File size: 82,194 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:24:37.949234Z"
},
"title": "Large-scale text pre-training helps with dialogue act recognition, but not without fine-tuning",
"authors": [
{
"first": "Bill",
"middle": [],
"last": "Noble",
"suffix": "",
"affiliation": {
"laboratory": "Centre for Linguistic Theory and Studies in Probability",
"institution": "University of Gothenburg",
"location": {}
},
"email": "bill.noble@gu.se"
},
{
"first": "Vladislav",
"middle": [],
"last": "Maraev",
"suffix": "",
"affiliation": {
"laboratory": "Centre for Linguistic Theory and Studies in Probability",
"institution": "University of Gothenburg",
"location": {}
},
"email": "vladislav.maraev@gu.se"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We use dialogue act recognition (DAR) to investigate how well BERT represents utterances in dialogue, and how fine-tuning and large-scale pre-training contribute to its performance. We find that while both the standard BERT pre-training and pretraining on dialogue-like data are useful, task-specific finetuning is essential for good performance.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "We use dialogue act recognition (DAR) to investigate how well BERT represents utterances in dialogue, and how fine-tuning and large-scale pre-training contribute to its performance. We find that while both the standard BERT pre-training and pretraining on dialogue-like data are useful, task-specific finetuning is essential for good performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Large-scale neural language models trained on massive corpora of text data have achieved state-ofthe-art results on a variety of traditional NLP tasks. Given that dialogue, especially spoken dialogue, is radically different from the kind of data these language models are pre-trained on, it is uncertain whether they would be useful for dialogue-oriented tasks. In the example from the Switchboard corpus, shown in Table 1 , it is evident that the structure of dialogue is quite different from that of written text. Not only is the internal structure of contributions different-with features such as disfluencies, repair, incomplete sentences, and various vocal soundsbut the sequential structure of the discourse is different as well.",
"cite_spans": [],
"ref_spans": [
{
"start": 415,
"end": 422,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "In this paper, we investigate how well one such large-scale language model, BERT (Devlin et al., 2019) , represents utterances in dialogue. We use dialogue act recognition (DAR) as a proxy task, since both the internal content and the sequential structure of utterances has bearing on this task",
"cite_spans": [
{
"start": 81,
"end": 102,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We have two main contributions. First we find that while standard BERT pre-training is useful, the model performs poorly without fine-tuning ( \u00a73.1). Second, we find that further pre-training with data from the target domain shows promise for dialogue, but the results are mixed when pre-training with a larger corpus of dialogical data from outside the target domain ( \u00a73.2). Example from the SWDA corpus (sw2827).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Dialogue acts:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "sd-Statement-non- opinion, sv-Statement-opinion, bd-Downplayer, b-Backchannel, x-Non-verbal.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "1 Background",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The concept of a dialogue act is based on that of speech acts (Austin and Urmson, 2009) . Breaking with classical semantic theory, speech act theory considers not only the propositional content of an utterance but also the actions, such as promising or apologizing, it carries out. Dialogue acts extend the concept of the speech act, with a focus on the interactional nature of most speech.",
"cite_spans": [
{
"start": 62,
"end": 87,
"text": "(Austin and Urmson, 2009)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dialogue Act Recognition",
"sec_num": "1.1"
},
{
"text": "DAR is the task of labeling utterances with the dialogue act they perform from a given set of dialogue act tags. As with other sequence labeling tasks in NLP, some notion of context is helpful in DAR. One of the first performant machine learning models for DAR was a Hidden Markov Model that used various lexical and prosodic features as input (Stolcke et al., 2000) . Most successful neural approaches also model some notion of context (e.g., Kalchbrenner and Blunsom, 2013; Tran et al., 2017a; Bothe et al., 2018b,a; Zhao and Kawahara, 2018) .",
"cite_spans": [
{
"start": 344,
"end": 366,
"text": "(Stolcke et al., 2000)",
"ref_id": "BIBREF19"
},
{
"start": 444,
"end": 475,
"text": "Kalchbrenner and Blunsom, 2013;",
"ref_id": "BIBREF11"
