446288623,
-0.0307452381,
-0.0031283337,
0.0192840006,
-0.0129809584,
-0.00990442,
-0.0205533821,
0.0009342277,
0.0213721432,
0.0225553717,
0.0180424117,
0.0087581491,
-0.0510995574,
-0.0184872765,
0.0212183613,
-0.0401931815,
0.0269517899,
-0.0325925089,
-0.0391454585,
0.0341096669,
-0.0438906252,
0.0215704404,
-0.0145107144,
0.0293651726,
0.0301884674,
0.0784737691,
-0.0770080537,
0.0443397909,
0.0035349077,
-0.0366827138,
0.0344867669,
0.0110176997,
0.0117004616,
0.0728459582,
0.0250874404,
0.0168235768,
-0.0914223269,
0.0170557592,
0.0181845333,
0.0330304094,
0.0725907981,
0.0800418407,
0.0238279663,
-0.0396688096,
-0.0593548864,
0.0512761883,
0.0761390179,
-0.0463660993,
0.0266676936,
0.0027631184,
-0.0282258485,
-0.0048861727,
-0.0276669078,
-0.0551414713,
-0.0229287054,
0.0052062464,
-0.0084753716,
0.0024990651,
-0.0728719458,
0.0114631364,
0.0364894234,
-0.0398639217,
-0.0267961361,
0.0682589039,
-0.0118815834,
0.0300633479,
0.0339924693,
0.0420314521,
-0.0340513475,
0.0035854639,
-0.0148701649,
0.0242081285,
0.021837227,
-0.024181474,
0.0862003714,
0.0184640996,
-0.0507779308,
0.0043503041,
-0.021517301,
0.0097441906,
-0.0455885865,
0.0373361371,
-0.0013891414,
0.0294372253,
0.0022290756,
0.0006909909,
0.0099292602,
0.0306019336,
0.0457375459,
-0.0035740132,
-0.0445926934,
0.0061888658,
-0.005345582,
-0.0283503719,
0.0141760232,
-0.0596238747,
-0.0022762171,
-0.0758733153,
0.0162719283,
-0.0292949341,
2.574e-7,
0.0647316873,
0.0470098741,
-0.0246110205,
-0.0714988634,
0.014286668,
0.0277294032,
-0.0323679931,
0.0014620214,
-0.0184452161,
-0.0429221354,
-0.0212665908,
0.067350179,
0.036700163,
0.0346236229,
-0.019186154,
-0.0502686203,
0.0072556557,
0.0169018041,
-0.0491184741,
0.0455635935,
0.007505713,
0.0755072683,
0.0476170555,
-0.0055501084,
-0.0107132597,
0.0087800538,
0.0724527314,
-0.0627811626,
-0.0105346758,
0.0023674071,
0.0625824332,
-0.0275301859,
0.0017094497,
0.0289281979,
-0.0228949003,
0.0091500347,
0.0074625085,
0.0213884041,
0.0188864283,
-0.0173726752,
-0.0149912778,
-0.0043307333,
0.0200320967,
-0.0484145842,
0.0183281396,
-0.0519852452,
-0.0247443598,
-0.0542036556,
-0.0540840812,
0.0410381444,
0.0477276482,
0.0018407092,
0.0214860234,
-0.0050049708,
0.0162031297,
-0.0212726053,
-0.0038194738,
-0.0545936525,
0.0120971715,
-0.060806714,
-0.0698281378,
-0.0464575328,
0.00772662,
0.0148943141,
0.043347396,
0.0335326865,
0.0017627832,
2.226564002e-34,
0.0438961312,
0.0112367254,
-0.0349532515,
-0.0513317697,
0.0323367678,
-0.0245504603,
-0.0267869197,
0.0194149464,
0.0245343018,
-0.0557844639,
0.0068481117
] |
http://arxiv.org/abs/2402.01967v2 | http://arxiv.org/pdf/2402.01967v2 | MasonPerplexity at Multimodal Hate Speech Event Detection 2024: Hate Speech and Target Detection Using Transformer Ensembles | 2024-02-03T00:23:36 | 2024-02-18T06:39:55 | Amrita Ganguly, Al Nahian Bin Emran, Sadiya Sayara Chowdhury Puspo, Md Nishat Raihan, Dhiman Goswami, Marcos Zampieri | cs.CL | cs.CL | The automatic identification of offensive language such as hate speech is
important to keep discussions civil in online communities. Identifying hate
speech in multimodal content is a particularly challenging task because
offensiveness can be manifested in either words or images or a juxtaposition of
the two. This paper presents the MasonPerplexity submission for the Shared Task
on Multimodal Hate Speech Event Detection at CASE 2024 at EACL 2024. The task
is divided into two sub-tasks: sub-task A focuses on the identification of hate
speech and sub-task B focuses on the identification of targets in text-embedded
images during political events. We use an XLM-roBERTa-large model for sub-task
A and an ensemble approach combining XLM-roBERTa-base, BERTweet-large, and
BERT-base for sub-task B. Our approach obtained 0.8347 F1-score in sub-task A
and 0.6741 F1-score in sub-task B ranking 3rd on both sub-tasks. | MasonPerplexity at Multimodal Hate Speech Event Detection 2024: Hate Speech and Target Detection Using Transformer Ensembles: The automatic identification of offensive language such as hate speech is
important to keep discussions civil in online communities. Identifying hate
speech in multimodal content is a particularly challenging task because
offensiveness can be manifested in either words or images or a juxtaposition of
the two. This paper presents the MasonPerplexity submission for the Shared Task
on Multimodal Hate Speech Event Detection at CASE 2024 at EACL 2024. The task
is divided into two sub-tasks: sub-task A focuses on the identification of hate
speech and sub-task B focuses on the identification of targets in text-embedded
images during political events. We use an XLM-roBERTa-large model for sub-task
A and an ensemble approach combining XLM-roBERTa-base, BERTweet-large, and
BERT-base for sub-task B. Our approach obtained 0.8347 F1-score in sub-task A
and 0.6741 F1-score in sub-task B ranking 3rd on both sub-tasks. | [
0.0306013972,
0.0734699667,
-0.0117246248,
0.06924095,
-0.0192928221,
0.0506669022,
0.0345904976,
0.0298860781,
-0.0082636522,
-0.076545164,
-0.0036816131,
-0.0176460594,
0.0010688795,
-0.024163926,
-0.0290933568,
-0.0618428737,
0.0035491153,
0.0241227448,
-0.0180370696,
-0.0255038943,
0.0556021594,
-0.0034015814,
-0.0130829178,
0.0399613008,
-0.0983049721,
0.0277670827,
0.0379376784,
0.0075420933,
0.0544046052,
0.0190207195,
0.023166487,
0.0472150333,
0.0171252545,
0.071628809,
0.000001834,
-0.0303514209,
-0.0137726394,
-0.007472557,
0.031060243,
-0.0690275803,
0.0232923608,
-0.0043468624,
-0.0104338694,
0.0241803713,
-0.0015290782,
0.014294344,
0.0758572817,
0.1166264713,
0.0529040135,
0.0537613221,
-0.0191816743,
-0.0452074781,
0.0286542419,
0.03347506,
-0.0009285329,
0.0268300232,
0.0311499815,
-0.0574034266,
-0.0724796727,
-0.0007887293,
-0.0301547535,
0.006504613,
-0.011076156,
0.0381428413,
-0.0117437998,
-0.0167668108,
-0.0427784361,
-0.019729197,
-0.0045558121,
0.0299191419,
-0.0159806628,
0.0337762572,
0.0121472096,
0.0257349219,
-0.0326945707,
0.030607596,
-0.0231240578,
0.013500982,
-0.0417405218,
0.0164196808,
-0.0334209837,
0.0163696799,
0.0140162203,
0.0514148884,
-0.0142561868,
0.043919269,
0.0207025427,
-0.0156269204,
-0.0035018856,
-0.0012160136,
-0.02497264,
0.0320516601,
0.0509076342,
-0.0076959976,
-0.0198738184,
-0.0018797218,
-0.0340020731,
-0.0851114839,
-0.0085521564,
-0.0651883706,
0.0505449623,
0.0333889239,
-0.0516873673,
0.0340247639,
-0.0035874078,
0.0816192925,
-0.0273300037,
0.0132580744,
-0.0092147077,
-0.0119404849,
-0.0547505692,
0.0215062201,
-0.0144199049,
-0.01347108,
-0.0179009624,
0.0094630523,
-0.0491730422,
-0.0265587214,
-0.0438579284,
-0.0392086394,
0.0253118537,
0.0239657052,
-0.0293034483,
-0.0003295743,
0.0119855041,
0.0517883897,
-0.0114144376,
-0.009398533,
-0.0682109818,
-0.0622063205,
0.0139281712,
0.0765305758,
-0.013090712,
-0.0497682914,
0.0183892567,
0.0517068729,
-0.007324391,
-0.0516606718,
-0.0130505105,
-0.0017084632,
-0.0704671517,
-0.0404481068,
0.0535834506,
-0.0491526835,
0.0345618464,
-0.0156539697,
-0.0035916762,
-0.0396075584,
0.0093903709,
0.0017416818,
-0.0222561248,
0.0921928361,
0.0411211364,
-0.0187661164,
0.0760965198,
0.0284832362,
-0.0233185329,
0.0749513879,
0.0122916084,
0.044947762,
0.0493624583,
-0.0092101973,
0.014287699,
0.0700604245,
-0.0131232291,
-0.0040073758,
-0.0629337803,
-0.0014658915,
-0.0407686941,
-0.007434505,
0.0172584429,
0.0409517325,
-0.0062609422,
-0.0685088336,
0.0657729656,
0.0596257709,
-0.0196467713,
0.0008472956,
0.0189095978,
-0.0072024642,
-0.0371587612,
-0.0125029795,
-0.0304558855,
0.0189554039,
-0.0368152969,
-0.0566023253,
0.0243098587,
0.0917666033,
-0.0303918365,
-0.0205230899,
-0.0074451976,
-0.0358018763,
0.0423808917,
-0.0524790958,
0.0562900901,
-0.019579472,
-0.0547869951,
0.0114158299,
0.0237523466,
-0.0105697438,
0.0034974464,
-0.0627572611,
-0.0076991799,
0.0522237495,
0.0111088231,
-0.0058704177,
-0.01216242,
-0.053926684,
0.0161912665,
0.0592412427,
-0.0409266651,
0.0139328334,
0.004385998,
-0.0506681614,
-0.0447402894,
0.042015627,
0.0171410758,
0.0173474364,
0.0525886007,
0.0002818596,
0.0022125798,
0.047789406,
0.0054941303,
0.0257076342,
0.0577850565,
-0.0137909958,
-0.0072060185,
0.0181194879,
0.0509993583,
0.0175213348,
-0.0319562629,
-0.0564988405,
0.0138207013,
-0.0374426432,
-0.0653962046,
0.0192398392,
-0.0045148204,
-0.0265157316,
0.0244520232,
-0.0735194981,
-0.0140959471,
0.1014779657,
-0.003621191,
0.0286917854,
0.0674994588,
-0.0395488068,
-0.0075721755,
-0.0410997085,
-0.0196647253,
-0.0377357565,
0.0418730713,
-0.0127375741,
0.069449909,
-0.0136091905,
-0.0212685782,
0.0133943753,
-0.0784679204,
0.0257922523,
0.0171944331,
-0.0720585212,
-0.0194002297,
-0.003056553,
0.0433356166,
-0.0107021462,
-0.0171132758,
-0.0175321121,
-0.0121303927,
-0.0633225664,
-0.0057002194,
-0.0391486622,
-0.0401034728,
-0.0101589262,
-0.0139620146,
0.0065539982,
0.0150283119,
0.018376939,
-0.03317222,
0.0433624312,
0.0225893464,
-0.0151023017,
-0.0614476576,
-0.0341438279,
-0.002550157,
-0.0093043586,
-0.0231350679,
0.0590458103,
0.010132228,
-0.0723463669,
0.0263641346,
-0.0441727452,
-0.013582184,
0.0074862274,
-0.0146887591,
-0.0614492707,
0.0221589766,
0.0280347653,
0.0384027176,
0.0418336429,
0.079574272,
0.0945062339,
-0.023229016,
-0.0159108713,
-0.008917965,
-0.0040690452,
-0.023551194,
0.0335829221,
0.0055855368,
-0.0385103859,
0.0080518108,
0.0763697326,
0.0223568603,
0.0264058318,
-0.0445937663,
0.0008786701,
-0.0401776917,
-0.028578138,
-0.002635601,
-0.0777948573,
-0.0245128553,
-0.0378672034,
-0.013488221,
-0.0536947586,
0.0370330401,
-0.0053593419,
0.0138825681,
-0.0171623006,
-0.0068398588,
-0.0124981366,
0.0451984257,
0.017049931,
0.0571481362,
-0.0537030026,
-0.0169379339,
0.001786679,
-0.0384226292,
0.0092937546,
-0.0397233367,
0.0226081815,
0.0284811892,
0.0049233949,
0.0116271451,
-0.0222421139,
-0.0592350475,
-0.038180165,
-0.0349473804,
-0.0468234196,
0.0013911682,
-0.0323429592,
-0.065652214,
0.0381511077,
0.0245201234,
-0.0388028882,
0.0022073532,
0.012571902,
0.0516594648,
0.045172371,
-0.0140928784,
0.0115265474,
-0.0149035444,
-0.0230570436,
-0.0422616377,
0.064204298,
0.01099745,
0.0071777138,
-0.0673863664,
-0.0062292409,
0.0315921605,
0.0204610713,
0.020240929,
-0.0460107625,
0.0316947326,
0.0135086337,
0.0275697298,
0.0886813179,
0.0012523013,
-0.02445958,
-0.0187832974,
0.0185799059,
0.0000848246,
-0.0035839914,
-0.0138580017,
-0.0507395566,
-0.0108684609,
-0.0223271493,
-0.0012926076,
-0.0571313351,
-0.035787791,
0.0323056281,
-0.0373051874,
0.0415989608,
-0.0012414668,
-0.0240048021,
-0.0321426168,
0.014031346,
0.0393670276,
-0.0142558487,
0.0236675721,
-0.0141554615,
0.0289677829,
-0.0258367062,
0.0177108943,
-0.0429393537,
-0.0161667578,
0.0138330916,
-0.0472473465,
0.0106644649,
0.0296163168,
0.0398944728,
-0.0487235188,
-0.0377749316,
-0.0054052924,
-0.0027565251,
0.0039039757,
-0.0608857274,
0.0719972625,
-0.0117597748,
-0.0422448851,
-0.0316652246,
0.0169384181,
0.0158147011,
-0.032548923,
-0.007869116,
0.0588646233,
-0.0082225259,
-0.0043343119,
-0.0524601676,
0.0474703573,
-0.0120914299,
0.0692797229,
0.0137196705,
-0.0469400436,
0.0291336812,
-0.0296782795,
0.026096493,
-0.0521577671,
-0.0553105026,
0.0099984584,
-0.0053148605,
-0.0576547422,
-0.0453911759,
0.0160867274,
0.0495631136,
0.041031003,
-0.0175937023,
-0.0307501424,
-0.048264835,
0.0238748118,
-0.0696494579,
0.0280539282,
0.0214380827,
0.0457497053,
0.0010283411,
0.0103672342,
-0.0038119857,
-0.0076930099,
-0.0391143896,
0.037507575,
-0.0602230318,
-0.0130757084,
0.0622934103,
0.0555705428,
0.0215262808,
0.0055598915,
0.0214568861,
-0.0418870077,
0.0568961315,
0.0565808266,
0.017693663,
-0.0004079857,
0.0257221647,
-0.0267920624,
-0.0310892966,
0.0021363373,
-0.0050865891,
0.0630329326,
0.034156315,
-0.0211977772,
-0.0478401519,
-0.0112001095,
-0.0179075934,
-0.0411119759,
0.0634470955,
-0.0120522352,
-0.0506042764,
-0.0325122699,
-0.0156729333,
0.027557021,
-0.0139898853,
-0.0250028484,
0.0081523079,
-0.0090840384,
0.0151422638,
-0.0402368791,
-0.0177763943,
-0.0299253296,
0.0173828937,
-0.0381841399,
0.0582332611,
0.0394255705,
-0.0013073456,
0.0028136387,
0.0229382496,
-0.0163160302,
0.040044155,
0.0487207212,
-0.0187837481,
0.012236679,
0.0634888783,
0.0129483808,
0.0132147716,
0.0270489305,
-0.0031111673,
0.0187149607,
0.0135992048,
-0.0641144067,
0.0041489578,
-0.0136241866,
0.0001181182,
0.0253850017,
0.0043584863,
-0.0106260404,
-0.0131137595,
-0.0007372758,
-0.0142312087,
-0.0129832001,
0.062517643,
0.0219367091,
-0.0024048735,
0.0077320831,
0.0419041514,
-0.0496602096,
0.0247561205,
0.0694372803,
-0.0372026227,
-0.0310310125,
-0.012537404,
0.0525645912,
-0.0075164582,
0.0093480656,
0.0095282979,
-0.0233985204,
-0.0075533087,
0.0659846589,
-0.0556317456,
0.0211510584,
0.0046665575,
0.0181146841,
0.0516661443,
-0.0010512554,
0.0207901169,
0.0128875133,
-0.0170579497,
0.009780651,
-0.0322604142,
0.0240241643,
-5.85230442e-33,
-0.0311147235,
-0.0401764885,
-0.0225264449,
0.0172460526,
-0.0075192419,
-0.0009418906,
0.0564032346,
0.0089005744,
0.0041814693,
0.0195550453,
0.0248709228,
-0.0112106325,
0.0106608057,
0.0133771077,
0.0033899811,
-0.0303890537,
0.0080822101,
-0.0204855092,
-0.0021891922,
-0.0014274737,
0.0233097412,
0.0193526763,
0.0678169802,
-0.0532913208,
-0.0134015586,
0.023085786,
-0.0280180965,
-0.0108978767,
0.0484838895,
0.0837706774,
-0.033794675,
0.053415738,
0.0453997664,
-0.0436248519,
0.017776696,
0.0470543094,
-0.0138566447,
-0.0113255195,
0.0060367477,
0.0485023223,
0.0134216389,
0.0162722953,
0.0335713066,
-0.0208022427,
-0.055261638,
-0.0608808175,
0.0442581214,
-0.0339038186,
0.0325501524,
-0.0510757938,
-0.067129381,
0.0387970582,
-0.035126783,
-0.0145523492,
-0.006712005,
0.0116673419,
0.0559745915,
0.0534609519,
-0.0650999025,
0.0198078174,
-0.0394103117,
-0.0389635116,
0.0078829406,
0.0156031633,
-0.0047328332,
0.0806383193,
0.0236029904,
-0.0047254371,
-0.0454977825,
-0.0114206774,
0.0286021009,
0.0687611848,
0.0785595849,
0.0704670697,
0.0187690221,
-0.0295448452,
-0.0565648265,
0.0341138169,
0.0534171239,
0.0011488453,
0.0331779569,
-0.0202895068,
-0.0306825656,
-0.006596359,
0.0046880064,
-0.0453348272,
-0.0313092805,
-0.0281047877,
0.0125723472,
-0.0079370243,
-0.0262978598,
0.0341524631,
0.0634969249,
-0.0234429687,
0.00922505,
0.0114193009,
-0.009249487,
0.0463038459,
0.0048391842,
0.0450728759,
-0.004334385,
-0.0246497225,
0.0352962017,
-0.0084794359,
0.0394629426,
-0.0038155264,
0.0483318269,
0.0317709818,
-0.0896266997,
-0.0059028636,
-0.0005493612,
0.0093838358,
-0.016112797,
0.0432816073,
-0.0009379561,
0.0298037678,
-0.0204372294,
-0.0350910872,
0.0194525011,
-0.0184852909,
-0.0074824775,
-0.0192186311,
-0.0093454653,
-0.0087381667,
0.0015359734,
-0.0128640169,
0.0051107793,
-0.0486413762,
0.0079629403,
-0.1111365035,
0.0151431318,
-0.0489526466,
2.602e-7,
0.0174291059,
0.0635903403,
0.0021250925,
-0.0349433497,
0.0034919269,
0.0166994762,
-0.0048187952,
0.0309695806,
0.0114209689,
0.003607197,
-0.006850672,
0.0352556966,
0.0106842509,
0.0462697819,
-0.0029818607,
-0.0404897407,
-0.0560257994,
-0.0123061407,
-0.0664035752,
0.0236966293,
0.0236663129,
0.0745288357,
0.0518186465,
-0.0591350906,
-0.0257866103,
-0.0224117804,
0.0411078706,
-0.0446576402,
-0.0193419438,
-0.0339490399,
0.0538961552,
-0.0061193104,
-0.0026158695,
0.046893917,
0.010202921,
0.014971341,
0.0479981862,
0.0263789129,
-0.0023771538,
0.0128364461,
0.0013529811,
-0.0018811695,
-0.037151508,
-0.0313420594,
0.0279291142,
-0.0253325906,
-0.0297126081,
-0.0499985181,
-0.0697064698,
-0.0078355223,
0.0520712882,
-0.0238436814,
-0.0163000356,
0.0008996932,
0.0511206351,
-0.0195999108,
-0.0122493105,
-0.0981004164,
0.0322357826,
-0.0208041128,
-0.0698090643,
-0.0226383433,
-0.015895417,
-0.0476143509,
0.0490797237,
0.047259286,
-0.0010274685,
2.244944647e-34,
0.0301834382,
-0.011556536,
-0.0168727227,
-0.0263432264,
0.0180109441,
0.0011047366,
-0.0323952585,
-0.0103295539,
0.000185746,
-0.0508850329,
0.0030044189
] |
http://arxiv.org/abs/2402.01963v1 | http://arxiv.org/pdf/2402.01963v1 | Improving Large-Scale k-Nearest Neighbor Text Categorization with Label Autoencoders | 2024-02-03T00:11:29 | 2024-02-03T00:11:29 | Francisco J. Ribadas-Pena, Shuyuan Cao, Víctor M. Darriba Bilbao | cs.LG | cs.LG, cs.CL, cs.IR, 68T50, 68T07, I.2.6; I.2.7; H.3.1 | In this paper, we introduce a multi-label lazy learning approach to deal with
automatic semantic indexing in large document collections in the presence of
complex and structured label vocabularies with high inter-label correlation.
The proposed method is an evolution of the traditional k-Nearest Neighbors
algorithm which uses a large autoencoder trained to map the large label space
to a reduced size latent space and to regenerate the predicted labels from this
latent space. We have evaluated our proposal in a large portion of the MEDLINE
biomedical document collection which uses the Medical Subject Headings (MeSH)
thesaurus as a controlled vocabulary. In our experiments we propose and
evaluate several document representation approaches and different label
autoencoder configurations. | Improving Large-Scale k-Nearest Neighbor Text Categorization with Label Autoencoders: In this paper, we introduce a multi-label lazy learning approach to deal with
automatic semantic indexing in large document collections in the presence of
complex and structured label vocabularies with high inter-label correlation.
The proposed method is an evolution of the traditional k-Nearest Neighbors
algorithm which uses a large autoencoder trained to map the large label space
to a reduced size latent space and to regenerate the predicted labels from this
latent space. We have evaluated our proposal in a large portion of the MEDLINE
biomedical document collection which uses the Medical Subject Headings (MeSH)
thesaurus as a controlled vocabulary. In our experiments we propose and
evaluate several document representation approaches and different label
autoencoder configurations. | [
0.0016808043,
0.006685629,
-0.008973754,
-0.0515752248,
0.0119729592,
0.0590328574,
0.059822876,
0.0569933765,
0.0345006995,
-0.0635081306,
-0.0540890656,
0.0295938477,
-0.0114606554,
0.0196609478,
0.0009108023,
-0.0137460511,
0.0255901311,
0.0089434301,
0.0181669742,
-0.029145319,
-0.0324356072,
0.0141004287,
-0.0156150712,
0.0244094599,
-0.0031035903,
0.0181058254,
0.0248647481,
-0.043285042,
-0.0265927445,
-0.0463929325,
0.0160925202,
0.0286036506,
0.0132117029,
0.0043216255,
0.0000018728,
-0.0150564248,
-0.0019231713,
0.0111351069,
0.0172547828,
0.0017425022,
0.0608726665,
-0.0725305304,
0.0404004194,
-0.0186535623,
-0.0039744019,
-0.0774900168,
0.0106063569,
0.0279305521,
0.041945409,
0.010424451,
-0.0292437896,
-0.0765660852,
0.0425894931,
-0.0066518397,
0.0279984307,
-0.0854970887,
0.0535834022,
-0.0052224891,
0.0104115847,
0.0269046836,
-0.0189157017,
0.0417415462,
-0.010475073,
-0.017710181,
0.0953111351,
0.0527500995,
0.0153771006,
-0.0292541999,
-0.000753576,
0.0230435971,
-0.0293625873,
-0.0050219647,
0.0105827712,
-0.0141532868,
-0.0577985793,
0.019776728,
-0.0047708717,
-0.0154700605,
-0.0079441611,
-0.0408616588,
0.0166221093,
0.0431408882,
0.0086485744,
-0.0076278863,
-0.0232247002,
-0.0000167896,
0.0262151454,
-0.0471878387,
0.0350952707,
0.002726166,
0.0072720512,
0.002438877,
0.034318015,
0.0350644328,
-0.0351441577,
-0.0086079407,
-0.0072895554,
-0.0741764605,
0.0433584899,
-0.0540283509,
-0.0142701063,
0.0003221927,
-0.0011916944,
0.022026455,
-0.0644373298,
0.0116557581,
-0.015425778,
0.0220149551,
-0.023042284,
-0.0152922347,
-0.0496750921,
-0.0209450461,
-0.0119627854,
0.0354597904,
0.0686807409,
-0.0504706688,
-0.0278204642,
-0.0154940728,
0.0095317196,
0.0122476211,
-0.0129592614,
0.0235430151,
-0.0562789887,
0.0014020469,
-0.0103760622,
0.0253050663,
-0.0393958613,
0.046512343,
-0.0563232377,
0.0607491657,
0.0444794036,
0.0228759311,
0.0770232528,
-0.013051942,
0.0105737066,
-0.0194115732,
-0.0035412887,
0.0049747205,
-0.0204993449,
-0.0503337644,
-0.0056377826,
0.067583777,
-0.0147844683,
-0.0501441993,
0.0048881322,
-0.008269175,
0.0285055581,
-0.0348728113,
-0.0220839418,
-0.0267307851,
-0.0634915531,
0.0695927367,
-0.0159813631,
0.0001263153,
0.0847416669,
0.0052451403,
0.0453410856,
0.0164296962,
-0.0415835083,
-0.0156190293,
0.0017327372,
-0.0274864342,
-0.0146294478,
-0.0289013125,
-0.0012617972,
-0.0508402884,
-0.0029694282,
-0.0204911362,
0.0149974879,
0.0734434053,
-0.0158239901,
0.0121835647,
-0.0259896889,
-0.0054356353,
-0.0515225604,
0.044611793,
0.0434745252,
0.0140420608,
-0.0055028573,
-0.0109858001,
-0.0390018634,
-0.0444304459,
0.0643846244,
0.0353198349,
0.0143586481,
0.0013852081,
0.0530246496,
-0.0123629598,
-0.0370425098,
0.004433854,
-0.008989145,
-0.0229478125,
-0.0397609733,
-0.0318172388,
0.077975221,
0.0161015745,
-0.0061815088,
0.03375737,
0.0608048774,
0.011623485,
0.011284804,
-0.0123353321,
-0.0056280014,
0.02054324,
0.0509634577,
-0.038667433,
-0.0440348722,
-0.0447399542,
0.0321559608,
-0.0238307491,
0.0103462879,
-0.0634464249,
-0.0162969306,
-0.0520712771,
-0.0496085845,
0.0382382199,
0.0303536765,
0.0224298574,
-0.0532808639,
0.0756342784,
0.0443060882,
-0.013933762,
0.0446835123,
0.0477479324,
-0.0479522459,
-0.006430516,
0.0362769961,
0.0267903488,
0.014775008,
0.0141908117,
-0.0259807035,
0.0012886686,
0.0449792333,
-0.0726967528,
-0.0344687514,
-0.0918466672,
-0.0161162503,
-0.0068541337,
0.017149508,
-0.0248128138,
0.0047203721,
-0.0134571632,
-0.0002389461,
-0.0299157538,
0.0043389741,
-0.0049601393,
-0.0666362494,
-0.0087796086,
-0.0258624144,
-0.039413251,
-0.0031234266,
-0.0331579112,
0.0427500904,
0.0202539321,
0.0620216988,
0.0121799232,
-0.0193939656,
0.0268395916,
0.0184753574,
0.0117303235,
0.0063767429,
0.0225702208,
-0.0174552053,
0.033899352,
-0.0022673004,
0.0157082267,
-0.059422452,
-0.0332881697,
-0.0207854398,
-0.0020614252,
-0.0344845764,
0.0142127834,
0.0291712247,
0.0370940343,
-0.0180750247,
-0.0502169393,
0.0344835483,
0.0155171454,
0.0035119988,
0.0284114499,
0.0097696511,
-0.0001343905,
0.0481957309,
-0.0372420363,
-0.0216765068,
0.0151488045,
0.0175120793,
-0.0464002863,
0.0110726496,
0.0014257091,
-0.0240823682,
-0.0423270501,
0.0213577207,
-0.0128001301,
0.0175212566,
-0.0357625037,
-0.0382573679,
0.0224406682,
0.0138139781,
0.0103717623,
0.0438819528,
0.0326890573,
-0.0406408384,
-0.0106630735,
-0.0592738912,
-0.0308258291,
-0.0181500614,
-0.0321561433,
0.0089363409,
0.0379834697,
0.018761402,
0.0088943243,
-0.0412071645,
-0.0296884272,
-0.0174393486,
0.0060841292,
-0.0653778613,
0.0047742641,
0.0113632577,
0.0347368568,
0.0291136373,
0.0362262651,
0.0951047316,
0.0461158007,
0.0179055724,
-0.0484299771,
-0.0062644635,
-0.0538376458,
-0.0098321512,
-0.01462314,
0.0416455418,
-0.0292241015,
-0.0111160064,
0.0170965809,
0.0554763637,
0.0349673927,
-0.0045665447,
-0.0247861333,
0.0524952114,
0.0085936869,
0.0222042315,
-0.0258645061,
0.002207262,
-0.0520514473,
0.003707418,
0.0110605247,
0.0011199448,
0.0252389442,
-0.0190934837,
0.0322593935,
-0.0233403295,
0.0796879083,
0.0128674423,
-0.0200698934,
-0.0174825583,
0.0377251953,
0.0285288636,
-0.0397179909,
-0.0283436831,
0.0103359362,
0.010800533,
0.0154854469,
0.0126782041,
0.0134658227,
0.0081250109,
0.0536277741,
0.0368449129,
-0.0013703631,
0.0392761752,
-0.004859644,
-0.0763637573,
0.0017761234,
-0.000478506,
0.1004149839,
0.0602107644,
-0.095852688,
0.0012958951,
-0.0484547727,
0.0260857865,
0.02473698,
0.043708019,
-0.0131952716,
-0.0046788915,
-0.0261640195,
0.0261940286,
-0.0067132683,
-0.0291502383,
0.0622372702,
-0.0449389666,
-0.0170181636,
0.0215409268,
0.0177711118,
0.000086264,
-0.0635559037,
-0.0122712953,
0.0000969571,
-0.0626486987,
-0.0192572363,
0.0302531067,
-0.0288262106,
0.0175400786,
-0.0149003724,
-0.0482608154,
0.0035497334,
-0.0086792847,
0.0521221645,
0.0294882879,
0.0247367304,
-0.0420862213,
-0.0181527995,
0.0070759999,
0.0368790887,
-0.0224772971,
-0.0243502837,
0.022732595,
-0.0148689141,
0.0156931635,
0.0282636974,
0.0127545213,
-0.084556967,
0.0197605845,
0.0085644638,
0.0587180182,
-0.0408916511,
-0.0551418662,
-0.0422474965,
-0.0470571294,
0.0120745087,
-0.0133956252,
-0.011169645,
0.0127904546,
0.0130535653,
0.0202484801,
0.0616409779,
0.0424253978,
-0.0211422909,
0.0364463553,
-0.0553644039,
-0.0808081105,
-0.0622725263,
-0.0183373578,
0.0406995304,
0.0175217595,
0.021371942,
-0.0066574365,
0.0206116643,
-0.0337204263,
-0.0435048714,
-0.0278360303,
0.0408123508,
-0.016313348,
0.0004493757,
-0.0296952352,
0.1017274484,
0.0015121722,
-0.0226953458,
-0.0633296892,
0.0293458719,
0.0042553507,
0.0145005882,
0.0097285099,
-0.0041166446,
0.0359582826,
-0.026963992,
-0.0695841834,
0.0724735186,
0.0505964458,
-0.1058029458,
-0.0021039369,
-0.0292031579,
-0.0649919584,
-0.0256452542,
0.0119245714,
-0.0592538863,
0.0582652725,
-0.0470018946,
0.0193659868,
-0.0222198758,
-0.0578679964,
-0.0019841052,
0.0237774402,
0.0617540851,
-0.0147954756,
0.0243505053,
-0.0499088019,
-0.0016130344,
0.0073494958,
-0.0204744302,
0.0118631823,
0.0097797187,
0.0609143972,
-0.017458966,
-0.0338319689,
0.0633364916,
0.0313991345,
0.0133509748,
0.0295878313,
0.0323175341,
0.0217616074,
0.0451154783,
-0.0048480364,
0.0220214743,
-0.0400727615,
0.005019452,
0.0308010448,
0.0294020753,
-0.0583980121,
0.0093903169,
0.032573279,
0.0444583073,
-0.0565351248,
0.0396894664,
-0.0562349707,
-0.0113134887,
-0.0297564678,
-0.0287505519,
-0.0232309699,
-0.0098221041,
0.0012095124,
-0.0231330581,
-0.0287757125,
-0.0131004881,
0.0170919932,
-0.0214834418,
-0.0369000994,
0.0124748368,
0.0029969926,
0.0070394096,
-0.0042988537,
0.0714061111,
0.024630392,
0.0006986859,
-0.0050850194,
-0.0501557365,
-0.0278227404,
0.0306556784,
0.0180629045,
0.0230991375,
-0.0131824948,
0.0247103386,
0.0292378087,
0.0104236584,
0.0790327936,
-0.0752704069,
-0.0482424982,
-0.012601993,
0.0622817129,
0.0179584846,
0.0241818167,
0.0505997576,
-0.0146053853,
0.0121672219,
-0.0319219753,
-0.0563331544,
0.0150189754,
-6.066508878e-33,
0.0154369166,
-0.0311054531,
-0.0279938988,
-0.0530602708,
0.0096431132,
-0.0035310532,
-0.0231447276,
-0.0340922885,
0.0596329756,
0.0339673981,
-0.0116646662,
0.0021935287,
-0.0108685084,
0.0203030631,
0.0550213344,
-0.0219257027,
0.0296085626,
0.0108305896,
0.0175592043,
0.0563797206,
0.0041896841,
0.0577802472,
0.0074138492,
-0.1120448261,
-0.0157207828,
0.0396489985,
0.0554897524,
-0.0303005297,
0.0647221506,
0.0542525873,
-0.0519304872,
0.0438710004,
0.0095945885,
-0.0186704285,
0.0211532917,
-0.0040551391,
-0.071508199,
-0.036988914,
-0.1228026003,
0.0217102244,
-0.0010822014,
-0.0408538468,
0.0279003121,
-0.0022550814,
-0.0060217939,
-0.0006221302,
0.057811819,
0.0027056693,
-0.0702078044,
-0.0809087828,
-0.0146212177,
0.0160486735,
-0.016410036,
0.0303212609,
0.0487918556,
-0.0307954066,
-0.0202622954,
0.0617813766,
-0.0287452377,
0.0175217967,
-0.0461469106,
0.0257614385,
-0.0108239939,
-0.0125393979,
0.0238112975,
0.0457244515,
0.0383992195,
0.0175659638,
-0.0550177619,
0.0505626798,
-0.0078782281,
-0.0407422408,
0.0229942705,
0.0482425429,
0.0382205062,
-0.1003741696,
-0.0429893322,
0.0512971468,
-0.0396569222,
0.0338259339,
0.008974527,
0.0026025854,
-0.0312136691,
-0.0510781519,
-0.0096731689,
0.0126771554,
-0.0456252135,
-0.0061115408,
0.0055371206,
-0.0141317723,
-0.0210986715,
0.0326880068,
-0.0004751292,
-0.0352709368,
-0.0230695829,
-0.0121587487,
0.0228121784,
0.0603789017,
-0.0228753816,
0.0018206291,
-0.0286667757,
-0.0306340531,
0.0208734572,
-0.0216855817,
0.0298481658,
-0.0046886043,
0.029329,
0.0079633975,
-0.0325395241,
-0.0359166712,
0.0547374822,
-0.1076688915,
0.0434888378,
0.0236712005,
-0.019487692,
0.0307202805,
0.0416964926,
0.0837719738,
0.0426745415,
0.0064155054,
0.0005115216,
-0.0390276574,
0.0459597223,
-0.0057874005,
-0.0448230058,
-0.026090879,
-0.0576084219,
-0.0034608366,
0.0714861527,
-0.0454715043,
0.0165952891,
0.0227456335,
2.732e-7,
0.0619124696,
0.0447761491,
-0.0119989179,
-0.0645474344,
0.0225845464,
0.0317878462,
0.0322694555,
0.0463296436,
-0.0201443117,
-0.022375999,
0.0527521446,
-0.0403266586,
0.0150489742,
0.0182839278,
-0.0561801493,
-0.0256730933,
-0.0344470255,
0.02019912,
-0.0433656015,
-0.0107834088,
0.0287321154,
0.0756482482,
-0.0279250313,
-0.0143857542,
0.0265277885,
-0.0136982873,
0.026218107,
0.033054702,
0.0547432154,
0.0365603007,
-0.0155348107,
0.0044840961,
0.0126324194,
0.0347447395,
0.001445647,
0.0040404629,
0.0325243995,
0.0229287222,
-0.0661729351,
-0.0180057772,
0.005444224,
-0.0095424308,
0.0170887001,
0.036046911,
0.044689998,
-0.0717772841,
-0.0373282433,
0.0346290059,
-0.0021788306,
0.0081481626,
0.0412397571,
-0.0289488304,
-0.021675257,
0.0285949577,
0.0374555364,
0.0344119221,
-0.0063808588,
-0.0761798397,
0.0556074902,
0.0189980324,
-0.0477326363,
0.0386753231,
0.0174854565,
0.0512281843,
0.0377516896,
-0.0089050727,
0.0067123584,
2.031288677e-34,
-0.0025173188,
0.0087364865,
-0.0257474296,
0.0123259546,
-0.0169629753,
-0.050851427,
-0.0138358586,
-0.00917143,
0.0045068092,
-0.0779294819,
-0.0181460772
] |
http://arxiv.org/abs/2402.01945v1 | http://arxiv.org/pdf/2402.01945v1 | A Case Study on Filtering for End-to-End Speech Translation | 2024-02-02T22:42:33 | 2024-02-02T22:42:33 | Md Mahfuz Ibn Alam, Antonios Anastasopoulos | cs.CL | cs.CL | It is relatively easy to mine a large parallel corpus for any machine
learning task, such as speech-to-text or speech-to-speech translation. Although
these mined corpora are large in volume, their quality is questionable. This
work shows that the simplest filtering technique can trim down these big, noisy
datasets to a more manageable, clean dataset. We also show that using this
clean dataset can improve the model's performance, as in the case of the
multilingual-to-English Speech Translation (ST) model, where, on average, we
obtain a 4.65 BLEU score improvement. | A Case Study on Filtering for End-to-End Speech Translation: It is relatively easy to mine a large parallel corpus for any machine
learning task, such as speech-to-text or speech-to-speech translation. Although
these mined corpora are large in volume, their quality is questionable. This
work shows that the simplest filtering technique can trim down these big, noisy
datasets to a more manageable, clean dataset. We also show that using this
clean dataset can improve the model's performance, as in the case of the
multilingual-to-English Speech Translation (ST) model, where, on average, we
obtain a 4.65 BLEU score improvement. | [
0.0266241785,
0.0377112404,
-0.0222555697,
0.0347471572,
-0.0252230596,
-0.0080862036,
0.035586495,
0.0116950311,
-0.0477727875,
-0.0200520363,
-0.0326770172,
-0.0263203233,
0.0382959768,
-0.0244885813,
0.0059252316,
-0.0430344865,
0.0313649476,
-0.0063656927,
-0.0099647474,
0.0005855804,
-0.025103109,
0.0468968526,
-0.0003474847,
0.0307898242,
0.0161658954,
0.0041776467,
0.0342601053,
0.0479976088,
0.034053769,
0.0119634811,
-0.0486416332,
-0.0385215953,
-0.0158965252,
0.0456062406,
0.0000020964,
-0.0346943438,
-0.0413391963,
0.0114635751,
0.0083972765,
-0.0033353125,
0.0387784466,
-0.0124663617,
-0.0215159226,
0.0728668496,
-0.0207105689,
0.006036147,
0.0320716463,
-0.0174665321,
0.0469276533,
0.0602836013,
-0.0342090465,
-0.022287149,
-0.0159561597,
-0.0430370569,
-0.0642101169,
-0.0339180157,
-0.0194332711,
-0.0226591565,
-0.0106553705,
0.0026366205,
-0.0227229949,
0.0718062446,
-0.0080831256,
0.0060360334,
-0.0035711299,
-0.0189634375,
-0.0026700918,
-0.0464639217,
-0.0027335421,
0.0682098046,
-0.0682526454,
-0.0059884675,
0.0016682917,
-0.0055028307,
-0.0049283607,
-0.0148236742,
-0.0246303286,
0.0481329449,
-0.0035642949,
0.0175215509,
0.0006363224,
0.0037779324,
0.056330882,
-0.0195742957,
-0.0626788437,
-0.0070669581,
0.022567343,
-0.0393575206,
0.0304536894,
0.0089486418,
-0.0046477197,
-0.0114274155,
0.0876384899,
-0.0132814674,
-0.0201106481,
-0.0002611915,
-0.0298086982,
-0.0606632344,
-0.0408978537,
-0.0541146994,
0.0228756331,
0.0754413828,
-0.0044524828,
0.0371590704,
0.0063281292,
-0.0145773571,
-0.0059690643,
0.0350606032,
-0.062547721,
-0.0424915105,
-0.0150827123,
0.0674883425,
-0.0541755408,
-0.0131362751,
-0.0724861473,
0.0096039735,
-0.0093835853,
-0.0087496005,
-0.0230382476,
0.0090812417,
0.0287080891,
-0.0199552551,
-0.0577609651,
0.0057012178,
0.0266078245,
0.0318931267,
-0.0007238304,
0.0219085347,
-0.0042229756,
-0.049444519,
-0.0205330756,
0.0026640557,
0.0315527134,
-0.045567289,
-0.0039636656,
0.0623396486,
-0.0014503245,
-0.0115697738,
-0.0356835946,
-0.0036937089,
-0.017435506,
-0.054505907,
0.0048395772,
-0.026425058,
0.0559443198,
-0.0224104226,
0.0517208949,
-0.0225380026,
-0.0175083037,
0.0072800033,
-0.031737458,
0.0666106492,
-0.0120517565,
0.0075910478,
0.0258536227,
-0.0086892508,
-0.004048469,
0.0688578114,
0.0834586918,
0.0045061521,
0.0337031819,
-0.0287283938,
-0.0424570516,
0.0731104016,
0.0174150933,
0.0415793508,
-0.0411664881,
-0.0122648394,
-0.0004520309,
0.0268616732,
0.0214526821,
0.0442380831,
-0.0101121916,
0.0110667292,
0.0602796488,
0.0251103863,
0.0478801765,
0.0385453962,
0.0273645967,
-0.008252644,
-0.0276446287,
-0.0002775926,
0.0319541469,
0.0078431871,
0.019415779,
-0.0320125446,
0.0646396577,
0.0806753263,
-0.0180394687,
-0.0203198772,
-0.0101431524,
0.0014567781,
0.0159569122,
0.0182744134,
0.0402036943,
0.0225561541,
-0.0403108075,
0.0325431898,
0.0092203328,
-0.0853245482,
0.0540390946,
-0.0176663585,
-0.0092413733,
0.0578627288,
0.0472393185,
0.02457981,
0.0223973226,
-0.0652934015,
-0.0118782902,
0.0201339275,
0.0203480925,
-0.0038103852,
0.0798527822,
-0.0649667755,
-0.0184334181,
0.0603099912,
-0.019118866,
0.0160269439,
-0.0428030007,
0.0269111525,
0.0242139772,
-0.0061284243,
-0.0597874932,
0.0306754615,
-0.0205423683,
-0.0040823985,
0.0700771585,
0.0454200283,
0.0616072752,
0.032825768,
-0.0063759503,
-0.0252585784,
0.0434417911,
-0.0416424461,
0.0135133946,
0.034302175,
-0.0017418439,
-0.0338408351,
0.0288752206,
-0.0658234283,
0.0179405175,
0.0725952983,
-0.0281583238,
-0.0724352226,
-0.0136484364,
-0.0185798798,
0.0317741148,
-0.0086198365,
0.0404821374,
0.0051119477,
0.0184667744,
-0.0266237799,
-0.0228189491,
0.0175288133,
-0.0053275307,
0.0242876559,
-0.0050049801,
0.0270791203,
-0.0798015445,
-0.0150554441,
0.084547624,
0.0053791325,
0.0305059962,
0.0587657355,
-0.0019776158,
-0.0069138189,
0.0000438867,
-0.0649389774,
0.0242679305,
-0.0074247317,
0.0116056725,
0.0488246232,
0.0643266663,
0.018773796,
-0.0085725337,
-0.0048687542,
-0.0091294786,
0.0037283618,
-0.0126036862,
-0.0288734045,
0.0098789781,
-0.0028368032,
-0.0205538347,
-0.0106920255,
0.0112122893,
0.0216873828,
0.0045399703,
-0.020069547,
-0.0416175537,
0.0537830368,
-0.0029274786,
0.0052666054,
0.0288162306,
-0.0598950796,
-0.0158788227,
-0.0034314487,
0.0074291895,
0.0332899392,
0.0002275022,
-0.0139330486,
0.0484077409,
-0.025366487,
-0.0599165224,
0.0278546568,
-0.0103092119,
0.020507006,
-0.024424186,
-0.0270130262,
0.0205369331,
0.0375131369,
0.00966202,
0.0026634946,
-0.0630837604,
-0.024617387,
-0.0140961921,
-0.0196440667,
-0.0205880478,
0.0188159738,
-0.0050987406,
0.0341485813,
-0.0346437767,
-0.0221963134,
0.0387628227,
-0.0638134554,
-0.0014502807,
-0.0659552738,
-0.0354809947,
-0.0066720694,
0.0004573068,
-0.0404498391,
0.0255704857,
-0.0683748946,
0.0122922575,
-0.0296480283,
0.0173818357,
0.0167889241,
-0.0215517581,
0.0387745351,
-0.0655912161,
0.0179123711,
-0.0178083871,
-0.023928741,
-0.0209646039,
-0.0015160547,
-0.0094485264,
-0.0191881545,
0.0375346802,
-0.024857644,
-0.0702515244,
0.030428594,
0.0041873786,
0.0723296553,
0.0191612039,
-0.0106016425,
-0.012034487,
0.0107653029,
-0.0116881784,
-0.0080810292,
0.0142535102,
-0.0470535383,
0.0203054212,
-0.0022067581,
0.0574050024,
-0.016784016,
-0.0265068263,
-0.0467065498,
0.0470714495,
0.0190258212,
-0.026323054,
-0.0001479061,
0.0147313671,
-0.0470237993,
-0.0063330783,
0.0821654275,
-0.0019924955,