},
{
"start": 476,
"end": 495,
"text": "Tran et al., 2017a;",
"ref_id": "BIBREF21"
},
{
"start": 496,
"end": 518,
"text": "Bothe et al., 2018b,a;",
"ref_id": null
},
{
"start": 519,
"end": 543,
"text": "Zhao and Kawahara, 2018)",
"ref_id": "BIBREF25"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dialogue Act Recognition",
"sec_num": "1.1"
},
{
"text": "Transfer learning techniques allow a model trained on one task-often unsupervised-to be applied to another. Since annotating natural language data is expensive, there is a lot of interest in transfer learning for natural language processing. Word vectors (e.g., Mikolov et al., 2013; Pennington et al., 2014) are a ubiquitous example of transfer learning in NLP. We note, however, that pre-trained word vectors are not always useful when applied to dialogue (Cerisara et al., 2017) .",
"cite_spans": [
{
"start": 262,
"end": 283,
"text": "Mikolov et al., 2013;",
"ref_id": "BIBREF15"
},
{
"start": 284,
"end": 308,
"text": "Pennington et al., 2014)",
"ref_id": "BIBREF16"
},
{
"start": 458,
"end": 481,
"text": "(Cerisara et al., 2017)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transfer learning for NLP",
"sec_num": "1.2"
},
{
"text": "BERT, a multi-layer transformer model (Devlin et al., 2019) , is pre-trained on two unsupervised tasks: masked token prediction and next sentence prediction. In masked token prediction, some percentage of words are randomly replaced with a mask token. The model is trained to predict the identity of the original token based on the context sentence. In next sentence prediction, the model is given two sentences and trained to predict whether the second sentence follows the first in the original text or if it was randomly chosen from elsewhere in the corpus. After pre-training, BERT can be applied to a supervised task by adding additional un-trained layers that take the hidden state of one or more of BERT's layers as input.",
"cite_spans": [
{
"start": 38,
"end": 59,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transfer learning for NLP",
"sec_num": "1.2"
},
{
"text": "There is some previous work applying BERT to dialogue. Bao et al. (2020) and Chen et al. (2019) both use BERT for dialogue generation tasks. Similarly, Vig and Ramea (2019) find BERT useful for selecting a response from a list of candidate responses in a dialogue. Mehri et al. (2019) evaluate BERT in various dialogue tasks including DAR, and find that a model incorporating BERT outperforms a baseline model. Finally, Chakravarty et al. (2019) use BERT for dialogue act classification for a proprietary domain and achieves promising results, and Ribeiro et al. (2019) surpass the previous state-of-the-art on generic dialogue act recognition for Switchboard and MRDA corpora. This paper aims to supplement the findings of previous work by investigating how much of BERT's success for dialogue tasks is due to its extensive pre-training and how much is due to task-specific fine-tuning.",
"cite_spans": [
{
"start": 55,
"end": 72,
"text": "Bao et al. (2020)",
"ref_id": "BIBREF1"
},
{
"start": 77,
"end": 95,
"text": "Chen et al. (2019)",
"ref_id": "BIBREF7"
},
{
"start": 265,
"end": 284,
"text": "Mehri et al. (2019)",
"ref_id": "BIBREF14"
},
{
"start": 548,
"end": 569,
"text": "Ribeiro et al. (2019)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transfer learning for NLP",
"sec_num": "1.2"
},
{
"text": "Fine-tuning vs. further in-domain pre-training We experiment with the following two transfer learning strategies (Sun et al., 2019) : further pretraining, in which the model is trained in an un-supervised way, similar to its initial training scheme, but on data that is in-domain for the target task; and single-task fine-tuning, in which the Whether or not the encoder model has undergone further in-domain pre-training, there remains a choice of whether to fine-tune during task training, or simply extract features from the encoder model without training it (i.e., freezing). Freezing the encoder model is more efficient, since the gradient of the loss function need only be computed for the task-specific layers. However, fine-tuning can lead to better performance since the encoding itself is adapted to the target task and domain. Peters et al. (2019) investigate when it is best to fine-tune BERT for sentence classification tasks and find that when the target task is very similar to the pre-training task, fine-tuning provides less of a performance boost. We note that there is some conceptual relationship between DAR and next sentence prediction, since the dialogue act constrains (or at least is predictive of) the dialogue act that follows it. That said, the discourse strucutre of the encyclopedia and book data that makes up BERT's pre-training corpus is probably quite different from that of natural dialogue.",
"cite_spans": [
{
"start": 113,