-0.0193640012,
-0.014739139,
0.0381528549,
0.0106800003,
0.0015946103,
-0.0143344607,
-0.0671064705,
-0.0428991057,
0.0402188972,
0.0461773872,
0.0188301243,
0.0223040897,
-0.0360949524,
-0.0875913724,
0.0497792363,
0.0524954982,
-0.0656716228,
-0.0505056269,
-0.009785898,
-0.0498213619,
-0.068297565,
0.0306316204,
-0.0308438838,
0.0202959683,
-0.0027438644,
0.0399576761,
0.0315206349,
-0.0528572612,
-0.0054871663,
-0.0387277901,
0.030453587,
0.0386915505,
0.0778090805,
0.0001002443,
-0.022249354,
0.0075439597,
-0.0120602343,
-0.0276145693,
-0.0266027693,
0.0167925172,
-0.006449277,
-0.0313923471,
0.0422251113,
0.0058065709,
-0.0605664738,
0.03337951,
0.0341953337,
0.0408244617,
-0.0338033549,
0.0195293911,
0.0288736727,
-0.0030738995,
0.0227656346,
0.0757486224,
0.0004574774,
-0.0590621121,
0.0271588825,
-0.0156777222,
-0.0278463867,
-0.0646515638,
0.005514144,
0.0150178317,
-0.0619777441,
-0.0463475585,
-0.0792915896,
-0.0505513549,
-0.0003417792,
0.0635437667,
0.031811107,
0.0202722549,
0.0298682507,
0.0187582355,
0.0087651601,
0.0200802535,
-0.0463715605,
0.0586749054,
-0.0599148385,
0.0191515069,
-0.0078727081,
-0.0295322835,
-0.0134125771,
0.0924879611,
-0.0822833404,
-0.0521438494,
0.0027447513,
0.0305130724,
0.0295582078,
0.0582493134,
-0.0351250991,
-0.0056609279,
0.0128361369,
-0.0118522691,
-0.0805245936,
0.0002780897,
0.0238980204,
-0.0270727295,
-0.0111219324,
0.0397071391,
-0.0187402051,
0.005632496,
0.0007429058,
0.0263256282,
-0.0205139648,
-0.057528507,
0.0026956904,
0.0472950302,
0.0387298539,
-0.0168941841,
-0.0327648483,
-0.0553808548,
0.0564677194,
0.0155494548,
0.0102008786,
-0.0686128587,
-0.002625312,
0.0574934781,
-0.0037721095,
-0.0052487324,
0.0285013672,
0.0084848329,
-0.0464633815,
0.0003818521,
0.0041299844,
-0.0375124104,
0.0388275273,
0.0177805964,
0.0270285476,
0.0030837981,
0.0151925869,
0.0347682089,
0.0188128613,
-0.018461043,
0.0218880791,
0.0093149906,
0.0383230411,
0.0338690989,
0.0202169083,
-0.0639228821,
-0.021945646,
-0.0232656281,
-0.0286985151,
0.0458973795,
-0.0151662445,
0.1245041415,
0.0135631161,
0.015523674,
-0.0410550907,
0.0363580436,
-0.003810185,
0.0135097485,
-0.0178478081,
-0.0383403115,
-0.0175156277,
0.033340387,
0.0522034168,
-0.0276331045,
0.0426029675,
-0.0213487688,
-0.0277286395,
-0.0299201012,
-0.0421009697,
-0.0263818055,
0.0097170249,
-0.0325335301,
-0.0317771509,
-0.0242087338,
-0.0386264771,
-0.0089230677,
-0.0121421702,
0.0146809453,
-0.0434478037,
0.0228315461,
0.0504499413,
-0.0433273315,
0.0380070135,
-0.0467246771,
-0.0323335193,
-0.0127004851,
-0.0037591618,
-0.0142192878,
-6.245142877e-33,
-0.0310490709,
0.0173492189,
0.0173091628,
-0.0408977717,
0.0006318598,
-0.0606814325,
-0.023336282,
-0.0548572987,
-0.0155607937,
-0.0059591429,
-0.0086324662,
0.0371331051,
0.0347862728,
0.0052783182,
0.0057633556,
0.0430263206,
0.111200273,
-0.0015977162,
-0.0413599797,
0.0268100128,
0.0837788656,
0.0550563447,
0.0608312897,
-0.0487267412,
-0.0095659923,
0.0039835763,
-0.0318365581,
-0.0121493889,
-0.0178599637,
0.0185545795,
0.0198908858,
0.0154748559,
-0.0025817435,
-0.0268487558,
-0.0338561833,
-0.0212788992,
-0.0795727372,
0.0453086197,
0.0182878207,
0.0779828653,
0.0088310419,
-0.0472875275,
0.0366925672,
-0.0186886396,
0.0022600833,
-0.0328018963,
0.009281992,
-0.0592233874,
-0.0319362655,
0.0039606658,
-0.0283875149,
0.0238975566,
-0.0513685234,
-0.0258306377,
0.059845455,
0.0731764734,
-0.0261388365,
-0.005028083,
-0.0810458884,
0.0222395174,
-0.0148098236,
0.0267550033,
0.0585134774,
-0.0001489521,
-0.0003123916,
0.0058830413,
0.0351870619,
-0.0169261191,
0.0840309486,
-0.0012441959,
0.0160695389,
0.0201102756,
0.0353329703,
0.0386031792,
0.0535804071,
-0.0626160726,
-0.0974432901,
0.0312799364,
-0.0286087822,
0.0110272113,
0.0030995724,
0.0178581178,
0.0061023762,
-0.0002132993,
-0.0278768595,
-0.0544813089,
-0.0004743492,
0.0013547727,
0.0287530441,
0.0347037539,
-0.0396157168,
-0.0145100867,
0.0700907111,
-0.0235049073,
0.0337005667,
-0.0279208366,
-0.0620899685,
-0.0345258079,
-0.0244037025,
0.0280732494,
-0.0308265593,
-0.0460668281,
0.0289733075,
0.0121061616,
0.0018697757,
-0.0178443622,
0.0332535356,
0.0272393096,
-0.0529564396,
0.0029727079,
-0.0787225142,
0.0436551794,
-0.0314325169,
0.047775507,
0.0260660071,
-0.0054177828,
0.0082193026,
0.0467514023,
-0.0006033197,
0.0180161092,
0.0117132962,
-0.0295671858,
0.0218974389,
0.0105229858,
-0.0566599779,
-0.0112585267,
-0.0079446239,
-0.0044816202,
0.0886782706,
-0.0425653867,
-0.0262730308,
-0.0124361999,
2.816e-7,
0.005074908,
0.0185699351,
-0.0143515803,
0.0048660943,
-0.0026597185,
0.0208992828,
-0.0470811427,
0.0240524262,
0.0019855213,
0.0567428842,
0.0041347393,
-0.0144931637,
0.1300660521,
0.0345544033,
-0.0819837898,
-0.0560788065,
-0.0487651527,
0.0184366647,
-0.0216641631,
-0.0197915938,
-0.0139544252,
0.1113031954,
0.011887406,
-0.0336806364,
0.0143627636,
-0.0341896005,
0.0591950044,
-0.0443558469,
-0.0244888011,
-0.0170332864,
0.0448088609,
0.0176463854,
-0.065367505,
0.0050931969,
-0.0554396696,
0.0490949899,
0.0561363958,
0.0832953602,
0.0137238093,
-0.0048449044,
0.011319953,
-0.0009543914,
-0.0304952674,
-0.0009351433,
0.0316582806,
0.0227812566,
-0.0057620695,
-0.0136736836,
-0.0200413857,
0.0214089025,
0.0396358222,
-0.023227701,
-0.000565165,
0.0068697147,
0.0287863836,
-0.0181934629,
-0.0075748665,
-0.0310121924,
-0.0075208349,
-0.0251001082,
-0.0365745574,
-0.0699800774,
-0.0274581835,
0.0020268825,
0.0029177708,
-0.015478082,
-0.0086508561,
2.186300336e-34,
0.0330425352,
-0.0175687745,
0.0164098535,
-0.0412754118,
0.043575611,
-0.0501365364,
0.0131970206,
-0.0149939889,
-0.0115568107,
-0.032238286,
-0.0436428078
] |
http://arxiv.org/abs/2402.01939v1 | http://arxiv.org/pdf/2402.01939v1 | A Morphologically-Aware Dictionary-based Data Augmentation Technique for Machine Translation of Under-Represented Languages | 2024-02-02T22:25:44 | 2024-02-02T22:25:44 | Md Mahfuz Ibn Alam, Sina Ahmadi, Antonios Anastasopoulos | cs.CL | cs.CL | The availability of parallel texts is crucial to the performance of machine
translation models. However, most of the world's languages face the predominant
challenge of data scarcity. In this paper, we propose strategies to synthesize
parallel data relying on morpho-syntactic information and using bilingual
lexicons along with a small amount of seed parallel data. Our methodology
adheres to a realistic scenario backed by the small parallel seed data. It is
linguistically informed, as it aims to create augmented data that is more
likely to be grammatically correct. We analyze how our synthetic data can be
combined with raw parallel data and demonstrate a consistent improvement in
performance in our experiments on 14 languages (28 English <-> X pairs) ranging
from well- to very low-resource ones. Our method leads to improvements even
when using only five seed sentences and a bilingual lexicon. | A Morphologically-Aware Dictionary-based Data Augmentation Technique for Machine Translation of Under-Represented Languages: The availability of parallel texts is crucial to the performance of machine
translation models. However, most of the world's languages face the predominant
challenge of data scarcity. In this paper, we propose strategies to synthesize
parallel data relying on morpho-syntactic information and using bilingual
lexicons along with a small amount of seed parallel data. Our methodology
adheres to a realistic scenario backed by the small parallel seed data. It is
linguistically informed, as it aims to create augmented data that is more
likely to be grammatically correct. We analyze how our synthetic data can be
combined with raw parallel data and demonstrate a consistent improvement in
performance in our experiments on 14 languages (28 English <-> X pairs) ranging
from well- to very low-resource ones. Our method leads to improvements even
when using only five seed sentences and a bilingual lexicon. | [
0.0235726275,
0.0675024167,
-0.0209570862,
0.0479352772,
-0.0649059042,
0.0093295863,
-0.0058069495,
0.0015661406,
-0.0520545207,
-0.0182761103,
0.0017060559,
-0.0381028615,
0.0555081703,
-0.0347107202,
0.0426427275,
-0.0331960768,
0.0261552352,
-0.0046264757,
-0.0972487777,
-0.0357188918,
-0.0010269637,
0.0426964648,
0.012629969,
-0.0256811548,
-0.0343296528,
0.0319432877,
-0.0080493102,
0.0533836894,
0.0372237749,
0.0092385691,
-0.0369670615,
0.0255660247,
-0.0250004213,
0.081569396,
0.0000018523,
-0.0403218307,
-0.0340077356,
0.0140309809,
0.0187132023,
0.0046586702,
0.0579585433,
-0.0937087163,
-0.0355477743,
0.0269166827,
-0.005237089,
-0.0378115997,
0.0704384521,
-0.0406373926,
0.0313214958,
0.0572727732,
-0.0368078612,
-0.0520023853,
0.0339152217,
-0.0169246346,
-0.0310316086,
0.038436126,
-0.0055336603,
-0.0044929283,
-0.0105338814,
0.0108046588,
0.0118869795,
0.0535463952,
0.0132905198,
0.0107036773,
-0.022205526,
-0.0098836897,
-0.0251392182,
-0.0472760238,
0.0025669448,
0.0445265546,
-0.035741318,
-0.0500346646,
-0.013935592,
-0.0321142487,
-0.0031214424,
-0.0078928219,
-0.0240639709,
0.0210061651,
-0.0027364374,
0.0400270745,
0.0903553367,
-0.0071531786,
0.0381758884,
-0.0100265564,
-0.0473997705,
0.0066139363,
0.0652504787,
-0.0345811583,
-0.022635214,
0.0539216325,
-0.0071727247,
-0.007935144,
0.0042308006,
-0.0325324163,
0.0050322004,
0.0435387343,
-0.050624311,
-0.128180474,
0.0283952896,
-0.0598327704,
0.0447386689,
0.0423430838,
-0.0057876194,
0.0093906121,
-0.03568976,
-0.0035869009,
-0.0057646562,
-0.0069025992,
-0.0961857513,
-0.0432660691,
0.0138974255,
0.0680080056,
-0.0538268872,
-0.0416934676,
-0.0598772801,
0.0286469236,
0.0020920879,
0.0440155864,
0.0353195444,
-0.0067397514,
0.0296278745,
-0.0573623478,
-0.0746311769,
-0.0108884908,
-0.0033470714,
0.0082348734,
-0.0170154702,
-0.0081708198,
-0.018864641,
-0.0850765705,
-0.0218841992,
0.006751684,
0.0327879004,
-0.0318291374,
0.0014159734,
0.0692395493,
-0.0073627499,
-0.0612055808,
-0.0097146071,
-0.0059396815,
-0.0408735499,
-0.0256657414,
-0.0124604069,
-0.0727047846,
0.0614038594,
0.0095585007,
0.0125765326,
0.0106037818,
0.007028623,
-0.0114049967,
-0.0358252749,
0.0675242916,
-0.0983372852,
-0.002897291,
-0.0133625893,
-0.0354732089,
0.0340555049,
0.0856402442,
0.0257908069,
-0.0037625267,
0.0154720573,
0.0140237911,
-0.0599762201,
0.0502962358,
0.0160644595,
0.0396990478,
-0.068186447,
-0.0030391812,
0.0163368043,
0.0469760858,
0.0279040672,
0.0695866719,
-0.0039574751,
0.0413961112,
0.0511144213,
0.0204887539,
-0.0150063941,
0.0417368561,
0.0357680358,
-0.0004735389,
-0.0259773079,
0.0190502349,
0.0019948457,
-0.0057241661,
0.0046810415,
0.0511393622,
0.024508791,
0.0832458884,
0.0081919702,
-0.0281829517,
-0.0163532551,
-0.004481487,
0.0167189911,
-0.0032553708,
0.0358784832,
0.0436645336,
-0.0128766717,
-0.0000917607,
0.0358652808,
-0.0740239918,
0.028245762,
0.024892116,
-0.0559739098,
0.0756013468,
0.0489541776,
0.0258759204,
-0.0610676669,
0.0197520424,
-0.0102453986,
0.0447574593,
0.0570798516,
0.0067910971,
0.0307346918,
-0.0486767627,
0.002130873,
0.0231787805,
-0.0063141142,
0.0455882289,
0.0122951549,
0.0322974548,
0.0351690538,
0.0051630745,
-0.0782671049,
0.021963628,
-0.0398366302,
0.0021834255,
0.045391351,
0.0057210056,
0.0328773819,
0.0334823504,
-0.0042644613,
0.0055361628,
0.041559469,
-0.0353375413,
0.0097164921,
0.0470593609,
-0.0216987208,
-0.0551621728,
0.0479843505,
-0.0365095474,
-0.0000458376,
-0.0312844105,
-0.012028438,
-0.0736834332,
-0.0243403204,
-0.0763790533,
0.0631326362,
-0.0617221519,
0.051477924,
0.0145377861,
0.0208818745,
-0.0149504635,
0.0273731798,
0.0080345385,
0.0322476327,
-0.0016034044,
-0.0216454472,
0.0157950241,
-0.0390314981,
-0.0068636672,
-0.0058874777,
0.0362090953,
0.0398839489,
0.0214659255,
-0.0131753571,
-0.0081589166,
-0.0301160384,
-0.073145777,
-0.0056301821,
-0.0206177235,
0.0412966236,
0.0319149159,
0.0957413092,
-0.0009237586,
0.0011149897,
-0.0281668417,
0.0570051372,
0.0310597625,
0.0244206525,
0.0177592523,
0.0124063827,
0.0010974251,
0.0051978254,
0.005875607,
0.0364837572,
-0.0036114811,
0.0331454054,
-0.0033429912,
-0.0423387177,
-0.0037453412,
-0.017835971,
0.0182342827,
0.0167658404,
-0.058548294,
-0.0115488674,
-0.0338450447,
0.0934844017,
-0.0129475612,
-0.0276860315,
-0.060456682,
0.0504104421,
-0.0149059556,
-0.0233195052,
-0.0087837838,
-0.019271953,
-0.0130610336,
-0.0349161364,
-0.0669682249,
-0.0056219138,
0.0574303679,
-0.0261958018,
-0.0372423902,
-0.0245604571,
-0.0082661659,
0.0002300429,
-0.0119522521,
0.0047665955,
0.0393008068,
-0.0116523309,
0.0533282645,
-0.0317707695,
-0.0456092544,
0.0771741793,
-0.0559615009,
-0.0076596071,
-0.0705368593,
-0.0280886889,
-0.0222130343,
0.0340851061,
0.0001131827,
0.0183627345,
-0.0373694748,
0.0213343203,
-0.032313399,
0.0679188594,
0.0163074546,
-0.0153690493,
0.0262628067,
-0.0853819028,
0.0188980475,
-0.0160331987,
-0.0269621965,
-0.0378380865,
-0.0585674495,
-0.0022197498,
-0.0331573375,
0.076702863,
-0.0401007161,
0.0000368173,
-0.0125053534,
-0.020027142,
0.0895494074,
0.0023462672,
0.0228740238,
-0.0065000262,
-0.0224075448,
-0.0103811258,
-0.0021619441,
-0.0022856665,
-0.0167962462,
-0.0138478763,
-0.0314179808,
0.0077720471,
-0.0462605581,
-0.016058227,
-0.0197750516,
0.029036656,
0.0096588777,
-0.0437319428,
-0.021882385,
-0.0078646801,
-0.0034615526,
-0.0405490808,
0.069445841,
0.0264614485,
-0.027808113,
-0.014251682,
-0.0118325679,
0.0195444822,
0.0223630238,
0.0246801116,
-0.0914918929,
-0.002913286,
0.0026105901,
0.0287090242,
0.009691433,
0.0120312274,
-0.046621263,
-0.1214836836,
0.0037391467,
0.0451653488,
-0.0245818757,
-0.0211689062,
-0.0112432968,
-0.0179329794,
-0.0468288511,
0.0181931052,
-0.0036964703,
0.0109800855,
0.0090942131,
0.0076227998,
0.0738550201,
-0.0481884703,
-0.0689572468,
-0.0324322283,
0.0029945204,
0.0143513363,
0.0331586972,
-0.02116617,
-0.0329550244,
0.0113075739,
-0.0289476626,
-0.0221444778,
0.0095025087,
-0.0008443774,
0.0237603411,
-0.0325201787,
0.0085810255,
0.0130223101,
-0.0251999684,
0.0084884204,
-0.0052897613,
0.0486221164,
-0.008154544,
0.0417657234,
0.0078862691,
-0.0397181138,
0.0133522805,
0.0272015985,
0.0092774192,
-0.0373394452,
0.0556105785,
0.0067155212,
-0.0206019003,
-0.071214579,
0.0381412879,
0.0008451801,
-0.0534465872,
-0.0548181087,
-0.08908654,
-0.05920076,
0.0104833953,
0.0451057963,
0.035567034,
0.0106080938,
-0.0386237465,
-0.0234813839,
-0.0419944488,
0.0219630729,
0.0043495479,
0.0046335682,
-0.0426966958,
-0.0144489696,
0.016776599,
-0.0306992158,
0.0147129297,
0.038064003,
-0.0674711913,
-0.0605519414,
0.0209974665,
0.0416668504,
0.0093510607,
0.0458113253,
-0.0386267342,
-0.0092049967,
0.0292302053,
0.0425009988,
-0.0044331094,
-0.006991589,
0.0089114793,
-0.0111076953,
0.0119054299,
0.0477808118,
-0.0023971593,
-0.0017286551,
0.0153742498,
0.0264756139,
-0.0353126563,
-0.0878728479,
0.0445056632,
0.0295962505,
0.0438303836,
-0.0821825042,
0.0219208263,
-0.0223922636,
0.0335399583,
-0.0284180306,
-0.0238128062,
-0.0201384034,
-0.0136238411,
0.0571975298,
-0.0148027306,
-0.0078811049,
0.0161861498,
-0.0254097823,
-0.023605315,
0.0156161441,
-0.0245024171,
0.0194937643,
-0.0029516392,
0.031386137,
0.0379716568,
0.0065183486,
0.0242062379,
-0.0071727908,
0.0241856184,
-0.0584854633,
-0.0205713063,
-0.00557158,
-0.0102707632,
-0.0111402217,
0.024779534,
-0.0113803931,
0.0145704765,
-0.0316704363,
0.0226410497,
0.0743435025,
-0.0206370577,
0.0877427012,
-0.0456010513,
-0.0067545539,
-0.0140058575,
0.0283261668,
0.0076575782,
-0.0238370989,
-0.0187775567,
-0.0079253977,
-0.0170208775,
0.0107513424,
0.0574624464,
-0.0006014484,
-0.0006905817,
0.0000378915,
0.0172839612,
-0.002916245,
-0.0207132176,
-0.0165972374,
0.0272142142,
-0.0387448855,
-0.0084822671,
-0.025446998,
-0.0716777444,
-0.0118373642,
0.0479937196,
0.0131455446,
-0.0044204681,
0.0112219658,
0.0314373337,
-0.041164875,
0.0255390219,
-0.0333737843,
0.0406447947,
-0.012981846,
-0.0381141305,
0.0005111874,
-5.643511277e-33,
-0.0040768497,
0.0185088292,
-0.029335469,
-0.0514712259,
0.0307846386,
0.0055806339,
-0.011888979,
-0.0121429469,
0.026790278,
0.0178064387,
-0.0261284858,
0.0292336233,
0.0337070599,
0.0140413325,
0.0260960273,
0.0579015166,
0.0476022512,
0.0163595043,
0.0109167397,
-0.0214741472,
0.0430379286,
-0.0037323737,
0.0956727937,
-0.0251368042,
0.0575941838,
-0.0513623394,
0.0046748249,
0.0008534832,
-0.0086375028,
-0.0030894326,
0.0244179815,
0.0490303822,
0.0022209468,
-0.0096970527,
-0.0357151404,
-0.0267527886,
-0.07610064,
0.0000772885,
0.0620104186,
0.0467434116,
-0.0525312535,
-0.0210575983,
0.0193674844,
-0.0463821627,
0.011134591,
0.0209251698,
-0.0177276116,
-0.0573611148,
-0.0229428094,
0.0234952848,
-0.0220878366,
0.0122882295,
0.026538875,
0.0074129179,
0.0669462681,
0.0156540461,
0.0162919424,
-0.0712984204,
-0.0077935262,
0.0210203771,
-0.0081777992,
0.0540148541,
0.0841565952,
0.0148636755,
-0.0186323822,
0.0023458651,
0.0291985236,
-0.0029803074,
0.0884730071,
0.030885689,
0.0075717084,
-0.0233458709,
0.0275696646,
0.0970793068,
0.0660309568,
-0.0337057523,
-0.0157699734,
0.0411475934,
0.0258808471,
-0.0354496837,
-0.0048446669,
-1.082e-7,
-0.0118851764,
-0.0304293558,
-0.0256279577,
-0.055470299,
-0.0077544865,
-0.0074282559,
-0.0200484637,
0.0059506092,
-0.0530094355,
0.0582887679,
0.0447005555,
-0.0040895054,
-0.0060593849,
0.0001856873,
-0.0140103465,
0.0072075804,
-0.0450564101,
0.0420843661,
-0.015733663,
-0.0410056598,
0.0322478041,
0.0514998026,
0.0014497242,
0.006610916,
0.0347622782,
0.0123314103,
-0.0442799479,
0.0187200811,
-0.0549830049,
0.0171827357,
-0.0372999385,
0.0212275106,
0.0421383195,
0.0189396739,
0.0161461793,
-0.0014968709,
0.0820778236,
-0.0075883945,
-0.022556169,
-0.0359732509,
-0.0003767844,
0.017249288,
-0.0362544544,
0.0243078079,
-0.0173545182,
-0.0243081097,
0.0417815149,
-0.0090528475,
-0.01489692,
0.0115097351,
2.651e-7,
0.0327467918,
0.031018097,
0.0125656277,
0.0093671726,
0.0041125915,
-0.0193623062,
-0.0295936391,
0.0245241709,
-0.0210108552,
0.0014306249,
0.0311539974,
-0.0068921265,
0.0576292202,
0.0271777716,
-0.097949408,
-0.0469227023,
-0.0329999849,
0.0538855828,
-0.0639432222,
0.0068580578,
0.0305068698,
0.1097424403,
0.0150409108,
-0.0172705632,
-0.0057149376,
0.0000719836,
0.0239983778,
-0.0189525709,
-0.0278162025,
-0.0000123329,
0.0288322177,
0.0327792726,
-0.0340276174,
0.0190155711,
-0.0656593814,
0.0329233259,
0.0151041951,
0.0268630385,
-0.0159538463,
0.035756059,
0.027735386,
-0.0473316833,
-0.027488919,
-0.0062359236,
0.038336318,
0.0045060595,
0.0179321785,
0.0042589242,
-0.0233035367,
0.0438607298,
-0.0061481646,
-0.0258945245,
-0.0431143679,
0.0114268474,
0.0683310851,
-0.0276552197,
-0.0468705148,
-0.0418889821,
-0.0291585159,
0.0196336098,
0.0139854439,
-0.0506380573,
-0.0004160125,
0.0336722471,
0.0182625391,
-0.0163886379,
-0.0016339059,
2.409744363e-34,
0.0257153194,
0.0017356211,
0.0058422447,
-0.017688347,
0.0076916465,
-0.0500000417,
0.017973166,
0.0315126851,
-0.0215652641,
-0.0291144084,
-0.0360824391
] |
http://arxiv.org/abs/2402.01935v1 | http://arxiv.org/pdf/2402.01935v1 | Code Representation Learning At Scale | 2024-02-02T22:19:15 | 2024-02-02T22:19:15 | Dejiao Zhang, Wasi Ahmad, Ming Tan, Hantian Ding, Ramesh Nallapati, Dan Roth, Xiaofei Ma, Bing Xiang | cs.CL | cs.CL | Recent studies have shown that code language models at scale demonstrate
significant performance gains on downstream tasks, i.e., code generation.
However, most of the existing works on code representation learning train
models at a hundred million parameter scale using very limited pretraining
corpora. In this work, we fuel code representation learning with a vast amount
of code data via a two-stage pretraining scheme. We first train the encoders
via a mix that leverages both randomness in masking language modeling and the
structure aspect of programming language. We then enhance the representations
via contrastive learning with hard negative and hard positive constructed in an
unsupervised manner. We establish an off-the-shelf encoder model that
persistently outperforms the existing models on a wide variety of downstream
tasks by large margins. To comprehend the factors contributing to successful
code representation learning, we conduct detailed ablations and share our
findings on (i) a customized and effective token-level denoising scheme for
source code; (ii) the importance of hard negatives and hard positives; (iii)
how the proposed bimodal contrastive learning boost the cross-lingual semantic
search performance; and (iv) how the pretraining schemes decide the downstream
task performance scales with the model size. | Code Representation Learning At Scale: Recent studies have shown that code language models at scale demonstrate
significant performance gains on downstream tasks, i.e., code generation.
However, most of the existing works on code representation learning train
models at a hundred million parameter scale using very limited pretraining
corpora. In this work, we fuel code representation learning with a vast amount
of code data via a two-stage pretraining scheme. We first train the encoders
via a mix that leverages both randomness in masking language modeling and the
structure aspect of programming language. We then enhance the representations
via contrastive learning with hard negative and hard positive constructed in an
unsupervised manner. We establish an off-the-shelf encoder model that
persistently outperforms the existing models on a wide variety of downstream
tasks by large margins. To comprehend the factors contributing to successful
code representation learning, we conduct detailed ablations and share our
findings on (i) a customized and effective token-level denoising scheme for
source code; (ii) the importance of hard negatives and hard positives; (iii)
how the proposed bimodal contrastive learning boost the cross-lingual semantic
search performance; and (iv) how the pretraining schemes decide the downstream
task performance scales with the model size. | [
-0.032918632,
0.0620871969,
-0.0260268338,
0.0287107248,
0.0050370567,
0.0226697717,
0.0013343777,
0.0096227899,
-0.0213420913,
-0.0110230753,
-0.0125543075,
-0.0043781516,
-0.0024992973,
0.0798680559,
0.0705239996,
-0.0208032485,
0.0303040463,
-0.0152003877,
-0.0251572635,
-0.002393228,
0.0170041136,
0.0333384797,
0.0277524237,
0.0558480062,
0.0142940171,
-0.0300456211,
-0.0020118654,
0.0306038465,
-0.0144127654,
-0.0314321257,
0.0212422591,
0.0195055567,
0.0410873368,
0.0233367495,
0.0000022147,
-0.0084209619,
-0.0109556178,
-0.0216443446,
0.0176537093,
0.0274699666,
0.0324997567,
-0.0481201448,
0.0036515179,
0.0030879173,
-0.0331225544,
-0.036715012,
0.1052849665,
-0.0378883481,
0.0599572062,
0.042566441,
0.0007662374,
-0.0230748244,
-0.0025888884,
-0.0065525859,
0.0199524835,
0.0225436781,
-0.0294403266,
0.0349877663,
-0.0271381848,
0.0226155035,
-0.0249024034,
0.0192432757,
0.044647634,
0.0276232287,
0.0265473761,
0.0473790616,
-0.0404307097,
0.011289469,
0.0170074981,
0.0458867624,
-0.0747984275,
-0.0608318225,
-0.0149701955,
-0.0201823618,
-0.0401334167,
-0.0497248843,
-0.0127784759,
0.0309642814,
0.0135477232,
0.0364953242,
0.0121223638,
-0.0011488637,
0.0417409353,
-0.0147735868,
-0.0387600511,
-0.0208654758,
0.0133482879,
0.0010289033,
-0.0304655712,
-0.0160863493,
-0.0219521467,
-0.0492611788,
-0.0000593357,
0.0264240503,
-0.0060986755,
-0.0119062057,
-0.0399227776,
-0.0528572984,
0.0151226083,
-0.0256243255,
0.00670616,
0.0587888956,
-0.0392776467,
-0.0028965347,
-0.0762550384,
0.041187983,
-0.0205372572,
0.0165078528,
-0.0119246216,
-0.021183271,
-0.052694872,
-0.0133363977,
-0.0477345362,
0.0534030125,
-0.0447367616,
0.0266518351,
-0.0233828947,
0.0223986749,
-0.0095218876,
0.0118497284,
-0.0366759114,
-0.004137916,
-0.060982123,
0.0122672115,
0.0498150475,
0.0162386429,
-0.0373394378,
-0.045404423,
-0.004866187,
-0.0429869629,
-0.0111940308,
0.030577546,
-0.0142837614,
-0.0385003835,
0.0295645464,
0.0639755651,
0.0015354904,
0.0290946476,
-0.0508261621,
-0.0296265595,
0.0234442391,
0.0005631426,
-0.0160288084,
-0.082331486,
-0.0430125669,
0.0253171436,
0.0110226376,
-0.1007831693,
0.0051088338,
-0.0124386866,
0.0338985212,
0.07742621,
-0.0462942831,
-0.0371057019,
0.0116506414,
0.002799107,
-0.0013548363,
0.0545832627,
0.0380640551,
0.0115225408,
-0.0134547139,
-0.0552468859,
0.01194786,
0.0160523336,
0.0356779322,
0.0197341889,
-0.0119708683,
0.0092180846,
0.0526979007,
-0.0066975793,
0.1015266553,
0.0335623808,
-0.0382140018,
-0.0249208678,
-0.0000056455,
-0.0026169317,
-0.0060380669,
0.0424159877,
0.0852482021,
0.02951972,
-0.0583242029,
-0.0190185346,
-0.0062469486,
0.0529082902,
0.0283069089,
0.0102731623,
0.0090436116,
0.0251070969,
-0.0315676518,
-0.0215268228,
0.0269396268,
-0.0240938086,
0.0199852884,
-0.0171712898,
0.0157332178,
0.0159115158,
-0.0503670648,
0.0356392562,
0.0326625742,
-0.062227305,
0.0131711178,
-0.057624612,
0.0036445486,
0.0231232774,
0.0457749665,
-0.0261053275,
0.0149417296,
-0.012764534,
-0.0259461682,
-0.0010112573,
0.0005653106,
-0.0016836841,
0.0480689816,
-0.0753683299,
-0.0296084974,
-0.0829815641,
-0.0272321086,
0.0584540777,
-0.0052824453,
-0.001857792,
0.0491256155,
-0.0228267573,
-0.0450786166,
0.0361856893,
-0.0147984391,
-0.014075594,
0.0316596441,
0.0148963835,
0.0211356897,
-0.0520600416,
-0.0546509922,
0.0124470033,
0.0141866161,
-0.0520482138,
0.0008924992,
-0.0397492275,
-0.0646077543,
-0.0695131868,
0.0064990679,
-0.0783949047,
0.0197546333,
0.0650682747,
-0.0154779814,
-0.0200623386,
0.075530991,
-0.0393548645,
0.0430668257,
-0.015112875,
-0.0093919914,
-0.0116569959,
0.0197171196,
-0.0143580372,
0.0269008558,
-0.0003409658,
0.0255482253,
-0.0005560662,
0.0731256977,
0.045101881,
0.0033300442,
-0.0099939369,
-0.0173250623,
-0.0226976089,
0.0232906789,
-0.0019872901,
0.0389690138,
-0.0057439776,
0.0007412538,
-0.0846308172,
0.0101068225,
-0.0355387256,
-0.0175822321,
0.0356061012,
0.0185971484,
0.0366432965,
-0.0341173671,
-0.047962945,
-0.0131295128,
-0.0302833468,
0.0208704006,
0.0114764618,
-0.0070950179,
-0.040600501,
-0.024104422,
-0.0053182878,
-0.0034109864,
0.0479493439,
-0.0395740233,
-0.0063740881,
-0.0649074987,
0.0880466029,
-0.0009349092,
0.0123703266,
-0.02031116,
-0.0547792688,
0.0233754683,
-0.0285873543,
0.056508746,
0.0221101101,
-0.0148538481,
-0.0130374096,
0.0622405931,
0.0142208645,
-0.0625545233,
0.0073432885,
0.0261283536,
-0.0245643016,
-0.0057047149,
-0.0436272733,
0.0253994036,
0.0813585669,
0.0269175358,
-0.0040522697,
-0.0710731223,
-0.0087010805,
-0.0013485629,
-0.0352809951,
-0.0205191094,
-0.0303479377,
0.0022885578,
0.0541967489,
-0.0413827188,
-0.0906304419,
0.0623350665,
-0.0233874787,
-0.0377272628,
-0.0229378194,
-0.0127938632,
-0.0314915441,
0.0443117209,
-0.0078660566,
0.0336453207,
0.0343930088,
-0.0463703163,
-0.0587361902,
-0.0022774208,
-0.0168129969,
0.0013392161,
-0.0059001208,
-0.0369561873,
0.0097723622,
-0.0009203405,
-0.0415422507,
-0.0204578619,
-0.0379405469,
0.0154462792,
0.0215330664,
0.0730011761,
0.0152232992,
0.0253961403,
0.0347017013,
0.0069081187,
0.0534534119,
0.0400623195,
-0.0412832685,
-0.0341318473,
-0.0077158394,
-0.00958037,
-0.0399758108,
0.0005346456,
-0.0306483582,
-0.0343008228,
0.0011368842,
0.0269848723,
0.026631061,
-0.073228091,
0.0286032297,
0.0296200644,
0.0159805622,
-0.0405688882,
0.0512346104,
0.064344503,
-0.0316316932,
-0.0519167446,
0.0816744491,
0.0200367272,
-0.0681295991,
0.0107635166,
-0.010365082,
0.0617373325,
0.0699689761,
-0.0084585845,
-0.0616503954,
-0.0350286625,
-0.0175579246,
0.0226725973,
0.0003349296,
0.0129051385,
0.0031936404,
-0.0849507004,
-0.0006971165,
0.039110817,
-0.0281874537,
-0.0670594797,
-0.0304369684,
-0.057183411,
-0.0140836947,
-0.0111053428,
0.006829638,
-0.0008219149,
0.0172911026,
0.0145506384,
0.017173579,
0.0126221757,
0.015116985,
-0.0489630699,
0.0134658087,
0.0903817266,
0.0062601301,
0.0175347179,
-0.0148128886,
0.0114053432,
-0.0355467238,
0.0204603672,
-0.0260021538,
0.0598086305,
0.007938168,
-0.0312729031,
0.0320646465,
-0.0074497703,
-0.0734690428,
-0.0074499086,
0.0166897215,
0.0379199162,
-0.0120007368,
0.0219469722,
-0.0132852904,
-0.0124509512,
0.0095650023,
0.0240260996,
0.0629997104,
-0.0148252398,
-0.0333465636,
-0.0590867102,
0.0472396314,
-0.0569852591,
-0.0124522801,
-0.0264678989,
-0.0023461445,
0.028894918,
-0.0418222062,
-0.0168273095,
0.0000595207,
0.054785084,
0.0125983153,
-0.0357784182,
-0.0094015729,
-0.0146518815,
-0.0802455693,
0.0244156532,
-0.0533449613,
0.0070121074,
-0.0189678539,
-0.0491765402,
0.0022216372,
-0.0323803239,
-0.0258177835,
0.0084361201,
-0.0764849707,
-0.0270776786,
-0.0472850241,
0.0430105776,
0.0108670359,
0.0499114506,
-0.0002161038,
-0.0558088459,
0.0053916397,
-0.0038149208,
-0.0511282384,
0.0147321876,
0.0470099114,
-0.0079998001,
0.0115618417,
0.0079060132,
-0.0586841628,
0.0021529933,
0.0332652666,
-0.007950169,
-0.0294411983,
0.0468082428,
0.0434403196,
-0.0109112114,
0.0335268863,
-0.004142514,
-0.0140468851,
-0.0134102292,
-0.0019111946,
0.0104878899,
-0.0158495028,
-0.0444862545,
0.0038134467,
0.0451623462,
-0.000312785,
0.0190006252,
0.0703752041,
0.006087563,
0.0281691141,
-0.0273977332,
0.0431812741,
0.0084355194,
0.0545630641,
0.0436064824,
0.0284164269,
-0.0137451235,
0.0133313695,
0.0088489624,
0.0295982957,
-0.0184420235,
0.0412994586,
0.0156644192,
0.0424418598,
-0.0007155117,
0.0075934627,
-0.0374536254,
-0.0417620651,
-0.0117361238,
-0.06403853,
0.0657174885,
0.0299804285,
0.0472291298,
0.0140339071,
0.0220115632,
-0.024617862,
0.0684936941,
-0.0044483999,
-0.0374660268,
0.0047888546,
0.0150995022,
-0.0314853564,
-0.0376601666,
0.0976524875,
-0.0043752054,
0.0246383026,
-0.0240950156,
0.02712618,
-0.0377342626,
0.0139362793,
0.0487838015,
0.0330038816,
0.0097218864,
0.0054695942,
-0.0084000099,
-0.0634202883,
-0.0097250426,
0.0019494611,
0.0060372096,
-0.0071123731,
0.0683204606,
0.0820947289,
-0.0186498109,
-0.0098730819,
-0.0163264498,
0.0226464607,
0.0312716551,
-0.0246960092,
-0.0508460402,
-6.279252784e-33,
-0.01148747,
-0.0048008673,
-0.0095834658,
-0.020137921,
0.0158931147,
0.015702229,
-0.0560435243,
-0.0076851016,
-0.0155487582,
-0.021909887,
0.0655681342,
0.0314832516,
-0.0106717441,
-0.0026969004,
0.0173299313,
0.0250233486,
0.0797319934,
0.024875015,
-0.0272626523,
0.0703960806,
-0.0004848328,
0.0282819401,
0.0160085987,
-0.0521770418,
-0.0353814252,
0.0051926067,
0.0263139829,
-0.0249374844,
0.0748301372,
0.0369217284,
-0.0237196609,
0.0357594304,
-0.0063660848,
-0.0272650626,
-0.0603472106,
-0.0705150515,
-0.031766586,
-0.018793894,
-0.0037578868,
0.0071124672,
-0.0081151249,
-0.0889898688,
0.0780718103,
0.00940926,
-0.0239171777,
0.001203989,
0.0282172263,
0.0110896369,
-0.0155671686,
-0.0482352823,
-0.0759490803,
0.0289425272,
-0.0522414148,
-0.0206307638,
0.0208114889,
0.0224477593,
0.0021515873,
0.0617050976,
-0.0128641911,
-0.0022985835,
-0.017998185,
0.0596367568,
0.0502627492,
0.0383816697,
0.0214348771,
0.0361096263,
0.0478629395,
0.0202491321,
0.0156286173,
0.0351631083,
0.0707317591,
0.011113517,
0.046762377,
0.036428567,
0.0564666204,
0.0068291132,
-0.06187075,
0.0106095094,
-0.0449242108,
0.0247171801,
0.0217284076,
0.0694017559,
-0.0210126452,
-0.0200860985,
-0.0339122601,
-0.0534078591,
-0.0227181576,
-0.0364014432,
0.0351616144,
0.0059743328,
-0.0132602174,
0.001856456,
0.040636301,
-0.0385511927,
0.0244185086,
0.0486223102,
-0.0380072892,
0.0158964507,
0.0163325705,
0.0205735322,
-0.0354312286,
-0.0340804197,
0.0435840823,
-0.0044743917,
0.0745564699,
0.0012222767,
0.0356921032,
0.0112894764,
-0.0629642159,
-0.0103639569,
-0.0512176566,
-0.0097157462,
-0.004194478,
0.0229485035,
0.0082449717,
0.0063968077,
0.0041116634,
0.0021212555,
-0.007000959,
-0.0314162411,
-0.042292282,
-0.0113929864,
-0.0111565897,
0.0601969995,
-0.057609003,
-0.0307202898,
-0.0650380924,
0.0236775968,
0.0494106412,
-0.0494011752,
-0.0872415453,
-0.0050476901,
2.725e-7,
0.0132196378,
0.0106830746,
0.0099351155,
0.0435314402,
0.0066808695,
0.0448997095,
-0.0173211228,
0.0088645574,
0.0120881433,
-0.0355282314,
-0.0231697503,
0.0273785945,
0.0417747013,
0.0377524383,
-0.0353176557,
0.0609254874,
-0.0099512534,
0.0102590984,
-0.049818825,
0.0362317413,
0.0477323271,
0.1086612418,
-0.0200726781,
-0.0072618299,
0.0611639209,
-0.029584622,
0.0589942485,
0.0017934956,
0.0362957157,
-0.0005339532,
-0.1095789373,
0.0319556333,
-0.0271950904,
-0.017916983,
-0.0157200303,
-0.0216018017,
0.0090551684,
0.0461999848,
-0.0213654693,
0.0049411026,
0.0419574603,
0.00460112,
-0.0224978458,
-0.0691312626,
0.0607979931,
-0.0213635992,
0.0014563769,
-0.0073557789,
-0.0292161182,
0.0420723557,
0.0029269464,
-0.0291572157,
-0.0271417256,
0.006983262,
0.0120441495,
0.0004484629,
-0.0422275923,
-0.0453510098,
-0.0464696139,
0.021455463,
-0.0033706829,
-0.0424489565,
-0.0699877366,
0.0072996151,
0.0467217453,
-0.0054490468,
0.0405597985,
2.728030351e-34,
0.0373331606,
-0.0358121693,
0.0140688699,
0.0164532773,
0.0294611976,
-0.031698823,
-0.0475475118,
0.0087119387,
0.0155628603,
-0.0111583313,
-0.0319582708
] |
http://arxiv.org/abs/2402.01931v1 | http://arxiv.org/pdf/2402.01931v1 | Digits micro-model for accurate and secure transactions | 2024-02-02T22:01:27 | 2024-02-02T22:01:27 | Chirag Chhablani, Nikhita Sharma, Jordan Hosier, Vijay K. Gurbani | cs.LG | cs.LG, cs.CL, cs.SD, eess.AS | Automatic Speech Recognition (ASR) systems are used in the financial domain
to enhance the caller experience by enabling natural language understanding and
facilitating efficient and intuitive interactions. Increasing use of ASR
systems requires that such systems exhibit very low error rates. The
predominant ASR models to collect numeric data are large, general-purpose
commercial models -- Google Speech-to-text (STT), or Amazon Transcribe -- or
open source (OpenAI's Whisper). Such ASR models are trained on hundreds of
thousands of hours of audio data and require considerable resources to run.
Despite recent progress large speech recognition models, we highlight the
potential of smaller, specialized "micro" models. Such light models can be
trained perform well on number recognition specific tasks, competing with
general models like Whisper or Google STT while using less than 80 minutes of
training time and occupying at least an order of less memory resources. Also,
unlike larger speech recognition models, micro-models are trained on carefully
selected and curated datasets, which makes them highly accurate, agile, and
easy to retrain, while using low compute resources. We present our work on
creating micro models for multi-digit number recognition that handle diverse
speaking styles reflecting real-world pronunciation patterns. Our work
contributes to domain-specific ASR models, improving digit recognition
accuracy, and privacy of data. An added advantage, their low resource
consumption allows them to be hosted on-premise, keeping private data local
instead uploading to an external cloud. Our results indicate that our
micro-model makes less errors than the best-of-breed commercial or open-source
ASRs in recognizing digits (1.8% error rate of our best micro-model versus 5.8%
error rate of Whisper), and has a low memory footprint (0.66 GB VRAM for our
model versus 11 GB VRAM for Whisper). | Digits micro-model for accurate and secure transactions: Automatic Speech Recognition (ASR) systems are used in the financial domain
to enhance the caller experience by enabling natural language understanding and
facilitating efficient and intuitive interactions. Increasing use of ASR
systems requires that such systems exhibit very low error rates. The
predominant ASR models to collect numeric data are large, general-purpose
commercial models -- Google Speech-to-text (STT), or Amazon Transcribe -- or
open source (OpenAI's Whisper). Such ASR models are trained on hundreds of
thousands of hours of audio data and require considerable resources to run.