"end": 131,
"text": "(Sun et al., 2019)",
"ref_id": "BIBREF20"
},
{
"start": 837,
"end": 857,
"text": "Peters et al. (2019)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transfer learning for NLP",
"sec_num": "1.2"
},
{
"text": "We perform experiments on the Switchboard Dialogue Act Corpus (SWDA), which is a subset of the larger Switchboard corpus, and the dialogue acttagged portion of the AMI Meeting Corpus (AMI-DA). SWDA is tagged with a set of 220 dialogue act tags which, following Jurafsky et al. (1997) , we cluster into a smaller set of 42 tags. AMI uses a smaller tagset of 16 dialogue acts (Carletta, 2007) . See Table 2 for details.",
"cite_spans": [
{
"start": 261,
"end": 283,
"text": "Jurafsky et al. (1997)",
"ref_id": "BIBREF10"
},
{
"start": 374,
"end": 390,
"text": "(Carletta, 2007)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 397,
"end": 404,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Data",
"sec_num": "2"
},
{
"text": "Preprocessing We make an effort to normalize transcription conventions across SWDA and AMI.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "2"
},
{
"text": "We remove disfluency annotations and slashes from the end of utterances in SWDA. In both corpora, acronyms are tokenized as individual letters. All utterances are lower-cased.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "2"
},
{
"text": "Utterances are tokenized with BERT's word piece tokenizer with a vocabulary of 30,000. To this vocabulary we added five speaker tokens and prepend each utterance with a speaker token that uniquely identifies the corresponding speaker within that dialogue.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "2"
},
{
"text": "We also experiment with three unlabeled dialogue corpora, which we use to provide further pretraining for the BERT encoder.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-training corpora",
"sec_num": "2.1"
},
{
"text": "The first two corpora are constructed from the same source as the dialogue act corpora. We use the SWDA portion of the un-labeled Switchboard corpus (SWBD) and the entire AMI corpus (including the 32 dialogues with no human-annotated DA tags that are not included in the DAR training set). In both cases, we exclude dialogues that are reserved for DAR testing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-training corpora",
"sec_num": "2.1"
},
{
"text": "We also experiment with a much larger a corpus (350M tokens) constructed from OpenSubtitles (Lison and Tiedemann, 2016). Because utterances are not labeled with speaker, we randomly assigned a speaker token to each utterance to maintain the format of the other dialogue corpora.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-training corpora",
"sec_num": "2.1"
},
{
"text": "The pre-training corpora were prepared for the combined masked language modeling and next sentence (utterance) prediction task, as described by Devlin et al. (2019) . For the smaller SWBD and AMI corpora, we generate and train on multiple epochs of data. Since there is randomness in the data preparation (e.g., which distractor sentences are chosen and which tokens are masked), we generate each training epoch separately. 1",
"cite_spans": [
{
"start": 144,
"end": 164,
"text": "Devlin et al. (2019)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-training corpora",
"sec_num": "2.1"
},
{
"text": "We use a simple neural architecture with two components: an encoder that vectorizes utterances (BERT), and single-layer RNN sequence model that takes the utterance representations as input. 2 At each time step, the RNN takes the encoded utterance as input and its hidden state is passed to a linear layer with softmax over dialogue act tags. 3 Conceptually, the encoded utterance represents the context-agnostic features of the utterance, and the hidden state of the RNN represents the full discourse context.",
"cite_spans": [
{
"start": 190,
"end": 191,
"text": "2",
"ref_id": null
},
{
"start": 342,
"end": 343,
"text": "3",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3"
},
{
"text": "For the BERT utterance encoder, we use the BERT BASE model with hidden size of 768 and 12 transformer layers and self-attention heads (Devlin et al., 2019, \u00a73.1) . In our implementation, we use the un-cased model provided by Wolf et al. (2020) . The RNN has a hidden layer size of 100.",
"cite_spans": [
{
"start": 134,
"end": 161,
"text": "(Devlin et al., 2019, \u00a73.1)",
"ref_id": null
},
{
"start": 225,
"end": 243,
"text": "Wolf et al. (2020)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3"
},
{
"text": "First, we analyze how pre-training affects BERT's performance as an utterance encoder. To do so, we consider the performance of DAR models with three different utterance encoders:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-training vs. fine-tuning",
"sec_num": "3.1"