Despite recent progress large speech recognition models, we highlight the
potential of smaller, specialized "micro" models. Such light models can be
trained perform well on number recognition specific tasks, competing with
general models like Whisper or Google STT while using less than 80 minutes of
training time and occupying at least an order of less memory resources. Also,
unlike larger speech recognition models, micro-models are trained on carefully
selected and curated datasets, which makes them highly accurate, agile, and
easy to retrain, while using low compute resources. We present our work on
creating micro models for multi-digit number recognition that handle diverse
speaking styles reflecting real-world pronunciation patterns. Our work
contributes to domain-specific ASR models, improving digit recognition
accuracy, and privacy of data. An added advantage, their low resource
consumption allows them to be hosted on-premise, keeping private data local
instead uploading to an external cloud. Our results indicate that our
micro-model makes less errors than the best-of-breed commercial or open-source
ASRs in recognizing digits (1.8% error rate of our best micro-model versus 5.8%
error rate of Whisper), and has a low memory footprint (0.66 GB VRAM for our
model versus 11 GB VRAM for Whisper). | [
-0.0192773826,
0.0677958876,
-0.0184211377,
0.0583912097,
0.0077522476,
-0.0097391643,
0.1009753495,
0.0003446039,
-0.0221332833,
0.0109538678,
-0.0434986539,
-0.0331923403,
0.0475062132,
0.0288935974,
0.0251441039,
-0.0297876112,
0.0085393908,
-0.0648949966,
0.0211931281,
-0.0377339795,
0.0068692267,
0.00438037,
-0.0139290653,
0.026668584,
0.0038382926,
-0.0143112065,
-0.0104557164,
-0.0205866713,
0.0273635201,
0.0201362763,
-0.0166429132,
-0.0075614005,
0.0418705232,
0.0874756873,
0.0000021851,
-0.0204203464,
-0.0174783003,
0.0185630023,
-0.0927870795,
0.0449883379,
0.0169783533,
-0.0059695961,
-0.0055128825,
0.0170096122,
0.0072965184,
-0.0262144189,
0.0344236679,
0.0604494959,
0.0523873419,
0.0218308214,
-0.0145096704,
-0.0388354324,
0.0069137784,
-0.0345439799,
-0.016413616,
-0.0278169569,
-0.0077146175,
-0.0170045272,
-0.0618408844,
0.0031379489,
0.0236907061,
0.0128955385,
0.0139863826,
0.0047817021,
-0.0075128512,
0.006307655,
-0.0260613672,
-0.0149913887,
-0.0514909923,
0.0521788634,
-0.0271001663,
0.0137004806,
0.0462110788,
-0.0075125913,
-0.0191489346,
-0.0055299192,
-0.019631993,
0.0046035619,
-0.0036270211,
-0.0398396589,
-0.0503385179,
0.0319086984,
0.0376969129,
-0.0107848216,
0.0248377789,
-0.0504326597,
0.0107198814,
-0.0124298809,
-0.038855914,
-0.0853212774,
-0.0049680751,
0.0016392953,
0.0589297637,
-0.0240788553,
0.0628532171,
-0.0155291492,
-0.0602525622,
-0.0531656221,
-0.0042019901,
-0.0338475853,
0.0414111428,
0.0512101501,
-0.0380766541,
0.0009116655,
-0.0181853231,
0.0492131189,
-0.0359722339,
-0.056826137,
0.0298173521,
0.0318021812,
0.0452531278,
0.0092409188,
-0.0442004353,
0.0563068762,
-0.0412022881,
-0.0443457514,
-0.0641050115,
-0.0115655195,
-0.0066270963,
0.0258845147,
0.0174903981,
0.0067888699,
0.0001731603,
-0.0621782914,
0.0468968265,
-0.0040515787,
-0.0239631757,
0.0380803384,
-0.0026906452,
-0.0633409843,
-0.0100488476,
-0.0046200864,
0.0212852545,
0.0014910122,
0.0095929001,
0.0708281547,
0.0510835797,
-0.0150412517,
-0.0474092811,
-0.0405885056,
-0.0120359799,
-0.0198143087,
-0.0313307866,
-0.0440863334,
-0.0043745036,
-0.0062881657,
0.0159572028,
-0.0010819935,
0.0295356121,
-0.0093187029,
0.0112561341,
0.0869256109,
0.002229837,
-0.0124720642,
0.0386212915,
-0.0171736274,
0.020401787,
0.0059303949,
0.0606399253,
-0.0273000989,
0.0288343951,
-0.0501485318,
0.0228527039,
0.0312822498,
0.0276228581,
-0.0055080685,
0.0070509389,
0.0039774938,
0.0251814555,
0.0133167403,
0.0402954668,
0.0581062809,
-0.0348686837,
-0.0615512356,
0.1047009975,
0.0929967314,
0.0136923548,
0.0286650546,
-0.0234401356,
0.0115598813,
0.0206674561,
-0.0730610639,
-0.0168580003,
0.0176955722,
-0.0189412106,
0.0245790165,
0.0613181964,
0.0061214883,
-0.0266571362,
0.0036653411,
-0.009476088,
-0.0540988371,
-0.0422785133,
0.0698216483,
0.005979626,
-0.0058920598,
-0.0218201354,
0.0681638196,
-0.0073110666,
-0.0666938052,
-0.0359101444,
-0.0664638281,
0.0328262448,
0.0860306099,
0.0244490542,
-0.0063876836,
0.0391010828,
-0.0608010255,
-0.0646926388,
0.0260322802,
0.011466871,
0.0003363903,
0.0394988507,
-0.0707391575,
0.0262574535,
0.0502658896,
-0.0083582513,
0.0323446617,
0.0078795785,
0.0246488247,
0.0291750152,
-0.0143277179,
0.013804798,
-0.0229181331,
0.0018285722,
-0.0733558014,
0.0685738102,
0.0191034954,
0.0359428637,
0.0213897098,
-0.007215715,
-0.0162926819,
0.0201794896,
0.0107023055,
0.0429131575,
0.0018235783,
-0.0558212139,
0.0541943163,
0.0551173054,
0.0175403897,
0.0302030481,
0.0376270413,
-0.0143648889,
-0.0533713847,
0.0081000244,
-0.0197181068,
0.0063133226,
-0.0556998476,
0.0067114015,
0.0219873488,
0.0117341001,
0.0091408575,
-0.038336698,
0.0103481719,
0.0028200066,
-0.0312470589,
-0.0018628074,
0.0279774237,
-0.0666634142,
-0.0081130927,
-0.0218628999,
-0.0243374892,
0.0665044188,
0.0150685627,
-0.0205096416,
0.0037043313,
0.0005768132,
-0.0694316253,
-0.046099063,
0.0069031003,
0.0276666936,
0.0012822017,
0.0478127524,
-0.0034550244,
-0.0559406616,
-0.0172069985,
-0.0129399328,
-0.0302813333,
-0.0274231713,
0.0037968452,
-0.0188639443,
-0.0266922191,
-0.0156121207,
-0.01890276,
0.0237655435,
0.0171842761,
-0.028151799,
-0.0123881018,
-0.0383209325,
-0.0111564128,
0.0551449209,
0.0292041954,
-0.0175401978,
-0.0140586039,
-0.0128915133,
0.0367943905,
-0.0051120464,
0.029382972,
-0.0269920584,
-0.0038354029,
-0.0319792107,
0.0043465546,
-0.0257159881,
0.0042447047,
0.0187708754,
0.0190422069,
-0.0296482891,
0.0309059564,
-0.046419166,
0.0370126069,
0.001713387,
0.0134968469,
-0.0336018912,
0.0088568619,
0.037770208,
-0.0167533066,
-0.0199658908,
0.0215141233,
-0.0128406836,
0.0203287974,
0.0022545834,
-0.0219204463,
0.0173729528,
-0.0435378142,
-0.0096570747,
0.0144906687,
-0.0229638703,
-0.018793257,
0.0403435938,
-0.030046571,
0.0190162826,
0.0190999117,
-0.0328984559,
0.0073567489,
0.0210305788,
0.0436352715,
-0.0355977304,
0.0542262681,
-0.0044963821,
-0.00207502,
-0.03296341,
-0.0636970028,
-0.0072948723,
-0.046387732,
-0.0738427639,
0.0356777199,
0.0134728905,
-0.0713815764,
-0.0384722948,
0.0460424908,
0.0608300865,
0.0818011761,
0.0094482591,
0.0087977061,
-0.0102523165,
0.0384598039,
-0.0116486885,
-0.0088297753,
-0.0096923131,
-0.0256064776,
0.0616915599,
-0.0020567283,
0.0822641477,
0.0240939278,
-0.0530798994,
-0.0294584148,
0.0421438329,
-0.0401239991,
-0.0172828715,
0.035549365,
0.0441618487,
-0.0699603409,
0.0018332452,
0.0641139671,
0.0003228007,
-0.0023631477,
-0.0253726579,
0.0016432843,
0.0108544128,
-0.0131728379,
-0.0162014589,
-0.0731161982,
-0.015062267,
0.056606736,
0.0487743281,
0.0071947705,
0.0190870743,
-0.0168828368,
-0.0926398411,
0.0336513445,
0.0283398218,
-0.0088761784,
-0.0298784245,
-0.0182322897,
-0.0124288369,
-0.0517091192,
0.0711154342,
0.011586179,
-0.032490287,
-0.0441676825,
0.0305801295,
-0.0489389226,
-0.0090692258,
0.0383064151,
-0.038340617,
0.0344340801,
0.0368951112,
-0.0101174191,
0.0092458455,
-0.0465062261,
-0.0325634032,
0.0258774068,
-0.0277531296,
0.0297990236,
0.0210211221,
0.0106197698,
-0.0065739946,
0.0207718294,
0.0554996058,
-0.0051738238,
0.0067837709,
0.0194855947,
0.008082022,
-0.0191870425,
0.022179516,
0.002584218,
0.0050802436,
-0.0416017994,
0.0403102823,
0.0004146335,
-0.0440091491,
-0.0170376971,
-0.0392509624,
0.0050700693,
-0.0240715109,
0.0024573491,
-0.0092300549,
0.0768448263,
-0.0864169449,
-0.0023144209,
-0.0141469175,
-0.0078119384,
0.0170514882,
0.0368719324,
-0.0317722633,
0.0947868228,
0.0492903031,
-0.0329457261,
-0.0111694429,
-0.0864283815,
0.0395366848,
-0.0689108521,
0.055967968,
0.008911659,
-0.0390698574,
-0.0332424194,
0.0275666565,
-0.1055278108,
-0.0296904482,
0.0243214648,
0.0259525646,
0.025610555,
0.0009786113,
-0.0399676599,
-0.0630553216,
0.0605163313,
0.0317851715,
-0.1096122712,
-0.0100577623,
0.0767171234,
-0.0210306663,
-0.0122777913,
0.0004805372,
-0.016831167,
0.0238851123,
0.013219716,
-0.057101272,
-0.0809238926,
-0.0478976816,
0.005221955,
0.0333599895,
0.0190075766,
-0.0192612428,
0.0382980481,
-0.0230957847,
0.0580335632,
0.0057776007,
0.0330554806,
-0.0014759898,
-0.052266635,
0.0077688284,
0.0224161744,
0.0062474832,
-0.0459341034,
0.0389232039,
-0.0319087915,
-0.0324026421,
0.0266650114,
0.0132583277,
0.0488581508,
0.0006135239,
-0.0447694361,
-0.0099041369,
0.0836942866,
0.0066599031,
0.0008670961,
-0.0042901896,
-0.0140092932,
-0.0118823731,
0.0125744101,
0.0008759794,
0.0171211455,
-0.0168835521,
-0.0604916066,
-0.0082231173,
0.0322263688,
0.0088339858,
0.0137402499,
-0.0387485474,
-0.0406345949,
0.018708095,
-0.0586457625,
-0.0126315709,
-0.0020467716,
-0.0024843642,
-0.0285247006,
0.0390948914,
-0.0171655323,
0.024152616,
0.0374117754,
-0.0091142254,
0.013134921,
0.0029128706,
-0.0088530155,
-0.0484330393,
-0.0200023949,
-0.0082112849,
0.0467276014,
-0.0398534536,
-0.0245453846,
-0.0418691859,
-0.0365357287,
0.0531860813,
-0.0176348761,
0.0267532319,
-0.0385207757,
0.0370079763,
-0.001523905,
-0.0487136133,
0.0184487924,
0.0130191278,
-0.0178599935,
-0.0032728103,
-0.0380403735,
-0.0343371779,
-5.909974173e-33,
0.0134561416,
0.0023501106,
0.0123865716,
0.0208626706,
0.0562743731,
0.0313859358,
0.0116144223,
0.0075711384,
0.0125910351,
-0.0286216792,
-0.0496017709,
0.0041553429,
0.043359898,
0.0343328305,
0.0010261111,
0.0259947348,
0.1264082491,
-0.0395398252,
0.0266372189,
-0.0045323758,
0.0303856879,
0.0378073528,
0.0314930342,
-0.0576431043,
0.001573664,
0.0106104473,
-0.0586231723,
-0.065658249,
-0.0153949643,
0.0329197496,
-0.0345836543,
0.0175848454,
-0.0436095074,
-0.0634837002,
-0.0003967783,
0.0382070802,
-0.0757451132,
0.028414024,
-0.0240450222,
0.035005372,
0.0356966257,
-0.0504145436,
0.0129509512,
-0.0320236757,
-0.0516731627,
0.0065387911,
-0.0159673803,
-0.0106958849,
-0.0114163477,
-0.0651036352,
0.0314338729,
0.0339381658,
0.0109902611,
-0.0145321134,
0.1085107028,
-0.0098577486,
-0.0248388089,
-0.028330924,
-0.0851866826,
0.0527854078,
0.0348551422,
0.0014655147,
-0.0093607772,
0.0061427769,
-0.0183068495,
0.0017401672,
0.072633706,
-0.0520416088,
0.0237633027,
-0.0278649498,
0.0153267197,
0.0351005048,
0.0611061007,
0.0217238199,
0.0714209005,
-0.0438915454,
-0.031947162,
-0.0227623936,
0.0310452189,
0.013007679,
0.0594730228,
0.002920307,
-0.0567668043,
0.0243132245,
-0.0361321718,
-0.0461163372,
-0.0523771793,
0.0334897377,
0.016490303,
0.034454301,
0.0149924466,
0.0257892869,
0.02534426,
-0.0597018301,
-0.0056223366,
-0.0176265854,
0.0492488407,
0.0120563805,
-0.0075802002,
-0.0011575965,
-0.0070749368,
-0.0260475501,
0.0076679387,
-0.0097715557,
0.02066979,
-0.0206111837,
0.0237851031,
0.0206644554,
-0.0939488411,
-0.0127607677,
0.0027663587,
0.0324822925,
0.0196301229,
0.0316662639,
0.021437237,
-0.0198113509,
0.0178306997,
-0.0037043551,
-0.0135210464,
-0.0286711585,
0.0050468743,
0.0132415788,
-0.0318369083,
0.0502078794,
0.0082465643,
-0.0061784894,
-0.0256242473,
0.0159522984,
0.0276058968,
0.0241226573,
-0.0018103381,
-0.0023563341,
2.911e-7,
-0.0827344507,
0.0367067382,
-0.0159005802,
-0.0067192372,
0.0209177174,
0.070292443,
0.0349411927,
0.0335983634,
-0.0296646729,
-0.0419505127,
-0.003117719,
-0.0078882333,
0.1082728207,
0.0249240734,
-0.0671870783,
-0.0470579118,
-0.0178238284,
0.0261066053,
-0.0052572447,
0.0321590789,
-0.0545037463,
0.1020068601,
0.0232588705,
-0.0354357325,
0.0360226147,
-0.0591688938,
0.0177142937,
-0.0193386953,
0.0049076006,
-0.0420892164,
0.0130251786,
0.0552735738,
-0.0087290797,
0.0288462024,
0.0019735717,
0.0124704288,
-0.0283715017,
0.0908795297,
-0.0094879977,
-0.0097400397,
-0.0189687647,
0.0112460712,
-0.0390241705,
-0.0051909685,
0.0042621251,
0.0113148009,
-0.0273668673,
-0.0240352191,
-0.0106523475,
-0.0183088388,
0.0019650187,
-0.0031964367,
0.02625563,
0.0106564006,
0.0283018611,
-0.047625605,
-0.0042078993,
-0.0186507236,
0.02943388,
0.0127513045,
-0.0382575504,
-0.003549058,
-0.0422167815,
-0.0145234838,
0.0319127962,
-0.001793082,
0.035050109,
2.361958681e-34,
0.0363283642,
-0.0670060217,
-0.010667867,
0.064764902,
0.0055047227,
-0.0197222363,
0.0476193912,
-0.0055419258,
0.0488157794,
-0.0524369925,
-0.0479292236
] |
http://arxiv.org/abs/2402.01920v1 | http://arxiv.org/pdf/2402.01920v1 | Preference Poisoning Attacks on Reward Model Learning | 2024-02-02T21:45:24 | 2024-02-02T21:45:24 | Junlin Wu, Jiongxiao Wang, Chaowei Xiao, Chenguang Wang, Ning Zhang, Yevgeniy Vorobeychik | cs.LG | cs.LG, cs.AI, cs.CL | Learning utility, or reward, models from pairwise comparisons is a
fundamental component in a number of application domains. These approaches
inherently entail collecting preference information from people, with feedback
often provided anonymously. Since preferences are subjective, there is no gold
standard to compare against; yet, reliance of high-impact systems on preference
learning creates a strong motivation for malicious actors to skew data
collected in this fashion to their ends. We investigate the nature and extent
of this vulnerability systematically by considering a threat model in which an
attacker can flip a small subset of preference comparisons with the goal of
either promoting or demoting a target outcome. First, we propose two classes of
algorithmic approaches for these attacks: a principled gradient-based
framework, and several variants of rank-by-distance methods. Next, we
demonstrate the efficacy of best attacks in both these classes in successfully
achieving malicious goals on datasets from three diverse domains: autonomous
control, recommendation system, and textual prompt-response preference
learning. We find that the best attacks are often highly successful, achieving
in the most extreme case 100% success rate with only 0.3% of the data poisoned.
However, which attack is best can vary significantly across domains,
demonstrating the value of our comprehensive vulnerability analysis that
involves several classes of attack algorithms. In addition, we observe that the
simpler and more scalable rank-by-distance approaches are often competitive
with the best, and on occasion significantly outperform gradient-based methods.
Finally, we show that several state-of-the-art defenses against other classes
of poisoning attacks exhibit, at best, limited efficacy in our setting. | Preference Poisoning Attacks on Reward Model Learning: Learning utility, or reward, models from pairwise comparisons is a
fundamental component in a number of application domains. These approaches
inherently entail collecting preference information from people, with feedback
often provided anonymously. Since preferences are subjective, there is no gold
standard to compare against; yet, reliance of high-impact systems on preference
learning creates a strong motivation for malicious actors to skew data
collected in this fashion to their ends. We investigate the nature and extent
of this vulnerability systematically by considering a threat model in which an
attacker can flip a small subset of preference comparisons with the goal of
either promoting or demoting a target outcome. First, we propose two classes of
algorithmic approaches for these attacks: a principled gradient-based
framework, and several variants of rank-by-distance methods. Next, we
demonstrate the efficacy of best attacks in both these classes in successfully
achieving malicious goals on datasets from three diverse domains: autonomous
control, recommendation system, and textual prompt-response preference
learning. We find that the best attacks are often highly successful, achieving
in the most extreme case 100% success rate with only 0.3% of the data poisoned.
However, which attack is best can vary significantly across domains,
demonstrating the value of our comprehensive vulnerability analysis that
involves several classes of attack algorithms. In addition, we observe that the
simpler and more scalable rank-by-distance approaches are often competitive
with the best, and on occasion significantly outperform gradient-based methods.
Finally, we show that several state-of-the-art defenses against other classes
of poisoning attacks exhibit, at best, limited efficacy in our setting. | [
0.0207368229,
0.0747973844,
-0.0076842471,
0.0405692682,
-0.0283909105,
-0.0042518978,
0.0054536057,
-0.0194822978,
0.025246013,
0.003350503,
-0.0365317985,
0.0168640781,
0.0206555407,
-0.0000797356,
0.0584575459,
-0.0136009092,
0.0279458221,
-0.0047411332,
-0.0148755405,
-0.0330947414,
0.0277041923,
-0.0244050957,
0.0256156269,
-0.0043271971,
0.0255956687,
-0.0233006813,
-0.0111755356,
0.0090675754,
0.0503192581,
-0.0365559347,
0.0201009139,
0.0211646445,
-0.0022269394,
0.0186726581,
0.0000023995,
-0.0204072706,
-0.0667761117,
0.007491874,
-0.0164593309,
-0.0020471632,
0.0205152649,
0.029536156,
0.013538071,
0.0447366461,
-0.0280571561,
0.0251505692,
0.012620097,
0.0472424068,
0.0484823287,
0.0630948693,
0.0015071635,
0.0112450849,
0.0193393566,
-0.060326308,
0.018056592,
-0.0118266875,
0.0187668186,
-0.0838233829,
-0.0183674749,
0.0186003614,
0.019079376,
-0.0252959989,
-0.0143227298,
0.0246167686,
0.0061217048,
0.0398417562,
0.0309950113,
0.0163380317,
-0.0194156747,
0.0495250039,
-0.0507656671,
0.0328549892,
0.0280555822,
0.0058631399,
0.0321757309,
0.0025444855,
0.0134243043,
-0.000471639,
-0.0114883622,
-0.0171323605,
0.0050236573,
0.0283983573,
0.0491059721,
0.0306961164,
-0.011460009,
0.0459523946,
-0.0117876036,
-0.0417949595,
0.0136407949,
-0.028807845,
-0.026559107,
-0.0443895273,
0.0071321907,
-0.0328757912,
-0.035813123,
-0.0095348982,
-0.0202300157,
-0.0115751028,
0.0048306487,
-0.0420029685,
0.0796239302,
0.0107277511,
-0.0045052473,
0.0143840872,
0.0217718836,
0.022731781,
-0.0308996048,
0.0437287986,
-0.0020162773,
0.0757252127,
0.0324085727,
0.0082409587,
0.0538134053,
0.0557105877,
-0.0134868743,
0.0336311832,
-0.0300852954,
0.0402477756,
-0.0870112479,
0.0635236427,
-0.0350717492,
0.0604586601,
0.0060452935,
0.0494098552,
-0.0512933545,
-0.0310575329,
-0.0132599603,
0.0063233753,
0.0171296038,
0.0326492228,
-0.0309364256,
-0.0242051315,
-0.0436379462,
-0.0261414275,
0.0080782995,
0.1220578775,
0.023480786,
-0.0296730641,
-0.0151989097,
-0.0889267996,
-0.0240224935,
-0.0031923938,
-0.045266334,
-0.0664607435,
-0.0001593593,
-0.0155433267,
-0.0136727039,
-0.0486036129,
-0.0023718867,
-0.0083862357,
-0.0600580573,
0.0832051709,
0.0721661896,
-0.0266667232,
0.0001716907,
0.0298101567,
-0.0802385136,
0.006486645,
-0.0031995615,
0.1189531684,
0.0361464582,
0.0011250589,
0.0204466619,
-0.0298977382,
0.0274015795,
0.008048703,
0.0095067201,
0.0363842249,
0.0157359093,
-0.00706496,
-0.0295289457,
0.0361086093,
-0.0381541923,
-0.0665052161,
0.1175346449,
0.0181848686,
-0.0466264039,
-0.0155217107,
-0.0122847082,
-0.0147171197,
-0.0502959862,
-0.0015928836,
-0.0680811778,
0.0117568821,
-0.0529846102,
-0.0459580533,
0.0148651982,
0.0003292983,
-0.0153157236,
-0.0221906491,
0.0354449823,
-0.0372394398,
0.0241324045,
0.0531568304,
-0.0065449616,
0.031614054,
-0.0057190838,
-0.0655915141,
0.0284853354,
-0.0089061093,
-0.002838335,
-0.0350141861,
0.002444834,
0.0675458312,
0.0622291155,
-0.034346763,
0.0292995702,
-0.0472521,
-0.0487029366,
-0.0488028154,
0.0301257242,
-0.0273850262,
0.0111699626,
-0.0053647864,
-0.0117184436,
0.0054103774,
0.0141643807,
0.0480970219,
0.0051959148,
-0.0246417914,
-0.0346262306,
0.0070690094,
-0.0024580567,
-0.0100847427,
0.0148842959,
-0.0497799553,
0.0814837143,
0.0523004867,
0.0461150669,
-0.0053033512,
0.014887156,
0.0204894934,
0.0205014423,
0.0262322798,
-0.0246656165,
-0.0148703037,
0.0118852211,
0.0064634802,
-0.0020629806,
-0.0193342641,
-0.0072495821,
0.097755298,
-0.0265580211,
0.0132878264,
0.02938799,
0.0056269243,
-0.0220164191,
0.0167981014,
-0.0046083611,
0.0224382058,
0.0132852979,
-0.0620284677,
0.0555120781,
-0.0197901744,
-0.0118455179,
-0.0348297954,
0.0032442431,
-0.0085955625,
-0.0509371981,
-0.0227296017,
-0.0249621756,
-0.0324581787,
0.0017521152,
0.0080902195,
-0.0258743856,
0.002875264,
-0.0179671813,
-0.0282034855,
-0.0036950684,
-0.0121868523,
0.0191369336,
-0.0193555839,
0.0073067769,
0.0188916214,
-0.0488246083,
-0.0136224385,
-0.0464330241,
-0.0407236367,
0.043572288,
0.0014431253,
0.0210707281,
0.0212377552,
-0.0167066138,
-0.0193533152,
-0.0199331939,
0.0549515858,
-0.0146838306,
0.0652046725,
0.0096615935,
-0.0068206247,
-0.0362788066,
0.0226381719,
-0.0238116086,
0.009837782,
0.0212520454,
0.0100932559,
0.0091401506,
0.0996614918,
0.0638548285,
0.0836685821,
0.0681115985,
0.0473477319,
-0.0535190739,
-0.0176326782,
-0.0500577986,
0.0016247228,
-0.0336424336,
0.0008326108,
0.0053271544,
0.0404557101,
0.0391309746,
-0.0154192979,
-0.0366207846,
0.0568570942,
0.0101094935,
-0.0384106636,
-0.0444836132,
0.034678787,
0.0021398382,
-0.0222944338,
-0.0222980119,
-0.0624592155,
0.106666334,
0.0076703974,
0.0591583364,
-0.0259192884,
-0.0362412259,
0.0355682559,
-0.0241250265,
0.0237982478,
0.0134669077,
0.0180958472,
0.0281431247,
-0.041802533,
-0.0160514675,
0.0816647336,
-0.0403604582,
-0.0105250319,
0.0093855122,
-0.0225603394,
-0.0007514649,
-0.0934446976,
-0.0410452262,
-0.1025107205,
0.0214182567,
-0.0250249244,
0.0246941075,
-0.0313334465,
0.010014833,
-0.0637006685,
0.0397777408,
-0.0011853533,
-0.0348160304,
-0.0365791209,
-0.020341184,
0.0110811563,
0.0426790379,
-0.0704539195,
0.0200816076,
-0.0091943359,
-0.0009406016,
0.0888109952,
0.0601010881,
-0.0039607855,
-0.0284010787,
0.0194692202,
0.0061074365,
-0.02527133,
0.0122329686,
0.0216287132,
-0.0139675094,
0.0135047305,
0.0020977673,
0.0223058369,
0.0242756791,
-0.086559996,
0.0289688036,
0.0442461558,
0.0343659297,
0.0043510124,
0.0080663124,
-0.0844662488,
-0.0006559779,
0.0026901055,
0.0946250856,
-0.0513835512,
-0.0355729088,
0.0051381518,
-0.0882770121,
-0.0018790542,
-0.0126458406,
-0.0103643443,
0.0162483584,
-0.0739900619,
-0.0115450155,
-0.0217519347,
-0.0198511947,
-0.0285243075,
0.0030473601,
-0.0010432254,
0.0045467857,
0.0242441203,
-0.0028906548,
0.0247781593,
-0.0450766087,
0.0255369488,
0.0954466909,
0.0087401308,
-0.0225435998,
0.0053552133,
-0.0302389748,
0.0016441371,
-0.0304353647,
0.0100910543,
0.070533745,
-0.0281298906,
-0.001825552,
0.0038817581,
0.0535336174,
0.0058253766,
0.0022546358,
0.0133719053,
0.0246385187,
-0.0041496684,
-0.0168998241,
-0.0096230851,
-0.0336803682,
0.0610305779,
0.0387910344,
-0.0327440947,
-0.0710463673,
0.0066982033,
-0.0830167904,
-0.0045105796,
-0.0060699238,
-0.0475415923,
-0.0214243196,
0.0151238292,
0.0024774093,
-0.0707479417,
0.0038695945,
-0.0313706324,
0.0037119274,
0.005837454,
-0.0246903542,
-0.0142627293,
0.0112860668,
-0.0434676558,
-0.0060320371,
0.0291135721,
0.0374285392,
-0.0524869412,
0.0041109449,
-0.0105198789,
-0.0432347432,
-0.0395205989,
-0.0384951718,
-0.0145335011,
-0.0620119162,
-0.0195247289,
0.0987021253,
0.0134842461,
0.0519319363,
-0.0375468805,
-0.030497076,
0.0503812283,
0.0485421047,
0.0125482799,
0.015487575,
0.0862051696,
-0.0105934469,
0.0004887601,
0.01608699,
-0.0314958431,
0.0168151502,
0.0785849318,
-0.0548483357,
-0.0254295208,
0.0552251451,
0.0099605024,
-0.0003594302,
0.0806489363,
0.0016187179,
0.0136901811,
0.0219151601,
0.0279421322,
-0.0022364152,
-0.0339019671,
0.0049848193,
-0.0291125793,
0.0124139627,
0.0299131554,
0.0017163288,
-0.0429758616,
0.0696963295,
-0.0186165497,
-0.0492844507,
-0.0131330322,
0.0285580568,
0.0255729612,
0.0498426482,
0.0181628987,
-0.0119463783,
0.0083651347,
0.0322789736,
0.0486934148,
-0.0077129002,
0.0063112145,
0.0196895264,
0.0043420689,
-0.0110858176,
0.0516212843,
-0.0387170836,
-0.0368667804,
0.0135873705,
-0.038838625,
-0.0228900816,
-0.0169306919,
-0.0077698785,
0.0053854901,
0.0135112591,
-0.0475185439,
-0.0450025573,
-0.0354246087,
-0.0163688529,
0.0001319014,
0.0196799543,
0.0193687603,
0.0437971652,
0.0600613579,
0.0057096309,
0.0361130126,
-0.0317123719,
-0.0403943621,
-0.0068173213,
-0.0051912959,
0.0005368579,
0.0389771238,
-0.0283234268,
-0.0674325749,
-0.0373000056,
-0.0563511811,
-0.0160047989,
0.0043370081,
0.0409792028,
-0.0089836568,
-0.0009868675,
0.0042447224,
0.0005351143,
0.0714979991,
-0.0747109279,
0.0095170941,
-0.0412976928,
-0.0801064149,
-0.0039185495,
-6.480785413e-33,
0.0646280572,
-0.008745119,
-0.0470756628,
0.0202695765,
0.028484771,
-0.0195548423,
0.0097077806,
-0.0259827022,
-0.0181875154,
-0.025585195,
-0.0127568366,
-0.0130579453,
-0.0121404594,
0.0201343317,
-0.0329137295,
-0.052532021,
-0.0049673975,
-0.0099257771,
-0.0529242121,
0.0377588198,
0.0907795578,
-0.0119150272,
-0.0087043224,
-0.0880533978,
0.0123217748,
0.0092144469,
-0.0221821181,
-0.002998136,
-0.0816916451,
0.0538308583,
0.0238752775,
0.0106337685,
-0.0200780556,
-0.0315760896,
0.0029610991,
0.0367348716,
-0.0284094885,
-0.0148183899,
-0.0381282717,
-0.0343664661,
-0.0066746352,
-0.04233969,
0.0236572232,
0.021635158,
0.0044922112,
-0.0104558375,
0.0040469943,
0.0091189072,
-0.0221312735,
0.0070291897,
0.0257093851,
-0.002584219,
-0.0377593078,
-0.0107675586,
0.0302233249,
0.0474995486,
0.0196775813,
0.0299837478,
-0.035202492,
0.059705928,
-0.0170837548,
0.027799353,
0.006110684,
-0.0107795466,
0.0110964356,
0.052818872,
0.0215781648,
0.0026038222,
0.0212985724,
0.0012224388,
-0.0389414132,
0.0422369391,
0.0529022627,
0.0476167239,
0.0353767835,
-0.0471051224,
-0.0157896094,
0.040905796,
0.0318859331,
-0.0621400811,
0.0523947142,
-0.0054150769,
-0.0760649368,
-0.0176632274,
-0.0158595331,
-0.0604986027,
-0.0532555543,
-0.0298759881,
0.0390052125,
0.0414469242,
-0.034503717,
-0.0103178108,
-0.0095123574,
-0.0589512698,
-0.0292431694,
0.0212189816,
-0.0117954006,
0.0020666984,
0.0012337654,
-0.0574690178,
-0.0256994348,
0.0319831483,
-0.0152743096,
-0.0690470785,
0.0053188489,
0.0170287509,
0.0329443626,
0.0208166186,
-0.0086587863,
-0.0177933238,
-0.0158815812,
0.032900162,
-0.004535771,
0.0306681953,
0.0146885943,
-0.00200468,
-0.0056541613,
0.018092962,
0.0105665512,
0.0285849832,
-0.0735350177,
0.0050933389,
-0.0437095352,
-0.0096377153,
0.0269123837,
0.0059014452,
-0.0102335718,
0.0406444296,
-0.0040528798,
-0.0436659269,
0.0076294476,
-0.0518882982,
3.108e-7,
-0.0393636003,
0.0162990335,
0.0053382949,
-0.0023850577,
0.0213206429,
0.1097561121,
0.0237875767,
0.0923370421,
0.0200521015,
-0.0057467436,
-0.0181763414,
-0.0154304337,
0.0130892759,
0.0256088171,
-0.0474953987,
-0.0043442668,
-0.0194334593,
0.0146447131,
-0.0353196263,
-0.0433077663,
0.0266361367,
0.0510904118,
0.0543223284,
-0.0034984408,
-0.0059793158,
0.0495834313,
0.023746971,
-0.0497964621,
-0.0016566837,
-0.0575791486,
0.0249793343,
-0.0306723341,
-0.0148484968,
0.0154971266,
-0.0565881729,
-0.0178869702,
-0.0512574837,
0.0079237893,
0.0384023264,
-0.0610642433,
-0.0208317097,
0.0034774877,
0.0367999151,
-0.032758899,
0.0614098385,
-0.0075720879,
-0.0067787855,
-0.0092986403,
-0.023988625,
-0.0248770043,
0.038626235,
-0.0013933334,
-0.0246068053,
-0.0249650385,
-0.0062608966,
0.0441530161,
0.0477556624,
-0.0835155919,
-0.0122022917,
-0.0004500694,
0.0112938154,
-0.0396596082,
0.0145459808,
0.0932234749,
-0.0055991854,
-0.0096427239,
-0.0273230225,
3.12504714e-34,
-0.022228945,
-0.0123033114,
-0.0019887534,
-0.0408907644,
0.0275916308,
-0.0028702002,
0.0072300415,
-0.0062714783,
0.0643299669,
-0.0123400996,
-0.020501703
] |
http://arxiv.org/abs/2402.01917v1 | http://arxiv.org/pdf/2402.01917v1 | Whispering in Norwegian: Navigating Orthographic and Dialectic Challenges | 2024-02-02T21:38:12 | 2024-02-02T21:38:12 | Per E Kummervold, Javier de la Rosa, Freddy Wetjen, Rolv-Arild Braaten, Per Erik Solberg | cs.CL | cs.CL | This article introduces NB-Whisper, an adaptation of OpenAI's Whisper,
specifically fine-tuned for Norwegian language Automatic Speech Recognition
(ASR). We highlight its key contributions and summarise the results achieved in
converting spoken Norwegian into written forms and translating other languages
into Norwegian. We show that we are able to improve the Norwegian Bokm{\aa}l
transcription by OpenAI Whisper Large-v3 from a WER of 10.4 to 6.6 on the
Fleurs Dataset and from 6.8 to 2.2 on the NST dataset. | Whispering in Norwegian: Navigating Orthographic and Dialectic Challenges: This article introduces NB-Whisper, an adaptation of OpenAI's Whisper,
specifically fine-tuned for Norwegian language Automatic Speech Recognition
(ASR). We highlight its key contributions and summarise the results achieved in
converting spoken Norwegian into written forms and translating other languages
into Norwegian. We show that we are able to improve the Norwegian Bokm{\aa}l
transcription by OpenAI Whisper Large-v3 from a WER of 10.4 to 6.6 on the
Fleurs Dataset and from 6.8 to 2.2 on the NST dataset. | [
0.0045909695,
0.075997062,
0.0003104131,
0.0540458784,
-0.0365151875,
-0.0258038826,
0.0253455918,
0.0115664341,
-0.0209578611,
-0.009776921,
-0.0465621948,
-0.0232626814,
0.0643258765,
-0.0068477979,
0.0346306749,
-0.0364520624,
0.0341832899,
-0.0904507339,
0.0055041867,
-0.0159949232,
-0.0241622515,
0.0200800709,
-0.059664838,
0.057875637,
0.017390348,
0.003592313,
-0.0151617387,
0.0081769144,
0.0202181209,
0.0281852167,
-0.0116305966,
-0.0294411741,
-0.0151297469,
0.0822433829,
0.0000019845,
-0.0120990425,
0.0184775684,
0.0039466475,
-0.0406043231,
0.0569703914,
0.0776642188,
0.0235103667,
-0.0159311332,
0.0106410831,
0.021473702,
-0.0306945425,
0.0681825429,
0.0128568709,
0.0176462959,
0.0394001268,
-0.0000043634,
-0.009821875,
0.0274345856,
-0.0100776413,
-0.057542298,
-0.0234850477,
-0.0069821202,
-0.0229877904,
-0.0313374363,
-0.0158113167,
-0.0053139087,
0.0842038468,
-0.0234487578,
-0.0144294947,
-0.016710097,
0.0129891802,
-0.0461334772,
-0.0257590916,
-0.0109270765,
0.0647351742,
-0.0024085061,
-0.0163333919,
0.0024877007,
-0.003928212,
-0.0462839007,
-0.0915009081,
0.0184310265,
-0.0297411811,
0.0206538104,
-0.0173263606,
0.0293038245,
0.0248922016,
-0.0053731529,
-0.0241680704,
0.020648092,
0.0027756449,
-0.006592277,
-0.0079444377,
0.0083018132,
-0.0323907472,
0.0349860489,
0.0117767034,
0.0040287892,
0.017826682,
-0.0234069452,
-0.0191796347,
0.0132067921,
-0.0605756529,
0.030012859,
0.0059712944,
0.0540673919,
0.0261879116,
-0.0941157416,
-0.0120421425,
-0.0024038868,
0.0209278595,
0.0021632258,
0.0037423326,
-0.0260986257,
-0.0113437036,
0.0434555784,
0.0166791063,
-0.0593756028,
-0.0405270271,
-0.0477879532,
-0.03919686,
-0.0165513717,
-0.0532020256,
0.0245169085,
0.0036665604,
0.0440008007,
-0.04531404,
-0.0705563575,
0.0044279988,
0.0237355828,
0.0349682122,
0.0598186031,
0.0520074032,
-0.0530992225,
-0.0198764037,
0.0283175968,
0.0316984579,
0.0441897847,
0.0265797116,
-0.0327438675,
0.0248681232,
-0.0212417226,
0.0012945209,
-0.009424231,
0.03487074,
-0.0204130486,
-0.0588097014,
-0.0426592492,
-0.0318523198,
0.0183774717,
-0.0310592484,
-0.0112637598,
-0.058794871,
0.0080810655,
-0.0340610854,
0.0184972845,
0.0473381467,
-0.0270937178,
-0.0078858249,
0.0528132506,
-0.0055609443,
-0.0131701417,
0.0297030378,
0.0537232235,
-0.0089599462,
0.0102021694,
-0.0421046689,
-0.0030912082,
0.0941366851,
0.0394132808,
0.0144503051,
0.0002767872,
0.0056912955,
0.0018846109,
-0.0053551365,
0.041588828,
0.0781416222,
0.0308235157,
-0.0280412249,
0.0332905054,
0.0623441003,
0.0542567335,
0.0286591556,
0.0357071236,
0.0505216233,
-0.0250050332,
-0.0357930996,
-0.0040365658,
0.0160615332,
0.0240441915,
-0.0050237887,
0.0448948145,
0.0278069451,
0.0231496282,
-0.0069614667,
0.0002813388,
-0.0022927141,
-0.0154192019,
0.0549609885,
-0.0046229502,
0.0146416016,
-0.0529266186,
0.0162110422,
0.037686266,
-0.0809157267,
0.0402718894,
-0.0210923497,
0.0178182516,
-0.007861062,
0.0027671854,
0.0088446932,
-0.0097418595,
-0.0530971661,
-0.0302330367,
0.0234355982,
0.0305009633,
-0.0136485081,
0.0598026328,
-0.0552019328,
0.0519845188,
0.0514554344,
-0.0231884141,
0.0733074173,
-0.0197474714,
0.0044461582,
-0.0175586138,
-0.0599707104,
-0.0508308262,
-0.0169024821,
-0.0334213451,
-0.014961591,
0.0275461879,
0.0162385404,
0.0690248683,
0.0253191385,
-0.031242229,
0.0268297382,
0.0233091917,
-0.052720543,
0.0009974069,
0.0202267691,
0.0090690926,
-0.0075456616,
-0.0013638652,
0.0478792861,
-0.0136068817,
-0.0139351701,
0.0227563251,
-0.0595400222,
0.0769669041,
-0.0142096803,
0.0487007871,
0.0071783797,
0.0008904519,
0.008655725,
0.0027970742,
0.0096086785,
-0.0338219106,
0.0129618188,
-0.0348153077,
-0.0148167787,
0.0110830823,
0.0098566739,
-0.0718536079,
-0.0210965667,
0.0592374764,
-0.0350494571,
0.0281598754,
-0.0220797136,
-0.0116753606,
0.0279385112,
-0.0180850048,
-0.0451640002,
-0.0089355074,
-0.0380273387,
0.0272879712,
-0.0163800512,
0.0669295713,
-0.0052469196,
-0.0193671975,
-0.0506383255,
0.0471598655,
-0.0147687374,
-0.0415889211,
-0.0547091775,
0.0016356017,
-0.0355738252,
-0.0177894272,
0.0095424028,
0.0068426421,
-0.0190284438,
-0.0380492434,
-0.0395311825,
-0.0911331326,
0.0741731972,
-0.0326826349,
0.0319159329,
-0.0057655619,
-0.0564624481,
-0.0088177593,
-0.0226294659,
0.0810580552,
0.0583415776,
-0.0645392463,
-0.0325597599,
0.0031066143,
-0.006902332,
-0.0603498966,
0.0090673259,
0.0028857461,
0.0148432544,
-0.0054682302,
0.0211671218,
-0.0190776419,
0.0051810211,
0.0594403669,
0.0238483734,
-0.0069789551,
-0.032235533,
-0.0218086801,
-0.025615098,
-0.019955894,
-0.0205808356,
-0.0064412975,
0.0641041026,
0.005126819,
0.0066019483,
0.0122219287,
-0.0512205884,
-0.0304285511,
-0.0472601466,
-0.0262553059,
-0.0156134646,
0.0333079621,
-0.0393982269,
0.0146544036,
0.0046917489,
-0.0642955378,
0.0524810404,
-0.0427290797,
0.0067638275,
-0.0131096002,
0.0201882459,
-0.0094378646,
-0.0058544232,
-0.0057370025,
-0.0403719619,
0.0117524276,
-0.0100732204,
-0.0529534854,
0.0205487646,
0.005910872,
-0.0385844372,
-0.018594278,
0.022362968,
-0.039119143,
0.0354997553,
-0.023989195,
0.0622371361,
-0.0420742743,
0.000443063,
0.002328408,
-0.0002735688,
0.0196413361,
0.0504079014,
0.0738450363,
-0.007287466,
0.0456110947,
0.0050918139,
-0.0445124321,
-0.0400326103,
0.0674227625,
0.0071613607,
0.0188685451,
0.0214611758,
0.0852401927,
-0.012564064,
0.011185565,
0.0328930505,
-0.0242950208,
0.0064964858,
-0.0007448242,
0.0169587508,
0.0728881061,
-0.0103624668,
0.0135943191,
-0.0597546883,
-0.0023248487,
0.0158421583,
0.0643343925,
0.0240551699,
0.0071891122,
-0.0215566549,
-0.0579378493,
0.0724019632,
0.0526426286,
-0.0391793065,
0.0010146126,
-0.0422560945,
-0.0266432557,
-0.0817698315,
0.0192257985,
0.032578215,
0.0251290072,
-0.047061082,
0.011433769,
-0.0243371893,
0.0052668359,
0.0131390383,
-0.0280666277,
0.0823885649,
0.0296618491,
0.024902897,
-0.0078648152,
-0.0886326879,
0.0599125586,
0.0159296691,
-0.0194378961,
-0.0196476039,
0.0283739995,
0.0314086415,
-0.0159717556,
-0.0390282609,
0.06304387,
-0.0055411044,
-0.0020511334,
0.0347950682,
0.0148203839,
-0.0327627473,
0.0137877045,
0.048590932,
0.0523368493,
0.0521848463,
-0.0185715966,
0.036735937,
-0.0191581585,
-0.0016086786,
-0.0124916313,
0.001321017,
-0.0087172342,
0.0108853299,
-0.0159548037,
0.0149868,
-0.0035462056,
-0.0017925622,
-0.0120771248,
0.0034317949,
0.0028978705,
0.0553481318,
0.0064540259,
0.0578256287,
0.0205603372,
-0.0148352804,
-0.0488346852,
-0.0153806051,
0.0081271091,
-0.0217884425,
-0.0157988947,
0.0739143416,
-0.011181823,
0.037370909,
-0.0079396479,
-0.1055241823,
-0.0677118748,
0.0307342466,
0.0237646159,
0.0129215168,
0.0539656021,
-0.0190442409,
-0.0211938843,
0.0027194857,
-0.0229958221,
-0.0492087938,
0.0037315865,
0.0263402313,
0.0257818289,
0.0247587245,
0.0409151651,
0.0126079284,
-0.000763279,
-0.0081376228,
-0.0283419862,
0.0089612715,
-0.0699521303,
-0.049548775,
0.0149244796,
0.0235151909,
-0.0335351489,
-0.0178960487,
-0.0744136572,
0.0016430276,
0.0171868037,
0.0212515052,
-0.0092955306,
0.0264258031,
0.0369217135,
0.031287808,
-0.0270036887,
0.0203623883,
-0.0662198067,
0.0136667434,
-0.0563623868,
0.0667410642,
0.0381374694,
0.0352182314,
0.0373525918,
0.0324206911,
-0.0487207025,
0.1041861475,
-0.0087111713,
0.0482680239,
0.0110023096,
-0.0152309323,
-0.02731934,
0.0251937564,
0.0017538255,
0.0025449812,
-0.0500287563,
-0.0216939487,
0.0155626964,
-0.0025259315,
0.0478920899,
0.0290552974,
0.0854365826,
-0.0146380775,
0.0285403226,
0.0216898341,
0.0448876582,
-0.0603919663,
-0.0225444641,
-0.0143557116,
-0.0078266812,
-0.0033208283,
0.0491300188,
-0.0148359481,
-0.0210485477,
0.0413106121,
0.066423282,
0.0391899124,
-0.0504348353,
-0.0563206114,
0.0535635911,
0.0257984325,
-0.0678909495,
0.0021796315,
0.0123465713,
-0.0375339091,
-0.0148991514,
0.0003581598,
0.0020886417,
-0.0050159087,
-0.0371683724,
-0.0023235842,
0.0107368715,
-0.0328469165,
-0.0512328893,
-0.041536022,
-0.0079211658,
-0.064984329,
0.033754725,
-6.298780684e-33,
-0.0100362143,
0.0151846381,
-0.0020061168,
0.0758177489,
-0.0115782851,
-0.0063623502,
-0.0090890015,
0.0042442996,
0.0047529927,
-0.0092663765,
-0.0192793757,
-0.0501352958,
0.0267010108,
-0.0091827214,
-0.0327403955,
-0.0038375859,
0.0505193397,
-0.0051581943,
0.0152832521,
0.0178560317,
0.0176677089,
0.0306757838,
0.0487068892,
-0.0617292225,
0.0391240679,
0.0146870567,
-0.0509083718,
-0.0066767843,
0.0137918284,
-0.0047138822,
-0.0153229441,
0.0361989215,
-0.0312324967,
0.0502781086,
-0.0426640697,
-0.0061283167,
-0.0719918013,
0.0458586924,
0.0257861577,
0.0778590962,
-0.0189797767,
-0.0267723873,
-0.0054144505,
-0.0428740382,
-0.0857818574,
-0.0036010381,
-0.0181749668,
-0.0059061362,
-0.0053918711,
-0.0090887221,
-0.0328611694,
0.0015086643,
-0.0314427726,
-0.0646251813,
0.0657931268,
0.0283299685,
-0.011330192,
-0.1465186626,
-0.1205214486,
0.0547972992,
0.0089105945,
0.0420668088,
-0.0089141196,
-0.0069561382,
-0.0823529512,
-0.000136956,
0.0669183508,
-0.0326744281,
-0.0022425153,
-0.0159053262,
0.0637897477,
0.0067571821,
0.0195647497,
0.012670517,
0.0326719396,
-0.0587899685,
-0.0096737267,
-0.0211476255,
0.0218049455,
-0.0032995627,
0.0361085311,
0.0309338644,
-0.0053909617,
-0.0246790126,
-0.0477966517,
-0.1136285514,
-0.0288287662,
0.0026666725,
0.0347752981,
-0.023088228,
0.0067429324,
-0.0152109638,
0.0936534405,
-0.0033322575,
0.0068154661,
-0.030122634,
0.0201960225,
-0.0200048462,
-0.0446778759,
0.0164760202,
-0.0045883525,
-0.0262869317,
0.035036236,
0.0516527742,
0.0315898582,
-0.0142351156,
0.0280931238,
0.0227764416,
-0.028019581,
0.0219173525,
-0.0143919969,
0.0392030664,
0.030136222,
0.0067173205,
0.0178211536,
-0.0453177504,
0.0283803977,
0.0333037712,
-0.0072062328,
0.017675044,
-0.0255760811,
0.0127871446,
-0.0152177205,
0.0110853165,
-0.0972824171,
-0.0024624646,
-0.0133916764,
-0.0360655561,
-0.0015385342,
0.0061449558,
0.0047136755,
-0.010787325,
2.638e-7,
-0.0231804363,
0.0230479222,
0.0297110043,
-0.0464534946,
0.029207889,
-0.0205433145,
0.0070731742,
0.0179792531,
0.00619571,
0.0597323366,
0.0123191066,
-0.0088235438,
0.0747278258,
0.022785563,
-0.0823948383,
-0.0492143594,
-0.0149003388,
0.0030215627,
-0.0649045855,
-0.0034218836,
0.0076079397,
0.0707663,
-0.0429601781,
-0.0161198787,
-0.0201157704,
0.0130428094,
-0.0123209395,
-0.0187028497,
0.0105729457,
-0.0033590973,
-0.003120438,
0.010152244,
0.0212153047,
-0.026868701,
-0.0199191831,
0.0042329822,
0.001985047,
0.0396956727,
0.0146377282,
-0.0603579916,
0.0231124703,
0.0616303794,
-0.0065877638,
-0.0454755165,
-0.0145661738,
0.0124644805,
-0.0695810243,
-0.0099369399,
0.0400039516,
0.0537271686,
0.0169286169,
-0.0226620808,
0.0546087623,
0.0224464685,
0.0348139629,
-0.0290194452,
-0.040509399,
0.00682142,
-0.0715751275,
-0.0002183864,
-0.0054539894,
-0.014969782,
0.0111225992,
0.0123299742,
-0.0298638027,
-0.0206856392,
0.0034285537,
1.990900114e-34,
0.0174043141,
-0.0158082712,
0.0157178529,
0.0274169892,
0.036634177,
-0.0268832054,
0.0443164371,
0.0468713865,
-0.0113103781,
-0.0292484798,
-0.0708543211
] |
http://arxiv.org/abs/2402.01916v1 | http://arxiv.org/pdf/2402.01916v1 | CoLe and LYS at BioASQ MESINESP8 Task: similarity based descriptor assignment in Spanish | 2024-02-02T21:36:03 | 2024-02-02T21:36:03 | Francisco J. Ribadas-Pena, Shuyuan Cao, Elmurod Kuriyozov | cs.IR | cs.IR, cs.CL, cs.LG, 68P20, 68T50, H.3.3; I.2.7 | In this paper, we describe our participation in the MESINESP Task of the
BioASQ biomedical semantic indexing challenge. The participating system follows
an approach based solely on conventional information retrieval tools. We have
evaluated various alternatives for extracting index terms from IBECS/LILACS
documents in order to be stored in an Apache Lucene index. Those indexed
representations are queried using the contents of the article to be annotated
and a ranked list of candidate labels is created from the retrieved documents.
We also have evaluated a sort of limited Label Powerset approach which creates
meta-labels joining pairs of DeCS labels with high co-occurrence scores, and an
alternative method based on label profile matching. Results obtained in
official runs seem to confirm the suitability of this approach for languages
like Spanish. | CoLe and LYS at BioASQ MESINESP8 Task: similarity based descriptor assignment in Spanish: In this paper, we describe our participation in the MESINESP Task of the
BioASQ biomedical semantic indexing challenge. The participating system follows
an approach based solely on conventional information retrieval tools. We have
evaluated various alternatives for extracting index terms from IBECS/LILACS
documents in order to be stored in an Apache Lucene index. Those indexed
representations are queried using the contents of the article to be annotated
and a ranked list of candidate labels is created from the retrieved documents.
We also have evaluated a sort of limited Label Powerset approach which creates
meta-labels joining pairs of DeCS labels with high co-occurrence scores, and an
alternative method based on label profile matching. Results obtained in
official runs seem to confirm the suitability of this approach for languages
like Spanish. | [
-0.0046707341,
0.0148998918,
-0.0121174995,
-0.0390500836,
-0.0257444959,
0.0328777991,
-0.0093495008,
0.0355358236,
0.0017561428,
-0.0501422472,
-0.0257091764,
0.0292087197,
-0.0327189416,
0.010213946,
-0.0009246151,
0.0051821722,
0.0646296293,
0.0526542626,
-0.0166621543,
-0.0187021494,
0.0006282871,
-0.0138994735,
-0.0178754274,
0.008350214,
0.0178460423,
0.0392877646,
0.0362468325,
-0.0419540815,
0.0425144136,
-0.0810257494,
0.0099340761,
0.0242940467,
-0.0337669775,
0.0359980762,
0.0000016919,
-0.0769627467,
-0.0754659846,
0.0451175906,
0.0338773429,
-0.0405750014,
0.0411979295,
-0.0131524745,
0.019783983,
-0.006135304,
0.0078749666,
-0.0524744987,
0.069011502,
-0.044334393,
0.0212209728,
0.0075638695,
-0.0230670273,
-0.0486810543,
-0.0063453983,
0.0114555955,
0.0401847139,
-0.0429912545,
0.0170544423,
-0.0326467641,
0.0720721111,
0.0073844926,
0.0528923236,
0.0180361755,
-0.0498734079,
-0.0011253481,
0.0249338765,
0.0265485886,
-0.0258008484,
-0.0291154198,
0.0425873883,
0.0087331431,
0.0276095886,
-0.0600608513,
0.0098977117,
0.0235353,
-0.0394450687,
0.0279167742,
-0.0162351876,
-0.0253494028,
-0.0111366324,
-0.0103790564,
0.0645451918,
0.0746486485,
0.0083304644,
0.0171556305,
-0.0009367439,
0.0284666363,
0.0355206057,
-0.0369165652,
0.0407269448,
-0.0118737761,
0.0362980478,
-0.0099673234,
0.0457331724,
0.0253845789,
-0.0019369279,
-0.0062746028,
0.0435840338,
-0.0608730204,
0.0828484744,
-0.0752187669,
-0.0644612834,
-0.0077434606,
-0.014194713,
0.0251484569,
-0.0171014331,
0.0177614372,
0.0104972459,
0.0584243238,
-0.0376313366,
-0.0459389426,
-0.0801052451,
-0.0102569079,
-0.0473847724,
0.0488773622,
0.0392937548,
0.0165381096,
-0.0638798848,
-0.0435055234,
-0.0367309488,
-0.0248394236,
0.0089621274,
0.0159722529,
-0.0207422823,
0.0439098403,
-0.0465435572,
0.023992369,
-0.0160270426,
-0.0241842251,
-0.0585140623,
-0.0386901796,
0.0334238745,
0.0221057776,
0.0117061306,
-0.0522346422,
0.0445033871,
0.0380781479,
0.0069408333,
-0.0497906767,
0.023909932,
-0.0281427372,
-0.0119472109,
-0.0128651308,
-0.0023783199,
-0.0660492405,
0.0478627719,
-0.0033265313,
0.0074302913,
-0.0721901879,
0.024116952,
0.0117593482,
0.0092614461,
0.0742405504,
-0.0577371009,
-0.0271740444,
0.0189745855,
0.0132097704,
0.0159769878,
0.0132748811,
-0.0189166646,
0.0006039264,
0.0179765429,
-0.0164121948,
-0.0172183327,
-0.0201342124,
-0.0135702677,
-0.019732086,
0.0152667221,
0.0094880126,
0.0622296743,
0.1023257524,
0.0074752239,
0.0425688699,
-0.0120581724,
-0.0079061724,
-0.0066991355,
0.0956098735,
-0.0041472632,
-0.0062011029,
-0.0131196762,
-0.0208233818,
0.0093170926,
-0.0897829235,
0.011963171,
-0.0206776932,
0.037181437,
0.0129621476,
0.0045954259,
0.0299816784,
-0.0081388419,
-0.0016629497,
-0.0109702721,
0.0410392024,
-0.035221938,
-0.0228710398,
0.0356205106,
-0.026050102,
0.0348197632,
0.0265958235,
0.0056377654,
-0.0234864503,
0.0024755171,
-0.0138423545,
-0.0463529006,
0.0426384062,
0.0381519683,
-0.0449230261,
-0.0574628264,
0.0398537032,
-0.0283545591,
0.0074200812,
-0.0150797758,
-0.0448599495,
-0.0129518798,
-0.0279480461,
-0.0129117053,
0.0562132448,
-0.0257180873,
0.0136254542,
-0.0501363501,
0.0622336455,
-0.0234780703,
-0.0563745275,
-0.0203447826,
0.0398537889,
-0.0234231651,
-0.0347961336,
-0.003499232,
-0.0064017167,
0.0360585079,
0.0372936167,
0.0173845254,
0.0200863983,
0.0674611628,
-0.0521818139,
0.0029302568,
-0.0017436135,
0.0380175784,
0.0219195895,
0.0586614981,
-0.0642159209,
0.0042634201,
0.0020055161,
0.0238919389,
-0.018854361,
-0.0046530562,
-0.0128030786,
0.011970887,
-0.0260554906,
0.0575137734,
-0.0284653232,
-0.0158978496,
-0.0596275665,
0.022844173,
0.0200323965,
0.051767949,
-0.0183525085,
-0.0811698362,
0.026778996,
-0.0619629249,
-0.002032998,
0.0438482799,
0.0495487712,
-0.0139481761,
0.0227570347,
0.0284111444,
0.0034163101,
0.0178825911,
-0.0611800514,
0.0118946237,
0.038784761,
0.0268827304,
0.0436337516,
-0.0216122214,
0.0157739501,
-0.0042114868,
-0.0472073257,
-0.0404403433,
0.0067498768,
0.0042131711,
0.0242373198,
0.0227460843,
-0.0002014185,
-0.0195853822,
0.0033648724,
-0.0431642234,
0.0142871868,
0.005991668,
-0.0656495839,
0.0171024688,
-0.0130830156,
-0.0420001671,
0.0031014201,
0.02809163,
-0.0202785581,
-0.0082378434,
0.036026217,
-0.0459088422,
0.0025728438,
0.0137845799,
0.0056397733,
0.0824380741,
0.0343377627,
-0.0234699883,
-0.0082586575,
-0.0395956486,
0.0318459719,
-0.0020285328,
-0.0751614124,
0.0647013336,
0.0134038981,
-0.0405455567,
-0.0157218929,
-0.0104422811,
-0.0190095939,
-0.0067812605,
-0.0466121919,
-0.0562607609,
0.0482299477,
-0.0064167813,
0.0524060391,
0.0065094861,
-0.0442637876,
0.0510020293,
-0.0244500823,
0.0224545822,
-0.0963429213,
-0.0063584899,
-0.0477722436,
0.0488407835,
-0.0057620504,
0.0350493416,
-0.0046968549,
0.0072446587,
0.0361182615,
0.0257340427,
0.002133644,
-0.0589445271,
-0.0114254626,
-0.0149793169,
0.0289996378,
0.0058066365,
-0.051203087,
-0.0029500122,
0.0517183617,
-0.0334882289,
0.0127250524,
0.0556474403,
0.0089572845,
-0.0554471537,
0.0661736727,
0.0430997275,
0.0067477943,
0.0056958362,
-0.0061853505,
0.0065789605,
0.0024589375,
0.0194670986,
-0.0359477736,
0.0086901849,
0.0170153193,
-0.0126515348,
0.0298013352,
-0.0152909467,
-0.009630356,
0.0246512592,
0.019219812,
0.0486132167,
-0.0344585329,
-0.0059067099,
-0.0321622789,
-0.0116221365,
0.0080002882,
-0.0392176062,
0.0508838668,
0.0713453814,
-0.0935581997,
-0.0170650538,
-0.0121073471,
-0.0203299113,
0.0403237566,
0.0510137416,
-0.0539057851,
0.0054851109,
0.0177250411,
0.0289645139,
-0.0220126584,
-0.020476291,
-0.0147729078,
-0.0551304482,
0.0248604044,
0.0077806446,
-0.0297379345,
-0.0184360612,
-0.0733159557,
-0.0001535386,
-0.0410831682,
-0.0465045869,
0.0013710387,
0.0040662447,
-0.0044105775,
0.0452031232,
-0.0037406599,
-0.041461803,
-0.0142350281,
-0.023951523,
-0.0309984721,
0.0141681824,
0.0496334881,
-0.0001990961,
-0.0561979078,
0.0001177651,
-0.0469194837,
-0.0484723747,
-0.0196357127,
-0.0010474628,
-0.0016442121,
-0.0083907731,
-0.0314317197,
-0.0223530903,
-0.0399121381,
-0.010991673,
0.0143749751,
0.0957527459,
-0.0320558995,
0.04307504,
-0.0530777089,
-0.0346169174,
-0.0420945697,
0.0164087657,
-0.0030667903,
0.0060705827,
0.0033316216,
0.0284573976,
0.0583200939,
-0.0020841064,
-0.0286944937,
0.027119644,
-0.1038714051,
-0.0390582941,
-0.0388504602,
-0.0311778355,
0.0418664925,
0.0317010581,
0.0497173481,
0.0287640709,
0.0139965871,
-0.0139115257,
-0.0445115715,
0.0251663029,
0.0164659191,
0.0194557756,
-0.0718797892,
-0.076104328,
0.0651011616,
0.0293665472,
-0.0466945469,
-0.0190880727,
0.0428846627,
-0.0364550874,
0.020877935,
-0.0055556074,
0.0167238936,
-0.0151653374,
-0.0232042484,
-0.0437468626,
0.035432782,
0.0602734834,
-0.0157775562,
-0.0288333427,
-0.0402932055,
-0.0521302149,
-0.0295013227,
-0.0206463914,
-0.0097760307,
0.0662938803,
-0.044439543,
0.0708580464,
-0.0012468084,
-0.0744641349,
0.0773240998,
0.0456355102,
0.067888543,
-0.0654543489,
-0.007030454,
-0.0293886308,
-0.0095737828,
0.0055687549,
-0.0063974545,
0.0281171463,
-0.0221151859,
0.0112312781,
0.0291179959,
0.0036494664,
0.0780555308,
-0.0135748656,
-0.0033852684,
0.0176872537,
0.0421421304,
0.0117630865,
0.0817525834,
-0.0432299525,
0.0082939146,
0.015020838,
-0.0109489337,
-0.0087566646,
0.0323582515,
-0.0874451548,
-0.026915716,
0.0365429111,
0.0201499816,
-0.0113502359,
0.0642014667,
-0.0144295301,
0.0353844725,
-0.0263909213,
-0.0367664248,
0.0481283255,
-0.0003684974,
-0.0044318303,
0.0172644779,
-0.042195797,
-0.0121316509,
-0.0015299481,
-0.0333405845,
0.0414879471,
0.0417941324,
0.0274647679,
-0.0108339405,
-0.0391016304,
0.0506732501,
0.0131388614,
0.0186903924,
0.0064221057,
-0.060109403,
0.0234432332,
-0.0028912479,
-0.0029137828,
0.0129190553,
-0.0384683833,
-0.0740185753,
-0.0603829287,
0.0092379618,
0.0932908356,
-0.002665099,
-0.025297543,
0.0273639578,
0.060046073,
0.0274405647,
0.0262621604,
-0.0426056013,
-0.0056775748,
0.0010706731,
-0.04639709,
-0.0222091563,
0.0319021456,
-5.861869271e-33,
-0.0013917132,
-0.056645114,
0.0183128286,
-0.0054888986,
0.0160520766,
0.0003321987,
-0.0027987605,
-0.0407916568,
0.0470141433,
0.0558745861,
0.0083307112,
-0.0491610467,
0.0184024032,
-0.0334791727,
0.0318533108,
0.0327315107,
-0.0053415778,
0.0297809523,
-0.0036240935,
-0.0185208973,
0.0400945097,
0.0270162802,
0.0649737045,
-0.1106944308,
-0.0095391423,
0.0443219505,
0.0358803198,
0.0204241648,
0.0548307672,
0.057621289,
-0.0333842486,
0.0317065232,
0.0022906449,
-0.0060938629,
0.028760856,
-0.0377058424,
-0.0491230488,
-0.0452327617,
-0.0165806003,
-0.0013344232,
0.0268406849,
-0.0505242497,
-0.0100573329,
-0.023751637,
-0.0679178089,
0.0574430898,
0.0459918939,
-0.0077626267,
-0.0485259891,
-0.0472807176,
-0.020674618,
0.0280857347,
0.0579454266,
0.0578576252,
0.0709835291,
0.0023054595,
-0.0529962219,
0.0376033485,
0.031484805,
0.0502007827,
-0.0280783437,
0.0473802239,
0.0090049161,
-0.0065810569,
0.0227565672,
0.051309552,
0.032459829,
-0.073285751,
0.0876911283,
0.0416686237,
-0.0094220378,
0.0025309832,
0.025799213,
0.0699984804,
0.0151081588,
0.0072027706,
-0.0269906167,
0.0639653951,
-0.0488150977,
-0.0050134407,
-0.0390527658,
0.0264485907,
-0.0204359032,
-0.0040326994,
-0.0308909342,
0.0559830852,
-0.0214778315,
-0.0241849273,
0.0108921826,
-0.0364884213,
-0.0095891748,
0.0648575723,
-0.0085278377,
0.0226851106,
-0.0040686224,
-0.0531013198,
0.0562962405,
0.0173224751,
-0.0386274867,
0.0178176537,
-0.038507428,
-0.0592642613,
0.0317268185,
0.0120439399,
0.0600300021,
-0.0343150087,
0.0281921998,
-0.0342469588,
-0.0275071301,
0.011285278,
-0.0014607792,
-0.0369660147,
-0.0155473016,
-0.0047655771,
-0.017099468,
0.0150042269,
0.0456778668,
0.0331614129,
0.0663488582,
0.0258967262,
-0.0045570536,
-0.0529317595,
0.0222603716,
0.0300664529,
-0.0353886671,
0.0282149687,
-0.0590607859,
0.0294112936,
-0.0163837802,
-0.0416849516,
0.0092013963,
-0.016889235,
2.791e-7,
0.0503513403,
0.0427117236,
0.0161641967,
0.0119005321,
0.0137135452,
-0.0310841631,
-0.0109829223,
0.0154756159,
-0.0501680411,
0.0268733017,
0.0396233127,
-0.017203141,
-0.0074354657,
0.0793868452,
-0.0351830348,
-0.0215670783,
-0.0257709194,
-0.0555315018,
-0.0377970785,
0.0113048172,
0.0406614803,
0.046136111,
-0.0132454699,
0.0041615632,
-0.0030380108,
-0.022284925,
0.0178281032,
-0.0399690047,
0.0084471172,
0.0174856316,
0.0066873627,
0.0090339137,
0.0244295206,
0.0178077351,
0.0049635922,
-0.0236992948,
0.0341877528,
-0.0206407439,
-0.0872119218,
-0.0096517839,
0.0254298337,
-0.0006485636,
0.0135081187,
-0.0254360531,
0.0020007493,
-0.0195260346,
-0.0398780964,
-0.0200948212,
-0.0530019514,
0.0254045241,
-0.0053192819,
0.0083808135,
-0.0661547855,
0.044322297,
0.0207695849,
0.0335953347,
-0.0364224054,
-0.0454732366,
0.0369508788,
0.0554639138,
0.000606198,
-0.0196920373,
-0.0214038454,
0.0105589256,
0.016271634,
-0.0365299918,
0.001996181,
2.202215887e-34,
0.0020615638,
-0.0043538124,
-0.0348146372,
0.0056453189,
0.0003087637,
-0.0219885111,
-0.0244458504,
0.0317488536,
-0.0211505238,
-0.0209542885,
0.0109578278
] |
http://arxiv.org/abs/2402.01912v1 | http://arxiv.org/pdf/2402.01912v1 | Natural language guidance of high-fidelity text-to-speech with synthetic annotations | 2024-02-02T21:29:34 | 2024-02-02T21:29:34 | Dan Lyth, Simon King | cs.SD | cs.SD, cs.CL, eess.AS | Text-to-speech models trained on large-scale datasets have demonstrated
impressive in-context learning capabilities and naturalness. However, control
of speaker identity and style in these models typically requires conditioning
on reference speech recordings, limiting creative applications. Alternatively,
natural language prompting of speaker identity and style has demonstrated
promising results and provides an intuitive method of control. However,
reliance on human-labeled descriptions prevents scaling to large datasets. Our
work bridges the gap between these two approaches. We propose a scalable method
for labeling various aspects of speaker identity, style, and recording
conditions. We then apply this method to a 45k hour dataset, which we use to
train a speech language model. Furthermore, we propose simple methods for
increasing audio fidelity, significantly outperforming recent work despite
relying entirely on found data. Our results demonstrate high-fidelity speech
generation in a diverse range of accents, prosodic styles, channel conditions,
and acoustic conditions, all accomplished with a single model and intuitive
natural language conditioning. Audio samples can be heard at
https://text-description-to-speech.com/. | Natural language guidance of high-fidelity text-to-speech with synthetic annotations: Text-to-speech models trained on large-scale datasets have demonstrated
impressive in-context learning capabilities and naturalness. However, control
of speaker identity and style in these models typically requires conditioning
on reference speech recordings, limiting creative applications. Alternatively,
natural language prompting of speaker identity and style has demonstrated
promising results and provides an intuitive method of control. However,
reliance on human-labeled descriptions prevents scaling to large datasets. Our
work bridges the gap between these two approaches. We propose a scalable method
for labeling various aspects of speaker identity, style, and recording
conditions. We then apply this method to a 45k hour dataset, which we use to
train a speech language model. Furthermore, we propose simple methods for
increasing audio fidelity, significantly outperforming recent work despite
relying entirely on found data. Our results demonstrate high-fidelity speech
generation in a diverse range of accents, prosodic styles, channel conditions,
and acoustic conditions, all accomplished with a single model and intuitive
natural language conditioning. Audio samples can be heard at
https://text-description-to-speech.com/. | [
0.0054173344,
0.0926796421,
0.0239351485,
-0.0016495313,
0.0134670287,
-0.0171659887,
0.0474012755,
-0.0261054989,
-0.0401919931,
-0.0300532915,
-0.0604372509,
-0.0449314117,
0.0128363641,
0.048421517,
0.0495172814,
-0.0887011215,
-0.0118581094,
-0.0375326276,
0.0073128343,
-0.0248162076,
0.0162317622,
0.026444573,
-0.0327421911,
0.036061734,
-0.0436108261,
-0.0004236565,
0.0057749604,
0.0347853675,
0.0463218726,
0.0218714885,
-0.0032153032,
0.0559173636,
0.0427614488,
0.0024564215,
0.0000020185,
-0.0199405253,
-0.0119030466,
-0.0102284588,
-0.0353982225,
0.0178568196,
0.0263276584,
0.0350494385,
-0.0428991057,
0.0120795118,
-0.0148709333,
-0.0095232464,
0.0484816357,
0.0340655781,
0.0745527819,
0.0727265924,
0.0051729782,
-0.0820307732,
0.002753909,
-0.0121828178,
0.0127237029,
0.0236189272,
-0.0206963941,
-0.0384149216,
-0.0023258296,
0.0023655945,
0.0349923968,
0.0404298194,
0.0101595968,
-0.0023434758,
0.0687672123,
0.0135535877,
-0.0444720536,
-0.0039670388,
0.0139229866,
0.0777913556,
0.0305676013,
-0.0177896526,
0.0226988625,
0.0400770418,
-0.0186414495,
-0.0528762601,
0.0247645415,
0.0253179315,
0.0019325233,
-0.0048147971,
-0.057711836,
0.0598762445,
0.0237849951,
-0.0360071845,
0.0347093381,
-0.0258527547,
0.0044404007,
-0.0198277608,
0.0007184733,
-0.0183629505,
-0.0445732512,
-0.0198431928,
0.042436488,
-0.0180123262,
-0.0274929721,
-0.0207994282,
-0.0540206842,
-0.028273182,
-0.0431674793,
-0.0237582456,
0.0388653725,
0.0376732014,
-0.0166889895,
-0.0086132148,
0.0059221839,
0.0156824663,
-0.0125202592,
-0.0235466789,
-0.0205022804,
0.0048882538,
-0.0213368814,
0.0056033614,
-0.0362572819,
-0.0251757111,
-0.0269225873,
-0.0026995342,
-0.0451993607,
-0.0249454565,
-0.0478088297,
0.0266603176,
-0.0024819635,
0.0317607746,
-0.074856028,
-0.0142544834,
0.0043289908,
-0.0208257455,
-0.0235292632,
0.023518065,
0.0036446212,
-0.0485814288,
-0.0013559314,
0.0405442677,
0.0354492441,
-0.0023174831,
0.0289401636,
0.0122192418,
-0.0105426554,
0.0039986572,
-0.037573427,
-0.0324886702,
-0.0290889721,
-0.0493851043,
-0.0245341957,
-0.0791670084,
0.028791545,
-0.0317891166,
-0.0127787944,
-0.0460510701,
-0.0092860013,
0.0051876795,
0.0180518795,
0.0392155088,
-0.0135319578,
0.0327905789,
0.0470134951,
-0.0043087061,
0.0075085498,
-0.0053878785,
0.0594536215,
-0.0012496784,
-0.0133138783,
-0.0425842628,
0.0400014669,
0.070054017,
0.032684911,
0.0129014384,
-0.0136297662,
-0.0354862362,
-0.0025516786,
0.060609512,
0.0789058283,
0.0163711887,
0.0025932395,
-0.0157652516,
0.0936448723,
0.0015181918,
0.0144481277,
0.0279399846,
0.0849406645,
0.0567522869,
0.0027439138,
-0.0229405873,
-0.0415156037,
0.0039751595,
0.0361008085,
0.0215991922,
0.0613366589,
0.0173516367,
-0.0362407714,
-0.0123970974,
0.0255545899,
-0.0138574475,
-0.0321512669,
0.0036325788,
-0.0378996804,
0.0497344658,
-0.0032791961,
0.0225262027,
0.0061394614,
-0.0605334267,
0.0301615745,
0.0107962321,
-0.0167031437,
0.0786157176,
-0.0261348598,
0.0127882846,
0.0634476617,
-0.0596715435,
0.0212897565,
0.055605676,
0.0160235763,
-0.0174823739,
0.0500222184,
-0.0553754419,
0.0179499686,
0.0148908151,
-0.001567003,
0.0260967948,
-0.0093464227,
0.0202843901,
0.0149682667,
-0.0429407842,
0.0047837095,
-0.006264382,
-0.0365815014,
-0.0755787417,
0.0354147181,
0.0182812624,
0.0044450001,
0.0567618273,
-0.0501274988,
0.0054479619,
-0.0012000164,
-0.0320013985,
0.0013864598,
0.0470482782,
-0.0804795548,
0.0019022224,
-0.0057458025,
-0.0523374565,
0.0093986746,
0.0089760339,
0.0066529252,
-0.0835636929,
0.0352083445,
-0.0438180938,
0.0256111044,
-0.0140705146,
0.0132785738,
0.0386686996,
0.0065467702,
0.0053385682,
0.026252823,
0.0432362109,
-0.027826868,
0.0204787739,
0.0337458178,
-0.0288322065,
-0.074275136,
0.0152783003,
-0.0356287025,
0.0058983504,
0.0446361825,
-0.0188298989,
-0.0217633229,
-0.0406725034,
0.0029127612,
-0.0436883233,
0.0039407476,
-0.0298686232,
0.0039945133,
0.003163826,
0.0091374004,
-0.019288484,
-0.0297290385,
-0.0942142531,
-0.0780673772,
-0.0086828396,
-0.0377808623,
-0.0165527575,
-0.0070420257,
-0.022800317,
-0.0188617595,
0.011753167,
0.0040264558,
-0.0066791116,
0.0053846901,
0.0114100715,
-0.0370897353,
-0.0113654099,
-0.0066135866,
0.0274464227,
0.0063164094,
0.0208848864,
-0.025402002,
0.0323507972,
0.0642531291,
0.0147873377,
0.0220549684,
-0.0438659973,
0.017734237,
0.0083157737,
-0.0437086746,
0.0116694858,
0.0078231459,
0.054675024,
-0.0141214216,
-0.0241374653,
0.025457371,
0.0013342587,
0.0436046943,
0.0238531195,
-0.0383100286,
0.0204767175,
-0.0006477604,
-0.0196841136,
0.0289987754,
-0.0207194779,
-0.0284936782,
0.0181812439,
-0.0062696058,
-0.0066329986,
0.0573720075,
-0.0520082228,
-0.0349727385,
-0.0098155867,
-0.0589014515,
0.0128221037,
0.0029860958,
-0.0125652337,
0.0476897545,
0.008325276,
-0.0029280637,
-0.0053998618,
-0.0030980487,
0.0589425229,
-0.0484802611,
0.0108215204,
-0.0272220485,
0.0183524042,
-0.004765904,
-0.0633698553,
-0.0393095501,
-0.0606167726,
-0.0314198956,
-0.0143227354,
0.0352344438,
-0.0455155894,
-0.0039955857,
0.0094146822,
0.0224710181,
0.0559731796,
-0.0237064213,
0.0091761593,
-0.030318331,
0.025916148,
0.0296463426,
0.0511166789,
-0.0334980264,
-0.0101304883,
0.0206391402,
-0.0070243063,
0.0273467321,
0.0507834293,
-0.0592233613,
-0.0042851795,
0.0051429337,
0.0112668769,
-0.0376452506,
0.0272917598,
0.0565926656,
-0.0422985107,
-0.0167310163,
0.1059156954,
-0.0058547147,
-0.0230286904,
-0.0031519914,
0.0066136732,
0.0765611976,
0.01069606,
0.0178073794,
-0.075390622,
-0.0339286327,
0.0303638279,
0.0860533342,
-0.0073590172,
-0.0048132394,
-0.0414675362,
-0.0893964916,
0.0515450016,
0.0372011103,
-0.0732966661,
-0.0402320474,
-0.0596992448,
-0.0348410495,
-0.0203348566,
0.0117283603,
0.0190031622,
-0.0051969285,
0.0051645911,
0.0466839224,
-0.0211473443,
-0.0589566529,
0.0250205267,
-0.0631713644,
0.1158649325,
0.0616266392,
0.0379261747,
0.0190674439,
-0.0370774195,
-0.0323493928,
0.0268010143,
-0.0046279924,
0.0198047571,
0.0756595507,
-0.0234876238,
-0.0107084671,
-0.0155747738,
0.106272392,
0.0128995599,
0.02118681,
0.0006432865,
-0.0207852162,
-0.0082328264,
0.0065121758,
0.0257099569,
0.0742976516,
0.0475995876,
0.052020669,
0.0290733501,
-0.0775046945,
0.006959151,
-0.0111897839,
0.0628573224,
-0.069847025,
-0.05920716,
0.0389460884,
-0.0183169432,
-0.0373651572,
-0.0628925562,
-0.0062936372,
-0.0251616146,
0.039930094,
0.0197784379,
-0.0465867817,
-0.0038932285,
-0.0056825443,
-0.0313810594,
0.0538250953,
-0.0652539358,
0.0567307621,
-0.0172314327,
-0.0195839833,
0.0324282609,
-0.0722812191,
0.0021573186,
0.0631245598,
-0.039541129,
-0.0379058868,
0.0211317874,
0.014230066,
0.0152524719,
0.1212079376,
-0.0142584993,
-0.0612634718,
0.0164486598,
0.0337631665,
-0.0046806936,
-0.0105896778,
0.0610523634,
0.0291496515,
-0.0186613798,
0.0165829565,
-0.0323982611,
0.0725485012,
0.0010977633,
-0.0106359785,
-0.0075622327,
-0.0461496264,
0.002436433,
0.0169531479,
0.0203934684,
-0.0320972167,
0.0656887367,
-0.0256661363,
-0.0262688845,
-0.0040510506,
-0.0143799875,
-0.0419325009,
-0.0059228148,
0.0338108353,
0.0488308482,
-0.003514729,
-0.0103856428,
0.0036110713,
0.0041084858,
-0.0265939943,
-0.030251421,
0.0172884669,
-0.0058657001,
0.0057491213,
0.0173740033,
-0.0288347919,
0.0634983331,
0.013214414,
0.0506770611,
0.0046876413,
-0.0052301399,
0.0182058737,
-0.0201409627,
-0.0169299562,
0.0165573154,
-0.0411816202,
-0.0018164632,
-0.0501018167,
0.0141192032,
0.036377117,
-0.0132801477,
0.0723549128,
-0.0109604737,
0.0390545353,
-0.0033543457,
0.017380422,
0.002118527,
-0.0344465151,
0.0095622633,
-0.013638298,
-0.0235559363,
0.0180154257,
0.0250085685,
0.0137765585,
0.0258912612,
-0.0530452356,
0.0244443137,
-0.0413488038,
-0.0106313433,
-0.0535849929,
0.005550595,
-0.0106149409,
-0.0616936497,
-0.0327405035,
-0.0545312501,
-0.0035806648,
0.0070674359,
-0.0134781338,
-0.0218413752,
0.060955856,
-0.036565844,
-0.044936154,
-0.0051942705,
-0.0057135527,
-0.0274876095,
0.0327614024,
-0.0178159215,
0.0098838462,
-6.144481255e-33,
-0.029232379,
0.0134430518,
-0.0020524568,
0.0090428367,
-0.0373498611,
-0.0546441674,
-0.0377182104,
0.0015863887,
-0.0012796576,
0.0028129807,
-0.0437919945,
-0.0130842058,
0.0239748042,
0.0603134036,
-0.038560763,
0.010949987,
0.0927215144,
-0.0061468468,
-0.0354296528,
0.0696656406,
0.0588053912,
-0.0044326056,
0.0304864049,
-0.0861223936,
-0.0012496052,
-0.0337755606,
-0.0306171719,
-0.0010789332,
0.0137207843,
-0.0012898843,
-0.0202084947,
0.0571698435,
-0.0237302221,
-0.064557299,
-0.0287757292,
0.0232216921,
-0.0596950389,
0.0136434026,
-0.0481103398,
0.0408236384,
-0.0050096004,
-0.0223970432,
0.0283594634,
-0.02227626,
-0.0380532704,
0.0111683998,
0.0604018122,
-0.0363228135,
-0.0015372681,
-0.0581453852,
-0.0406281203,
0.0239022132,
-0.027262412,
-0.0719642639,
0.021094894,
0.0758177936,
-0.0177488048,
-0.049993977,
-0.0707858801,
0.048383452,
0.0015505195,
-0.005042349,
0.0354528427,
0.0704590231,
-0.0152492272,
0.0335440226,
0.0433204323,
-0.0064267502,
-0.0094277998,
0.0023371719,
0.0335325934,
0.0165886339,
0.0124204028,
0.0811486319,
0.0150933024,
-0.022953378,
0.0132547757,
0.0227431282,
0.0185420923,
-0.0622064397,
-0.0026879215,
0.0040163565,
-0.02709019,
-0.0154288234,
-0.0811091214,
-0.0461150371,
0.0077359402,
0.0120895859,
0.0297617763,
-0.0081719887,
-0.0354633406,
0.0014333058,
0.0427606665,
-0.0068826447,
-0.0037863564,
-0.0350508541,
-0.0569300652,
0.0234546699,
-0.0325922854,
-0.0048162541,
-0.0153814033,
-0.0217772182,
0.090953134,
0.054766614,
0.0228236206,
-0.0065975771,
-0.0041950536,
0.045111306,
-0.0384044349,
0.0213159975,
-0.0069737155,
0.0548297167,
-0.0015363053,
-0.0331447236,
0.0355971046,
-0.0079480214,
-0.0051512765,
0.044869177,
0.014648525,
0.0249475781,
0.035810329,
-0.0351940654,
0.0042334087,
0.0115574822,
-0.0353428982,
0.0461034589,
0.0029218455,
-0.0627929866,
0.0268075913,
-0.0454406328,
0.0023468784,
-0.029288752,
2.726e-7,
-0.0650834441,
-0.0141309239,
0.0387633219,
0.0008342033,
0.0355010666,
0.0416091233,
0.0214731675,
0.0267391521,
-0.0164675564,
-0.0343235731,
0.0108960578,
-0.006429771,
0.1120272428,
0.0058195377,
-0.0953683779,
-0.0245515667,
-0.0991769731,
0.0338732824,
-0.0412931517,
-0.0241194554,
0.0211501122,
0.0789488852,
0.0089726625,
-0.0434758514,
0.0463070273,
-0.01538203,
0.0358082913,
-0.0065218941,
0.013456461,
-0.0172315463,
0.0389903821,
-0.0089113656,
-0.0338139422,
0.0220330618,
-0.0353631265,
0.0137739265,
0.0201367885,
0.0990368128,
0.0418238305,
-0.0130795026,
-0.0273565557,
0.0378544033,
-0.0215133671,
0.0196960233,
0.037992388,
0.0020063631,
-0.0235000923,
-0.020426631,
-0.0468028784,
0.0064687631,
-0.0244827364,
-0.0140422406,
-0.0203527194,
0.0028038532,
0.0200827774,
0.0186928678,
-0.0121797919,
-0.0223118588,
-0.0323448144,
0.0237553325,
-0.0424983799,
-0.0198636744,
0.0003133649,
0.001105862,
0.0174683519,
-0.0307579096,
-0.0005969335,
2.040195113e-34,
0.0080409758,
0.0048551797,
0.0353670008,
-0.0237682257,
-0.011437959,
0.0085509885,
0.0211406015,
0.0014685787,
0.06069085,
-0.0392482989,
-0.0298910197
] |
http://arxiv.org/abs/2402.01878v2 | http://arxiv.org/pdf/2402.01878v2 | LiPO: Listwise Preference Optimization through Learning-to-Rank | 2024-02-02T20:08:10 | 2024-05-22T18:51:02 | Tianqi Liu, Zhen Qin, Junru Wu, Jiaming Shen, Misha Khalman, Rishabh Joshi, Yao Zhao, Mohammad Saleh, Simon Baumgartner, Jialu Liu, Peter J. Liu, Xuanhui Wang | cs.CL | cs.CL, cs.LG | Aligning language models (LMs) with curated human feedback is critical to
control their behaviors in real-world applications. Several recent policy
optimization methods, such as DPO and SLiC, serve as promising alternatives to
the traditional Reinforcement Learning from Human Feedback (RLHF) approach. In
practice, human feedback often comes in a format of a ranked list over multiple
responses to amortize the cost of reading prompt. Multiple responses can also
be ranked by reward models or AI feedback. There lacks such a thorough study on
directly fitting upon a list of responses. In this work, we formulate the LM
alignment as a \textit{listwise} ranking problem and describe the LiPO
framework, where the policy can potentially learn more effectively from a
ranked list of plausible responses given the prompt. This view draws an
explicit connection to Learning-to-Rank (LTR), where most existing preference
optimization work can be mapped to existing ranking objectives. Following this
connection, we provide an examination of ranking objectives that are not well
studied for LM alignment with DPO and SLiC as special cases when list size is
two. In particular, we highlight a specific method, LiPO-$\lambda$, which
leverages a state-of-the-art \textit{listwise} ranking objective and weights
each preference pair in a more advanced manner. We show that LiPO-$\lambda$ can
outperform DPO variants and SLiC by a clear margin on several preference
alignment tasks with both curated and real rankwise preference data. | LiPO: Listwise Preference Optimization through Learning-to-Rank: Aligning language models (LMs) with curated human feedback is critical to
control their behaviors in real-world applications. Several recent policy
optimization methods, such as DPO and SLiC, serve as promising alternatives to
the traditional Reinforcement Learning from Human Feedback (RLHF) approach. In
practice, human feedback often comes in a format of a ranked list over multiple
responses to amortize the cost of reading prompt. Multiple responses can also
be ranked by reward models or AI feedback. There lacks such a thorough study on
directly fitting upon a list of responses. In this work, we formulate the LM
alignment as a \textit{listwise} ranking problem and describe the LiPO
framework, where the policy can potentially learn more effectively from a
ranked list of plausible responses given the prompt. This view draws an
explicit connection to Learning-to-Rank (LTR), where most existing preference
optimization work can be mapped to existing ranking objectives. Following this
connection, we provide an examination of ranking objectives that are not well
studied for LM alignment with DPO and SLiC as special cases when list size is
two. In particular, we highlight a specific method, LiPO-$\lambda$, which
leverages a state-of-the-art \textit{listwise} ranking objective and weights
each preference pair in a more advanced manner. We show that LiPO-$\lambda$ can
outperform DPO variants and SLiC by a clear margin on several preference
alignment tasks with both curated and real rankwise preference data. | [
0.0531882308,
0.0537143461,
0.004115866,
0.0208240841,
0.003173643,
0.0275976788,
-0.0146578075,
-0.0168117881,
0.0481293425,
-0.023909891,
-0.0437877737,
-0.0353589132,
0.0240782537,
-0.003620751,
0.0076677175,
-0.0602567941,
0.0335837528,
-0.0444404855,
0.0174498204,
-0.0202754997,
0.0185364578,
0.010679272,
-0.0165769346,
0.0289904568,
0.0795024112,
0.0047524967,
-0.0087985136,
-0.0017479215,
0.0767538473,
-0.0273249131,
0.0274480991,
0.0494439974,
-0.0004943777,
0.0286785513,
0.000001915,
-0.0420474075,
-0.028773563,
-0.0263542645,
-0.0158271641,
-0.0192421861,
0.0273703672,
0.0222099107,
-0.0212656539,
0.0273755118,
-0.0248662457,
0.0161193199,
0.0530955046,
-0.0130238133,
0.0579069331,
0.0578827895,
0.0160456263,
-0.0104244007,
0.0089285281,
-0.0465931632,
0.0204376262,
-0.0184959285,
0.0092146667,
-0.0359245539,
0.0201889705,
0.0164090246,
-0.0064925202,
0.0109360367,
-0.0114078308,
-0.0056514335,
0.0083581572,
0.0360769182,
-0.0137531003,
-0.0159898754,
-0.0434322655,
0.0944091529,
-0.0314203389,
-0.0174033046,
0.0075922529,
-0.0204616096,
0.0177608058,
-0.0164216012,
0.0071587614,
-0.0208489243,
0.0098455669,
0.0267292112,
0.0184247103,
0.009964237,
0.0240199398,
0.0230391771,
-0.0290623549,
0.098860167,
0.0009566912,
-0.0592178106,
0.0503572077,
0.0040951381,
-0.0234846584,
-0.0399797037,
0.0162091199,
0.0088766804,
-0.0793621987,
0.0240043644,
-0.0182365105,
0.0067041814,
0.0165308695,
-0.0341347978,
0.0687414631,
0.0543355308,
0.0137534104,
0.0192420725,
0.0079013258,
0.0895937681,
-0.0234169364,
0.0377018824,
-0.0743761435,
0.0492502153,
-0.0083814068,
-0.0260727722,
0.0220884755,
0.0422766246,
-0.0359765142,
-0.0223727468,
-0.0595292076,
0.0277871732,
-0.0387963355,
0.0621705763,
-0.0012430324,
0.005057442,
-0.0411353372,
0.0043160296,
-0.0697430596,
0.0359538905,
-0.0011510997,
0.014429206,
0.0236821044,
0.0300945453,
0.0221582837,
0.0080831684,
-0.0353752635,
-0.0231510419,
0.0009403843,
0.0805254951,
-0.0229254272,
0.0116256047,
0.0253777243,
0.0213589594,
-0.0398278944,
-0.0424186997,
-0.0455388203,
-0.0658118427,
0.0248187892,
-0.0134662185,
-0.0104567092,
-0.0236195996,
0.0097304769,
0.0255844109,
-0.0295071602,
0.0813037902,
0.0605990253,
-0.0234911721,
0.0154299289,
-0.0336310789,
0.0076082502,
0.0281020775,
0.0423727259,
0.0441171862,
0.0077264826,
0.0090302108,
-0.0240349695,
0.0430590212,
0.0206351876,
0.0212927572,
-0.0310666934,
0.0185140315,
0.001460592,
0.0426743366,
0.0457331352,
0.0238360353,
-0.0052797226,
-0.025700504,
0.0651303828,
0.0394642651,
0.0252058841,
-0.0524042286,
0.0006409966,
0.0305604693,
-0.0314471051,
0.0256077722,
-0.0449317247,
0.0332543477,
-0.0325073749,
0.0344259702,
0.0410778858,
0.0166697502,
-0.0088095926,
-0.0654853806,
0.0164782628,
-0.0111039318,
0.0129629197,
0.0524843149,
0.0065933261,
-0.039621491,
-0.0688751787,
-0.0074388292,
0.0235007014,
-0.0184447393,
-0.022479387,
-0.0400253236,
0.0334699973,
0.1147840917,
0.025633147,
-0.0281142984,
-0.0534694977,
-0.0080145393,
-0.0603888705,
0.0443011001,
0.0318694599,
-0.0442681871,
0.009395808,
-0.0327479094,
-0.0468532294,
0.0568367988,
-0.0240755882,
0.0558486357,
-0.0093951449,
-0.0147600053,
-0.0161360689,
0.0073377257,
-0.0519800857,
0.0300434381,
0.0083653769,
-0.0268455856,
0.0224976931,
0.0277350713,
0.0553038456,
0.0040849391,
0.0395473689,
0.0011246047,
0.0066907289,
-0.0035681778,
0.0177589916,
-0.0386975147,
-0.0412207395,
-0.0365614481,
0.0430599526,
-0.0095218606,
-0.0127693638,
0.0992756858,
-0.0562890172,
-0.016975658,
-0.0624416769,
-0.0031474663,
0.002178072,
-0.0226732902,
0.0190825481,
0.0042050793,
-0.0395867191,
-0.0999337882,
0.0577741675,
-0.0197060294,
0.0428553224,
0.0357915089,
0.0060247364,
-0.0201845057,
-0.0969049633,
-0.0269611571,
0.0464098714,
-0.0046337219,
0.0212852713,
0.0512532443,
0.0534989014,
-0.0298998412,
-0.0253974851,
-0.0648082197,
0.0156803075,
-0.0141528426,
-0.0099974135,
0.0066371164,
-0.0078122285,
-0.0062619969,
-0.0301487949,
-0.0366904587,
-0.04040806,
-0.0038025156,
0.018646853,
0.0470772572,
0.0077631189,
-0.0135817397,
-0.0744190365,
-0.0186409932,
-0.0101682404,
-0.0019865371,
-0.0355865806,
0.0169309936,
-0.0615440756,
0.0383331627,
-0.0604257919,
0.0131208878,
-0.0054687727,
0.0021013152,
0.0111929607,
-0.0076933173,
0.0372604392,
0.1075476706,
0.04322768,
0.0215923823,
0.0613338239,
0.0346484371,
-0.0946125835,
0.0077254861,
-0.0554025173,
-0.0088913543,
-0.0398436971,
0.0058889762,
0.0396032184,
0.0361404903,
-0.0009322519,
0.0120107653,
-0.0807369798,
0.0082770428,
0.0014602975,
-0.0326001421,
-0.0697690919,
0.0062327185,
-0.0494573191,
0.0570644997,
-0.0215541646,
-0.0581177659,
0.0953770503,
-0.0066754115,
0.0204703975,
-0.0897710547,
-0.0024922681,
-0.0396093279,
-0.0152646229,
0.026779322,
0.0353014804,
-0.0522972159,
-0.0246532727,
-0.0819750577,
-0.0098776324,
0.0199585054,
-0.0483424664,
0.0455335677,
-0.0321139842,
0.0193021931,
-0.0274019782,
-0.075584963,
-0.0540119782,
-0.0258198287,
0.002113431,
-0.0128344083,
-0.0053021531,
-0.0186512135,
0.0034998534,
-0.0002140997,
0.0283958632,
0.0580966473,
-0.0008417435,
0.0008025464,
-0.0249138866,
0.013972233,
0.0797341391,
-0.0016134867,
0.0042577023,
-0.0088380808,
0.0108551439,
0.0086390534,
0.0751914307,
-0.0014857737,
0.019356804,
0.0332510956,
0.0777708218,
-0.0575648248,
0.0009770364,
0.0343544818,
-0.0118649593,
0.0070810034,
0.0147138573,
0.0833966061,
0.0355069488,
-0.0796940103,
0.0043069622,
-0.0036764562,
0.088168852,
0.0015159347,
0.0594847761,
-0.0433817096,
0.010638956,
0.0184015837,
0.0678887069,
-0.0070652491,
0.0110963741,
-0.015859507,
-0.1304039657,
0.015126205,
0.012438694,
-0.0272822529,
0.001909979,
-0.0236605871,
0.0088145696,
-0.0250345878,
0.0173424613,
0.0219211709,
0.0005334057,
-0.0016473203,
0.0118163638,
0.0071576228,
-0.0173390862,
0.0009595779,
0.00343137,
0.0344797857,
0.1121385098,
0.0427526422,
0.0194794666,
-0.0267692618,
0.0004546695,
0.0035199637,
-0.0560307056,
-0.0202168915,
-0.0404695794,
-0.0286273845,
0.0138209173,
-0.0248887353,
0.0840952024,
-0.0241975076,
0.0017230096,
-0.0156653468,
0.0375008434,
-0.0129170222,
0.0211012457,
-0.0152206421,
0.0025506208,
0.0225730482,
0.0531571843,
-0.0123540601,
-0.0851832032,
-0.0133861769,
-0.0562338792,
-0.0276636779,
-0.0494589843,
-0.054952126,
0.0086690895,
0.0033839447,
0.0126627265,
-0.0696912333,
-0.0375819169,
-0.0109985871,
0.0200766772,
0.0046951575,
0.0061104205,
-0.0150235668,
-0.0260809362,
0.0169526525,
-0.0119759571,
0.0085004512,
0.0270476844,
-0.0901769102,
-0.0151173407,
0.0121834315,
-0.0350954868,
-0.0562381521,
-0.019802127,
-0.0282309502,
-0.0733117759,
-0.0052582794,
0.0799826086,
0.0221726559,
0.0341822132,
0.011259743,
-0.0223921034,
0.0511988774,
0.0391182899,
0.0128648207,
-0.0078281565,
0.0139214313,
0.0472833738,
-0.0130568752,
0.0257267784,
-0.0477720425,
0.0021533505,
0.0041406672,
0.0349490233,
-0.0240728185,
-0.0426319502,
0.0213474333,
-0.002398514,
0.0207030028,
-0.0445543863,
-0.0474956445,
-0.0425888225,
0.0403255112,
0.009021922,
-0.0360361673,
-0.0041934038,
-0.0227645896,
0.0368586518,
-0.0222034026,
0.0096456008,
-0.0432253107,
0.0246521253,
0.023711903,
-0.0415503904,
-0.0023897563,
0.0288858395,
0.0161631331,
0.040728718,
0.0313503779,
-0.0041942196,
0.0464053452,
0.0210405625,
0.08061748,
0.0159580614,
0.0316679925,
0.048146192,
-0.0049086753,
0.0000031094,
0.0518997833,
-0.0453868955,
-0.0483940765,
-0.040781416,
-0.0070159207,
0.0173316263,
-0.0028208382,
0.044073876,
-0.0056786737,
0.0006882354,
-0.0301459022,
-0.0042726109,
-0.0127640991,
-0.0102347191,
0.0136304116,
0.033112634,
0.0115276873,
0.0267254859,
0.0757916197,
0.0390880667,
-0.0193001796,
-0.02434255,
0.0223748926,
-0.0557857789,
-0.0216467846,
-0.0192567948,
0.0411087796,
-0.0513314083,
-0.020087434,
-0.0427017063,
-0.0306465141,
0.0040693646,
0.0165687967,
0.0178213939,
-0.0187108107,
0.0280635245,
0.0209213682,
-0.0506375395,
0.0207628943,
-0.0498641953,
-0.0265736859,
-0.0353799649,
-0.048536215,
-0.0516353585,
-5.919940896e-33,
-0.0565120727,
-0.0388888344,
0.0210388117,
0.0179141574,
-0.0172952786,
-0.0279110391,
0.0207786635,
0.0107997563,
0.0123531576,
-0.0281609111,
-0.0062931734,
-0.0374152176,
0.0193849616,
0.0062505919,
0.0278809648,
-0.0088509293,
0.0116188806,
-0.014493118,
-0.0678008497,
-0.0093548968,
0.0395367928,
0.0061214096,
-0.0105166398,
-0.0314879119,
0.0663732663,
0.0110484222,
-0.0158682819,
0.0152443303,
-0.0140556898,
0.0815994963,
0.0022175401,
0.0098781427,
0.0036522781,
-0.0423676185,
0.0048149643,
0.0563904643,
-0.0544446968,
-0.0234368909,
0.0118747754,
-0.0056421403,
-0.0084982887,
-0.0197908413,
0.0084105767,
0.0068543549,
-0.0368567705,
-0.0102496706,
0.022828266,
-0.0255032107,
-0.0414057933,
0.0161780268,
-0.0452066734,
0.0041021449,
-0.0197321586,
-0.0166341234,
0.0350464322,
0.0960440561,
-0.0141700841,
0.0814580694,
-0.0470202193,
0.0486585498,
-0.0315444395,
0.052680824,
0.0946192294,
0.0303246658,
-0.0354096107,
0.0427584425,
-0.0046683247,
-0.0437712744,
0.0364205725,
-0.0321386941,
0.0097805262,
0.0243191347,
0.013659982,
0.0187871791,
0.000680424,
0.0012736341,
-0.0337914862,
0.0700134635,
0.0125019215,
-0.0137178032,
0.030641418,
0.0392879099,
-0.0538502075,
-0.027504554,
-0.0447344296,
-0.0220179148,
-0.038576819,
-0.0357830822,
0.0419396609,
0.0200985111,
-0.0389663801,
-0.020746991,
0.0447910242,
-0.0271660462,
-0.0385673717,
-0.0115454067,
-0.0175274406,
0.0257343892,
0.009910075,
0.0216315258,
-0.0859688297,
-0.0164484642,
-0.0286007002,
-0.0107427957,
-0.0301585831,
0.0140114222,
0.0103621939,
-0.0136951022,
-0.0422708802,
0.0283157714,
0.0014247658,
-0.0064340606,
-0.0225196909,
0.0732437819,
0.0271250065,
-0.0279464815,
0.010237176,
0.0648084283,
-0.0038203311,
-0.0262696911,
-0.0258115213,
-0.0248901583,
-0.0383448303,
0.0071035055,
-0.0263775736,
0.0183301102,
-0.0071380297,
0.0415461436,
-0.0212040748,
-0.0113316327,
-0.013297827,
0.0124692004,
2.529e-7,
-0.0041071102,
0.0363703854,
0.0143950814,
0.0116785876,
-0.0172823817,
0.0443642177,
-0.004839214,
0.054808449,
0.0112494081,
0.0075191311,
0.0197109133,
-0.0109824389,
0.0310441088,
-0.0127011947,
-0.0836151913,
-0.0138121611,
-0.0649203658,
0.019923877,
-0.0438589826,
-0.0259792618,
0.0250555035,
0.0714086741,
-0.0030424492,
-0.0085320696,
0.0075226123,
0.0579088591,
0.0440067761,
-0.0405712873,
-0.0181781407,
-0.0034966667,
0.0377419367,
-0.0296548381,
-0.0269633997,
0.0066251145,
0.0040902928,
0.0348699875,
0.012336175,
0.0631841719,
0.0055322172,
-0.0780529007,
-0.0179705229,
0.0370058864,
0.0161026083,
-0.0736686736,
0.035483472,
-0.0109560657,
0.0008539052,
-0.0009918716,
-0.0275796745,
-0.015420856,
0.0140644154,
-0.0300732367,
-0.0293236654,
0.0126775783,
0.0217008423,
0.0030900063,
0.0116394348,
-0.0422786325,
-0.0196825154,
-0.0252012871,
-0.0167657416,
-0.0349687189,
-0.0213233177,
0.0341079533,
-0.0185832325,
-0.0476799048,
-0.0341084637,
2.239505919e-34,
0.0034545024,
-0.012799643,
0.0117904032,
0.0119541818,
0.0024761956,
0.0235667042,
-0.0133936675,
0.0356343761,
0.0563076437,
-0.0699888319,
-0.0272013489
] |
http://arxiv.org/abs/2402.01874v1 | http://arxiv.org/pdf/2402.01874v1 | The RL/LLM Taxonomy Tree: Reviewing Synergies Between Reinforcement Learning and Large Language Models | 2024-02-02T20:01:15 | 2024-02-02T20:01:15 | Moschoula Pternea, Prerna Singh, Abir Chakraborty, Yagna Oruganti, Mirco Milletari, Sayli Bapat, Kebei Jiang | cs.CL | cs.CL, cs.AI, cs.LG, cs.RO | In this work, we review research studies that combine Reinforcement Learning
(RL) and Large Language Models (LLMs), two areas that owe their momentum to the
development of deep neural networks. We propose a novel taxonomy of three main
classes based on the way that the two model types interact with each other. The
first class, RL4LLM, includes studies where RL is leveraged to improve the
performance of LLMs on tasks related to Natural Language Processing. L4LLM is
divided into two sub-categories depending on whether RL is used to directly
fine-tune an existing LLM or to improve the prompt of the LLM. In the second
class, LLM4RL, an LLM assists the training of an RL model that performs a task
that is not inherently related to natural language. We further break down
LLM4RL based on the component of the RL training framework that the LLM assists
or replaces, namely reward shaping, goal generation, and policy function.