},
{
"text": "\u2022 BERT-FT -pre-trained + DAR fine-tuning \u2022 BERT-FZ -pre-trained, frozen during DAR \u2022 BERT-RI -random init. + DAR fine-tuning BERT-FT is more accurate than BERT-RI by several percentage points on both DA corpora, suggesting that BERT's extensive pre-training does provide some useful information for DAR (Table 3). This performance boost is much more pronounced in the macro-averaged F1 score, 4 which is explained by the fact that at the tag level, pretraining has a larger impact on less frequent tags (see Figure 1 in the supplementary materials) .",
"cite_spans": [],
"ref_spans": [
{
"start": 508,
"end": 548,
"text": "Figure 1 in the supplementary materials)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Pre-training vs. fine-tuning",
"sec_num": "3.1"
},
{
"text": "The BERT-FZ performs very poorly compared to either BERT-FT or BERT-RI, however. It is heavily biased towards the most frequent tags, which explains its especially poor macro-F1 score (Table 3 ). In SWDA, for example, the model with a frozen encoder predicts one of the two most common tags (Statement-non-opinion or Acknowledge) 86% of the time, whereas those two tags account for only 51% of the ground truth tags. BERT-FT is much less biased; it predicts the two most common tags only 59% of the time.",
"cite_spans": [],
"ref_spans": [
{
"start": 184,
"end": 192,
"text": "(Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Pre-training vs. fine-tuning",
"sec_num": "3.1"
},
{
"text": "Next, we assess the effect of additional dialogue pre-training on BERT's performance as an utter-ance encoder. 5 Sun et al. (2019) has reported that performing additional pre-training on unlabeled in-domain data improves performance on classification tasks. We want to see if BERT can benefit from pre-training on dialogue data, including from data outside the immediate target domain.",
"cite_spans": [
{
"start": 111,
"end": 112,
"text": "5",
"ref_id": null
},
{
"start": 113,
"end": 130,
"text": "Sun et al. (2019)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Impact of dialogue pre-training",
"sec_num": "3.2"
},
{
"text": "For each of the target corpora (SWDA and AMI-DA), we compare four different pre-training conditions: The in-domain corpus (ID), consisting of the AMI pre-training corpus for the AMI-DA model and the SWBD pre-traning corpus for the SWDA model; the cross-domain corpus (CC), consisting of both the AMI and SWBD pre-training corpora; and finally the OpenSubtitles corpus (OS). As before, we experiment with both frozen and fine-tuned models at the task training stage.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Impact of dialogue pre-training",
"sec_num": "3.2"
},
{
"text": "We performed 10 epochs of pre-training on the in-domain models and 5 epochs of pre-training on the cross-domain models so that the total amount of training data was comparable. The OpenSubtitles models were trained for only one epoch but with much more total training time.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Impact of dialogue pre-training",
"sec_num": "3.2"
},
{
"text": "In the fine-tuned condition, additional pretraining offers a modest boost in overall accuracy and a substantial boost to the macro-F1 scores, with the cross-domain corpus providing the largest boost. In the frozen condition, only the very large OpenSubtitles corpus is helpful, suggesting that when adapting BERT to dialogue, the size of the corpus is more important than its quality or fidelity to the target domain. Still, pre-training provides nowhere near the performance improvement achieved by fine-tuning on the target task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Impact of dialogue pre-training",
"sec_num": "3.2"
},
{
"text": "A key aspiration of transfer learning is to expose the model to phenomena that are too infrequent to learn from labeled training data alone. We show some evidence of that here. Pre-trained BERT-FT performs better on infrequent dialogue acts than BERT-RI, suggesting it draws on the extensive pre-training to represent infrequent features of those utterances. Indeed, a simple lexical probe supports this explanation: in utterances where the pre-trained model is correct and the randomly initialized model is not, the rarest word is 1.9 times rarer on average than is typical of corpus as a whole. Table 3 : Comparison of macro-F1 and accuracy with further in-domain (ID), cross-domain corpus (CC), and OpenSubtitles (OS) dialogue pre-training, for the frozen (FZ) and fine-tuned (FT) conditions. BERT-RI uses a randomly initialized utterance encoder with no pre-training but with fine-tuning.",
"cite_spans": [],
"ref_spans": [
{
"start": 597,
"end": 604,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "4"
},
{