Finally, in the third class, RL+LLM, an LLM and an RL agent are embedded in a
common planning framework without either of them contributing to training or
fine-tuning of the other. We further branch this class to distinguish between
studies with and without natural language feedback. We use this taxonomy to
explore the motivations behind the synergy of LLMs and RL and explain the
reasons for its success, while pinpointing potential shortcomings and areas
where further research is needed, as well as alternative methodologies that
serve the same goal. | The RL/LLM Taxonomy Tree: Reviewing Synergies Between Reinforcement Learning and Large Language Models: In this work, we review research studies that combine Reinforcement Learning
(RL) and Large Language Models (LLMs), two areas that owe their momentum to the
development of deep neural networks. We propose a novel taxonomy of three main
classes based on the way that the two model types interact with each other. The
first class, RL4LLM, includes studies where RL is leveraged to improve the
performance of LLMs on tasks related to Natural Language Processing. L4LLM is
divided into two sub-categories depending on whether RL is used to directly
fine-tune an existing LLM or to improve the prompt of the LLM. In the second
class, LLM4RL, an LLM assists the training of an RL model that performs a task
that is not inherently related to natural language. We further break down
LLM4RL based on the component of the RL training framework that the LLM assists
or replaces, namely reward shaping, goal generation, and policy function.
Finally, in the third class, RL+LLM, an LLM and an RL agent are embedded in a
common planning framework without either of them contributing to training or
fine-tuning of the other. We further branch this class to distinguish between
studies with and without natural language feedback. We use this taxonomy to
explore the motivations behind the synergy of LLMs and RL and explain the
reasons for its success, while pinpointing potential shortcomings and areas
where further research is needed, as well as alternative methodologies that
serve the same goal. | [
0.088161923,
0.0477959961,
-0.0102302954,
0.0122089144,
-0.075115703,
0.0362131558,
-0.0065155597,
-0.0196479391,
0.0048462413,
-0.0414205417,
-0.038505964,
-0.0122755095,
-0.0134218447,
0.0275242366,
0.0053940271,
-0.0595471002,
0.0444785468,
-0.051820159,
0.0398731716,
0.0058261738,
-0.006273597,
-0.007070912,
-0.0004697234,
0.0495515838,
-0.0039626304,
0.0365277715,
-0.0253689736,
-0.0034056904,
0.0152103631,
-0.0148107465,
0.0249868203,
-0.0018275196,
0.0652505904,
0.0845576972,
0.0000019433,
-0.0690973178,
0.0175119322,
-0.0184118655,
-0.0130317574,
-0.0248285104,
0.0546554215,
-0.0404331125,
-0.0072350376,
0.0001403248,
-0.0438623428,
-0.0229581706,
0.0353571512,
0.0701558217,
0.0199052133,
0.1117284074,
0.0326675065,
-0.0541559272,
0.0365738608,
-0.0454990678,
0.0177089609,
-0.033908654,
0.0385659337,
-0.01548866,
0.0305108689,
0.0115017425,
-0.0333692357,
0.0109506715,
0.0576929674,
-0.023371717,
-0.0326909237,
0.0575963259,
0.0060762558,
-0.0373684652,
-0.019515289,
0.0440362319,
-0.0357443579,
0.0097640883,
0.0238966346,
-0.0177059416,
0.0030493489,
0.0020068337,
-0.0152794654,
0.0612496026,
0.0370403193,
0.0383044854,
0.0233381875,
-0.0376500562,
0.0557536036,
0.0383012332,
-0.0222787652,
0.060544841,
-0.0241789389,
-0.0693964958,
0.0469573624,
0.0158332232,
0.0124187414,
-0.0797015429,
-0.0172413848,
0.0503165349,
-0.0545777157,
0.0381780788,
-0.0145917097,
0.0197106935,
0.020443894,
-0.0161165111,
0.0262145307,
0.0203423817,
0.0129637858,
0.0387537181,
-0.0339567326,
0.0492952578,
-0.022424629,
0.0461301394,
-0.0019297083,
0.0120483311,
-0.0500194281,
-0.0069248686,
-0.0297674537,
0.0431782007,
-0.0236024298,
-0.0307785962,
-0.0453750305,
0.0287115462,
-0.0021178653,
0.0449769162,
0.0089871418,
0.0195869915,
-0.035511326,
-0.0131755061,
-0.0934945121,
0.0196591746,
-0.032178387,
0.0336296447,
0.0193050299,
-0.0260809846,
0.027271688,
0.0050528836,
0.0060025565,
-0.0224835072,
0.0198282339,
0.0803912506,
-0.0041570421,
-0.0223996285,
-0.0412656143,
-0.0209286157,
-0.0179491919,
-0.0680110976,
0.0315042511,
-0.071065411,
0.0126542551,
-0.0501021408,
-0.0142673831,
-0.0056281746,
0.0122055719,
0.0047237449,
-0.0283865482,
0.0646413043,
0.019481089,
-0.0005184129,
0.0167503096,
-0.0107125612,
0.0181129146,
0.0467136875,
0.042594932,
0.0200871956,
-0.0140728829,
0.0270337779,
-0.0011515853,
0.0351288728,
-0.017471524,
0.0117296092,
-0.0208040532,
-0.0206162129,
-0.0335306823,
0.060741052,
0.0364333205,
0.0492174141,
-0.0290763024,
0.0162285808,
0.0558238961,
0.0674986467,
0.0252131559,
-0.0051474259,
0.0616700538,
0.0174482316,
-0.025002718,
-0.0048905867,
-0.0594448112,
0.0107652275,
-0.0038471387,
0.0137684001,
-0.0136342645,
-0.0060644452,
-0.0207154341,
-0.0630935282,
0.018145198,
-0.0241137035,
0.0621797256,
0.0372063033,
0.0541980751,
0.0152623886,
-0.0585478432,
-0.0482770205,
0.0268743932,
0.0285121128,
0.0230989363,
-0.0715397373,
0.0325373448,
0.0561514124,
0.0387891233,
-0.053536877,
-0.0121260174,
-0.0336725414,
-0.0154209491,
0.0978960991,
0.0351009406,
-0.0120338099,
-0.0249699876,
-0.0309097134,
-0.0328667983,
-0.0342606455,
0.0012265618,
0.0606975183,
0.0175566059,
0.0086660488,
-0.0243957397,
0.0608218163,
-0.0491109565,
0.0360121168,
-0.0022679812,
-0.0418968759,
0.0197643116,
0.0062999087,
0.0295605883,
-0.0017140249,
-0.0256293099,
0.0081613427,
0.0064763329,
-0.0151028279,
-0.0412420258,
-0.0401213802,
-0.0596943162,
-0.0204087086,
0.0268513318,
0.0149670374,
0.0018772618,
0.0330014899,
-0.0367757566,
-0.0295705684,
0.0191036053,
-0.0298346803,
0.0511486232,
-0.0183315724,
-0.0111309504,
0.0082003176,
-0.0255358573,
-0.0460570864,
0.0503948405,
-0.0083347186,
-0.018966943,
0.0163801424,
0.0195646081,
-0.0395019911,
-0.0351934768,
-0.0144092878,
0.0315907225,
0.0089426357,
0.0173589289,
0.0657082573,
0.0639281422,
-0.0482332073,
-0.0387021154,
-0.0203002151,
-0.0132593522,
-0.0440466218,
-0.0019140503,
-0.0125707574,
-0.0077817366,
-0.0160214137,
-0.0187241361,
-0.0312790163,
-0.006399943,
0.0362818874,
0.0042180689,
0.0438560843,
0.0361942053,
-0.0105917538,
-0.0505976602,
-0.0254969448,
0.0083658788,
-0.0090189371,
0.007343872,
-0.008039494,
-0.0488010161,
0.0189981759,
-0.0235965718,
0.038238585,
-0.0141617842,
-0.0494140685,
0.0017736581,
-0.0401436277,
0.0210098885,
0.0757691637,
0.0001249938,
0.0384030752,
0.051302392,
-0.0117399208,
-0.0618102662,
0.0574841015,
-0.0633936152,
-0.0667508915,
-0.0855063424,
0.0062101004,
0.0346950591,
0.0636900216,
-0.0267895143,
0.0133685032,
-0.0353064276,
0.0206327029,
-0.0020371631,
-0.07301227,
-0.0098458957,
-0.015437129,
-0.0417743661,
0.0348217003,
-0.0298901983,
-0.0593743026,
0.0853657573,
0.0177394804,
-0.0119071677,
-0.0459456667,
-0.04301681,
-0.0524199829,
0.0172029939,
0.0163263008,
0.0265927296,
-0.0254831091,
-0.0497824028,
-0.0326362625,
-0.0451397263,
-0.0017376257,
0.0056917928,
0.0719845444,
-0.0236774944,
0.0153841479,
-0.0009011055,
-0.0316813178,
-0.0585562848,
-0.0116145657,
-0.0457524136,
0.0328171477,
-0.0165663995,
-0.0397778191,
-0.014909639,
-0.0326503441,
0.044685334,
0.0283912383,
0.0023810444,
-0.0215712283,
-0.0435254164,
-0.0194114465,
0.0495787486,
0.0110194134,
0.0090024779,
-0.0135880141,
0.0506935157,
0.0102772005,
0.0153047321,
-0.008121388,
-0.0283596814,
0.0287899468,
0.0297625568,
-0.0654321909,
-0.0006507723,
0.0442558005,
0.0138191404,
-0.0912943259,
-0.016091723,
0.1023341194,
0.0214261729,
-0.0361895449,
-0.0118560679,
-0.013305027,
0.0311648548,
0.0135233132,
0.0451377667,
-0.0427453555,
-0.0246416852,
0.0095528523,
-0.0026752679,
-0.0132929347,
-0.028362805,
0.0072487695,
-0.1248011813,
0.0137401391,
-0.0196473971,
-0.0446993299,
-0.023649903,
-0.0136460196,
-0.0065428191,
0.0227430612,
0.04875236,
-0.0007571977,
-0.0120436819,
-0.0471131541,
0.0148010198,
-0.0328695998,
-0.0275971498,
-0.006106704,
-0.0762745515,
0.0503164232,
0.116703108,
0.0270456653,
-0.0276896134,
0.0169554502,
0.012516004,
0.0208763871,
0.019753037,
-0.0184877329,
-0.0023118297,
-0.0172842462,
-0.0107947299,
-0.0010322419,
0.0568584763,
-0.0030145552,
-0.0163693111,
-0.0040003299,
0.0746868774,
-0.02464821,
-0.0238698889,
-0.0042685699,
0.0115440609,
0.0276001375,
0.0514961705,
0.0008175609,
-0.0619194657,
-0.0838684514,
-0.0152344434,
-0.0553979576,
-0.0520361252,
-0.0317340381,
-0.0138352262,
-0.0220813137,
0.0294718593,
-0.0516447909,
-0.088997364,
0.0194125213,
-0.0215742849,
0.0061464249,
-0.0031032942,
0.0238979738,
0.0060140071,
-0.0043969434,
0.0087121185,
-0.0508916341,
0.0265637618,
-0.0404055566,
-0.0165396165,
0.0046998989,
-0.0368212499,
-0.0117523083,
-0.0242438149,
-0.0568568893,
-0.0686738044,
-0.018381374,
0.0514767244,
0.006875949,
0.0610506684,
-0.0112275491,
-0.073805958,
0.0640347451,
0.0205905475,
-0.0208866615,
0.0214272607,
0.0241613276,
0.0487983823,
0.0058780205,
0.0488616638,
-0.0546622314,
0.0142889507,
0.0181281045,
0.0306044426,
-0.0113182943,
-0.0032746927,
0.0151742436,
-0.0288913641,
0.0249217991,
-0.0059394483,
-0.0182535555,
-0.0258721393,
-0.0087767029,
0.004829057,
-0.0335889049,
-0.0183532517,
-0.0103195505,
0.0340024084,
0.0272467807,
0.0034897286,
0.0152486237,
0.0220728572,
0.007985685,
0.0012622998,
0.0188355297,
0.0335820019,
0.0129485857,
0.0218452103,
0.0442406498,
0.0006580412,
0.02721522,
0.002649582,
0.0372424275,
0.020469401,
0.0490618199,
0.0515257269,
0.002195749,
-0.03002543,
-0.0387265235,
-0.0564604513,
-0.075647302,
0.0547356941,
0.0419265404,
0.0715045258,
0.0102123143,
0.0072919815,
-0.0706349909,
-0.0169331487,
-0.0109230354,
-0.0311585497,
-0.024892142,
-0.0258947015,
0.0506805591,
0.0146201709,
-0.016366655,
0.0050921002,
0.0334312469,
0.0267479643,
0.058548063,
0.0067329574,
0.0138360187,
-0.0292494409,
-0.0353233516,
0.0237173345,
0.0226675235,
-0.0167802386,
-0.00888249,
-0.0164051447,
-0.0489187241,
-0.0135159194,
0.0368396305,
-0.0065644658,
-0.0226867478,
0.0491440296,
0.02203588,
-0.004150515,
-0.0045627784,
-0.0448660888,
-0.0066374084,
-0.0083997631,
-0.058661975,
-0.0090953456,
-5.86539906e-33,
-0.0060989349,
-0.0075040492,
0.0515688621,
-0.0194012467,
-0.0540778749,
0.0094605945,
-0.0050844112,
-0.0131646227,
-0.0000920906,
0.0373887531,
-0.0064922022,
0.0020565977,
0.0128739309,
0.0247290675,
0.0249863155,
-0.0510876998,
0.0447841175,
-0.0238194223,
-0.0098247966,
0.0062168394,
0.0050477409,
0.024510039,
0.0389219783,
-0.0120287677,
0.0290016625,
0.0241069496,
0.0042646704,
0.0532256961,
-0.016897982,
0.0828647241,
-0.0119658327,
0.0165665187,
0.0134873688,
-0.0322838873,
-0.0315340459,
0.0910368711,
-0.0841399133,
-0.0397004411,
0.0316437967,
0.0110754389,
-0.0818091556,
-0.0301448982,
0.0559375025,
0.0097736521,
-0.0013692417,
0.0151241655,
0.0221828278,
0.0003526298,
-0.018816499,
-0.0069221966,
-0.0481507741,
-0.0134957787,
0.0025084086,
-0.0140321404,
0.0383695662,
0.00996402,
-0.0172746927,
0.0283366963,
-0.0268638805,
0.0207380261,
0.0044996445,
0.0212711394,
0.0375633799,
0.0621248893,
0.0192968659,
0.0604178645,
0.0112215485,
-0.0151331434,
-0.0398043282,
-0.0162785947,
0.0623192154,
0.0333846323,
0.0061298516,
0.040836174,
0.0134757757,
0.0192545094,
-0.0539506339,
-0.0263653528,
0.0391931534,
0.0661045909,
0.0265170522,
0.0112570859,
-0.0397360697,
-0.027460238,
-0.0735108033,
0.0352838598,
-0.0272518359,
-0.0353065953,
0.0465565845,
0.0484535359,
0.0143571263,
-0.0151035003,
0.0281144958,
-0.0364779793,
-0.0185863227,
0.0296415128,
-0.0079575833,
0.0311345495,
-0.0097662657,
0.0020131369,
-0.0894229785,
-0.015765829,
-0.0082173254,
-0.0353567451,
-0.0251706168,
0.0130952355,
0.0107444618,
0.0545395724,
-0.0411515422,
0.0260110553,
-0.0087775541,
0.0388593152,
0.0198596232,
0.0156935267,
0.0458962582,
-0.0079710037,
-0.0110183908,
0.0470464677,
0.0252489168,
-0.0598891452,
-0.0098926015,
-0.0059943851,
-0.0193341803,
0.0118646994,
-0.0278468411,
0.0109381573,
-0.0417090692,
0.0261376873,
0.0062527927,
-0.0612831265,
0.0055325758,
0.0113163134,
2.707e-7,
-0.0190118868,
0.0536346957,
0.0359666385,
0.0406823754,
0.0335251614,
0.0330382772,
-0.0276572257,
0.034408275,
0.0033905134,
0.0142051522,
0.0307542942,
0.045989614,
0.0153877242,
-0.0178512279,
-0.071794264,
-0.0007293123,
-0.0720159113,
-0.0010610585,
-0.0738762021,
-0.0313616842,
0.0058863345,
0.0742748007,
0.0278216787,
-0.0297538769,
0.0043597445,
-0.0285028163,
0.0562929846,
-0.0678774118,
-0.0127518447,
0.0202006083,
0.0072155618,
0.0398757532,
-0.0069323424,
-0.008792636,
0.0225411002,
0.0342015661,
-0.0525080077,
0.0704488307,
0.0193913896,
-0.0238856692,
-0.020072151,
0.0036615271,
0.0084179835,
-0.0336834379,
0.0403905734,
-0.0801749304,
-0.0160948504,
-0.0172746908,
0.037353497,
-0.054831069,
-0.0053233169,
-0.0361018218,
-0.0230587777,
-0.0037832076,
0.0116177322,
0.0175456759,
-0.0040054452,
-0.0998618901,
0.0079447199,
0.0187876392,
-0.0224223919,
-0.0152798183,
-0.0674348772,
0.0461291745,
-0.0027818212,
-0.072517924,
-0.0170049779,
2.296669154e-34,
0.0124402512,
-0.0049360869,
-0.0005121336,
0.0040800329,
0.0019412228,
0.0569647364,
-0.0309660025,
0.0341942646,
0.0605677143,
-0.0761023164,
-0.0189933404
] |
http://arxiv.org/abs/2402.01869v2 | http://arxiv.org/pdf/2402.01869v2 | InferCept: Efficient Intercept Support for Augmented Large Language Model Inference | 2024-02-02T19:47:57 | 2024-05-30T04:18:03 | Reyna Abhyankar, Zijian He, Vikranth Srivatsa, Hao Zhang, Yiying Zhang | cs.LG | cs.LG, cs.CL, cs.DC | Large language models are increasingly integrated with external environments,
tools, and agents like ChatGPT plugins to extend their capability beyond
language-centric tasks. However, today's LLM inference systems are designed for
standalone LLMs. They treat each external interaction as the end of LLM
generation and form a new request when the interaction finishes, causing
unnecessary recomputation of already computed contexts, which accounts for
37-40% of total model forwarding time. This paper presents InferCept, the first
LLM inference framework targeting augmented LLMs and supporting the efficient
interception of LLM generation. InferCept minimizes the GPU resource waste
caused by LLM interceptions and dedicates saved memory for serving more
requests. InferCept improves the overall serving throughput by 1.6x-2x and
completes 2x more requests per second compared to the state-of-the-art LLM
inference systems. | InferCept: Efficient Intercept Support for Augmented Large Language Model Inference: Large language models are increasingly integrated with external environments,
tools, and agents like ChatGPT plugins to extend their capability beyond
language-centric tasks. However, today's LLM inference systems are designed for
standalone LLMs. They treat each external interaction as the end of LLM
generation and form a new request when the interaction finishes, causing
unnecessary recomputation of already computed contexts, which accounts for
37-40% of total model forwarding time. This paper presents InferCept, the first
LLM inference framework targeting augmented LLMs and supporting the efficient
interception of LLM generation. InferCept minimizes the GPU resource waste
caused by LLM interceptions and dedicates saved memory for serving more
requests. InferCept improves the overall serving throughput by 1.6x-2x and
completes 2x more requests per second compared to the state-of-the-art LLM
inference systems. | [
-0.0080692405,
0.0564683117,
-0.0154030649,
0.0846805573,
-0.054136306,
-0.0011869727,
-0.0012012827,
0.0139261372,
0.0758722723,
-0.0359691009,
0.0001735781,
-0.0193311982,
-0.0041517718,
0.0072414083,
0.0100586405,
0.0249483567,
0.0460668914,
-0.0510679334,
-0.0334163122,
-0.0215320121,
-0.0139693636,
0.0273620486,
0.0090172924,
0.0134918988,
0.0175675116,
-0.0379273891,
-0.0267699677,
0.015701307,
-0.0103528872,
-0.002034846,
-0.0074216947,
0.0232942626,
0.0240567364,
0.0686835274,
0.000002066,
-0.0550249368,
-0.0400287323,
0.0230532177,
-0.0455186404,
0.032163512,
0.0112105198,
0.0045106784,
-0.0259512514,
0.0300408248,
-0.0162515808,
0.0263365488,
0.0398932211,
-0.013735421,
0.0150254294,
0.0469965078,
-0.005642774,
0.0157289021,
-0.0059972289,
-0.020376144,
-0.0039346209,
-0.0537778512,
0.0165478252,
-0.0308255516,
0.037028145,
-0.0006979796,
0.0091481684,
0.0179843977,
0.0302207265,
-0.0124128815,
-0.0624920987,
0.0187863521,
-0.0165296905,
0.006125052,
-0.0352680571,
0.0074177021,
0.0304710194,
-0.030598009,
0.0202319324,
0.0133988634,
-0.0302640907,
0.0051192553,
-0.0424174666,
0.0231193341,
-0.0221989788,
0.0001868759,
0.0785926357,
-0.0244338792,
0.0215737503,
0.0171466488,
-0.0523042008,
0.0751204863,
0.0608254187,
0.024182694,
-0.0686758384,
0.0149466377,
-0.0125213386,
-0.0709369555,
-0.0061443928,
0.0240376107,
-0.0616708994,
0.0139624095,
-0.0301473066,
-0.0030933388,
0.0223781224,
-0.0475488976,
0.0362496115,
0.0682550147,
-0.0356436633,
0.0491539948,
-0.0257378761,
0.0146802589,
-0.0480528325,
-0.0239550322,
-0.0187328327,
-0.0258135907,
0.032567922,
0.013844504,
-0.0251777489,
0.0268142633,
-0.0715523884,
-0.0129178492,
-0.0442652963,
0.0723777935,
0.0017565917,
0.0049591986,
-0.0110676447,
-0.015572696,
-0.0144125344,
0.0049724733,
-0.0022865771,
0.0223924965,
-0.0038164887,
-0.0195014738,
0.0022334214,
-0.1406887472,
-0.0216429587,
0.0291421097,
0.0011194984,
-0.0326669179,
0.0168876853,
0.1079300269,
0.0118106399,
-0.0402121581,
-0.02569695,
-0.0494159758,
0.0536787212,
-0.0676774457,
0.0016830143,
-0.0826531798,
0.0297500249,
0.0341447927,
-0.0185819604,
-0.0131067857,
-0.0146606322,
0.000778174,
-0.0177064575,
0.0298356656,
-0.0321691521,
0.0039341473,
0.0141196297,
0.0082337148,
-0.0277132336,
0.1145367846,
-0.0147475237,
0.0197529159,
0.0411417894,
0.0187566802,
0.0082978345,
0.0880628973,
0.0198768545,
0.0347167775,
-0.0367920361,
0.0017207698,
0.0224091355,
-0.0022196611,
0.0122985356,
0.020899877,
-0.026788367,
-0.0129345013,
0.0411960557,
0.0655719116,
-0.0043572029,
-0.0044204169,
0.0243084151,
0.0223292615,
-0.0180471968,
-0.0047668261,
-0.0622986145,
0.0184398759,
-0.0522423834,
0.0025857857,
0.0566609912,
0.0718509033,
-0.0070158499,
0.0067735319,
-0.0075993137,
0.0026961451,
-0.0387091748,
0.0225480963,
0.0584629662,
-0.008913964,
0.0020958609,
0.0533081591,
0.0793288648,
-0.032956779,
-0.0621093884,
-0.0704452321,
0.0341133177,
0.0526872911,
0.0543758012,
-0.0080269538,
-0.0638349131,
0.0369613431,
-0.0067035495,
0.0702444538,
0.0162141304,
0.0125068799,
0.037540853,
-0.0107130948,
-0.0081379963,
0.0154703259,
0.0342181213,
0.0483648591,
0.057766702,
-0.0022197324,
-0.0155368736,
-0.0267525446,
-0.0059086001,
0.0105729466,
-0.043633204,
-0.025464382,
-0.0044356599,
0.0064442721,
0.0304992963,
-0.0078655463,
0.0233228318,
-0.0186020881,
0.0358140245,
-0.0101544838,
0.0103377597,
0.032788001,
0.0011273978,
-0.0340960659,
0.0470306426,
0.0334516913,
-0.0115480255,
-0.0303145908,
-0.0443693176,
-0.0149787366,
0.0423918888,
-0.0203241017,
0.1124476492,
-0.0516501591,
0.0128328577,
0.0069934758,
0.0087608416,
0.0146047212,
0.0137991039,
-0.0029625345,
-0.0068945684,
0.0496724844,
0.0102493363,
-0.0074528232,
-0.0115173599,
-0.0084634433,
0.0251695942,
0.0075226617,
0.0106924605,
-0.0314631127,
0.0252518486,
-0.0123358965,
0.0228728782,
-0.0958994329,
0.0315936394,
-0.0252811629,
0.0499246866,
0.0324935764,
0.0656173378,
-0.0209973212,
-0.0049545863,
0.0273501482,
0.0233096164,
0.025778288,
0.0571503378,
-0.0147704836,
0.006528697,
-0.0769201443,
-0.0014064763,
0.0006984372,
-0.0057120048,
0.0532699786,
-0.0249826889,
-0.0625087693,
-0.051019866,
-0.0018738738,
-0.032835912,
0.0364047922,
-0.0334011838,
-0.0203105174,
0.0059639285,
-0.0591004416,
-0.00418441,
0.0390500501,
-0.0215593912,
-0.0049100486,
0.0087348288,
-0.0395043641,
-0.0835037902,
-0.0076705692,
-0.0123731568,
-0.0377632901,
-0.0195837878,
0.000115991,
-0.0391285755,
0.0838130787,
-0.0616595298,
-0.0185980536,
0.0124985278,
-0.0184883811,
0.0264832657,
-0.018763449,
-0.0151574109,
0.0150726512,
-0.0219457038,
0.0392750278,
0.0182534829,
-0.0860928297,
0.0453019291,
-0.0694226697,
-0.0059499796,
-0.0194105152,
-0.0372657366,
-0.0274994932,
0.0704548061,
-0.0032402789,
0.0154867964,
-0.045684915,
-0.0268991999,
0.0142328851,
0.0412605219,
0.0202493276,
-0.0306198671,
0.0634319037,
-0.0642718449,
0.0353930108,
-0.0349275246,
-0.0336027108,
-0.0369832255,
0.0410584323,
-0.0488189124,
-0.0103706,
-0.0204754248,
-0.0333841108,
0.002272578,
0.0181818809,
-0.015172312,
0.0559098981,
-0.0289215744,
0.0173429586,
-0.042945005,
-0.0273022242,
0.0194693953,
0.0029568714,
-0.0069083078,
-0.0351356529,
0.0085762842,
0.088821061,
0.0024657,
-0.0461775437,
-0.0300091971,
-0.0078735212,
0.0647830665,
-0.0069004288,
0.0031781788,
0.0526326075,
0.0206146874,
-0.0944629312,
-0.0167102125,
0.0789119527,
0.0083594536,
-0.0644952357,
-0.0139572574,
-0.0224748291,
0.0516219996,
0.016987117,
0.0510682166,
-0.1193758845,
-0.0322855525,
0.0437000617,
-0.0115762521,
-0.0158176888,
-0.0142615316,
-0.0211462956,
-0.1149637774,
0.058522664,
0.0031905817,
-0.0068082334,
0.0256472323,
-0.0480639078,
0.0198225174,
0.0400196575,
0.0090199681,
0.0091836741,
0.0147513868,
-0.0081080096,
-0.0306478497,
-0.0062986305,
0.0049893097,
-0.0227847602,
-0.0465496667,
0.0472529121,
0.0092836088,
0.0184134003,
0.0156465415,
0.0016502673,
0.0582020245,
0.0094192587,
-0.0190706216,
-0.0605960675,
0.0028162489,
-0.0049275476,
0.0107747214,
-0.0162576865,
0.0221298989,
-0.0168233197,
-0.009857038,
-0.0077155801,
0.0814208388,
-0.051455386,
0.0196015798,
-0.0007749547,
-0.0128402663,
0.0117731718,
0.056771405,
0.0149203045,
0.0095124543,
-0.0474494696,
-0.0637708381,
-0.0773967579,
0.0056355507,
-0.013572678,
0.0018062543,
-0.0041388255,
-0.051485341,
0.0077561289,
0.0523970388,
0.0410310738,
0.030282354,
0.0028258797,
0.0705953538,
-0.0036152911,
-0.020046575,
-0.0489309505,
0.0135871759,
-0.0606877618,
0.0204558596,
-0.0209725257,
0.0026121358,
0.0414347574,
-0.0332470089,
-0.0115547478,
0.0570540316,
-0.0744454339,
-0.1021610871,
-0.0119183958,
0.0384434946,
0.0341943316,
0.0781181231,
-0.0418999828,
-0.0127307149,
0.000247381,
0.0511402898,
-0.0551311634,
0.0315970331,
0.0034419335,
0.0311100539,
0.0057773897,
0.0304404274,
-0.0419288315,
-0.0099909054,
0.0135433553,
0.0268980004,
0.0091671431,
-0.060828425,
-0.0287207644,
0.0586518906,
0.0000578073,
-0.0297873542,
-0.0272542518,
-0.0149677545,
0.0486033373,
-0.0526712872,
0.0063528614,
-0.0349743553,
0.027456075,
-0.0520479679,
-0.0067932573,
-0.0039452123,
-0.0062172371,
0.0242011696,
0.0102802543,
-0.0367733948,
0.0182542931,
-0.0342895314,
-0.0459941402,
0.070601061,
-0.0178549532,
0.0245698821,
0.0033922272,
0.023459537,
0.0675234199,
-0.0544493273,
0.0298566259,
0.0229477808,
-0.0158015341,
-0.0150215495,
0.0106771747,
-0.033524245,
-0.0724131316,
-0.003281574,
0.0670893714,
0.0712969974,
0.0207295995,
0.0680508465,
-0.0343460627,
-0.0120380474,
-0.0103837214,
0.0398008712,
-0.0093628597,
-0.0249728989,
0.0304844845,
0.015351966,
-0.0115640694,
0.0256281961,
0.0362121798,
-0.0143649941,
0.0520006008,
0.0107492972,
-0.0305048581,
-0.0331050232,
-0.0659085438,
-0.0028334809,
0.0053990898,
-0.0193141717,
-0.0142863793,
-0.0375544541,
-0.0587174743,
-0.0289296787,
0.0058748382,
-0.0509796813,
-0.020655999,
0.0014648867,
0.0635188594,
0.0497764051,
-0.0451490618,
-0.0411954634,
0.031694036,
-0.0189366248,
-0.014862149,
0.0299622528,
-6.438972345e-33,
-0.0298844017,
-0.0419988036,
0.0587290004,
-0.0079534007,
-0.0158289149,
-0.0048795519,
-0.0321151018,
-0.0291595887,
0.0295469873,
-0.024122607,
-0.0047526257,
0.0239403117,
0.0335396864,
0.0061726286,
-0.0082030883,
0.0010212142,
0.0086687533,
0.0167750008,
-0.001168675,
0.021029057,
0.037154112,
0.0245605651,
0.0418873653,
-0.0505445488,
-0.0271126069,
-0.0305052772,
0.0436237417,
0.0199213233,
-0.0519263372,
0.0332291052,
-0.0196176041,
0.0127485255,
0.0124830455,
-0.0291576758,
0.0109008988,
-0.0129742129,
-0.0366264991,
-0.0164616574,
0.0556810685,
0.0062585478,
-0.0354276448,
-0.0264364369,
-0.0314407498,
0.0378216244,
-0.0037098273,
-0.0031841104,
-0.0299447048,
-0.011689079,
0.0010799934,
0.0086685875,
-0.037434198,
0.035205733,
0.0408691727,
0.0530858003,
0.0470422879,
-0.0194565058,
0.0055261529,
0.0063282945,
-0.0209486466,
0.0030464081,
0.0000692446,
0.0237066038,
0.0814652443,
0.03222882,
0.0133611197,
0.0349797346,
0.0362555161,
-0.0082300957,
-0.0189397596,
0.0066685551,
0.0080757979,
0.0405830406,
-0.0032567661,
0.0103542795,
0.0065733669,
0.0089932792,
-0.0451888405,
0.0177711304,
0.0704461485,
0.0015211549,
0.0504021905,
-0.0078946641,
-0.0019610559,
-0.0279080737,
-0.0054725176,
0.0026022447,
-0.035231974,
-0.0241439436,
0.0146804983,
0.0090286806,
0.0030298482,
-0.0081663355,
0.0538417883,
-0.0387910195,
-0.0237364639,
-0.020645529,
-0.0028615389,
-0.0312746055,
0.0156641584,
0.0446112342,
0.0135019403,
-0.0270733386,
-0.0196861196,
-0.0462535322,
0.0150655014,
0.0058590285,
0.0057252413,
0.0078407731,
-0.0401780829,
0.0164862666,
-0.0667129308,
0.0372482799,
0.0046980195,
0.0132381273,
0.0557987131,
0.0008576741,
0.0299569089,
0.0481955893,
0.0291449595,
0.0404652357,
-0.0375884473,
-0.0159623604,
0.0042230762,
0.0011403721,
-0.0445471704,
0.008755235,
-0.0467663594,
0.0179445352,
0.0181017555,
-0.0626605004,
-0.0210121851,
0.0117349261,
2.889e-7,
0.0141235702,
0.0280273464,
0.0179555733,
0.0482513309,
0.0160151999,
0.0053084521,
0.0121672256,
-0.0142288348,
0.0562950261,
-0.0153663978,
-0.0198672172,
0.0221468396,
0.0615995377,
0.01439723,
-0.0586261339,
0.0042271535,
-0.0409184769,
-0.0302085411,
-0.0555386245,
-0.029164562,
-0.0065643354,
0.0714075342,
-0.0083655966,
-0.0238742009,
0.0435857065,
-0.0142627535,
-0.0047051874,
-0.0542065725,
-0.0457202531,
0.0149266506,
0.0288043059,
0.0148642529,
0.0105473604,
0.0031898287,
-0.0046978667,
0.0125162173,
-0.071470499,
0.0426296592,
-0.0061542811,
0.0031685559,
-0.0015303829,
-0.0695942044,
0.0036594367,
-0.0250540301,
0.0804267302,
0.0217043646,
-0.0083929347,
-0.0407358259,
0.0128029827,
-0.0697357431,
-0.0566580184,
-0.0170525163,
0.0257548839,
0.036708571,
0.0176585913,
-0.0464383513,
-0.0178238321,
-0.0474924967,
-0.0467854515,
-0.0120642297,
-0.0542123355,
-0.0830480307,
-0.0403526798,
0.0199014489,
-0.0033780828,
-0.0437063426,
-0.0029709246,
3.458533739e-34,
0.0278124399,
0.0332253724,
0.0016451573,
0.0997254401,
0.0361453965,
-0.015654739,
-0.0044136737,
0.0215456579,
0.0140331658,
-0.0916674137,
-0.0366946422
] |
http://arxiv.org/abs/2402.01867v1 | http://arxiv.org/pdf/2402.01867v1 | Leveraging Large Language Models for Structure Learning in Prompted Weak Supervision | 2024-02-02T19:45:39 | 2024-02-02T19:45:39 | Jinyan Su, Peilin Yu, Jieyu Zhang, Stephen H. Bach | cs.LG | cs.LG, cs.CL | Prompted weak supervision (PromptedWS) applies pre-trained large language
models (LLMs) as the basis for labeling functions (LFs) in a weak supervision
framework to obtain large labeled datasets. We further extend the use of LLMs
in the loop to address one of the key challenges in weak supervision: learning
the statistical dependency structure among supervision sources. In this work,
we ask the LLM how similar are these prompted LFs. We propose a Structure
Refining Module, a simple yet effective first approach based on the
similarities of the prompts by taking advantage of the intrinsic structure in
the embedding space. At the core of Structure Refining Module are Labeling
Function Removal (LaRe) and Correlation Structure Generation (CosGen). Compared
to previous methods that learn the dependencies from weak labels, our method
finds the dependencies which are intrinsic to the LFs and less dependent on the
data. We show that our Structure Refining Module improves the PromptedWS
pipeline by up to 12.7 points on the benchmark tasks. We also explore the
trade-offs between efficiency and performance with comprehensive ablation
experiments and analysis. Code for this project can be found in
https://github.com/BatsResearch/su-bigdata23-code. | Leveraging Large Language Models for Structure Learning in Prompted Weak Supervision: Prompted weak supervision (PromptedWS) applies pre-trained large language
models (LLMs) as the basis for labeling functions (LFs) in a weak supervision
framework to obtain large labeled datasets. We further extend the use of LLMs
in the loop to address one of the key challenges in weak supervision: learning
the statistical dependency structure among supervision sources. In this work,
we ask the LLM how similar are these prompted LFs. We propose a Structure
Refining Module, a simple yet effective first approach based on the
similarities of the prompts by taking advantage of the intrinsic structure in
the embedding space. At the core of Structure Refining Module are Labeling
Function Removal (LaRe) and Correlation Structure Generation (CosGen). Compared
to previous methods that learn the dependencies from weak labels, our method
finds the dependencies which are intrinsic to the LFs and less dependent on the
data. We show that our Structure Refining Module improves the PromptedWS
pipeline by up to 12.7 points on the benchmark tasks. We also explore the
trade-offs between efficiency and performance with comprehensive ablation
experiments and analysis. Code for this project can be found in
https://github.com/BatsResearch/su-bigdata23-code. | [
0.0348171033,
0.0657009855,
0.0219299179,
0.0494458936,
-0.0062683374,
0.0381098092,
0.0223118011,
-0.0067883334,
0.0603666082,
-0.0188408494,
-0.0850049779,
-0.0257268008,
-0.0059198188,
0.0180689972,
0.014195174,
-0.0137318512,
0.0205763597,
-0.0494585373,
0.0165878236,
-0.0078836204,
-0.0117366994,
-0.0117653739,
0.0168712232,
-0.0097077591,
-0.0451425239,
-0.0101420004,
0.0164439697,
-0.0002722032,
-0.0139392363,
-0.0191677194,
0.050375618,
0.0851847529,
0.0220472012,
0.0189751089,
0.000002025,
-0.0068847891,
-0.0063915146,
-0.0045137201,
0.0112631749,
0.0321169235,
0.008397107,
-0.0422932506,
0.0168107394,
-0.0058948961,
-0.0108306659,
-0.0503126681,
0.0523738861,
0.0782629848,
0.0742745548,
0.0657489747,
-0.0013827847,
-0.0564999729,
-0.0093270438,
-0.0714740306,
0.0236034039,
-0.0198836606,
0.0692197531,
0.0102044521,
-0.0323858708,
0.0370015688,
-0.0095035378,
0.0447462313,
0.0100157419,
-0.0339873321,
-0.009966542,
0.0291796532,
-0.040827848,
-0.0412745886,
0.0271816645,
0.0978940874,
0.0190406963,
-0.0176709425,
0.0262897667,
-0.0021756811,
-0.0310816113,
0.0189773478,
-0.033936739,
-0.0100803468,
0.0085419537,
0.0352746435,
0.0272924174,
0.0278208964,
0.0396002047,
-0.008677952,
0.0208980963,
0.0478314869,
0.0267353673,
-0.0496546179,
0.0013359318,
0.0130561534,
-0.0335422717,
-0.0348577835,
-0.0329455771,
-0.0259543695,
-0.0413397066,
0.0334067792,
-0.0475508198,
-0.053729523,
0.0110648824,
-0.0104454691,
-0.0041928473,
-0.030489184,
-0.0652602017,
0.0890047699,
-0.0478125177,
0.0344740003,
-0.0125533687,
0.0418022089,
-0.0243736655,
-0.0001629076,
-0.0283439532,
-0.0142013151,
-0.0055000666,
0.0280632172,
-0.0011300085,
-0.007829044,
-0.0218792353,
-0.018038379,
-0.0298382472,
0.0324106328,
0.0073567182,
-0.0087834047,
-0.0305086803,
0.0378589183,
0.0258868709,
-0.008805234,
-0.0097982381,
-0.0020459841,
-0.0459269807,
-0.0187746231,
-0.012097192,
0.0423705094,
0.0269475449,
-0.0406910814,
0.0428978167,
0.0376240164,
0.0155660994,
-0.0394907258,
-0.0625665113,
-0.0269523095,
-0.0521268845,
-0.0460441411,
0.0145606557,
-0.0685060173,
0.0328684933,
-0.001128321,
0.0096028596,
-0.0342774875,
0.0246908832,
0.0126011828,
-0.085983485,
0.0798769966,
-0.0069434647,
-0.001872876,
-0.0109227011,
0.0041064196,
0.0033240612,
0.0488467626,
-0.0434816256,
0.0297291819,
0.0404184908,
-0.0131708076,
0.0073804311,
0.0349789374,
0.0141976904,
0.0435955338,
-0.0195809696,
-0.0014562842,
0.0492479578,
-0.0158003867,
0.0421318412,
0.0657321364,
-0.0158347636,
-0.0034662243,
0.0335271917,
0.011388381,
0.006432245,
-0.0247452203,
0.0522588305,
-0.0080818906,
0.0173770655,
0.0284407623,
-0.0639946237,
-0.0227001011,
-0.0353112854,
0.0005504213,
-0.0109698167,
-0.0193739031,
-0.050186187,
-0.0265850909,
0.010552126,
-0.0172334351,
-0.0324891955,
0.0120338378,
0.0660982803,
0.0566917025,
-0.0141245956,
0.0069171567,
0.0730984882,
-0.0129469633,
-0.0026463452,
-0.0199363288,
0.0403495207,
0.0397898592,
0.0386807993,
-0.0523633063,
-0.0194744729,
0.0053898739,
0.0765644908,
0.0124039948,
-0.0006022658,
-0.0318545699,
0.0326975361,
-0.0428755097,
-0.0440346524,
-0.0106494725,
0.0200986918,
0.0133336056,
0.0103778318,
-0.0279702581,
0.0377095416,
-0.0135164345,
-0.0305981748,
0.0534061715,
-0.0005106857,
-0.0326198153,
0.014028999,
0.0161700621,
0.0305683035,
0.0099809496,
-0.0277411118,
0.003085667,
0.050965894,
-0.0655062944,
0.010128309,
-0.0513670966,
-0.0701443106,
-0.0484603234,
0.0342712849,
-0.0026151063,
-0.0489417203,
-0.0014863529,
-0.0550984666,
-0.0604848042,
-0.0135205425,
0.0032315431,
0.008826185,
-0.0260324851,
-0.0081941485,
0.0033021518,
-0.0065328777,
-0.0539094023,
0.0651071817,
0.0101937437,
0.0437640287,
0.052750662,
0.0151722087,
0.0033121782,
0.017407883,
-0.0363898613,
0.0004941117,
-0.0141110299,
-0.007493082,
0.0325001478,
0.0501944497,
-0.0459140874,
-0.0303028785,
-0.0490283668,
-0.0079152016,
-0.0416755415,
-0.0047443127,
0.033743538,
0.0011319215,
0.0217037052,
-0.039627254,
-0.0192583073,
-0.0311677959,
-0.0300734174,
0.0290354453,
0.0031965454,
-0.0034564519,
-0.0615785681,
-0.0555156432,
-0.0225114506,
-0.0128839668,
0.0353692994,
-0.0058192266,
-0.0762995332,
-0.0451317467,
0.0652329773,
-0.0133863883,
0.0188602544,
-0.0097103352,
-0.018546313,
0.0177242234,
-0.0420426913,
0.0861244202,
0.0070758024,
-0.0565101877,
0.0067279236,
0.0513433889,
-0.0175911356,
-0.0127884708,
-0.023302367,
-0.0228772257,
-0.0087010618,
-0.009967776,
-0.0038900776,
-0.0077837091,
0.0400411375,
0.0195258632,
-0.0003634891,
-0.0488742143,
-0.0116926152,
-0.0048864167,
-0.0016350274,
-0.05315689,
0.0042760083,
-0.0189154595,
-0.0241886731,
-0.0392579883,
-0.0474258214,
0.0321871527,
-0.0415140912,
-0.0153573034,
-0.0076415213,
0.0031150398,
-0.0470704064,
0.0633815378,
0.0426613279,
0.0093791829,
0.0265098661,
-0.027956441,
-0.0550583787,
0.0154596716,
0.0042614974,
-0.010732621,
-0.03332242,
-0.0509998575,
0.0355118029,
0.0131254038,
-0.0402409583,
-0.0306631997,
-0.032916598,
-0.0262845047,
0.0191100053,
0.0109748608,
-0.0020725406,
0.0092638843,
-0.002145899,
0.0337393768,
0.0659578517,
0.0069941911,
0.0279412232,
-0.0521875881,
0.0336657651,
0.0544953682,
-0.0613569953,
-0.0373372287,
-0.0300152339,
0.0093149794,
0.0463306978,
0.024985116,
0.0073781572,
-0.0300777145,
0.0244093314,
0.0115366252,
0.0483398102,
-0.0132637955,
-0.0188130233,
0.0286160037,
-0.0244691595,
-0.0000297396,
0.1023795009,
0.0258647967,
-0.0530018881,
-0.0178648587,
0.024094129,
0.0459890477,
0.0583101772,
0.0075717936,
-0.047618635,
-0.0498036966,
0.0108182132,
0.0182588715,
0.0174340438,
-0.0380303413,
0.0216501877,
-0.1011630297,
0.0168753359,
0.003688012,
-0.0044067339,
-0.0395011306,
-0.0772238895,
0.0303693693,
-0.0019153417,
-0.0224316996,
-0.0141622536,
0.009296556,
0.0095991641,
0.0275640022,
-0.0206089355,
-0.0626884326,
0.011229719,
-0.0420035943,
0.063866064,
0.0565597154,
0.005827643,
-0.0117399907,
-0.011876463,
0.0642929524,
-0.0365582705,
-0.0465791561,
-0.0318368375,
0.0478073694,
0.0015980004,
-0.0354841687,
-0.0054707881,
0.0719612837,
-0.0582930632,
-0.0010035023,
-0.001126016,
0.0826802626,
-0.0192162246,
-0.0027198512,
-0.0037592126,
-0.0189717095,
0.0355268642,
0.0862457901,
0.0462009534,
-0.04268419,
0.0169587079,
0.0309383571,
0.002917178,
0.0131426789,
-0.0304036383,
0.0314515084,
-0.0451692417,
0.0027325416,
-0.0651322529,
-0.0230990164,
0.0769621506,
0.0097799283,
0.0068411166,
0.015021137,
-0.0124127837,
-0.0537045896,
-0.0243493374,
0.0526409745,
-0.0492148623,
0.012957627,
-0.089338094,
-0.049479451,
0.0490984879,
0.0123416828,
-0.0096718874,
0.0404128246,
-0.0714529008,
-0.0535524152,
0.0148981279,
0.0614771843,
0.0067872568,
0.0359815061,
-0.0145491334,
-0.0254224278,
0.0164495632,
0.0368436016,
-0.0459277034,
0.0053466181,
0.0146185197,
-0.0271552559,
-0.0015377626,
0.0379233621,
-0.0391117819,
0.0293202456,
0.0713852942,
0.0500619821,
0.004545806,
0.0187432207,
0.0319437385,
0.0405612923,
0.0998900831,
-0.0173667371,
0.0200570226,
-0.0653752461,
-0.0081422832,
-0.0230281744,
-0.0609692223,
-0.0462763868,
0.048474174,
0.0268548224,
0.0014417303,
-0.0124807674,
0.002353688,
-0.0022297648,
0.0718231425,
0.0013052529,
0.0045145261,
0.0125144031,
0.0028344067,
0.0510079451,
-0.0076260194,
-0.0108369989,
0.0544828363,
0.0333021432,
0.0001579312,
-0.0145449191,
-0.0014898651,
0.0466236137,
-0.0017697723,
-0.0604935177,
0.0092607616,
-0.0383421928,
-0.0253683645,
-0.0164900254,
0.0242980123,
0.0358006172,
-0.0526692495,
0.0433374457,
-0.0223827362,
-0.0301993098,
-0.0309889503,
0.0112799248,
0.0410877913,
-0.0036661648,
0.0564971119,
0.0281117037,
-0.0330978334,
0.075341478,
0.1052043661,
0.0293538701,
0.043909125,
0.0015356886,
0.0056290962,
-0.0356070101,
0.013585248,
0.0333527885,
0.0090919137,
-0.0484135859,
0.0208526887,
-0.0078422921,
-0.0609192327,
-0.0110711548,
-0.051449772,
-0.0487248264,
0.042924054,
0.0513687097,
0.0333941244,
0.0195311233,
0.0107038571,
-0.0181065947,
-0.0191041194,
0.0478391424,
-0.092720136,
-0.0031912869,
-6.982686236e-33,
0.0181752872,
-0.0407641754,
-0.0094561735,
-0.0108004855,
0.0342976265,
0.0412752181,
-0.0259325374,
0.0147809787,
0.0169343129,
0.0427535661,
-0.0213521402,
-0.0177443977,
0.0186307635,
0.0342193991,
0.0080254804,
-0.022110017,
0.0566144846,
-0.0151814362,
0.0134562626,
0.049474895,
0.0353990197,
0.0454032347,
0.0411536582,
-0.1193335056,
0.0420893617,
-0.004889776,
-0.0121773314,
0.0218078718,
0.0237700436,
0.0315491073,
-0.030719053,
0.0194698945,
0.0095162047,
-0.0232492946,
-0.0073612146,
0.0441765934,
-0.1223859712,
-0.0133408941,
-0.0132080531,
-0.0156768803,
-0.0750214085,
-0.0273357593,
0.0012545872,
-0.0178526137,
0.0472065546,
0.0088206548,
0.0346228518,
-0.0346226543,
-0.0460079722,
-0.0085123098,
-0.0433877744,
0.0379293822,
0.0053386209,
-0.0027678676,
0.0249231923,
0.0316930935,
-0.0166807026,
0.0315082632,
-0.0201057009,
0.0078722127,
-0.0258057751,
0.0530812405,
-0.0039024428,
0.0426898263,
0.0084970044,
0.052641619,
0.0140008805,
0.0157416482,
0.0305206031,
-0.028337447,
-0.0047538271,
0.0661457479,
0.0282234773,
-0.0014621174,
-0.0084517952,
0.023742171,
-0.07594686,
-0.0258039199,
0.009460995,
0.0146264983,
-0.0129341539,
0.0025271068,
-0.0239811242,
-0.0390883982,
-0.0394112244,
-0.0110207042,
-0.0437752269,
-0.0034296825,
0.0191588122,
0.0301579833,
-0.0595355295,
-0.0239688195,
0.0150760384,
-0.0304337218,
-0.0296638794,
0.0140004018,
-0.0022429843,
0.0154391853,
-0.0035658027,
-0.0142936092,
-0.0626299307,
-0.0444974639,
0.0612455904,
-0.0365984812,
0.0263957661,
-0.0167837273,
0.095147498,
0.014805872,
-0.0567496493,
-0.0256158933,
-0.0111900913,
-0.0088049565,
-0.0017559911,
0.0089720525,
0.0159469321,
0.0050160848,
0.0080989264,
0.0543579347,
0.0736961588,
0.0340921208,
0.0019960753,
-0.0216271169,
-0.006764872,
0.023296535,
-0.0155147808,
-0.0170534998,
-0.0278939996,
-0.020216886,
0.0474584177,
-0.0481453165,
0.0080884295,
-0.0142054623,
2.863e-7,
0.022084998,
0.0336668864,
-0.0023339025,
-0.018932946,
0.0067145838,
0.0349396281,
-0.027306864,
-0.0009013325,
-0.0269972347,
-0.0513740033,
0.0298563391,
-0.0210685544,
0.0518151373,
-0.0058354284,
-0.0957642198,
0.0075799525,
-0.0581155457,
-0.0398748927,
-0.0504930653,
-0.0154083706,
0.0181235876,
0.1177421957,
0.0214196108,
-0.0218127109,
-0.0177150276,
-0.0478070378,
0.0348411463,
-0.0116524035,
-0.0203525163,
0.0065049361,
0.0292279068,
-0.0191267803,
0.0019098648,
0.0015663153,
-0.0170736779,
0.0026968135,
0.034756396,
0.0597900786,
0.0237597656,
0.0277390964,
0.0192999821,
0.0197769534,
0.0069922935,
0.015641395,
0.0953293741,
-0.0526951663,
-0.038685184,
-0.0086672008,
-0.006493961,
-0.0114484364,
0.0074162991,
-0.0074955137,
-0.0247295052,
0.0244952086,
0.0280133467,
-0.0047859084,
-0.0406558663,
-0.0496606044,
0.0128816254,
0.0111407153,
-0.0433730148,
-0.0149044264,
0.0010635655,
0.0253336821,
-0.0014437514,
-0.0487509593,
0.0172089878,
3.24548021e-34,
0.0384731553,
-0.0195578281,
0.0069743553,
-0.0189218502,
0.0517158359,
-0.0539924018,
-0.1015945226,
0.036066439,
-0.0072976663,
-0.0774532706,
0.0070410208
] |
http://arxiv.org/abs/2402.01865v2 | http://arxiv.org/pdf/2402.01865v2 | What Will My Model Forget? Forecasting Forgotten Examples in Language Model Refinement | 2024-02-02T19:43:15 | 2024-06-20T06:25:39 | Xisen Jin, Xiang Ren | cs.LG | cs.LG, cs.CL, stat.ML | Language models deployed in the wild make errors. However, simply updating
the model with the corrected error instances causes catastrophic forgetting --
the updated model makes errors on instances learned during the instruction
tuning or upstream training phase. Randomly replaying upstream data yields
unsatisfactory performance and often comes with high variance and poor
controllability. To this end, we try to forecast upstream examples that will be
forgotten due to a model update for improved controllability of the replay
process and interpretability. We train forecasting models given a collection of
online learned examples and corresponding forgotten upstream pre-training
examples. We propose a partially interpretable forecasting model based on the
observation that changes in pre-softmax logit scores of pretraining examples
resemble that of online learned examples, which performs decently on BART but
fails on T5 models. We further show a black-box classifier based on inner
products of example representations achieves better forecasting performance
over a series of setups. Finally, we show that we reduce forgetting of upstream
pretraining examples by replaying examples that are forecasted to be forgotten,
demonstrating the practical utility of forecasting example forgetting. | What Will My Model Forget? Forecasting Forgotten Examples in Language Model Refinement: Language models deployed in the wild make errors. However, simply updating
the model with the corrected error instances causes catastrophic forgetting --
the updated model makes errors on instances learned during the instruction
tuning or upstream training phase. Randomly replaying upstream data yields
unsatisfactory performance and often comes with high variance and poor
controllability. To this end, we try to forecast upstream examples that will be
forgotten due to a model update for improved controllability of the replay
process and interpretability. We train forecasting models given a collection of
online learned examples and corresponding forgotten upstream pre-training
examples. We propose a partially interpretable forecasting model based on the
observation that changes in pre-softmax logit scores of pretraining examples
resemble that of online learned examples, which performs decently on BART but
fails on T5 models. We further show a black-box classifier based on inner
products of example representations achieves better forecasting performance
over a series of setups. Finally, we show that we reduce forgetting of upstream
pretraining examples by replaying examples that are forecasted to be forgotten,
demonstrating the practical utility of forecasting example forgetting. | [
0.0012165937,
0.0572829731,
-0.0153131885,
-0.0240155607,
0.016841419,
0.0387094989,
0.0148867276,
-0.0130482763,
-0.0016735751,
-0.0328582674,
-0.0082760779,
-0.0591686219,
-0.001427224,
0.0406184867,
-0.0027206016,
-0.0299265534,
0.0602912791,
-0.0411642492,
-0.0361475796,
-0.0301985927,
0.0082716839,
0.0111590298,
0.0092787826,
0.0614788942,
0.0022202027,
-0.0360081457,
-0.0410388969,
-0.0158991776,
-0.0451538302,
-0.0057564047,
0.0484797768,
0.025852656,
0.047185041,
0.0266755223,
0.0000018708,
-0.0206825975,
-0.0483036749,
-0.018327374,
-0.0570599698,
0.0207418557,
-0.0068509695,
-0.0149408439,
0.0061842706,
0.0329004116,
-0.0268036295,
-0.0207659341,
0.0427694991,
0.0448939502,
0.0769235119,
0.0664168149,
0.0203501023,
-0.0165974088,
0.0160734542,
-0.0145547641,
0.0270698462,
-0.0378096886,
0.0077851182,
-0.042500969,
-0.0739386752,
-0.0467123203,
-0.0204317495,
0.0172982663,
-0.0026900945,
0.0330590978,
-0.021392459,
0.0296794195,
0.0140205352,
-0.0027019987,
-0.0275096949,
-0.0039177132,
-0.0178528745,
-0.0443645529,
0.0044211876,
-0.0008044723,
-0.0111689689,
-0.0254331976,
-0.0458419956,
-0.0128557421,
0.0225467961,
0.0284826178,
0.0310132075,
-0.0534898266,
0.0327355713,
0.0168974958,
-0.0046054195,
0.0795376003,
-0.0058799442,
-0.0370377786,
-0.0441427492,
0.0220162198,
-0.0021811686,
-0.0113789495,
0.0024029792,
0.0303618908,
-0.0259220619,
0.0511665754,
-0.0297293365,
-0.0325151347,
-0.0434754081,
-0.0319789872,
0.1102292538,
0.0511578396,
-0.0094227269,
0.0182289351,
-0.020209929,
0.0406544097,
0.0291887,
0.0099818781,
-0.0062028929,
0.0346606299,
-0.0893123522,
-0.0293282662,
-0.068275854,
0.0410896465,
0.0538314767,
0.0451024659,
-0.0243144669,
0.0033480639,
-0.0437852219,
0.0275657233,
0.0109769525,
-0.0193768423,
-0.0852321833,
-0.011431966,
-0.0270359144,
0.0195734352,
0.0075960113,
-0.0137176551,
0.0035536333,
-0.0726072416,
-0.0075670681,
-0.0503217801,
-0.0187482163,
0.0061789597,
0.0072138184,
0.0500755906,
0.0209724139,
-0.0208090134,
-0.0645647049,
-0.0605737567,
-0.004478388,
-0.0405281447,
0.0350541361,
0.005586084,
-0.0209998302,
0.001219265,
0.0262320861,
-0.0337440744,
-0.0134762693,
-0.0104472525,
-0.0059754769,
0.0566322133,
0.0281858519,
-0.0169446766,
0.0041192751,
-0.0227318034,
0.0197035503,
0.0795788839,
-0.0014817379,
0.0018429819,
0.0174952075,
-0.014640661,
0.0087353997,
0.0352422595,
0.0579777323,
0.011453242,
-0.0590657815,
-0.0063177384,
0.0160999522,
0.0054324577,
0.0244996864,
-0.0349698886,
-0.0037712243,
0.0256377924,
0.0496794507,
-0.0109410379,
0.0164309982,
0.0332886279,
0.0408321582,
-0.0235650744,
-0.0107102357,
0.0139401918,
-0.022900356,
0.0068039484,
-0.0132012013,
0.0195726082,
-0.0160853565,
0.0297675375,
-0.0587044135,
-0.0438868888,
-0.0001592965,
-0.0190031026,
-0.0210081078,
0.0001477936,
0.0852417871,
0.0001976743,
-0.0485420972,
0.1212483793,
0.041777581,
-0.0213297643,
0.0386824235,
-0.0591074005,
0.0330017544,
0.0501086302,
0.0214007013,
-0.0116342101,
-0.0494732261,
-0.0075145117,
-0.0174460709,
-0.0369386598,
0.0800356269,
-0.0039703869,
-0.0043753292,
-0.0610013753,
-0.0420721211,
0.0208152868,
0.0402864181,
0.0497560203,
-0.0234486312,
0.016188452,
0.0137079097,
-0.0203858744,
-0.0644664839,
0.0226003472,
-0.0198628996,
-0.0354249924,
-0.0141842589,
-0.0198596176,
0.0486083552,
-0.0308827143,
0.0074843336,
0.0153952651,
0.0450324155,
-0.0001216318,
-0.0276964474,
0.0315375701,
-0.1076873392,
-0.0444899797,
0.0317961574,
0.0288986638,
0.0351233892,
0.0087509984,
-0.056145031,
-0.0691078752,
0.0672002733,
-0.0140835177,
0.0718933344,
-0.0343534984,
0.0379716568,
-0.0163902827,
-0.008972575,
0.0020546613,
0.0204473045,
-0.0147671252,
0.0213221554,
-0.0694061965,
-0.0154520879,
-0.0153958695,
-0.0093294196,
-0.030687606,
0.0188463144,
-0.0147281606,
0.0342998467,
-0.0240623727,
-0.0296117496,
-0.0498291217,
-0.0003747186,
-0.0308083221,
-0.0227635168,
-0.0422508754,
0.0037462288,
-0.0033371961,
-0.0164880566,
-0.0002108845,
-0.0422285013,
-0.0101028625,
0.0120078782,
-0.0119639467,
0.0172529295,
-0.0140714869,
0.0218582228,
-0.00117038,
0.0137324641,
-0.0242231619,
-0.01098332,
0.052648697,
-0.0376625173,
0.0097307023,
-0.0888664722,
0.0082482817,
-0.0399308875,
0.029490592,
-0.0112552214,
-0.023706276,
-0.029234089,
-0.0376482978,
0.012159694,
0.055076912,
-0.0310997665,
0.03979408,
0.0475847758,
-0.0111114299,
-0.0538725518,
0.0130654993,
-0.0243909601,
-0.0768575743,
-0.0339625925,
0.01003691,
-0.0286031086,
0.0776208788,
-0.003071405,
-0.0054217251,
-0.0545102246,
0.01330694,
-0.0597271957,
-0.0140246786,
-0.0388318375,
-0.007012316,
-0.0269076154,
0.0434057824,
0.0098856855,
-0.0690532923,
0.0704097748,
0.0132114366,
-0.0222434439,
-0.0381546617,
-0.0075010834,
-0.0028747909,
0.0330951065,
0.0191633366,
-0.0099494355,
0.0179268457,
-0.0294680074,
-0.0405854248,
-0.0372781903,
0.0248572454,
-0.0272877775,
0.0115731498,
-0.0224881303,
-0.0179907233,
0.0266092643,
-0.0009006757,
-0.0268201474,
0.0038240391,
-0.0317735672,
-0.0129504008,
0.0015166267,
0.0043304465,
-0.0330310278,
0.0138466582,
-0.0155374575,
0.040003676,
-0.0295463651,
-0.0144969504,
0.0032042591,
-0.0252378751,
0.0220475346,
-0.0055452096,
-0.0014676132,
-0.0906945467,
0.0155577045,
-0.0191428829,
0.0857134238,
-0.0141762905,
-0.0352562405,
0.0275059249,
0.0225657336,
-0.020519834,
-0.0501079261,
0.0674273074,
0.0198681448,
-0.1031786501,
-0.0599553101,
0.0857118368,
0.0266954415,
-0.0319562592,
0.0031222424,
0.0059959674,
0.0482985303,
0.0169895515,
0.0270275678,
-0.0184748657,
-0.0147771807,
0.0388397574,
0.0435543507,
-0.0018386131,
-0.0181810707,
-0.0172263887,
-0.1033422872,
0.0209480301,
0.0405598804,
-0.0148625495,
0.0081408815,
-0.0223347023,
-0.037412852,
-0.0294769835,
0.0319550931,
0.0334475562,
-0.0220323093,
0.0500814915,
0.0805038139,
-0.0482366718,
-0.063557595,
0.036406368,
0.0121450461,
0.0211220495,
0.0750340521,
-0.0138670839,
-0.0030285777,
-0.0245158169,
0.0513254143,
-0.0366457067,
0.0530151576,
-0.0056046806,
0.0584571771,
0.0151407104,
-0.0199693516,
0.0183638688,
0.0242074672,
0.0042431569,
0.0164348725,
0.0019604121,
0.0926771089,
-0.0166742448,
-0.037031021,
0.037618991,
-0.0257361643,
0.0015335173,
0.0379579552,
0.017340675,
-0.0169843528,
-0.047364261,
-0.0088197971,
0.017313268,
-0.0261297021,
0.0106854159,
0.0299758743,
-0.0393335484,
-0.0101016099,
-0.0199241284,
-0.0433407463,
0.0163582619,
0.0747733638,
0.0188320111,
-0.0080658626,
0.0067941071,
0.034267921,
-0.0188744124,
0.0334770083,
-0.0070417216,
-0.0070577795,
-0.0465962179,
0.0158618875,
0.037519604,
-0.0052258754,
-0.0270789992,
-0.0152898626,
-0.0662371144,
-0.0522369966,
0.083909288,
0.0407872126,
0.0300019272,
0.0649086908,
0.0176841356,
-0.029423343,
0.0489888601,
0.0309885722,
-0.0086795781,
0.0226905514,
0.0712602511,
-0.0115911784,
0.0082306564,
0.0390977897,
-0.044056572,
0.0080230953,
0.0239699297,
-0.017532533,
-0.0060140365,
0.0220829938,
0.0235954616,
-0.0056770942,
0.038978599,
0.0007895433,
-0.0285266284,
-0.0145486966,
0.0349875093,
0.0640345365,
0.002060391,
-0.0295520723,
0.0267464574,
0.012951022,
-0.0022272395,
0.0045440258,
0.0174271073,
0.006770852,
0.0060506668,
-0.0098079834,
-0.0198508296,
0.0381634161,
0.0390953422,
0.0422274172,
0.0191988256,
-0.016531989,
-0.0118112722,
0.00186273,
0.0237756427,
-0.0021927482,
0.0559591763,
0.0006426573,
0.0579575896,
-0.0473211072,
-0.0486328863,
-0.0682886392,
-0.071667254,
-0.0538866036,
0.0191437099,
0.0175777636,
-0.0431183465,
0.0311917495,
-0.037205819,
0.0011185628,
0.0091261063,
0.0047308924,
-0.0170630533,
-0.0442617498,
0.0225415844,
-0.0048847119,
-0.0405190922,
-0.0001341982,
0.0531164743,
-0.0064376676,
0.0885874107,
0.0174547844,
0.0241817944,
-0.0341321789,
0.0244188085,
0.065827325,
0.0127488459,
-0.0105647435,
-0.021484077,
-0.0226584114,
-0.054415904,
0.0019572859,
-0.0041330466,
0.0151281022,
-0.0313015766,
0.0097238515,
0.052834712,
-0.0384449996,
0.03204225,
-0.0242847167,
-0.0105363829,
0.0037018808,
0.0140535152,
-0.0361338742,
-6.169899484e-33,
-0.0431489833,
0.0284951013,
-0.0008455359,
-0.0308208298,
0.0022493491,
0.0381894708,
-0.0098408712,
-0.0213603489,
0.0124593219,
0.0042251698,
-0.0312851444,
0.0125744445,
0.0307576731,
0.0267128535,
0.0625797436,
-0.0242640004,
0.0697325468,
-0.041581247,
-0.0141763045,
0.02815843,
0.0660879016,
0.0808729231,
0.098586753,
-0.0181656126,
0.0299040433,
0.011367606,
-0.0437643304,
0.0077361781,
-0.0340455696,
0.0420531482,
0.0045572938,
0.0172818918,
0.0051870067,
-0.0530652814,
-0.0760412067,
-0.0674537718,
-0.0424161963,
-0.045169428,
-0.017221706,
0.0368582606,
-0.0067628869,
0.0301858876,
0.0298708137,
-0.0046264222,
-0.0406298414,
-0.0006633973,
0.0155361397,
-0.0303111281,
-0.0525765419,
-0.0140799507,
0.0068595284,
-0.0033339669,
-0.0077992221,
-0.0207379293,
0.0326948613,
0.0409172662,
0.0031028218,
-0.0053030574,
-0.0735203698,
0.0263955276,
0.042360954,
0.0455637127,
0.0642876923,
0.0199003741,
0.0016433926,
0.0324822553,
0.0515729003,
0.0112290829,
0.0309843179,
-0.0363246016,
0.0723849013,
0.0067211469,
0.0197778512,
-0.016485367,
0.0777786523,
-0.0189149398,
0.0039182459,
0.0059723384,
0.0246250387,
0.0701082349,
0.0499334559,
0.0149117224,
-0.0356084816,
-0.051518064,
-0.0439030491,
-0.0490015261,
-0.0235357638,
-0.0116364397,
0.0197262932,
0.0269396417,
0.0036754238,
0.0048868977,
0.0268891025,
-0.0203315467,
-0.0251233075,
0.0112645142,
-0.0128378049,
-0.0178232957,
-0.0239817034,
0.032188423,
-0.0437305756,
-0.0632184148,
-0.0205742735,
0.0100122578,
-0.0055411207,
0.0045255716,
0.0330579467,
0.0617688298,
-0.079964228,
-0.000744242,
-0.0265215673,
0.0050795656,
0.0323274955,
0.0129074585,
0.0680876374,
-0.0148342447,
0.0001585269,
0.0581306964,
0.0501816906,
-0.0337461792,
-0.0327591263,
0.0021871957,
0.0098901913,
0.0331197977,
-0.0258791558,
-0.0162994098,
-0.0435640477,
-0.028447073,
0.070363,
-0.0276925173,
-0.0037239357,
0.012138186,
2.604e-7,
-0.0122118685,
0.0323061906,
-0.0182401557,
-0.0176438726,
0.0282232966,
0.0672071353,
-0.0687259808,
0.0062902267,
0.0041271923,
0.0433943197,
0.0491593331,
-0.0319319926,
0.0609053001,
0.017845396,
-0.0461802334,
-0.0106729837,
-0.0951068401,
0.021125894,
-0.0449365601,
-0.019424824,
0.0064426353,
0.1298285723,
0.0628936738,
-0.0132882269,
0.0291314218,
-0.0386441126,
0.0368298255,
-0.0042794528,
-0.0100806095,
-0.0119627761,
0.0216225311,
0.052110903,
-0.0086761583,
-0.0047099008,
-0.0406998284,
-0.0089667346,
-0.0469485037,
0.0120439585,
0.0015587511,
0.0028971091,
0.0070543732,
0.0223304797,
-0.0389093049,
-0.0374450795,
0.0391078778,
-0.0761728361,
-0.0221193787,
-0.0387819968,
-0.0077887499,
-0.0174113419,
0.0087571451,
0.0214301217,
0.0158217773,
0.0355798192,
-0.0226262137,
0.0075498135,
-0.0110299001,
-0.0676515475,
-0.0030265138,
0.03851711,
-0.0189907681,
-0.0307202619,
-0.0420518108,
0.0255945846,
0.0321787,
-0.1022562534,
0.0155620174,
2.547728574e-34,
0.0714161396,
-0.0080192592,
0.0059566405,
0.059817154,
0.0136595443,
-0.0273131113,
0.0325956158,
0.0505790859,
0.0245206859,
-0.0225599371,
-0.0375544466
] |
http://arxiv.org/abs/2402.01858v3 | http://arxiv.org/pdf/2402.01858v3 | Explaining latent representations of generative models with large multimodal models | 2024-02-02T19:28:33 | 2024-04-18T03:54:39 | Mengdan Zhu, Zhenke Liu, Bo Pan, Abhinav Angirekula, Liang Zhao | cs.LG | cs.LG, cs.AI, cs.CL, cs.CV | Learning interpretable representations of data generative latent factors is
an important topic for the development of artificial intelligence. With the
rise of the large multimodal model, it can align images with text to generate
answers. In this work, we propose a framework to comprehensively explain each
latent variable in the generative models using a large multimodal model. We
further measure the uncertainty of our generated explanations, quantitatively
evaluate the performance of explanation generation among multiple large
multimodal models, and qualitatively visualize the variations of each latent
variable to learn the disentanglement effects of different generative models on
explanations. Finally, we discuss the explanatory capabilities and limitations
of state-of-the-art large multimodal models. | Explaining latent representations of generative models with large multimodal models: Learning interpretable representations of data generative latent factors is
an important topic for the development of artificial intelligence. With the
rise of the large multimodal model, it can align images with text to generate
answers. In this work, we propose a framework to comprehensively explain each
latent variable in the generative models using a large multimodal model. We
further measure the uncertainty of our generated explanations, quantitatively
evaluate the performance of explanation generation among multiple large
multimodal models, and qualitatively visualize the variations of each latent
variable to learn the disentanglement effects of different generative models on
explanations. Finally, we discuss the explanatory capabilities and limitations
of state-of-the-art large multimodal models. | [
-0.0093267867,
0.1004525274,
-0.0167845413,
0.0371152945,
-0.0041598906,
0.0169627368,
0.0682639703,
-0.0237111524,
-0.0657062382,
0.0118406704,
0.0173191726,
-0.021863902,
-0.0061326842,
0.0503996015,
0.0639756545,
-0.0773427039,
-0.0128344633,
-0.0514315851,
-0.0241694357,
-0.0193124153,
0.025593441,
0.0022090187,
-0.0264059696,
0.0540026985,
-0.0393535309,
-0.0463720597,
-0.0430952832,
-0.0020425557,
0.0031656232,
0.0046740067,
-0.0571487732,
0.0105013093,
0.0794118419,
-0.0063148369,
0.0000018983,
0.0042784512,
-0.0024941862,
0.0092936074,
-0.0179221183,
0.0275307614,
0.0076290127,
-0.0231387038,
0.0199341234,
-0.0217544809,
-0.0168924965,
-0.0041460423,
0.0952283517,
0.0408803597,
0.0499862507,
0.0064159259,
-0.0014778483,
-0.0364749357,
-0.002122167,
-0.0108928066,
0.0538913496,
-0.0209358316,
-0.0257566199,
0.0362054855,
-0.007584264,
0.0288066212,
-0.0800597593,
-0.0057878536,
0.0255465414,
-0.0408863015,
0.0291927215,
0.0336934775,
0.0627193525,
-0.0533682033,
0.0252552126,
0.07330098,
-0.0019829576,
-0.0035592713,
-0.007165581,
-0.014468574,
-0.0042261388,
0.0959551483,
-0.0813003927,
0.0176160373,
0.0190349855,
0.0149906259,
-0.0181255061,
-0.0022197301,
0.0415581092,
-0.0042098593,
-0.0157706346,
-0.0059830453,
0.0189359728,
0.0270182714,
-0.0831336752,
-0.0022592198,
-0.0562546402,
-0.0289987363,
-0.0175792146,
0.0297993049,
-0.0176790114,
-0.0031584189,
0.0208263546,
-0.0273575392,
-0.0303773619,
0.0306054223,
0.0322096609,
0.0423899479,
-0.0483295433,
0.0211579744,
0.0111653227,
0.0552850887,
-0.0611535646,
-0.020252306,
-0.0265859254,
-0.0102931056,
-0.0496377274,
-0.027884135,
0.0207446925,
-0.0172580667,
-0.0396644883,
-0.0452178568,
-0.0198281817,
0.0386344455,
-0.0177250691,
-0.0081666643,
0.0156774223,
0.0610879809,
-0.0300579667,
-0.0055729281,
-0.0404556282,
0.0497868881,
-0.076730296,
0.0223576725,
0.031255316,
-0.0071506901,
0.0009979523,
0.0335286707,
0.0316861421,
-0.0202296264,
-0.0218018498,
0.0838569626,
-0.014660676,
0.0137428716,
-0.0113613456,
-0.0386166908,
-0.0212965272,
-0.0603011139,
0.0510467403,
-0.0085076578,
0.029605329,
0.0096675837,
-0.002067605,
-0.0284684058,
-0.0081055239,
-0.0196654219,
-0.0148068033,
0.0934632793,
0.0015832918,
0.0093289837,
0.0237486009,
-0.0032478056,
-0.0163242184,
0.0194261037,
0.0163614377,
0.0303601995,
0.0086660841,
-0.0563608408,
0.0304828119,
-0.005569262,
0.0193847567,
-0.0140508479,
0.0120755974,
-0.0331099778,
-0.0310949367,
0.0395251475,
0.0531792417,
0.0054492652,
-0.0262730233,
-0.0030501874,
0.0602180213,
0.0105864061,
0.0491232835,
0.0195102785,
0.0552002564,
-0.0145432372,
-0.028799545,
0.0272267479,
-0.0259374604,
0.09151683,
0.0040789708,
0.0139802732,
0.06150534,
0.0288900342,
-0.024536185,
-0.0447321646,
0.0226564128,
-0.0001669175,
-0.0284042861,
0.0174495541,
0.0632392541,
-0.028952755,
-0.027496608,
-0.013643262,
0.0262528304,
-0.0252898093,
-0.0117730089,
-0.0316450521,
0.0533843562,
0.0678964928,
0.0483567752,
0.006188611,
0.0064129364,
-0.0053695017,
-0.0013724013,
0.0151674757,
-0.0356366932,
-0.0028626868,
0.0145080648,
-0.0319009982,
-0.0570216738,
-0.0603059605,
0.0167122465,
0.01655926,
-0.0202246215,
-0.0560857207,
-0.0123762367,
-0.0383425504,
0.0393704847,
0.007100089,
-0.0039225258,
-0.013263152,
-0.0250802059,
0.0071361265,
0.0438545831,
0.0240418855,
0.0160701536,
0.0204922985,
0.014220207,
-0.0315224826,
-0.0048154043,
0.0356159955,
-0.0191690661,
-0.0363613442,
-0.0173233971,
-0.0839919895,
0.0037111102,
-0.0212574061,
-0.0411948822,
-0.049885489,
0.0594326369,
-0.0279272236,
0.0032203561,
-0.0451127291,
-0.0300203227,
0.0351222418,
0.0381935239,
-0.0145772686,
0.061461661,
-0.0346684605,
-0.0189657137,
0.0235036183,
0.049776312,
0.0212856121,
0.0118069788,
0.0416736677,
0.0107563445,
-0.0105576841,
-0.010280489,
0.0103316391,
-0.0389451385,
0.006816728,
0.0146625033,
-0.0366798639,
-0.0230267644,
0.0022028834,
-0.019938862,
0.0229229331,
-0.0249911658,
0.0242459495,
0.0058222828,
-0.0178352278,
0.0475917049,
0.0308602098,
0.0091738179,
0.0227574706,
-0.0224950872,
-0.0073891995,
-0.024602795,
0.0129074305,
-0.011887745,
0.1189760119,
0.0183232632,
0.0071324054,
-0.0400354527,
-0.0225505419,
-0.0257604383,
-0.0103131831,
-0.015905397,
-0.0365966186,
0.0545763969,
-0.037130177,
0.0574264526,
0.0463556051,
0.0022136581,
0.0873637646,
0.0211840943,
-0.0106748696,
-0.0400856622,
-0.0261697602,
-0.0362637229,
-0.0135938264,
0.0074469871,
-0.0622665584,
-0.0438148677,
0.056948442,
0.0251921657,
-0.0158194751,
-0.0505391136,
0.0632358789,
0.0307387989,
-0.0415462144,
-0.0261493269,
-0.0167160947,
0.0145445606,
0.0595788695,
-0.0477406196,
-0.0708219111,
0.0452377275,
-0.0219870191,
0.0098758377,
0.0548972264,
0.0128858257,
-0.0083843488,
0.0815258697,
0.0256919619,
0.0334012434,
0.0083490396,
0.0117912488,
-0.0410209894,
0.0253221951,
0.0630396381,
0.0100768469,
-0.0032605303,
-0.0512992144,
-0.0185288358,
0.0294834841,
0.0024036057,
-0.010649953,
-0.0164568406,
0.0102841156,
0.0082006799,
0.0251353942,
-0.003797018,
-0.0225964673,
0.0051644235,
0.0338668488,
-0.0475071371,
-0.0190678407,
-0.0048831729,
-0.0177362189,
0.0437307134,
0.0393613391,
0.0169242043,
0.0118650468,
-0.0898896456,
0.0063594589,
0.0055752238,
-0.0144999949,
-0.0097255679,
-0.0689474866,
0.0298494119,
-0.026079163,
0.0537302755,
-0.0015488981,
0.0261959098,
0.0607774593,
-0.0253574383,
0.0164674632,
0.0544103421,
-0.0037527354,
-0.0053915563,
0.0330032445,
0.0047967918,
0.012228894,
0.0127446409,
0.013367991,
-0.0411899462,
-0.0342909805,
-0.0327656157,
-0.0215671062,
0.0322964229,
0.0035204329,
0.0140513172,
-0.0679897889,
0.0707981586,
-0.0011294412,
-0.0165313315,
-0.047111053,
-0.0119432947,
-0.02944047,
-0.0138137378,
-0.0330527835,
-0.0039852145,
0.021169886,
0.0324001797,
0.0339080989,
0.037727911,
-0.0494438931,
-0.0037914885,
-0.0269606393,
0.0222011935,
0.0619426295,
0.0211249869,
0.0054974426,
-0.0505600385,
-0.0205100998,
0.0166643728,
0.0300378539,
0.0080137048,
0.0578903221,
0.0336902142,
-0.0241244398,
-0.0446949266,
0.0722024515,
-0.0065544993,
-0.0331715345,
-0.0524218231,
0.053547021,
-0.0355951935,
-0.0161027405,
-0.028585026,
0.0194683019,
-0.0046080388,
0.0263529271,
0.0363641605,
-0.0413371101,
-0.0555777177,
-0.0418957323,
0.0129568921,
-0.0198585838,
-0.003245139,
0.0015060836,
0.0519154258,
-0.0081140045,
-0.0411647409,
0.0086703794,
0.0127605526,
0.0255968217,
-0.0090166973,
-0.009117458,
-0.050802011,
-0.0133277634,
-0.0210819114,
0.0536727495,
0.036867775,
-0.0049626846,
0.017194489,
-0.0308588184,
0.0093833534,
-0.0863204673,
-0.024340231,
0.0074026375,
-0.0613906384,
-0.0316488072,
-0.0269527696,
0.0469812639,
0.0241726432,
0.0256320611,
0.0416273475,
-0.073875308,
-0.0113606509,
0.0111886356,
-0.020523509,
-0.0072528636,
-0.0097333621,
0.0407625288,
0.0008362515,
0.0085129077,
-0.0217007548,
0.027739238,
-0.0107186176,
0.0159522407,
-0.0263728593,
-0.0200821199,
0.0308318604,
-0.0311980825,
0.0345020965,
-0.0386415906,
-0.0231056884,
-0.0582616776,
-0.0244226884,
-0.0119316848,
-0.0000338299,
-0.0464245155,
0.0060326904,
0.0354514197,
-0.0131250359,
0.0006221621,
-0.024645349,
0.0243262853,
0.0286428593,
-0.0467433669,
-0.0023927672,
0.0251319166,
-0.0177650359,
-0.0268767346,
0.0386657305,
0.0191567186,
0.0211143941,
-0.0565385632,
0.0147235766,
0.015530563,
0.1208138391,
0.0108039435,
-0.0020119112,
-0.0263764802,
0.0015503907,
-0.0691561028,
-0.0286402442,
-0.0155540546,
0.0351506583,
0.0459121428,
0.025732778,
0.0288915001,
0.0019149195,
-0.0031567297,
0.0062623713,
0.1006463617,
0.052380804,
-0.026903566,
0.052781906,
0.0070734839,
0.0042358795,
-0.044331342,
0.0640239939,
0.0426562354,
0.0028925368,
-0.0446073264,
0.0037376052,
-0.0201931428,
-0.0208998807,
0.0277472176,
0.0363690406,
0.0133238146,
0.0502010323,
-0.0269361064,
-0.040865507,
-0.0649204925,
0.0370216705,
-0.0299410112,
0.0079034856,
0.0236309003,
0.0523128472,
-0.0163520314,
-0.0560563691,
-0.0390461981,
0.0241381284,
-0.0302502271,
0.0216247533,
0.0005089928,
-6.298963621e-33,
-0.0203178525,
-0.0768122002,
0.0215194039,
-0.0428585708,
-0.0196605753,
-0.0071577593,
-0.046806246,
-0.0158625823,
0.0086199539,
0.0346017927,
0.0462050773,
-0.0203458294,
-0.0169900171,
0.0433303975,
0.0280692037,
-0.0058241049,
0.0363827273,
0.012914611,
-0.0006508611,
0.0380733646,
0.0522366464,
0.0077568083,
0.0157666355,
-0.0556774959,
0.0187539551,
-0.0187398251,
-0.0487003624,
-0.0004349065,
-0.0177180767,
0.0567620434,
-0.0734456703,
0.0047003967,
0.0225548223,
-0.0434685126,
-0.0137515459,
0.008358351,
-0.056327872,
0.0112074064,
0.0216571279,
-0.0139818145,
-0.0972533822,
-0.0587512366,
0.0101948874,
-0.0184117034,
0.0115523823,
-0.0443186425,
0.0500932336,
-0.0080005415,
0.0037072005,
-0.0137056122,
-0.0667602494,
-0.0034767776,
-0.0698219463,
-0.021135224,
-0.0035354833,
0.0398734584,
0.0250322428,
0.0414001569,
-0.0008131603,
0.0213528313,
0.019050505,
-0.0237806085,
0.1026607454,
0.0363081358,
0.0321408026,
0.0497503467,
0.0151026035,
-0.0484641716,
-0.0887780413,
-0.0055002947,
-0.0342935473,
0.0927504152,
0.0557577983,
0.0783475339,
0.025430629,
0.0450438447,
-0.0451112315,
0.0367992967,
-0.0634390935,
0.0145722721,
0.0558307581,
0.035225749,
0.018223064,
-0.0055257934,
-0.0466471426,
-0.0246796664,
-0.0456560999,
-0.0256291553,
0.0288861711,
-0.0214337092,
-0.0566048138,
-0.0040325243,
0.0493641309,
-0.0734537095,
0.0082675423,
0.001862303,
-0.0121252853,
0.0346752703,
0.0072559691,
-0.0161844008,
0.0157757346,
-0.0123564983,
0.0566995442,
-0.0143975457,
0.0340288132,
0.0197403822,
0.0076460247,
0.0113340495,
0.0093033267,
-0.01056235,
-0.0717853755,
-0.0427158475,
-0.0287304539,
-0.0302835628,
0.0260394905,
0.0025747432,
-0.0662750825,
-0.0130082536,
0.0147458846,
0.0409292169,
0.0176614188,
-0.0385547504,
-0.0247929469,
0.0081154518,
-0.0268430151,
0.014080965,
-0.0312861651,
-0.0833459496,
0.0586189739,
-0.0617741868,
-0.0310412962,
0.0515431874,
2.507e-7,
0.0471080355,
0.0195970684,
0.0054801456,
-0.0054092309,
0.0550242513,
0.0314686298,
0.0106074726,
-0.0209348164,
-0.0065584215,
-0.0430461913,
0.0136019103,
0.0616104193,
0.0299187042,
0.023177227,
0.0405679122,
0.0070529338,
-0.0131171057,
-0.022200333,
-0.0398671217,
0.0249909014,
0.0834008604,
0.0526375249,
-0.0370508656,
-0.0177319534,
0.0714898482,
-0.0038112891,
0.1002236158,
-0.0213017855,
0.0013601052,
-0.0325506106,
0.0050214306,
-0.0349591188,
0.016747199,
-0.0000732709,
-0.0029604351,
-0.0560379736,
0.0281271618,
0.037822973,
-0.014322063,
-0.0382173657,
-0.017192835,
-0.0009070861,
0.0031885784,
-0.0602972209,
0.0666440278,
-0.0366711617,
-0.0613614991,
-0.0087426258,
-0.0285630822,
-0.0157265831,
0.0001763388,
-0.0254578199,
0.0122892754,
-0.0025802797,
-0.026466934,
-0.0184176657,
0.0050887917,
-0.043974977,
-0.0187870525,
0.0147769405,
-0.0710917562,
-0.0008928879,
-0.039209567,
-0.0436029881,
0.0814081207,
0.0121074319,
-0.0666487217,
1.819766733e-34,
0.0130936587,
0.015503672,
0.0009339685,
-0.0121873543,
0.0416980237,
-0.02178986,
-0.0402932614,
-0.0157242846,
0.0217405371,
-0.0714119747,
-0.0330949873
] |
http://arxiv.org/abs/2402.01841v1 | http://arxiv.org/pdf/2402.01841v1 | COMET: Generating Commit Messages using Delta Graph Context Representation | 2024-02-02T19:01:52 | 2024-02-02T19:01:52 | Abhinav Reddy Mandli, Saurabhsingh Rajput, Tushar Sharma | cs.SE | cs.SE, cs.AI, cs.CL | Commit messages explain code changes in a commit and facilitate collaboration
among developers. Several commit message generation approaches have been
proposed; however, they exhibit limited success in capturing the context of
code changes. We propose Comet (Context-Aware Commit Message Generation), a
novel approach that captures context of code changes using a graph-based
representation and leverages a transformer-based model to generate high-quality
commit messages. Our proposed method utilizes delta graph that we developed to
effectively represent code differences. We also introduce a customizable
quality assurance module to identify optimal messages, mitigating subjectivity
in commit messages. Experiments show that Comet outperforms state-of-the-art
techniques in terms of bleu-norm and meteor metrics while being comparable in
terms of rogue-l. Additionally, we compare the proposed approach with the
popular gpt-3.5-turbo model, along with gpt-4-turbo; the most capable GPT
model, over zero-shot, one-shot, and multi-shot settings. We found Comet
outperforming the GPT models, on five and four metrics respectively and provide
competitive results with the two other metrics. The study has implications for
researchers, tool developers, and software developers. Software developers may
utilize Comet to generate context-aware commit messages. Researchers and tool
developers can apply the proposed delta graph technique in similar contexts,
like code review summarization. | COMET: Generating Commit Messages using Delta Graph Context Representation: Commit messages explain code changes in a commit and facilitate collaboration
among developers. Several commit message generation approaches have been
proposed; however, they exhibit limited success in capturing the context of
code changes. We propose Comet (Context-Aware Commit Message Generation), a
novel approach that captures context of code changes using a graph-based
representation and leverages a transformer-based model to generate high-quality
commit messages. Our proposed method utilizes delta graph that we developed to
effectively represent code differences. We also introduce a customizable
quality assurance module to identify optimal messages, mitigating subjectivity
in commit messages. Experiments show that Comet outperforms state-of-the-art
techniques in terms of bleu-norm and meteor metrics while being comparable in
terms of rogue-l. Additionally, we compare the proposed approach with the
popular gpt-3.5-turbo model, along with gpt-4-turbo; the most capable GPT
model, over zero-shot, one-shot, and multi-shot settings. We found Comet
outperforming the GPT models, on five and four metrics respectively and provide
competitive results with the two other metrics. The study has implications for
researchers, tool developers, and software developers. Software developers may
utilize Comet to generate context-aware commit messages. Researchers and tool
developers can apply the proposed delta graph technique in similar contexts,
like code review summarization. | [
0.0442181714,
0.0270102676,
-0.0154196471,
-0.0271027163,
-0.0228288323,
-0.0303540658,
-0.0242032371,
0.0086571388,
-0.0298377369,
-0.0236988179,
0.0154936956,
0.0029455589,
0.0226634499,
0.0256281458,
-0.0133691533,
-0.0569967851,
0.0212005079,
0.0073813903,
0.014040201,
0.0659391582,
-0.0032157807,
0.0145668071,
0.0230326857,
0.0428925678,
-0.0213784948,
-0.0429988354,
-0.000873205,
0.0575680509,
0.0600776821,
-0.0071486207,
-0.0172534529,
0.0634322912,
-0.0432302281,
0.0341477096,
0.0000021327,
-0.0223361719,
-0.0133831678,
-0.0039764554,
-0.0096433591,
0.0460882224,
-0.0035521337,
-0.0460895598,
-0.031655062,
0.0141023928,
-0.0207210761,
0.0164272897,
0.0734395981,
-0.0095068347,
0.1187804341,
0.035686519,
0.0088389609,
-0.0024242904,
-0.0188718382,
0.0188361034,
0.0333734564,
0.0249015037,
-0.0486168899,
0.0554529913,
-0.015163905,
0.0342750959,
0.0125651993,
-0.001004569,
0.0109456256,
-0.0275331326,
0.0505262502,
0.0342393145,
0.0300638154,
-0.0212997552,
0.022805484,
0.0166134741,
-0.0260166898,
-0.0342211053,
0.0128437672,
-0.0283030737,
-0.0376196615,
-0.0103640528,
-0.0624717511,
0.0027388677,
0.0049356152,
-0.0188416,
-0.0628404915,
-0.0174812451,
-0.0010770364,
-0.0679668412,
-0.0109998249,
-0.0312611088,
-0.0049068173,
0.0261719432,
-0.0636945814,
-0.0172180571,
-0.0175417289,
-0.0617438257,
-0.0270445067,
0.0215387139,
-0.0191519056,
-0.0102807945,
0.0039785588,
-0.0767425448,
0.0762275383,
-0.0222690273,
-0.0204241518,
0.0051423889,
-0.0222428218,
-0.0249061994,
-0.0202204529,
0.072708942,
0.0321670435,
-0.0052409428,
0.0387776159,
0.0034641228,
-0.0571975335,
-0.0009734032,
-0.0176963005,
-0.0104760658,
0.0490329377,
0.0355990864,
-0.0097502638,
0.0703036413,
-0.0025174145,
0.0703464821,
0.0333421938,
-0.0488822423,
-0.0687364787,
0.0232291538,
0.0553181395,
-0.0417956784,
-0.0266959406,
-0.0522377975,
-0.0033706152,
0.0118999137,
-0.0326581486,
-0.0156481173,
-0.0148946233,
0.0007035433,
-0.0351261087,
0.0597504899,
0.0476224497,
0.0449232757,
-0.0004670632,
-0.0368919782,
-0.0318082869,
0.0121490648,
-0.0684033483,
-0.066320233,
-0.0495570675,
0.032424964,
0.0085055409,
-0.0768209323,
-0.0255497918,
-0.013799875,
0.0204469319,
-0.0063601574,
-0.094034262,
0.0305214785,
-0.000630943,
0.0271096285,
-0.0554623678,
-0.0185193643,