"text": "In spite of that, the representations learned through pre-training are simply not performant without task-specific fine-tuning, suggesting that they are fundamentally lacking in information that is important for the dialogue context. We should note that this is in stark contrast to many other non-dialogical semantic tasks, where frozen BERT performs on par or better than the fine-tuned model (Peters et al., 2019) .",
"cite_spans": [
{
"start": 395,
"end": 416,
"text": "(Peters et al., 2019)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "4"
},
{
"text": "By performing additional pre-training on a large dialogue-like corpus (OpenSubtitles), we were able to raise the performance of the frozen encoder by a small amount. This deserves further investigation. Bao et al. (2020) find that further pre-training BERT on a large-scale Reddit and Twitter corpus is helpful for response selection, but given the unimpressive results with subtitles, it remains an open question how well the text chat and social media domains transfer to natural dialogue.",
"cite_spans": [
{
"start": 203,
"end": 220,
"text": "Bao et al. (2020)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "4"
},
{
"text": "There is also abundant room to investigate how speech-related information, such as laughter, prosody, and disfluencies can be incorporated into a DAR model that uses pre-trained features. Stolcke et al. (2000) showed, for example, that dialogue acts can have specific prosodic manifestations that can be used to improve dialogue act classification. Incorporating such information is crucial if models pre-trained on large-scale text corpora are to be adapted for use in dialogue applications.",
"cite_spans": [
{
"start": 188,
"end": 209,
"text": "Stolcke et al. (2000)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "4"
},
{
"text": "For details, see the finetuning example from Hugging Face.2 We have experimented with LSTM as the sequence model, but the accuracy was not significantly different compared to RNN. It can be explained by the absence of longer distance dependencies on this level of our model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Other work has shown that DAR benefits from more sophisticated decoding, such as conditional random field(Chen et al., 2018) and uncertainty propagation (Tran et al., 2017b).4 We report both accuracy (which is equal to microaveraged or class-weighted F1) and macro-F1, which is the unweighted average of the F1 scores of each class.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "In-domain pre-training is sometimes referred to as finetuning, but we reserve that term for task-specific training on labeled data.6 Kozareva and Ravi (2019)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "A c k n o w le d g e ( B a c k c h a n n e l) S ta te m e n to p in io n S e g m e n t ( m u lt iu tt e r a n c e ) A b a n ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "annex",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "How to Do Things with Words: The William James Lectures Delivered at Harvard University in 1955",
"authors": [
{
"first": "John",
"middle": [
"L"
],
"last": "Austin",
"suffix": ""
},
{
"first": "James",
"middle": [
"O"
],
"last": "Urmson",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John L. Austin and James O. Urmson. 2009. How to Do Things with Words: The William James Lec- tures Delivered at Harvard University in 1955, 2. ed., [repr.] edition. Harvard Univ. Press, Cambridge, Mass. OCLC: 935786421.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "PLATO: Pre-trained Dialogue Generation Model with Discrete Latent Variable",
"authors": [
{
"first": "Siqi",
"middle": [],
"last": "Bao",
"suffix": ""
},
{
"first": "Huang",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Fan",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Hua",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Haifeng",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "85--96",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.9"
]
},
"num": null,
"urls": [],
"raw_text": "Siqi Bao, Huang He, Fan Wang, Hua Wu, and Haifeng Wang. 2020. PLATO: Pre-trained Dialogue Genera- tion Model with Discrete Latent Variable. In Pro- ceedings of the 58th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 85-96, Online. Association for Computational Linguistics.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Conversational analysis using utterance-level attention-based bidirectional recurrent neural networks",
"authors": [
{
"first": "Chandrakant",
"middle": [],
"last": "Bothe",
"suffix": ""
},
{
"first": "Sven",
"middle": [],
"last": "Magg",
"suffix": ""
},
{
"first": "Cornelius",
"middle": [],
"last": "Weber",
"suffix": ""
},
{
"first": "Stefan",
"middle": [],
"last": "Wermter",
"suffix": ""
}
],
"year": 2018,
"venue": "Proc",
"volume": "",
"issue": "",