0.0160120837,
0.0497144684,
0.0303859171,
-0.0422327481,
0.055147849,
-0.0243587308,
-0.0155031579,
0.0411480553,
0.0096808178,
-0.0014475969,
0.0329956003,
-0.0379846208,
0.009393448,
0.0225925092,
0.002175282,
-0.0321240462,
0.0121368812,
-0.0002080999,
0.0247932933,
0.032585267,
-0.0011764499,
0.0678370371,
-0.0323235802,
-0.0149554331,
-0.0146696027,
0.0712150484,
0.0172373876,
0.0283680111,
-0.0274983365,
0.0155253354,
0.0064589046,
-0.0132783279,
-0.0183273386,
0.0035994309,
-0.0552664995,
-0.0445473939,
0.0017316674,
0.0111283027,
-0.0118183335,
0.0448631793,
0.0399281457,
-0.0660520718,
-0.0565877333,
-0.0037168013,
0.0320680812,
0.0578167848,
0.055528339,
-0.0312827751,
-0.0499501377,
-0.0226316061,
-0.0102022793,
-0.0109048141,
0.0612684488,
0.0004036661,
-0.0227526538,
-0.0401141606,
-0.0166803449,
-0.0300075002,
-0.0136341164,
0.0279459357,
-0.0102627138,
-0.003865683,
0.0414174572,
-0.0237255432,
-0.0287842378,
-0.0009310725,
0.0891869515,
0.0390556976,
0.0141848344,
-0.0361306593,
0.0111460835,
0.0362067185,
-0.0392894223,
0.000635849,
0.0454562381,
0.0308393501,
-0.0532727875,
0.0573952161,
0.0045645339,
0.0208763853,
-0.0010922272,
-0.0205430593,
0.003391949,
-0.0258134604,
-0.0522720776,
-0.0145863555,
0.0342999846,
-0.0181143433,
0.0324557647,
0.0271169953,
0.0220808648,
-0.0342009328,
-0.0330615304,
-0.0294255279,
0.0071489732,
0.0293540712,
-0.0160361603,
-0.0305978116,
0.0497404337,
0.0348065831,
0.0257662367,
-0.0159898587,
0.0349317789,
-0.0327965729,
0.0166631602,
0.014437844,
0.0287465006,
-0.0013235474,
-0.0333683267,
-0.0831103697,
-0.0479872264,
-0.0108539723,
0.0083891368,
0.033714205,
0.0378557332,
-0.0036326777,
-0.0012689404,
-0.0227067228,
-0.0014574853,
-0.0060105636,
-0.0126404036,
0.0306082517,
-0.004270588,
-0.0277321134,
-0.0496591032,
-0.0086948732,
-0.028473286,
0.059469495,
0.034465991,
-0.0570522882,
-0.0800487101,
0.0483591482,
-0.0091151018,
-0.0303982571,
0.0043467428,
-0.0351640582,
0.0128263449,
-0.0184949655,
0.0962092429,
-0.0006473335,
0.020008212,
0.0191036183,
0.0346062072,
0.0098035801,
-0.0518451072,
-0.0197682995,
0.0126521196,
-0.0516200922,
0.0260474347,
-0.0319573544,
0.079320848,
-0.0035546301,
0.0272249691,
-0.010055583,
-0.0473167561,
-0.0046051512,
-0.0030045116,
0.0555984005,
0.0490346476,
-0.0452536419,
0.0035018791,
0.0074475366,
-0.015819639,
-0.0699634999,
0.0311029777,
-0.0497824065,
0.0067889481,
-0.0150902085,
-0.0127354944,
-0.060882654,
0.0812881663,
-0.0217704494,
0.010358518,
0.0096382163,
-0.0099187838,
0.009981107,
0.0051753358,
0.0246533472,
0.0200222265,
0.0102968682,
-0.0337757617,
-0.008746637,
-0.0013169547,
0.0437977053,
0.0188714396,
0.0034805983,
-0.0338591598,
-0.0283615757,
0.0186567549,
-0.0259658676,
0.0514910705,
0.0806170478,
-0.01982628,
0.0227165818,
0.0532292165,
-0.0455456227,
-0.0155252032,
-0.0821745992,
-0.0270997211,
-0.0478360727,
-0.01861188,
-0.0192152541,
-0.0207180362,
0.0699539259,
0.0796614513,
0.0187953021,
-0.0526227616,
-0.0758200884,
-0.0242947005,
-0.0363406055,
-0.044417385,
-0.0032226592,
0.0356505699,
-0.0180775803,
0.0205236673,
0.0694158748,
0.0201150682,
-0.0724322945,
0.0295695756,
0.0016324894,
0.0726292655,
0.026720589,
0.0166658629,
-0.1251511276,
-0.0290981326,
-0.0297419447,
-0.0008350058,
0.0191604681,
-0.0061049787,
-0.0054534986,
0.0078457789,
0.0303834341,
0.0197495148,
-0.0112504736,
-0.0095554311,
-0.0614771098,
0.0145300804,
-0.0006907082,
-0.0723070502,
-0.0040072282,
0.025565682,
0.0309664719,
-0.0190780535,
-0.0320006981,
0.0174790602,
0.0145060094,
-0.0387106761,
-0.0081245378,
0.0414226502,
0.0112318154,
0.0399695821,
0.0245233364,
0.0491770543,
-0.0092872903,
0.0212114342,
-0.0182627477,
0.0339936018,
0.0112969987,
-0.0071853427,
0.0391077548,
-0.0513752028,
-0.0242229365,
-0.0379390903,
0.0158656798,
0.0290714633,
-0.0443068109,
-0.035863284,
-0.0113844099,
0.0013180214,
0.0081798024,
0.0152573055,
0.0765465721,
-0.0131403375,
-0.0415605158,
-0.0455828644,
0.0546147116,
0.0050561358,
0.0072833011,
-0.0838269293,
-0.0085159056,
-0.0209110472,
-0.0737943649,
-0.0195763074,
-0.0181098357,
0.0374235325,
0.0412467234,
-0.041866228,
0.003051135,
-0.0207781196,
-0.0635436848,
-0.0047117253,
-0.0091492282,
0.0169172268,
-0.0139244646,
-0.012703958,
0.0048127938,
0.065401651,
-0.0032976205,
-0.0107850069,
-0.0351701714,
-0.0318396166,
-0.0290508922,
0.030239325,
0.0232807901,
0.0015636027,
-0.0274421461,
-0.0314661227,
-0.0547675975,
0.0068361624,
0.0021486206,
-0.0005721223,
0.0160212051,
-0.0270150229,
-0.0131176626,
-0.0479388498,
-0.0584921204,
-0.0167406313,
-0.0132231666,
-0.0097135659,
-0.0257273465,
-0.0521520264,
-0.0044359877,
0.0028674137,
0.0190022197,
0.0573282093,
-0.004183786,
0.0040895361,
-0.0143559733,
0.0088167209,
0.0056619225,
-0.0039440384,
0.0369291268,
-0.0204722732,
-0.0604376346,
0.013247733,
0.0649362653,
-0.0563891158,
0.0067478227,
0.0019342239,
0.0167241041,
0.01329662,
0.0457846224,
0.0249194168,
-0.0113343252,
-0.0249493364,
-0.0313608237,
0.0171874613,
0.0156251192,
-0.0048782961,
0.0258108694,
0.0314436592,
-0.006814112,
0.0262947213,
0.0370395072,
-0.0797992051,
-0.0098115085,
-0.008109672,
-0.032918755,
0.0284240562,
0.0972596258,
0.0780984089,
-0.0215302482,
0.0132754399,
-0.0391251221,
0.0768531337,
0.0116972728,
-0.0221238751,
-0.0171382409,
-0.0252457075,
-0.0024487106,
-0.0105341151,
0.0413079411,
-0.0570134968,
-0.0097194817,
-0.0613252297,
-0.0392635874,
-0.0271244943,
0.0000817469,
-0.0055143377,
-0.0288397204,
0.0157663077,
-0.0615491457,
-0.0669981912,
-0.0016445045,
0.0565614477,
0.0227542166,
0.0458348282,
-0.013381199,
0.0053013675,
0.0686373785,
0.0315666348,
-0.0335553885,
0.0121736713,
-0.0041457205,
-0.0198829398,
-0.0125397379,
-0.0468391962,
-5.151183019e-33,
-0.0170579664,
-0.0586838797,
-0.0416935757,
0.0514819436,
-0.0252275616,
-0.0193015914,
-0.0719857365,
-0.0075025633,
-0.0034853132,
-0.0371260457,
0.0090087447,
0.0377474949,
0.001775121,
-0.0890785456,
0.0521504469,
0.0557382144,
0.0951205865,
0.0591158904,
-0.0395925604,
0.1065222695,
0.0149216847,
-0.0141916983,
0.0087680919,
-0.0515523963,
0.0243772548,
-0.0072341841,
0.0187517069,
-0.0168996733,
0.0521415398,
0.0233133081,
0.0337316804,
-0.0210647695,
-0.0051330868,
-0.002989497,
-0.0444848351,
-0.0498402528,
0.0075106621,
-0.0336746052,
0.0026338289,
-0.0021035126,
0.0534028672,
-0.0495193787,
0.0095710391,
0.0036561992,
0.0115773566,
-0.0323746391,
0.0214198437,
0.0408813506,
-0.0406413674,
-0.0363008305,
-0.0233749803,
0.0417285711,
-0.0579287708,
0.0571397915,
-0.0290430542,
0.0288363937,
0.0406577401,
0.0125938579,
-0.0259476192,
0.0155907972,
0.0338177197,
0.0135578457,
0.0457712077,
-0.0007213512,
-0.01812323,
0.005526681,
0.0112362886,
0.0627364293,
0.0081689432,
0.0480029471,
0.030212177,
0.0601490922,
0.0621556789,
0.0365174748,
-0.0238537155,
-0.0190247577,
0.0423144139,
0.0356892459,
-0.0243067611,
0.0313580148,
-0.0047575352,
0.0507535636,
-0.0357740074,
-0.0048954287,
-0.0060187005,
-0.0228663683,
-0.0011664998,
-0.053055428,
0.0159551799,
-0.0338115804,
-0.0083118733,
0.0401098542,
-0.0093598254,
-0.0138293952,
0.0621263832,
0.0483546071,
-0.0642585531,
0.0408699587,
-0.0321095735,
-0.0551761463,
0.0240529422,
0.0104728388,
0.0283326805,
0.0556630455,
0.079238236,
0.0161256026,
0.0385338217,
0.0166690852,
-0.0522084832,
0.0263756346,
-0.0115600545,
-0.0108760204,
-0.0900517032,
-0.0755989254,
-0.0314562209,
-0.0112311644,
0.0183771942,
0.0634968951,
0.0024589077,
0.0231128,
-0.0353565961,
0.079577975,
-0.0001771897,
0.0486277603,
-0.0240896381,
-0.0115407296,
0.034579549,
0.0131762624,
0.0514834076,
-0.0186842475,
-0.049955193,
-0.0112130567,
2.704e-7,
0.0029390519,
0.0666766316,
0.047622703,
-0.0112148067,
0.0111728283,
0.0500807427,
-0.0199807584,
-0.0107114483,
0.0439293683,
0.0442620553,
-0.016584361,
0.0052766982,
0.0176300928,
0.0565721169,
-0.010164774,
-0.0003123479,
-0.0635469258,
-0.024134893,
-0.0810760409,
0.0160241798,
0.0994361416,
0.0007022698,
0.0146387815,
0.022451384,
0.0223422926,
-0.0013093384,
-0.022121679,
-0.0153663773,
0.0222929977,
0.0657292232,
0.0083823446,
-0.0412059538,
-0.0027086472,
-0.0204358194,
-0.0403865427,
-0.074325785,
-0.0277830083,
0.03980783,
-0.0286787134,
0.0194875654,
0.0376376323,
0.0030148786,
0.0029417537,
-0.0326391049,
-0.0027146516,
-0.0042875158,
-0.0010318443,
-0.0220160745,
-0.0244114827,
-0.0254511088,
0.0150795048,
0.0209297147,
-0.0115262661,
0.0038110935,
-0.0219702125,
0.0082778269,
-0.023466751,
-0.0386078618,
-0.0510018617,
0.0301979594,
0.0159755945,
-0.0219582245,
-0.0084162001,
0.0057038926,
0.0207994934,
0.0264073461,
0.0386726558,
2.470285537e-34,
0.0708188936,
-0.0320866071,
0.0278872531,
-0.0291375332,
0.0189369842,
-0.021530902,
0.047261551,
0.0208428688,
0.0178454313,
0.0182296112,
-0.0232298784
] |
http://arxiv.org/abs/2402.01830v2 | http://arxiv.org/pdf/2402.01830v2 | PiCO: Peer Review in LLMs based on the Consistency Optimization | 2024-02-02T18:49:26 | 2024-04-20T09:09:38 | Kun-Peng Ning, Shuo Yang, Yu-Yang Liu, Jia-Yu Yao, Zhen-Hui Liu, Yu Wang, Ming Pang, Li Yuan | cs.CL | cs.CL, cs.AI, cs.LG | Existing large language models (LLMs) evaluation methods typically focus on
testing the performance on some closed-environment and domain-specific
benchmarks with human annotations. In this paper, we explore a novel
unsupervised evaluation direction, utilizing peer-review mechanisms to measure
LLMs automatically. In this setting, both open-source and closed-source LLMs
lie in the same environment, capable of answering unlabeled questions and
evaluating each other, where each LLM's response score is jointly determined by
other anonymous ones. To obtain the ability hierarchy among these models, we
assign each LLM a learnable capability parameter to adjust the final ranking.
We formalize it as a constrained optimization problem, intending to maximize
the consistency of each LLM's capabilities and scores. The key assumption
behind is that high-level LLM can evaluate others' answers more accurately than
low-level ones, while higher-level LLM can also achieve higher response scores.
Moreover, we propose three metrics called PEN, CIN, and LIS to evaluate the gap
in aligning human rankings. We perform experiments on multiple datasets with
these metrics, validating the effectiveness of the proposed approach. | PiCO: Peer Review in LLMs based on the Consistency Optimization: Existing large language models (LLMs) evaluation methods typically focus on
testing the performance on some closed-environment and domain-specific
benchmarks with human annotations. In this paper, we explore a novel
unsupervised evaluation direction, utilizing peer-review mechanisms to measure
LLMs automatically. In this setting, both open-source and closed-source LLMs
lie in the same environment, capable of answering unlabeled questions and
evaluating each other, where each LLM's response score is jointly determined by
other anonymous ones. To obtain the ability hierarchy among these models, we
assign each LLM a learnable capability parameter to adjust the final ranking.
We formalize it as a constrained optimization problem, intending to maximize
the consistency of each LLM's capabilities and scores. The key assumption
behind is that high-level LLM can evaluate others' answers more accurately than
low-level ones, while higher-level LLM can also achieve higher response scores.
Moreover, we propose three metrics called PEN, CIN, and LIS to evaluate the gap
in aligning human rankings. We perform experiments on multiple datasets with
these metrics, validating the effectiveness of the proposed approach. | [
0.0370526165,
0.0550009944,
0.009637747,
0.0351690762,
-0.0233413,
-0.0104037179,
0.0164211039,
-0.0075669242,
0.023680279,
-0.0551023521,
-0.0911118686,
-0.0144555178,
-0.0089194458,
-0.0183018316,
0.0176899601,
0.0032664174,
0.0042148437,
-0.0360985547,
0.0155781191,
-0.0362402387,
-0.0292991884,
-0.0356132388,
0.0021992952,
0.0379545689,
0.0812863037,
-0.0153495735,
0.0469613895,
-0.0396504216,
0.0158842374,
0.0167794563,
0.0144589646,
0.0768679082,
0.0015817386,
0.019485496,
0.0000020987,
-0.0655055419,
-0.0109698856,
0.0201233495,
-0.0114773829,
0.0364177898,
-0.0087634362,
0.0411746204,
0.0098416647,
0.0080803381,
-0.0616352446,
-0.0151406741,
-0.0142384535,
-0.0351997316,
0.0966652781,
0.0269772317,
0.0487467311,
-0.0075382506,
0.0262326319,
-0.0385844857,
0.0076048835,
-0.0556267984,
0.0255051851,
0.044145871,
-0.004220834,
0.0510191247,
-0.0149616478,
0.0188577455,
0.0144621879,
-0.0402006097,
0.0469006673,
-0.0063605988,
-0.0523391217,
0.0157820266,
0.0122936089,
0.0829698518,
0.0106481314,
-0.0080976719,
-0.0053094183,
-0.0092930505,
0.0088188434,
0.0364878289,
-0.0597363636,
0.0208948385,
0.0407016389,
-0.0115134446,
-0.0228533652,
0.0234170947,
0.0299648214,
0.0153092965,
0.017519081,
0.1119101197,
0.0127824275,
-0.0292374808,
0.020838581,
0.0042268685,
-0.0284855608,
-0.0464803912,
-0.0168419778,
0.0218862388,
-0.0478766263,
0.006339598,
0.0382472351,
0.0029112338,
0.0431451835,
-0.0307727885,
0.0503339395,
-0.0287230238,
0.0353870764,
0.0644338876,
-0.0216052476,
0.0654655918,
0.0114334989,
0.0424294546,
-0.0417385511,
-0.0101939822,
-0.0729185641,
0.0002956105,
0.0345055945,
0.0347812399,
0.0029447041,
0.0294246078,
-0.0723873377,
-0.0029110613,
-0.0652778,
0.019808121,
0.0513374172,
0.0077863429,
0.0149184521,
-0.007441822,
-0.0590469129,
0.0249745566,
-0.0024041298,
0.0141954077,
-0.0219897497,
-0.0210751221,
0.0241606794,
0.0221625008,
-0.0224944279,
-0.0088926181,
0.0181467272,
0.0670824945,
0.0030818169,
0.0193665083,
0.0414328575,
-0.02141078,
0.0153236967,
-0.0712631717,
-0.0206089523,
-0.0810882449,
0.0289374962,
0.018129006,
-0.0014174609,
-0.0319331624,
0.0054467195,
0.0154296095,
0.0058113127,
0.0268198363,
0.0124347666,
0.0102353403,
0.0341023393,
0.0258856416,
-0.0318583138,
0.0263073817,
0.0112053677,
-0.0081965178,
0.0392443836,
-0.0114062838,
0.012234848,
0.0354519449,
0.0248364061,
0.0063687186,
0.011409671,
0.0659662187,
0.0126139307,
0.0344888791,
0.0008781723,
0.0302235521,
0.0016462291,
-0.0341982692,
0.0467186347,
0.0462489389,
0.0187695771,
-0.0025060605,
0.0220059212,
0.0289015733,
0.0278639551,
0.0064165466,
-0.0376770981,
0.0342810564,
-0.0069463169,
-0.0127715487,
0.0474770181,
0.0053749327,
-0.0108200852,
-0.0540056005,
-0.0121690035,
-0.0010307264,
-0.0489789508,
0.0233027283,
0.059309449,
-0.0275256112,
0.0132403262,
0.0647142157,
-0.0076176482,
-0.0509059653,
0.0298358966,
-0.040124502,
0.0250273552,
0.0250221957,
0.0522828735,
-0.0234696958,
-0.0041477885,
-0.0082645323,
-0.0100991996,
0.0244937353,
0.023010999,
-0.0320962518,
-0.0118725765,
-0.015298944,
-0.0430252925,
0.0185408629,
-0.0143361883,
0.0032513971,
0.0096260188,
-0.0248067081,
0.0085176025,
-0.0767237842,
-0.0563460402,
0.0150352027,
-0.0330295935,
-0.0094895968,
0.0140365018,
-0.0159324631,
0.0485005975,
0.0259925183,
-0.0027806568,
0.0339462012,
0.0346001983,
-0.0268565975,
-0.0447383411,
-0.0512780994,
-0.0179051459,
0.009759862,
0.0352824777,
-0.0632311478,
-0.0278063007,
0.0512639359,
-0.0318757705,
-0.0460495949,
0.0138123026,
-0.0111679314,
0.0056062425,
-0.0394697152,
-0.0006878378,
-0.0277925003,
-0.0659158379,
-0.0637326315,
0.0610919669,
-0.0192755088,
0.073617883,
0.0315416194,
0.000725446,
0.0284155402,
-0.0104197916,
-0.0364199504,
-0.0109926667,
0.0055410215,
0.047818955,
0.0493692197,
0.020273352,
-0.0548240207,
-0.0431243256,
-0.0804598033,
0.0078412807,
-0.0166616179,
0.0208131243,
0.0201786757,
-0.0371937715,
0.02252266,
-0.0010153034,
-0.0643424839,
0.010946244,
0.011695534,
-0.0052348385,
0.0232503228,
0.0452190936,
0.0061109881,
-0.0459089279,
-0.0266858246,
-0.0087517248,
0.0319887251,
0.0227541104,
0.0389705151,
-0.0380635262,
0.0526745208,
-0.0655158237,
0.0203067567,
-0.0065112039,
-0.0298291463,
0.0140853822,
0.0049992213,
-0.0163962394,
0.0588350408,
-0.0010638707,
-0.0505922735,
0.0643750578,
0.0119542787,
-0.0688626766,
-0.015556776,
-0.0098620383,
-0.0535063185,
-0.0121583771,
-0.0222481992,
0.0242870469,
0.0434209593,
0.0173243638,
-0.0464035347,
-0.018423358,
0.0242407285,
-0.0492853299,
-0.0163484141,
-0.0450795665,
0.0133557785,
-0.001497948,
0.0296872444,
-0.0094432915,
-0.0625339076,
0.064117603,
-0.0440300256,
0.008522721,
-0.0122207487,
0.0068643861,
-0.0913566351,
0.042889677,
0.017905185,
0.0064273742,
-0.056925416,
-0.0029535051,
-0.0592785589,
0.0183313861,
0.0478205755,
-0.0624144375,
0.01737497,
-0.041995924,
0.0057334583,
-0.0063261818,
-0.0435095318,
-0.0068574091,
-0.0215301551,
-0.017069906,
0.0420544893,
0.0139338858,
-0.0130843613,
-0.0119567104,
0.0034175799,
0.0136475153,
0.0130520528,
-0.0151141668,
-0.0000534492,
-0.0187841244,
0.0254172049,
0.0307494346,
0.0148144457,
-0.0077332119,
-0.033544194,
0.0044773403,
0.0521863103,
0.0409161486,
0.0128395976,
-0.0060480456,
-0.0228692144,
0.0432817675,
-0.0436294526,
-0.0265082587,
0.0210289899,
0.0187493134,
-0.0338539965,
0.020527225,
0.1255189031,
0.0278942734,
-0.0662153512,
-0.0252532158,
-0.0299397502,
0.0056603318,
0.0460315011,
0.0258172993,
-0.0413923338,
-0.0062321988,
-0.0311092362,
0.0459901839,
0.0078043505,
-0.0317712873,
0.0077902982,
-0.1571540385,
0.0216763765,
-0.006350406,
-0.0158102382,
0.0131714093,
-0.032497365,
0.0192199834,
-0.0363174081,
-0.0212621409,
0.0239136051,
0.0050597647,
0.0187445469,
0.0120969051,
-0.0537441038,
-0.0563004091,
0.080611892,
-0.0380964018,
0.0895831212,
0.0392696001,
0.0089094741,
-0.0315266624,
-0.0647729337,
0.0478648543,
-0.0341543369,
-0.0101849996,
-0.0047376482,
0.0167532172,
-0.0266746413,
0.008895373,
-0.0361149386,
0.0124208769,
-0.0088698883,
0.0291840751,
0.0284665152,
0.0796449557,
0.0062628384,
0.004098583,
-0.0041438863,
-0.0057807099,
-0.0215198584,
0.0762308612,
0.0058920202,
-0.0583757944,
0.0038779036,
-0.0121046165,
-0.0065627676,
-0.000729869,
-0.0612625554,
0.0380005389,
0.012286718,
0.0007253852,
-0.0604591705,
0.0081174644,
0.0197776202,
0.0004346049,
0.004876269,
0.0469114184,
0.0194300245,
-0.0032564206,
-0.0206619445,
0.0145009225,
-0.0455005169,
0.0104061589,
-0.083544746,
0.0000667706,
0.0694375634,
0.0107184527,
-0.0398943163,
0.0119099459,
-0.0591303706,
-0.0373607352,
-0.0369895324,
0.0756739751,
0.0220240224,
-0.0110703064,
-0.0160497371,
-0.0984744579,
0.0212284755,
0.0021989685,
-0.0159074254,
0.012148425,
-0.0015647558,
-0.000813943,
-0.0212095305,
0.0252052564,
-0.0798287988,
0.0005546036,
0.0339447372,
0.0225267801,
-0.0156676453,
-0.0738975853,
0.0147894844,
0.0205130503,
0.0255783778,
0.0155814765,
0.0116440766,
0.0007462339,
0.0239603668,
-0.0210960377,
-0.0335342288,
-0.0247413795,
-0.0118786218,
0.0332955867,
-0.0348252878,
0.0365657471,
0.0116746947,
-0.0033344869,
0.0331191793,
-0.054627642,
0.039900817,
0.0319838896,
-0.0102090817,
-0.0079039969,
0.003188533,
-0.0290823374,
-0.0324937552,
0.0016497911,
0.0528481677,
0.0125340996,
0.0535944328,
-0.0055724727,
0.0162337516,
0.0253394395,
0.0311311074,
-0.0786139667,
-0.0451264642,
-0.0355993584,
0.0466857478,
0.042523969,
0.0726926029,
0.0395807847,
0.0227125399,
0.000199574,
-0.0390275046,
-0.0182152502,
-0.0088986503,
-0.0014120734,
0.0389354713,
0.0348457173,
0.002189029,
0.059418831,
0.0787692368,
0.0505215637,
-0.0186323542,
-0.0514289513,
-0.0031033107,
-0.0441881642,
-0.0117926979,
-0.0340836085,
0.0147140911,
-0.0196122304,
0.0109929834,
-0.0548928082,
-0.0488487259,
-0.0057207854,
0.0757692158,
-0.0058038658,
0.0035930383,
0.0237360224,
0.0621866994,
-0.0319496617,
-0.0507003888,
-0.0651691034,
-0.0027761278,
-0.031130394,
-0.0023271658,
-0.0201540664,
-6.205647002e-33,
-0.0464428775,
-0.0662492812,
0.0014830992,
0.0055170087,
0.0033577278,
0.0161057804,
0.0020500142,
0.0159903709,
-0.0097162938,
-0.018980654,
-0.0038507127,
-0.0116121918,
-0.0043786094,
0.0035018977,
0.0491648912,
-0.0114426408,
0.0042981775,
-0.0314837657,
-0.0436043441,
0.0194872897,
-0.0038941228,
-0.0109853316,
0.0309896041,
-0.0605518892,
0.0744153187,
-0.0264762659,
-0.0055556418,
0.037502449,
-0.0258699451,
0.0664695203,
0.005566421,
0.0199430399,
0.039422106,
-0.058926139,
-0.0048469105,
-0.0327584185,
-0.0347788818,
-0.0224494562,
0.020214051,
-0.0199912246,
-0.042400144,
0.0016190861,
0.0624975897,
0.0108673396,
-0.0788951442,
-0.0200325958,
0.0512993038,
-0.0090391347,
0.0025927078,
-0.0167583842,
0.0216953233,
0.0174540728,
-0.0192948915,
-0.0046980241,
-0.0276378449,
0.0306387264,
-0.0236268211,
0.0331135988,
-0.0383020975,
0.0178849399,
-0.018200811,
0.0648820549,
0.0952792019,
0.0048716278,
-0.0086483639,
0.0605866276,
-0.0030489822,
0.0069332938,
0.0168402288,
-0.0878228173,
-0.0058929236,
0.0572056063,
0.0817590356,
0.0768503398,
-0.0369462594,
0.0414259546,
-0.083501108,
0.0644744039,
0.031454131,
0.0210023075,
0.0105861099,
0.0118608726,
-0.0226291884,
-0.0121955751,
-0.0782017708,
-0.036066696,
-0.0296240319,
-0.0145231113,
0.0331957713,
-0.0147090377,
-0.0288785622,
-0.0521916747,
0.0293638892,
-0.0602653548,
-0.0295988228,
0.0124425963,
0.0129672578,
0.059070427,
0.0134934168,
0.0230424106,
-0.07488437,
0.0106975846,
-0.0165611356,
0.0218396299,
0.0179265384,
0.0011562171,
-0.0087708682,
-0.0161850676,
-0.0484269448,
-0.0045082457,
-0.0404319018,
-0.0374844037,
-0.0270587467,
0.0235837922,
0.0231783818,
-0.0232969392,
0.0026874032,
0.0574748479,
-0.0300223958,
0.0191968866,
-0.0052012294,
0.0269403141,
0.0120809162,
-0.0206066109,
-0.021400135,
0.0195204932,
0.0246842522,
0.0210794378,
0.0068536219,
0.0195202213,
0.0039375946,
-0.0104770232,
2.779e-7,
0.0476195253,
0.0470154136,
-0.0249300916,
-0.0007489338,
0.0307309162,
0.0348447748,
-0.0381831937,
0.0178878643,
0.0480699465,
-0.0388571173,
-0.0023880838,
-0.0177899711,
0.0410113707,
0.01734595,
-0.0841937661,
-0.0032265678,
-0.0677175745,
0.0006838472,
-0.0086649433,
-0.0251260102,
0.0537974723,
0.0434065312,
-0.0073306123,
0.0041843778,
-0.0514600314,
-0.0683056489,
0.0656253248,
-0.0540108457,
-0.0610908866,
0.0363121964,
0.0365780368,
0.0227915533,
0.0159510709,
0.0331771672,
0.0022290612,
0.0191760026,
0.0603367426,
0.0868380591,
0.0199083705,
-0.0059333206,
0.0067169457,
-0.0215743612,
0.0206570886,
-0.0572740324,
0.0602428839,
-0.0350413956,
-0.0487769656,
-0.0294366647,
-0.0045635528,
-0.046451278,
0.0341951773,
-0.0209565256,
-0.0177479126,
0.0103205899,
-0.0029545526,
-0.0160525888,
-0.0163766462,
-0.0254115146,
-0.0427921042,
-0.0330599248,
-0.0397541672,
-0.0969805568,
-0.003496031,
-0.0087046809,
-0.0276535563,
-0.0435568951,
0.0112500284,
2.469251699e-34,
0.0351574421,
0.0009247425,
0.0367223024,
-0.012997116,
0.025919877,
0.022201689,
-0.0428787805,
-0.0020326015,
0.0058398792,
-0.078567639,
0.0078292852
] |
http://arxiv.org/abs/2402.01629v1 | http://arxiv.org/pdf/2402.01629v1 | Position Paper: Generalized grammar rules and structure-based generalization beyond classical equivariance for lexical tasks and transduction | 2024-02-02T18:44:37 | 2024-02-02T18:44:37 | Mircea Petrache, Shubhendu Trivedi | cs.CL | cs.CL, cs.LG, stat.ML | Compositional generalization is one of the main properties which
differentiates lexical learning in humans from state-of-art neural networks. We
propose a general framework for building models that can generalize
compositionally using the concept of Generalized Grammar Rules (GGRs), a class
of symmetry-based compositional constraints for transduction tasks, which we
view as a transduction analogue of equivariance constraints in physics-inspired
tasks. Besides formalizing generalized notions of symmetry for language
transduction, our framework is general enough to contain many existing works as
special cases. We present ideas on how GGRs might be implemented, and in the
process draw connections to reinforcement learning and other areas of research. | Position Paper: Generalized grammar rules and structure-based generalization beyond classical equivariance for lexical tasks and transduction: Compositional generalization is one of the main properties which
differentiates lexical learning in humans from state-of-art neural networks. We
propose a general framework for building models that can generalize
compositionally using the concept of Generalized Grammar Rules (GGRs), a class
of symmetry-based compositional constraints for transduction tasks, which we
view as a transduction analogue of equivariance constraints in physics-inspired
tasks. Besides formalizing generalized notions of symmetry for language
transduction, our framework is general enough to contain many existing works as
special cases. We present ideas on how GGRs might be implemented, and in the
process draw connections to reinforcement learning and other areas of research. | [
0.0392301641,
0.0756968483,
-0.0138693918,
-0.01390744,
-0.0460077226,
0.0220192634,
-0.0232174285,
-0.0123608643,
-0.0682135001,
-0.0345088765,
-0.0281938966,
-0.0738408715,
0.0227052737,
-0.0072229188,
0.0442385003,
-0.0831846818,
0.0670870841,
-0.0311838873,
-0.0425461754,
-0.0135362828,
-0.0221878625,
-0.0468884297,
-0.044097349,
0.0175256766,
-0.0181069691,
0.0307588,
0.0092412103,
-0.0304291043,
0.0220311116,
0.018414814,
-0.0188113451,
0.0156426039,
-0.0089907059,
0.0493928082,
0.0000014858,
-0.0579925738,
0.0288532265,
-0.019214496,
0.0191663858,
-0.0203573238,
0.0001919663,
-0.0615929477,
-0.0316608511,
-0.0272272881,
-0.0283540282,
-0.0233520363,
0.102475211,
-0.0406909138,
0.024154257,
0.07640104,
-0.0153677706,
-0.0876271501,
0.0285211038,
-0.0208319128,
0.0408218205,
-0.000606097,
0.0358802713,
-0.0473345369,
-0.0011549249,
-0.0425307862,
-0.0478041023,
-0.0024197602,
0.0372484289,
0.0390509106,
0.0192140993,
0.0485431775,
-0.0273860451,
-0.0101291426,
-0.016952185,
0.0479345359,
0.0104688788,
-0.0424895324,
-0.003088577,
0.0220632721,
-0.0255957544,
0.0334665254,
-0.0326506793,
0.0293965992,
-0.0070326254,
0.0660904124,
0.0467528775,
-0.056987904,
0.0180467311,
-0.0496489815,
-0.0203600619,
0.0393571667,
0.0430233181,
-0.0441776961,
-0.0047414051,
0.0372283757,
-0.0437165834,
-0.0154397758,
-0.0127803916,
0.0192485303,
0.001520324,
-0.022389682,
7.767e-7,
-0.0448414087,
0.0069591198,
-0.0296070371,
0.0564315505,
0.0242137723,
0.0339365602,
0.0355112925,
-0.0290632397,
0.0626933798,
-0.0407594703,
0.0596501529,
-0.0851067528,
0.0110926768,
-0.0582629852,
-0.0319027267,
-0.0304131601,
0.0104533136,
0.0336880572,
-0.003531449,
-0.0386678576,
0.0239348542,
-0.0200923085,
0.0321090631,
0.0306769684,
-0.0114592034,
-0.0526209064,
-0.0132953711,
0.0344170481,
-0.0041022021,
-0.0096134692,
0.0075153424,
0.0079990812,
0.0007747656,
-0.0330652334,
-0.0071020075,
-0.0012039253,
0.0285644438,
-0.0300937835,
0.0840868279,
0.0244305301,
-0.0048796628,
-0.0145557476,
0.0225643907,
-0.0211072024,
-0.0240671914,
0.0786982477,
-0.0146532832,
-0.010579369,
0.0108084474,
-0.0171373654,
-0.0451482832,
-0.009123574,
-0.0160400029,
-0.009008076,
0.0351173095,
0.084781602,
0.0277252439,
-0.0177757535,
-0.0379241519,
-0.0077107488,
0.067805782,
0.007722178,
-0.0544318072,
-0.0132928649,
-0.0275200233,
0.0466643088,
0.034926001,
0.0113645196,
0.0229189228,
-0.0360320844,
-0.065534234,
0.0426837727,
0.0241881069,
0.0273426678,
0.0638546869,
-0.0156372748,
-0.0567573123,
0.0421724953,
0.0584340282,
0.0170694329,
0.0481074564,
0.0674937218,
-0.0913641378,
-0.0182581358,
0.0631065667,
-0.0577083491,
0.0052381638,
-0.0280246977,
0.0354282111,
-0.0782852992,
0.0272878855,
-0.0128743742,
-0.0173948891,
-0.0159584638,
0.0022120648,
0.0349817537,
-0.041483745,
0.0544591211,
0.0511147454,
0.0038401191,
0.0233380292,
0.0132132825,
-0.0358324721,
0.0206837934,
-0.0043538297,
0.0370001346,
-0.0128238183,
0.0003155102,
-0.0512048714,
-0.0165757518,
0.0085494556,
0.0258615762,
0.0548446439,
0.0403796732,
0.0063230721,
0.0001613257,
-0.033373978,
-0.0039706808,
0.0299753584,
0.0552149229,
0.0053557991,
-0.0377551988,
-0.0289745238,
0.0037818006,
-0.0167421717,
0.0015406328,
0.040394064,
0.0151067134,
-0.0278980136,
0.0180132072,
0.0088175051,
0.0262014568,
-0.0423584133,
0.0090319989,
0.0595665947,
0.009051905,
-0.0445367359,
0.0107721873,
0.0734213516,
-0.0700469986,
-0.0305526517,
0.0004603107,
-0.0624869056,
0.0370552018,
0.0260736961,
0.0106403092,
-0.0445045009,
0.0926456302,
-0.0607066341,
0.0719234347,
-0.0065083732,
-0.0064127701,
0.022049848,
0.0042550946,
-0.0296581164,
0.034738481,
0.0243508648,
0.0256509222,
-0.0086571341,
-0.0735659599,
-0.0153534487,
-0.0325928889,
0.0013579194,
0.0480276793,
0.0058808145,
-0.0131885661,
0.0163426064,
0.0050307922,
0.0032660139,
-0.0392882451,
-0.0013397982,
-0.0458261184,
-0.0217000023,
0.0689581558,
0.0457980335,
0.0343349427,
0.0114570307,
0.0068357103,
-0.0053958339,
0.0225813184,
0.0179608557,
0.0091538494,
0.0467399508,
-0.0050843544,
-0.0655247495,
-0.0127900233,
-0.054063689,
-0.0478732474,
0.0599770583,
-0.0162760448,
0.0572686754,
-0.0692477301,
-0.0097239008,
0.0181580763,
0.0457010381,
-0.0105063934,
-0.0536960214,
0.0202506464,
-0.0057829032,
0.0012623743,
0.0857021958,
-0.0178733822,
-0.0383004695,
0.0124084726,
-0.0145804416,
-0.0089364257,
-0.018687835,
-0.0856332257,
0.0102209235,
-0.0243879184,
0.0104552666,
0.0266002566,
0.0961943045,
-0.0058784448,
-0.0563976914,
-0.0686601922,
0.0141195478,
-0.0783368871,
-0.001576546,
0.0055384999,
-0.0369988829,
-0.0338254087,
-0.0168553926,
-0.0358114801,
-0.0271520969,
0.08246883,
0.026217442,
-0.0156331286,
-0.056375999,
-0.0042075594,
-0.056848295,
0.0546280965,
0.042602364,
0.0118541224,
0.0386033766,
-0.0128066046,
-0.0529947616,
0.0114657171,
0.0579470657,
0.0390523076,
-0.0227109473,
-0.0568362251,
-0.0095519256,
-0.0058479984,
-0.0484750643,
-0.0191731453,
0.0243335441,
0.003019334,
0.0062792734,
-0.0038983468,
-0.055006206,
-0.0007724152,
0.020292379,
-0.0289157946,
0.0363624506,
0.0088544097,
-0.0194832273,
-0.0292727333,
-0.0004998387,
-0.0077248253,
-0.0073912395,
0.000974356,
-0.0113737723,
-0.0346870497,
0.0156469047,
-0.0005369584,
0.0508645698,
0.0133417798,
0.1030962318,
0.037204273,
-0.0179953314,
-0.0529341251,
-0.006997366,
0.0082881385,
-0.0358641222,
-0.0320035815,
0.0426951833,
0.0361408442,
0.0089134369,
0.0534045771,
-0.0359709561,
0.0097033177,
0.0127092209,
0.0098916348,
-0.0275095291,
0.0013651632,
0.0068777581,
-0.02449001,
0.0060132686,
-0.0362243317,
0.0315425098,
-0.0128801279,
-0.0318984389,
0.0460068844,
-0.0845902115,
-0.0577135608,
-0.0044332668,
0.0290107168,
-0.0606132597,
-0.0285219289,
0.0260285642,
-0.047963731,
0.0117125669,
0.0341250561,
-0.0120537467,
-0.0253696255,
-0.0202790927,
-0.0586316846,
0.0085415384,
0.0898967162,
-0.0115339449,
-0.0093941046,
-0.0702503696,
0.0212374218,
0.021251861,
-0.00941379,
0.0062194313,
-0.045385249,
0.0429988839,
-0.0291966312,
-0.0334310792,
0.0663403198,
-0.0406437069,
-0.0267532524,
-0.0359671786,
0.0968567729,
0.0306922551,
-0.0568191931,
-0.0301521588,
-0.0060265646,
-0.0015265368,
0.0687761307,
0.0318720415,
-0.0470908359,
-0.0446071625,
-0.0652661994,
0.0212278273,
-0.0535375029,
-0.0238204747,
0.0227815378,
0.0427051708,
-0.0481399409,
-0.1052199826,
0.0101071512,
0.0093070371,
0.0156541951,
0.0254741516,
-0.047991436,
0.0367509276,
-0.0385691412,
0.0196830146,
-0.0420427807,
0.0027616716,
-0.0004786187,
-0.0444137864,
-0.0011494518,
0.0407162569,
-0.0277431644,
-0.0541891195,
-0.0097020185,
-0.005173123,
-0.0873089507,
0.0478213951,
0.0365007184,
0.0210275482,
0.0128681632,
0.0040813643,
0.0141755911,
0.0294541027,
0.0269744731,
0.0155726979,
0.04953086,
0.0540301651,
-0.0466092378,
0.035380438,
0.0050542937,
-0.0177561548,
-0.0009692211,
0.0277258828,
0.0342874788,
0.0021997585,
0.0055279806,
0.0768493712,
-0.0007206742,
0.0386033691,
0.0019808656,
-0.0044997525,
-0.004244714,
0.0106837591,
0.0017717741,
0.0274510272,
-0.0487013832,
0.0085347071,
0.0291165449,
0.0237795655,
-0.0162471756,
-0.0332735069,
0.0365870856,
0.0547873825,
-0.0149328364,
0.0059055192,
0.0130813904,
-0.0028862206,
-0.0015339968,
0.0507534258,
0.0353921466,
0.0432671532,
-0.0520166494,
0.0131701743,
-0.0244759582,
-0.0377531163,
0.0836780742,
0.009754899,
-0.0674702749,
-0.0214301255,
-0.0660071895,
0.0210306887,
-0.0253725797,
-0.0226285737,
0.0587520897,
0.0114949141,
-0.0115426928,
-0.0799898431,
0.0046051247,
0.0171608441,
0.0176498629,
-0.0312262364,
0.0327051356,
0.0130020566,
0.0197506044,
0.0613893867,
-0.0378264524,
0.0781407207,
-0.0219641607,
0.0926699266,
-0.0172220785,
0.028600391,
-0.0354780778,
0.0106116449,
-0.0068620723,
-0.0029353679,
0.0144822784,
-0.0071531753,
-0.0410766341,
-0.044476904,
-0.0306156892,
0.032451462,
0.0454985425,
-0.006303688,
0.0334246345,
0.0491738766,
-0.0389136337,
-0.0542242192,
0.0117496476,
-0.0441636667,
0.0418360941,
-0.0147706782,
0.001825845,
-5.348273381e-33,
0.0067325281,
0.0374877006,
0.0209208895,
0.0552318357,
-0.0236172993,
-0.0535275117,
0.0193465669,
-0.0089905467,
0.0007784055,
0.0187954903,
0.0108965617,
0.0055303052,
0.0009096882,
0.0169631261,
0.0396482646,
-0.011818843,
0.0044131521,
-0.0058532776,
-0.0065834466,
-0.032579761,
-0.0036412787,
0.008827934,
0.1129468158,
0.023185892,
-0.0075846734,
0.0506692342,
-0.0044616172,
0.0308596585,
-0.039863728,
0.0290894192,
-0.0120385159,
0.0203144737,
0.0087086596,
-0.0174883399,
-0.0425934382,
0.0077010323,
-0.0386349075,
-0.0758835748,
0.0121364528,
0.0197679456,
-0.0355661735,
-0.0442858711,
0.0593550578,
-0.051221583,
-0.0300099663,
-0.0158189051,
-0.0032921657,
0.0006659545,
0.0085456725,
0.0295592919,
-0.0410983339,
-0.0230208039,
-0.0324402414,
-0.0264699049,
0.019009294,
-0.0060311737,
-0.0355291069,
-0.0431376509,
0.031490434,
0.0612071007,
-0.0002116914,
0.0462340377,
-0.0044493605,
0.0162012372,
0.027537616,
-0.0039579165,
0.0481684208,
-0.0546280965,
-0.0453184508,
0.0223432127,
0.0328612328,
0.0189863518,
0.011552141,
0.0137758078,
0.0527575091,
-0.0030922501,
-0.0594130605,
-0.0279077925,
-0.0526828989,
0.0670028329,
-0.0172920413,
0.0520642325,
-0.0268645715,
-0.0296850633,
-0.0295996703,
-0.003454518,
-0.0050581414,
-0.0022003241,
0.011357612,
-0.0018571323,
-0.0394690111,
0.0113744643,
0.0158034544,
-0.0288950056,
-0.0218885578,
0.0331498608,
0.0035169011,
0.0530658253,
-0.0059880554,
-0.0252425298,
-0.0295300595,
-0.022823805,
0.0287796762,
0.0457392484,
0.0071375882,
0.0129161458,
0.0238608755,
0.0665460974,
0.0007419938,
-0.0348891802,
-0.0313367918,
-0.0195662174,
0.0257440265,
-0.0149982041,
0.0396978222,
0.0408303924,
-0.0093157757,
0.0280777216,
0.0015783074,
0.0151578905,
-0.0417222641,
0.0557159558,
-0.0109524457,
0.0876373276,
0.001957169,
-0.0411439799,
0.0146856606,
-0.0016688624,
0.0794304162,
-0.031074727,
-0.0192517657,
0.0294688791,
2.431e-7,
0.0377614163,
0.0212421268,
0.042028401,
0.0294776708,
0.0109178126,
0.002691593,
-0.0183670409,
0.0301424786,
0.0024448882,
0.0119720027,
0.0286521483,
0.015078553,
-0.0247635208,
0.0131769674,
-0.0543094799,
0.0080436366,
-0.0650274754,
0.0192085225,
0.00813461,
-0.0116595663,
-0.0246182159,
0.0284690056,
0.0131164156,
-0.0137064299,
0.001859642,
-0.0031654164,
0.0489684232,
-0.0327565037,
0.0538447239,
0.0095633687,
0.0278377719,
-0.0534342453,
0.0288698934,
0.035283003,
0.0002764277,
0.0080051478,
0.007191685,
-0.0358758233,
-0.0236803014,
0.0067420541,
-0.0208757836,
-0.0258122012,
-0.0121387802,
-0.0498646162,
0.0055023874,
-0.0559361316,
0.0053384565,
0.0053700963,
-0.0176118985,
0.0251265951,
0.035095796,
-0.0061711823,
0.0055839927,
-0.0206047408,
0.0414335616,
0.0585722551,
-0.0147033827,
-0.0346136354,
-0.0146924602,
0.0282220449,
-0.0313318372,
-0.0144434953,
0.0039286013,
0.0240721367,
0.0057102446,
-0.0584826134,
-0.0559178032,
2.040837502e-34,
-0.002956657,
-0.019230023,
0.0443740748,
0.0124433981,
-0.0247466061,
-0.0177473966,
-0.029418109,
-0.0021261906,
0.0271127336,
-0.0525489375,
-0.0458194911
] |
http://arxiv.org/abs/2402.01829v1 | http://arxiv.org/pdf/2402.01829v1 | Predicting ATP binding sites in protein sequences using Deep Learning and Natural Language Processing | 2024-02-02T18:42:39 | 2024-02-02T18:42:39 | Shreyas V, Swati Agarwal | q-bio.BM | q-bio.BM, cs.CL, cs.LG | Predicting ATP-Protein Binding sites in genes is of great significance in the
field of Biology and Medicine. The majority of research in this field has been
conducted through time- and resource-intensive 'wet experiments' in
laboratories. Over the years, researchers have been investigating computational
methods computational methods to accomplish the same goals, utilising the
strength of advanced Deep Learning and NLP algorithms. In this paper, we
propose to develop methods to classify ATP-Protein binding sites. We conducted
various experiments mainly using PSSMs and several word embeddings as features.