"pages": "996--1000",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chandrakant Bothe, Sven Magg, Cornelius Weber, and Stefan Wermter. 2018a. Conversational analysis us- ing utterance-level attention-based bidirectional re- current neural networks. Proc. Interspeech 2018, pages 996-1000.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A Context-based Approach for Dialogue Act Recognition using Simple Recurrent Neural Networks",
"authors": [
{
"first": "Chandrakant",
"middle": [],
"last": "Bothe",
"suffix": ""
},
{
"first": "Cornelius",
"middle": [],
"last": "Weber",
"suffix": ""
},
{
"first": "Sven",
"middle": [],
"last": "Magg",
"suffix": ""
},
{
"first": "Stefan",
"middle": [],
"last": "Wermter",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chandrakant Bothe, Cornelius Weber, Sven Magg, and Stefan Wermter. 2018b. A Context-based Approach for Dialogue Act Recognition using Simple Re- current Neural Networks. In Proceedings of the Eleventh International Conference on Language Re- sources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Unleashing the killer corpus: Experiences in creating the multi-everything AMI Meeting Corpus. Language Resources and Evaluation",
"authors": [
{
"first": "Jean",
"middle": [],
"last": "Carletta",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "41",
"issue": "",
"pages": "181--190",
"other_ids": {
"DOI": [
"10.1007/s10579-007-9040-x"
]
},
"num": null,
"urls": [],
"raw_text": "Jean Carletta. 2007. Unleashing the killer corpus: Experiences in creating the multi-everything AMI Meeting Corpus. Language Resources and Evalu- ation, 41(2):181-190.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "On the effects of using word2vec representations in neural networks for dialogue act recognition",
"authors": [
{
"first": "Christophe",
"middle": [],
"last": "Cerisara",
"suffix": ""
},
{
"first": "Pavel",
"middle": [],
"last": "Kr\u00e1l",
"suffix": ""
},
{
"first": "Ladislav",
"middle": [],
"last": "Lenc",
"suffix": ""
}
],
"year": 2017,
"venue": "Computer Speech & Language",
"volume": "47",
"issue": "",
"pages": "175--193",
"other_ids": {
"DOI": [
"10.1016/j.csl.2017.07.009"
]
},
"num": null,
"urls": [],
"raw_text": "Christophe Cerisara, Pavel Kr\u00e1l, and Ladislav Lenc. 2017. On the effects of using word2vec representa- tions in neural networks for dialogue act recognition. Computer Speech & Language, 47:175-193.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Dialog acts classification for question-answer corpora",
"authors": [
{
"first": "Saurabh",
"middle": [],
"last": "Chakravarty",
"suffix": ""
},
{
"first": "Raja",
"middle": [],
"last": "Venkata Satya Phanindra Chava",
"suffix": ""
},
{
"first": "Edward A",
"middle": [],
"last": "Fox",
"suffix": ""
}
],
"year": 2019,
"venue": "ASAIL@ ICAIL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Saurabh Chakravarty, Raja Venkata Satya Phanindra Chava, and Edward A Fox. 2019. Dialog acts clas- sification for question-answer corpora. In ASAIL@ ICAIL.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Semantically Conditioned Dialog Response Generation via Hierarchical Disentangled Self-Attention",
"authors": [
{
"first": "Wenhu",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Jianshu",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Pengda",
"middle": [],
"last": "Qin",
"suffix": ""
},
{
"first": "Xifeng",
"middle": [],
"last": "Yan",
"suffix": ""
},
{
"first": "William",
"middle": [
"Yang"
],
"last": "Wang",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "3696--3709",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1360"
]
},
"num": null,
"urls": [],
"raw_text": "Wenhu Chen, Jianshu Chen, Pengda Qin, Xifeng Yan, and William Yang Wang. 2019. Semantically Con- ditioned Dialog Response Generation via Hierarchi- cal Disentangled Self-Attention. In Proceedings of the 57th Annual Meeting of the Association for Com- putational Linguistics, pages 3696-3709, Florence, Italy. Association for Computational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Dialogue Act Recognition via CRF-Attentive Structured Network",
"authors": [
{
"first": "Zheqian",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Rongqin",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Zhou",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Deng",
"middle": [],
"last": "Cai",
"suffix": ""
},
{
"first": "Xiaofei",
"middle": [],
"last": "He",
"suffix": ""
}
],
"year": 2018,
"venue": "The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR '18",
"volume": "",
"issue": "",
"pages": "225--234",
"other_ids": {
"DOI": [
"10.1145/3209978.3209997"
]
},
"num": null,
"urls": [],
"raw_text": "Zheqian Chen, Rongqin Yang, Zhou Zhao, Deng Cai, and Xiaofei He. 2018. Dialogue Act Recognition via CRF-Attentive Structured Network. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR '18, pages 225-234, New York, NY, USA. Association for Computing Machinery.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "4171--4186",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1423"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Un- derstanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Switchboard SWBD-DAMSL Shallow-Discourse-Function Annotation Coders Manual",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Jurafsky",
"suffix": ""
},
{
"first": "Liz",
"middle": [],
"last": "Shriberg",
"suffix": ""
},
{
"first": "Debra",
"middle": [],
"last": "Biasca",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Jurafsky, Liz Shriberg, and Debra Biasca. 1997. Switchboard SWBD-DAMSL Shallow-Discourse- Function Annotation Coders Manual.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Recurrent Convolutional Neural Networks for Discourse Compositionality",
"authors": [
{
"first": "Nal",
"middle": [],
"last": "Kalchbrenner",
"suffix": ""
},
{
"first": "Phil",
"middle": [],
"last": "Blunsom",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the Workshop on Continuous Vector Space Models and Their Compositionality",
"volume": "",
"issue": "",
"pages": "119--126",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent Convolutional Neural Networks for Discourse Com- positionality. In Proceedings of the Workshop on Continuous Vector Space Models and Their Compo- sitionality, pages 119-126.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "ProSeqo: Projection Sequence Networks for On-Device Text Classification",
"authors": [
{
"first": "Zornitsa",
"middle": [],
"last": "Kozareva",
"suffix": ""
},
{
"first": "Sujith",
"middle": [],
"last": "Ravi",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "3894--3903",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1402"
]
},
"num": null,
"urls": [],
"raw_text": "Zornitsa Kozareva and Sujith Ravi. 2019. ProSeqo: Projection Sequence Networks for On-Device Text Classification. In Proceedings of the 2019 Confer- ence on Empirical Methods in Natural Language Processing and the 9th International Joint Confer- ence on Natural Language Processing (EMNLP- IJCNLP), pages 3894-3903, Hong Kong, China. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "OpenSubti-tles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles",
"authors": [
{
"first": "Pierre",
"middle": [],
"last": "Lison",
"suffix": ""
},
{
"first": "Jorg",
"middle": [],
"last": "Tiedemann",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 10th International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pierre Lison and Jorg Tiedemann. 2016. OpenSubti- tles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In Proceedings of the 10th International Conference on Language Resources and Evaluation, page 7.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Pretraining Methods for Dialog Context Representation Learning",
"authors": [
{
"first": "Shikib",
"middle": [],
"last": "Mehri",
"suffix": ""
},
{
"first": "Evgeniia",
"middle": [],
"last": "Razumovskaia",
"suffix": ""
},
{
"first": "Tiancheng",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Maxine",
"middle": [],
"last": "Eskenazi",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "3836--3845",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1373"
]
},
"num": null,
"urls": [],
"raw_text": "Shikib Mehri, Evgeniia Razumovskaia, Tiancheng Zhao, and Maxine Eskenazi. 2019. Pretraining Methods for Dialog Context Representation Learn- ing. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3836-3845, Florence, Italy. Association for Computational Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Distributed Representations of Words and Phrases and their Compositionality. NIPS Proceedings",
"authors": [
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Kai",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Greg",
"middle": [
"S"
],
"last": "Corrado",
"suffix": ""
},
{
"first": "Jeff",
"middle": [],
"last": "Dean",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor- rado, and Jeff Dean. 2013. Distributed Representa- tions of Words and Phrases and their Compositional- ity. NIPS Proceedings, page 9.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Glove: Global Vectors for Word Representation",
"authors": [
{
"first": "Jeffrey",
"middle": [],
"last": "Pennington",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "1532--1543",
"other_ids": {
"DOI": [
"10.3115/v1/D14-1162"
]
},
"num": null,
"urls": [],
"raw_text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global Vectors for Word Representation. In Proceedings of the 2014 Con- ference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "To Tune or Not to Tune? Adapting Pretrained Representations to Diverse Tasks",
"authors": [
{
"first": "Matthew",
"middle": [
"E"
],
"last": "Peters",
"suffix": ""
},
{
"first": "Sebastian",
"middle": [],
"last": "Ruder",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019)",
"volume": "",