We used 2D CNNs and LightGBM classifiers as our chief Deep Learning Algorithms.
The MP3Vec and BERT models have also been subjected to testing in our study.
The outcomes of our experiments demonstrated improvement over the
state-of-the-art benchmarks. | Predicting ATP binding sites in protein sequences using Deep Learning and Natural Language Processing: Predicting ATP-Protein Binding sites in genes is of great significance in the
field of Biology and Medicine. The majority of research in this field has been
conducted through time- and resource-intensive 'wet experiments' in
laboratories. Over the years, researchers have been investigating computational
methods computational methods to accomplish the same goals, utilising the
strength of advanced Deep Learning and NLP algorithms. In this paper, we
propose to develop methods to classify ATP-Protein binding sites. We conducted
various experiments mainly using PSSMs and several word embeddings as features.
We used 2D CNNs and LightGBM classifiers as our chief Deep Learning Algorithms.
The MP3Vec and BERT models have also been subjected to testing in our study.
The outcomes of our experiments demonstrated improvement over the
state-of-the-art benchmarks. | [
0.0344392136,
0.0306766666,
0.0094103292,
0.0133898519,
-0.0004442873,
-0.0079038218,
0.0059286659,
-0.0025828348,
0.0109776314,
-0.028023785,
-0.011132801,
-0.0370095856,
0.0090619111,
0.0245538168,
0.0120735699,
0.0310304444,
0.0644310191,
0.0510475859,
-0.0324825943,
-0.0060323239,
-0.0103296693,
0.0158838294,
0.020035198,
0.0500290394,
-0.0005123001,
-0.0182281882,
-0.0204558652,
-0.0291383322,
0.0473852605,
-0.0211766642,
-0.0207342617,
0.0448283702,
0.0221095756,
0.0524308905,
0.000001734,
-0.0612012073,
-0.0380830877,
-0.026256483,
-0.0313463844,
0.0718609393,
-0.0591700189,
-0.0203013271,
0.0052705221,
-0.012211117,
0.0015148476,
-0.0388636105,
0.0796069205,
0.0478522554,
0.0079890024,
0.032610245,
0.0054884213,
-0.0203149542,
0.0498502515,
-0.0348808803,
0.0470595509,
-0.0435661674,
0.072200343,
-0.0247099791,
0.0071937125,
-0.0031939284,
-0.0052175815,
0.0351881161,
-0.0290778223,
0.006367371,
0.0353104696,
0.0546209961,
-0.0451578088,
-0.0126393,
-0.0022025767,
0.0601195693,
-0.0392980948,
-0.0237833019,
-0.0424888879,
0.0137676997,
-0.0458227806,
0.0197386555,
-0.0111277765,
-0.0105860205,
-0.0234129522,
0.0674505308,
-0.0438269302,
-0.0329426751,
0.0357205942,
0.0182395503,
0.0343882367,
0.0077016773,
-0.0243142974,
-0.0519215651,
0.0201705974,
0.0332855508,
-0.0403811634,
-0.0064239567,
0.0504310355,
-0.0332427137,
0.0261572264,
0.0065221461,
0.0248423889,
-0.0064216689,
0.0224936102,
-0.1211936772,
0.0643000305,
-0.0093064168,
-0.0306018218,
0.0131666688,
-0.0767414868,
-0.0353485495,
-0.0250198636,
0.0287173484,
0.0158537272,
0.036152713,
-0.0640847757,
0.0161596462,
-0.0244884919,
0.0306156389,
0.0201295484,
-0.0257995557,
-0.0531268083,
-0.0452522524,
-0.04353429,
-0.0147780702,
-0.0126340389,
0.1002491415,
0.0093516437,
0.0466653593,
-0.0235722549,
0.0242960453,
-0.0261060335,
-0.0003601442,
-0.0608305372,
-0.0439386144,
0.0118508013,
-0.0170023963,
0.0388297066,
0.0018394461,
0.0182387587,
0.0858283415,
0.0223304462,
-0.0353318602,
-0.0380709954,
-0.0323293395,
-0.0338847265,
-0.0011639387,
-0.0159852151,
-0.057349775,
0.0254075266,
0.0091953147,
-0.0196282864,
0.0153499283,
0.0403048694,
0.0074049071,
-0.0429090187,
0.0945703164,
0.0084237643,
-0.0205612164,
0.0380665213,
0.0025903517,
0.0022138648,
0.0000265803,
0.0128988819,
0.0208956506,
0.0183506049,
-0.0180924553,
0.0163077135,
-0.0134663889,
-0.0179123841,
-0.0589569286,
0.0150160687,
-0.023500571,
0.0197740365,
0.0446857512,
-0.0128166685,
-0.0059257573,
0.0411922894,
-0.0268742051,
-0.0033705425,
0.035942588,
0.019819878,
0.069200851,
0.0365614705,
0.0350999348,
0.0200482849,
-0.0379827432,
0.0297464933,
0.0199909452,
0.0036147786,
0.0138184838,
0.0523759797,
-0.0257342793,
0.0004538377,
-0.0203189105,
-0.0411094353,
-0.010408733,
0.0083695818,
-0.044806961,
0.0250702575,
-0.0239199866,
-0.0728857741,
0.0591456816,
0.0189590566,
-0.0242705066,
0.0019384626,
0.0196800046,
-0.0094906706,
0.0643992051,
0.034080144,
-0.0367815793,
0.0099260816,
-0.0139070237,
-0.0210530981,
0.005706504,
0.0070652589,
0.0111369658,
0.0055103875,
-0.0485508665,
-0.0090062609,
-0.0346139781,
-0.0156240119,
0.0190794487,
-0.0365760401,
0.0196383167,
0.0117294025,
-0.0330141112,
-0.0129128825,
0.0093546445,
0.0409177765,
-0.025208395,
0.0257445388,
0.0171620455,
0.033594422,
0.0201410744,
-0.0130038587,
0.0301083382,
0.0445427969,
-0.0516204946,
-0.0254863072,
-0.0064765653,
-0.0097258193,
-0.0199999008,
0.0262946375,
-0.0193113815,
0.0076845875,
0.0311190058,
-0.0337094404,
-0.0216356553,
0.026644282,
-0.0435741171,
0.0495761298,
-0.0236868281,
-0.0280491207,
0.0240219664,
0.0511635579,
-0.019387139,
0.0593380108,
0.0249603502,
0.0383252017,
0.0304501895,
0.0034885451,
0.0558931418,
-0.0075689089,
-0.0211462025,
-0.0171973184,
-0.0338619947,
-0.003418369,
0.0116687035,
-0.023287721,
0.0013179261,
0.0283996724,
-0.0793900713,
-0.0011698271,
0.0195601694,
0.0068200766,
-0.0365888886,
-0.0550146252,
0.0000013148,
-0.0003305733,
-0.0509886071,
0.0232208744,
-0.0691083968,
-0.0070754578,
0.0215448141,
-0.0022094199,
0.0107501652,
-0.0418741442,
-0.0078874948,
-0.0752572864,
0.0672778562,
0.0043766066,
-0.0865006968,
0.0235910509,
0.018997062,
-0.0075214338,
0.0565338247,
0.0453979857,
-0.0379070938,
0.0139925722,
-0.0136532001,
0.0111108655,
0.0160108563,
0.0309377257,
0.0287827067,
0.0354598127,
0.047860872,
-0.0109416824,
0.0069606416,
-0.0517480485,
0.0221336074,
0.0045451871,
-0.0026657379,
-0.0178572051,
0.0562105887,
0.029573027,
-0.0053732772,
-0.0067544589,
-0.0278289877,
0.0148764458,
-0.011208402,
-0.0282341186,
0.0073195784,
-0.0189795475,
0.0124354837,
0.0191554166,
-0.0422683246,
0.0708749741,
-0.0205549151,
0.020562185,
-0.0539572351,
-0.024829777,
-0.0177951865,
0.0078311982,
0.014223502,
0.0555595979,
-0.0259235296,
0.0041232705,
0.0478154384,
0.0292868875,
0.0099194422,
-0.0311992783,
0.0095088333,
0.0271259677,
0.0194650088,
-0.004534171,
-0.0184286125,
-0.0225572046,
0.0108243944,
-0.039454814,
0.0269745737,
0.0052442197,
-0.0502907783,
-0.0567430928,
0.0330031663,
-0.0198766496,
-0.0160486139,
-0.0017512785,
0.0133697111,
-0.0400568135,
0.0152922105,
0.0261372905,
0.0039297417,
-0.0145082269,
0.0094909826,
0.0163300503,
0.0543608442,
-0.0560020879,
0.0471445471,
-0.0298101697,
0.0061032972,
0.0261429083,
-0.0224436857,
0.0035812135,
-0.0150613673,
0.0232101381,
-0.0306672864,
-0.0242073555,
0.1079699993,
0.0779947266,
-0.0488275141,
0.0128639899,
0.0032758338,
-0.0422169827,
-0.0234976131,
0.0039966139,
-0.0552402511,
-0.0475542098,
-0.0324709974,
0.0182458889,
0.0032234876,
-0.0116412453,
0.0188273937,
-0.0611068532,
0.0333581455,
0.0035095157,
0.0174156316,
-0.0066954968,
-0.0764231086,
-0.0340165198,
-0.0181044061,
-0.0195313618,
-0.0755661577,
0.0074806139,
-0.0013179922,
0.0532077588,
-0.0500236712,
-0.0436625853,
0.0685804263,
-0.0674788728,
-0.0191304907,
0.072421439,
0.0437126867,
-0.01077097,
-0.0293188486,
-0.0333262682,
0.0140797021,
0.012001439,
-0.0222601444,
0.007535601,
-0.0059706257,
-0.031371519,
-0.0063112988,
0.006891001,
-0.0637270063,
-0.0262400452,
-0.0394084975,
0.0970875472,
-0.026554361,
-0.0231494345,
-0.0420139916,
0.0580646656,
-0.0424392521,
0.0860494077,
0.0187739469,
-0.0296407677,
-0.0248193592,
-0.0235779863,
0.0660561621,
-0.0734142959,
0.000548016,
-0.0012793267,
-0.030745307,
-0.0182519294,
-0.0323643647,
-0.0211399011,
0.0727653876,
-0.0048011639,
0.0139273526,
0.0388352759,
0.0093892682,
-0.0371235795,
-0.0748264715,
0.0667666569,
0.0655726567,
0.0290921833,
-0.0254489128,
-0.0771713257,
0.08057753,
0.0053749806,
-0.0506289788,
0.0641677752,
-0.019016074,
-0.0829171613,
-0.0587867945,
0.0408641398,
0.0096967993,
-0.0010778414,
-0.0225379989,
-0.0189159717,
0.0111307083,
0.0787355006,
0.0461116843,
-0.0282358378,
0.0105263609,
-0.0150056761,
-0.0140042109,
-0.0182538144,
-0.0076680728,
0.0169620458,
-0.015335368,
0.0356999226,
0.0385481827,
-0.038725473,
0.0336465612,
-0.0427046977,
0.0919448808,
-0.0099848164,
-0.087276198,
-0.042806346,
0.0204320047,
0.0342002586,
0.0366003588,
-0.035319075,
-0.0109447762,
0.0402422398,
-0.0075744488,
-0.0347182602,
0.0223774649,
-0.0018458623,
-0.0230427813,
-0.0023263958,
0.0582955256,
0.0359753147,
-0.0250917971,
0.014045964,
-0.0288619492,
-0.0166911166,
-0.0288629252,
0.0221624188,
-0.0061584702,
0.0081819268,
-0.0417097174,
-0.012420373,
0.0639306605,
-0.0308402069,
0.0272539444,
-0.0034789103,
-0.0004598101,
-0.0377687812,
0.0157990828,
0.0598708093,
-0.0684696361,
-0.0148230037,
-0.0300483834,
-0.0176950414,
-0.0399125367,
-0.0115394285,
-0.0413768552,
-0.0136757409,
0.0732605755,
0.0404446349,
-0.0383612886,
-0.018001087,
0.0464233793,
0.0379805416,
0.0530644618,
-0.0088731889,
-0.1058631316,
-0.0044116727,
0.0341755077,
-0.0299715251,
0.0343756601,
-0.0501764603,
0.0011762013,
0.0009193065,
-0.0303422064,
0.0234609284,
0.0062837852,
-0.0214688815,
0.0407543592,
0.0409326591,
0.025780838,
0.0278353859,
-0.026723465,
-0.0046850983,
0.0380505063,
-0.0245645866,
-0.0290919431,
-0.0435854271,
-5.7660404e-33,
0.0426901542,
-0.0523991883,
-0.0179938935,
0.0066009942,
0.0285725333,
0.037917275,
-0.0553015545,
-0.0035641436,
-0.0249083918,
0.061074052,
0.0359899588,
-0.0573833473,
0.0141124055,
-0.027890671,
-0.0095812362,
-0.027282143,
0.0485401228,
0.0175148901,
0.0408809707,
-0.0367840491,
0.0303925667,
0.0789545178,
0.0780703574,
-0.0236413553,
-0.0090955347,
0.0429975577,
0.0179808978,
-0.0311283264,
0.0052899937,
0.071069479,
-0.0220235512,
0.0745189413,
-0.0228547696,
-0.0300992038,
0.0021079215,
0.0434585959,
-0.0558748655,
0.0088512804,
-0.0151332729,
0.0142074805,
-0.0304998048,
0.0132680843,
0.0382954106,
0.0314631835,
-0.0320181809,
0.0282690991,
0.0556664579,
-0.0083260722,
-0.0312824883,
-0.0223802011,
-0.0590316653,
0.0149635496,
0.0219229367,
-0.0212282222,
0.0288203377,
-0.0073005706,
0.0346317589,
0.0154883824,
-0.02032586,
0.0078136465,
0.0260573998,
0.0637701377,
0.0375652835,
-0.0112097617,
-0.0154655054,
0.0659138411,
0.0042139357,
-0.0216434952,
-0.0027942623,
0.0005166779,
-0.0104643274,
0.0589138195,
0.0638504773,
-0.040744137,
0.0320213251,
0.003824967,
-0.0889139399,
0.0389448069,
-0.0005088645,
0.0246141162,
-0.0156975184,
0.0170643087,
-0.0302754622,
0.008957603,
0.0053484705,
0.0332087651,
-0.02096669,
-0.0446476042,
0.012213056,
-0.0549990609,
-0.0370903462,
0.076482892,
0.0099695269,
-0.0498856716,
-0.0503458604,
0.0262198281,
0.0259806048,
0.003132595,
-0.0066513694,
-0.0154588744,
-0.0140921688,
-0.0339230485,
0.0129903434,
-0.0465514362,
0.0014233992,
-0.0295728389,
0.0201815926,
-0.0052100644,
-0.0077872053,
0.0105980169,
-0.0273185875,
-0.0466088131,
0.0084065422,
0.0926650688,
0.0247971881,
-0.0071263411,
0.0012784458,
0.0788657665,
-0.0001769882,
0.0319939926,
0.0274616834,
-0.0319545306,
0.0002520301,
0.0253423713,
-0.0289237313,
-0.0220443867,
-0.0484509803,
-0.0100533441,
-0.0112299416,
-0.0838966295,
-0.0156913102,
0.0066707996,
2.805e-7,
0.0605329499,
0.0433924049,
0.0186149236,
-0.0227170698,
-0.0414357409,
0.0233414751,
0.0088612465,
0.0458496064,
-0.0544040613,
-0.0150674935,
0.0170467626,
-0.011412981,
0.0345865674,
0.0080995485,
-0.0679158792,
0.0116144093,
-0.0639185905,
0.0111662848,
-0.0628911331,
-0.0015795396,
-0.0159318019,
0.1341896802,
-0.0266517513,
0.0014491545,
0.0744069442,
-0.0630402863,
0.0314689763,
-0.013818766,
0.015189304,
0.0025817219,
0.0408877693,
0.0034549653,
0.0089330152,
-0.0121378684,
-0.0904816538,
-0.0091328612,
0.0247677304,
-0.0204314925,
-0.020812748,
-0.0506479815,
0.0309815444,
0.0030159345,
0.0269597005,
-0.0070891585,
0.0404972471,
-0.0183360688,
-0.0564778857,
-0.0219985489,
0.0269379653,
0.0027389345,
0.0552636683,
-0.0326775461,
0.0092479726,
-0.0093895784,
-0.0122591443,
-0.0379552953,
-0.0194625705,
-0.066629447,
-0.0008760664,
0.0206645876,
-0.0188901406,
0.0544167757,
-0.0345539413,
-0.0199964829,
0.085639663,
-0.0209323969,
-0.0120442109,
2.137856882e-34,
0.0289331377,
-0.026184693,
-0.003719958,
0.0123313051,
0.0010827264,
-0.0043249149,
-0.0125196511,
0.0147746224,
-0.0326649658,
-0.0515917875,
-0.0337434448
] |
http://arxiv.org/abs/2402.01622v3 | http://arxiv.org/pdf/2402.01622v3 | TravelPlanner: A Benchmark for Real-World Planning with Language Agents | 2024-02-02T18:39:51 | 2024-06-23T08:50:17 | Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, Yu Su | cs.CL | cs.CL | Planning has been part of the core pursuit for artificial intelligence since
its conception, but earlier AI agents mostly focused on constrained settings
because many of the cognitive substrates necessary for human-level planning
have been lacking. Recently, language agents powered by large language models
(LLMs) have shown interesting capabilities such as tool use and reasoning. Are
these language agents capable of planning in more complex settings that are out
of the reach of prior AI agents? To advance this investigation, we propose
TravelPlanner, a new planning benchmark that focuses on travel planning, a
common real-world planning scenario. It provides a rich sandbox environment,
various tools for accessing nearly four million data records, and 1,225
meticulously curated planning intents and reference plans. Comprehensive
evaluations show that the current language agents are not yet capable of
handling such complex planning tasks-even GPT-4 only achieves a success rate of
0.6%. Language agents struggle to stay on task, use the right tools to collect
information, or keep track of multiple constraints. However, we note that the
mere possibility for language agents to tackle such a complex problem is in
itself non-trivial progress. TravelPlanner provides a challenging yet
meaningful testbed for future language agents. | TravelPlanner: A Benchmark for Real-World Planning with Language Agents: Planning has been part of the core pursuit for artificial intelligence since
its conception, but earlier AI agents mostly focused on constrained settings
because many of the cognitive substrates necessary for human-level planning
have been lacking. Recently, language agents powered by large language models
(LLMs) have shown interesting capabilities such as tool use and reasoning. Are
these language agents capable of planning in more complex settings that are out
of the reach of prior AI agents? To advance this investigation, we propose
TravelPlanner, a new planning benchmark that focuses on travel planning, a
common real-world planning scenario. It provides a rich sandbox environment,
various tools for accessing nearly four million data records, and 1,225
meticulously curated planning intents and reference plans. Comprehensive
evaluations show that the current language agents are not yet capable of
handling such complex planning tasks-even GPT-4 only achieves a success rate of
0.6%. Language agents struggle to stay on task, use the right tools to collect
information, or keep track of multiple constraints. However, we note that the
mere possibility for language agents to tackle such a complex problem is in
itself non-trivial progress. TravelPlanner provides a challenging yet
meaningful testbed for future language agents. | [
0.0491794348,
0.0750429332,
-0.0291515887,
0.0200484619,
-0.0356252715,
-0.0240140017,
-0.0033487077,
-0.0134654129,
0.0079831695,
0.013187727,
0.0196026936,
-0.0164085664,
-0.0294032637,
0.0619069003,
-0.0269714613,
-0.0570622161,
0.019117061,
-0.0596472025,
0.0202031583,
0.0148008596,
-0.0237265192,
0.0264881589,
-0.0362810306,
0.0434962548,
-0.0182148684,
0.016081728,
-0.068237491,
0.0088760452,
0.0275345091,
-0.052220352,
-0.0133050466,
0.0359563939,
0.0526456423,
-0.0005009914,
0.0000021376,
-0.0027943463,
-0.0265844688,
-0.0416168161,
-0.0378114879,
-0.0453651175,
0.0054558963,
-0.0031242645,
-0.0004420168,
0.0173946917,
-0.0533830598,
0.0145152453,
0.0083946772,
0.0459491834,
0.0473596156,
0.0804247409,
-0.0107157687,
-0.0040869908,
-0.0247179996,
-0.0025285033,
-0.0107476916,
-0.0546614081,
-0.013467093,
-0.006991019,
0.018572595,
-0.0050181327,
-0.0082245087,
-0.0308679119,
0.0257746745,
-0.0223024022,
-0.0491810665,
0.0181922596,
-0.0817413852,
-0.0216361377,
-0.0186344944,
0.0720721558,
0.0025009841,
-0.0251600668,
0.0372942276,
-0.0029167675,
-0.0540439412,
-0.0183583908,
-0.0208522342,
0.0095156636,
0.0148837054,
0.0383505933,
0.0278924536,
0.0087135052,
0.027557293,
0.0229459293,
-0.0232445151,
0.1087199599,
-0.0078622457,
-0.0164074507,
0.0199322011,
-0.0055300277,
0.0365244411,
-0.0991892815,
0.0336031653,
0.0570674166,
-0.0832952783,
0.0153645733,
0.0251334868,
0.0462014563,
-0.0070839291,
-0.00476838,
0.0623199977,
0.065623872,
-0.0030783722,
0.0499192365,
0.0328854062,
0.0580673814,
0.002622009,
0.0112958513,
-0.0393566079,
0.0030760383,
-0.0411623232,
0.0387487411,
-0.060982246,
-0.0102294656,
-0.0805346146,
0.0000049044,
0.0021275864,
0.034681499,
-0.0100466302,
0.0519651808,
0.0051518013,
-0.0017013239,
0.0022598673,
-0.0500427112,
-0.0767327175,
-0.0324290395,
-0.0708526969,
0.0150236171,
-0.0050733648,
-0.0093429917,
0.0050286921,
0.0174386725,
0.024744004,
-0.0276963636,
-0.0272505395,
0.0628519654,
-0.0030831904,
0.0028993057,
-0.0650720894,
-0.0630387217,
0.0261937194,
-0.1162840351,
0.0005951907,
-0.0697969347,
0.036950659,
-0.006636905,
-0.0477917828,
-0.064541921,
0.0179403927,
-0.0004066853,
0.0068129995,
0.0363368727,
0.044189997,
0.0148953144,
-0.0001253608,
0.0238226,
0.0121387634,
0.0302914325,
0.0409804061,
0.0771790296,
0.0191107374,
0.0459832698,
0.0293283165,
0.0506880656,
0.0033845755,
0.0252318326,
0.0069757905,
0.0201713815,
-0.0079984376,
0.0213163532,
-0.0315594487,
0.0390119739,
-0.0286151208,
0.0233101081,
0.0387059227,
0.0344399661,
-0.0206265058,
0.0049255756,
0.0627277121,
0.012613113,
-0.0345543884,
-0.0500944629,
-0.0337977819,
0.0249754153,
0.0009805656,
-0.010703722,
-0.0092829624,
0.0241818894,
-0.0115947872,
-0.0514103845,
0.013623612,
-0.0253109131,
0.0544214621,
-0.0222039912,
0.0212062635,
-0.0446652062,
-0.0509000085,
0.0186339207,
-0.0059109181,
-0.0322017781,
-0.0063779876,
-0.019399453,
0.0044589583,
0.0894802213,
0.0323902965,
-0.0062832041,
-0.0423689447,
0.0064908424,
0.0249940529,
0.0162368976,
0.0771197528,
-0.0169256739,
0.0147754401,
-0.053464707,
-0.027748948,
0.0330320112,
0.0092236307,
-0.0064566354,
0.0449227579,
0.0483716875,
0.0030564042,
-0.0292758085,
-0.0061276131,
0.0017380813,
-0.0244362354,
-0.030752996,
-0.0358036458,
0.0136233531,
0.038938418,
0.0002024867,
-0.0177612677,
-0.0118834758,
0.0457243249,
-0.0181318019,
-0.0004276009,
0.043975573,
-0.0171890203,
0.0187814534,
0.0224971939,
-0.0103666559,
-0.0404415131,
0.0493136346,
-0.0115328617,
-0.0520694256,
-0.0043809465,
0.0216866378,
0.0120767914,
0.0000472916,
0.036542993,
0.0177377611,
-0.0172415581,
0.0007348304,
0.0677289814,
0.0186107233,
-0.0304818172,
0.0403297879,
0.0354198888,
0.0442447662,
-0.0361223593,
-0.0002856131,
0.0294724517,
0.0345204026,
0.0021607687,
0.0274862088,
0.0191232823,
-0.0381444246,
-0.011671016,
-0.0747793317,
0.0019945861,
-0.0304996483,
-0.0057544429,
-0.009390343,
-0.0143369259,
-0.0107844649,
0.0008660723,
-0.0307458788,
0.0655307397,
-0.0142979352,
-0.0300730318,
0.0184938665,
0.0163100138,
-0.0330482535,
-0.0391799547,
0.0194274858,
-0.0189090651,
-0.0303146001,
0.0247916691,
-0.0076239891,
-0.0297616255,
-0.0051131533,
-0.0593211204,
0.030246418,
-0.0118479226,
-0.0571393296,
-0.0058362372,
-0.0309787225,
-0.0100044608,
-0.0026660904,
-0.0009383779,
-0.0057372754,
0.0916566551,
-0.0467372425,
-0.0739631727,
0.0595888495,
-0.0337333903,
0.019597644,
-0.0865654126,
-0.0576814525,
-0.0139823481,
0.0130863125,
-0.02953724,
0.0537535921,
-0.0608222149,
0.0116040334,
0.0071255062,
-0.0233450793,
0.0121539654,
0.0267292447,
-0.0257663522,
0.1280207932,
-0.0216510277,
-0.0021180292,
0.0300712343,
-0.0122030694,
-0.0119104581,
-0.0070331888,
-0.0065337745,
-0.0332657993,
0.0826661289,
0.0290851742,
0.0519464053,
-0.0071583879,
0.0434382372,
0.0103539992,
-0.0080274018,
-0.0413348787,
-0.0233331528,
0.0316390879,
-0.0650597885,
0.026459584,
0.002802422,
-0.0746628121,
-0.0345913209,
-0.0028021692,
-0.0537323244,
-0.0095534241,
-0.0011835404,
-0.0748262927,
-0.0312709548,
-0.0221103914,
0.0165190399,
0.078214176,
0.0087400945,
-0.0084290272,
-0.0092630889,
-0.0225600861,
0.0724230334,
0.0822802857,
0.0053902562,
0.0319076441,
-0.0008606127,
0.010149505,
0.0211978611,
-0.0033841401,
0.0003326196,
0.0493112057,
-0.0008999893,
0.0005050905,
0.0064066015,
0.0543487296,
0.042684447,
-0.0448717996,
-0.0445390232,
0.0801310465,
0.0437917486,
-0.0447896831,
0.0193931703,
-0.0647375882,
0.03059108,
0.0213691127,
0.0738213733,
-0.0474286042,
-0.0249065813,
0.0017782567,
-0.0297643226,
-0.0223199613,
-0.0534750074,
-0.0237395801,
-0.0846912265,
0.061494872,
-0.0040523671,
-0.0214453414,
0.0308790244,
-0.0598414838,
-0.0157975852,
-0.0059183114,
0.0039485726,
-0.0266618542,
0.0372768156,
-0.0357041247,
-0.0117373979,
-0.0378768295,
-0.0261597186,
0.0300765727,
-0.0436291657,
0.0768365934,
0.0510857701,
0.0696186274,
-0.0107772704,
-0.0384832919,
0.0033428862,
0.0639236271,
0.0241317227,
-0.0228962228,
0.0019311522,
0.0343323946,
-0.0113236895,
0.0189809557,
0.0101501569,
-0.0809029266,
-0.0367968418,
0.0470753163,
0.1022895128,
-0.013410598,
0.0067618624,
-0.0083009545,
0.0213758517,
0.0328104384,
0.0552408881,
-0.0457991958,
-0.0324303433,
-0.0665856898,
-0.0644069687,
-0.0158345737,
-0.0809905306,
-0.0196370762,
0.0226799473,
0.013250282,
0.0202298947,
-0.0501136482,
-0.0735764354,
0.050266955,
0.0250405893,
0.0070396271,
-0.004814059,
0.0348438807,
-0.0093144029,
-0.0043133027,
-0.0104972543,
-0.0318955742,
0.0196591076,
-0.0386450887,
-0.007425426,
-0.03576703,
-0.0159363393,
-0.0271796323,
0.0215253662,
-0.0529875755,
-0.0653466806,
0.03775746,
0.0190442093,
0.0254308973,
0.0105639501,
-0.006538352,
-0.0077103442,
-0.0102538234,
-0.0252467841,
-0.0319657885,
-0.0033655127,
0.0099150203,
0.0193371847,
0.0029028123,
0.0581529699,
-0.0526902415,
0.0392392576,
-0.0009404137,
0.0439078808,
-0.0784245059,
0.0091300774,
0.0211828724,
0.0125606786,
-0.0379985534,
0.0144606074,
0.0016237571,
-0.0211616643,
0.0306400731,
-0.0148064969,
-0.0262986384,
-0.0261168126,
-0.0388056636,
0.0139673026,
-0.0070054517,
0.0269996822,
0.0066822846,
-0.0075581353,
-0.053958483,
0.0122559946,
0.0066706366,
0.0238573328,
-0.0051536364,
0.0337518528,
0.0000191914,
0.0085135857,
0.0305470936,
-0.0073563508,
0.0457312725,
-0.0214917511,
0.0288690273,
0.0181735456,
0.0032650977,
-0.0476685464,
-0.0328558907,
-0.0316557139,
-0.0838306323,
0.0667961985,
0.0258031934,
0.0518269837,
-0.0173008479,
0.0083825206,
-0.0758344531,
0.0089711472,
-0.0000842651,
-0.046526432,
0.0194311999,
-0.0235611089,
0.0367248841,
0.0217537805,
0.0016533181,
0.0169496182,
0.028183559,
-0.0080173351,
0.0847825706,
0.0364481322,
-0.0237896964,
-0.0492474474,
-0.0057326509,
-0.0269785132,
0.0006437706,
-0.0179647282,
0.0066124764,
-0.0410113558,
-0.0280952957,
0.0148891415,
0.0315463021,
-0.0288012289,
-0.0412878096,
-0.0004416124,
-0.0179374982,
-0.0086082686,
-0.0251746885,
-0.0353078358,
-0.0536536053,
-0.0478092469,
-0.0133678988,
-0.0201849323,
-6.415796004e-33,
0.010413982,
-0.0363595895,
0.0952955857,
-0.0340451002,
0.0188288894,
-0.0293268748,
-0.0736946538,
-0.0035610255,
-0.0327788256,
0.0288683772,
-0.0421547703,
-0.0011818481,
0.0265395604,
0.0091157155,
0.0175300054,
-0.0072541577,
0.022126954,
0.0186775047,
-0.0446229801,
0.0025697083,
0.0506010838,
-0.032560125,
0.0118740816,
-0.0562806353,
-0.0284069832,
-0.0159692578,
-0.0300219469,
0.0557557493,
0.0142577039,
0.0576629117,
-0.0011058976,
0.0071625719,
0.0154630784,
-0.0064432067,
-0.0004393338,
0.0019720534,
-0.0294542853,
-0.0673362687,
0.0143482937,
0.008379628,
-0.0405514464,
-0.0222268645,
-0.0254101045,
0.0002210391,
-0.0070971418,
0.045794189,
0.0181132946,
0.0007046384,
0.0109264087,
0.0795247704,
-0.0621108562,
0.0069047003,
-0.0235263202,
0.0179116875,
0.0636593848,
0.0121827321,
0.0050323987,
-0.0109827304,
-0.0324093103,
0.0376379266,
-0.0312936604,
-0.0068978691,
0.032777071,
-0.0427426025,
-0.031830959,
0.0706050694,
0.0149203688,
0.012655477,
-0.0364536345,
-0.0001243278,
-0.0087524122,
0.0790834054,
-0.0282307696,
0.0610614344,
-0.02492458,
0.0007463022,
-0.0259321649,
0.0266843271,
0.0228339862,
0.0234577823,
0.0742530376,
0.0204358678,
-0.044707492,
-0.0019619872,
-0.0384748094,
0.0139173688,
-0.0190310515,
-0.0235775337,
0.021507198,
0.0564968176,
0.0260809343,
0.0297690798,
0.0417075306,
-0.0101336399,
-0.00533727,
0.0553569272,
-0.0062670987,
-0.0075325272,
0.0042152391,
0.0568252131,
-0.0857898742,
-0.0277084894,
-0.0265136305,
0.0022887886,
0.0362453423,
0.0345073752,
0.004751821,
0.0305024218,
-0.0112312175,
-0.0027641335,
-0.0589168295,
0.0132990777,
-0.0251721106,
-0.0195842665,
0.0375858359,
-0.0340076573,
-0.0105151366,
-0.017879704,
0.033582028,
-0.0497710705,
-0.0008523571,
-0.0366522633,
0.0122462679,
0.0082879439,
-0.0211858507,
0.0139830355,
-0.0439592414,
0.0118010882,
0.0384083577,
-0.035912808,
0.0181710366,
0.0251886658,
2.756e-7,
-0.0018485378,
0.0575905032,
-0.0073253587,
0.0719552338,
0.0099592535,
0.0192765277,
-0.0151940705,
-0.0113025205,
-0.0637804791,
0.03261986,
0.0386469886,
0.0419118851,
0.0233730525,
0.0625202134,
-0.0516744256,
0.0609779581,
-0.006023902,
-0.0430836193,
-0.0846031234,
-0.0014989674,
0.0537710749,
0.078104116,
0.0279120319,
-0.0104385493,
0.0061441194,
-0.0329984166,
0.000833623,
-0.0821045265,
0.0000652717,
0.0288163722,
-0.0116019659,
0.0534455217,
-0.0140897259,
-0.0015277241,
-0.0318834409,
0.0028284804,
-0.0204416942,
0.080169782,
0.0061650262,
-0.0355630778,
0.0016525819,
-0.041662246,
-0.0180143788,
-0.0509307273,
0.0333418287,
0.0426972061,
0.0011051841,
0.0115371449,
-0.0020958837,
0.0056619449,
-0.0238982793,
0.0045873942,
-0.000889406,
-0.0069631692,
0.0095256744,
0.0021447972,
-0.0008672136,
-0.0720503256,
0.0199714638,
0.0517354161,
-0.0011470708,
0.0076355417,
-0.0579346642,
0.0264341626,
-0.0054149395,
-0.0571460202,
0.0066752913,
2.193755771e-34,
-0.0086347051,
0.0406272821,
-0.0269940775,
0.0085502053,
0.0593731925,
0.0194008015,
0.0048898649,
0.0280598812,
-0.0036024724,
-0.0466766208,
-0.0304977354
] |
Subsets and Splits