"issue": "",
"pages": "7--14",
"other_ids": {
"DOI": [
"10.18653/v1/W19-4302"
]
},
"num": null,
"urls": [],
"raw_text": "Matthew E. Peters, Sebastian Ruder, and Noah A. Smith. 2019. To Tune or Not to Tune? Adapt- ing Pretrained Representations to Diverse Tasks. In Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019), pages 7-14, Florence, Italy. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Deep dialog act recognition using multiple token, segment, and context information representations",
"authors": [
{
"first": "Eug\u00e9nio",
"middle": [],
"last": "Ribeiro",
"suffix": ""
},
{
"first": "Ricardo",
"middle": [],
"last": "Ribeiro",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Martins De Matos",
"suffix": ""
}
],
"year": 2019,
"venue": "Journal of Artificial Intelligence Research",
"volume": "66",
"issue": "",
"pages": "861--899",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eug\u00e9nio Ribeiro, Ricardo Ribeiro, and David Martins de Matos. 2019. Deep dialog act recognition using multiple token, segment, and context information representations. Journal of Artificial Intelligence Re- search, 66:861-899.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech",
"authors": [
{
"first": "Andreas",
"middle": [],
"last": "Stolcke",
"suffix": ""
},
{
"first": "Klaus",
"middle": [],
"last": "Ries",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "Coccaro",
"suffix": ""
},
{
"first": "Elizabeth",
"middle": [],
"last": "Shriberg",
"suffix": ""
},
{
"first": "Rebecca",
"middle": [],
"last": "Bates",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Jurafsky",
"suffix": ""
},
{
"first": "Paul",
"middle": [],
"last": "Taylor",
"suffix": ""
},
{
"first": "Rachel",
"middle": [],
"last": "Martin",
"suffix": ""
},
{
"first": "Carol",
"middle": [],
"last": "Van Ess-Dykema",
"suffix": ""
},
{
"first": "Marie",
"middle": [],
"last": "Meteer",
"suffix": ""
}
],
"year": 2000,
"venue": "Computational Linguistics",
"volume": "26",
"issue": "3",
"pages": "339--373",
"other_ids": {
"DOI": [
"10.1162/089120100561737"
]
},
"num": null,
"urls": [],
"raw_text": "Andreas Stolcke, Klaus Ries, Noah Coccaro, Eliza- beth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteer. 2000. Dialogue Act Modeling for Au- tomatic Tagging and Recognition of Conversational Speech. Computational Linguistics, 26(3):339-373.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "How to Fine-Tune BERT for Text Classification?",
"authors": [
{
"first": "Chi",
"middle": [],
"last": "Sun",
"suffix": ""
},
{
"first": "Xipeng",
"middle": [],
"last": "Qiu",
"suffix": ""
},
{
"first": "Yige",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Xuanjing",
"middle": [],
"last": "Huang",
"suffix": ""
}
],
"year": 2019,
"venue": "Chinese Computational Linguistics",
"volume": "",
"issue": "",
"pages": "194--206",
"other_ids": {
"DOI": [
"10.1007/978-3-030-32381-3_16"
]
},
"num": null,
"urls": [],
"raw_text": "Chi Sun, Xipeng Qiu, Yige Xu, and Xuanjing Huang. 2019. How to Fine-Tune BERT for Text Classi- fication? In Chinese Computational Linguistics, Lecture Notes in Computer Science, pages 194-206, Cham. Springer International Publishing.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "A Hierarchical Neural Model for Learning Sequences of Dialogue Acts",
"authors": [
{
"first": "Ingrid",
"middle": [],
"last": "Quan Hung Tran",
"suffix": ""
},
{
"first": "Gholamreza",
"middle": [],
"last": "Zukerman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Haffari",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "428--437",
"other_ids": {
"DOI": [
"10.18653/v1/E17-1041"
]
},
"num": null,
"urls": [],
"raw_text": "Quan Hung Tran, Ingrid Zukerman, and Gholamreza Haffari. 2017a. A Hierarchical Neural Model for Learning Sequences of Dialogue Acts. In Proceed- ings of the 15th Conference of the European Chap- ter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 428-437, Valencia, Spain. Association for Computational Linguistics.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Preserving Distributional Information in Dialogue Act Classification",
"authors": [
{
"first": "Ingrid",
"middle": [],
"last": "Quan Hung Tran",
"suffix": ""
},
{
"first": "Gholamreza",
"middle": [],
"last": "Zukerman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Haffari",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "2151--2156",
"other_ids": {
"DOI": [
"10.18653/v1/D17-1229"
]
},
"num": null,
"urls": [],
"raw_text": "Quan Hung Tran, Ingrid Zukerman, and Gholamreza Haffari. 2017b. Preserving Distributional Informa- tion in Dialogue Act Classification. In Proceed- ings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2151-2156, Copenhagen, Denmark. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Comparison of Transfer-Learning Approaches for Response Selection in Multi-Turn Conversations",
"authors": [
{
"first": "Jesse",
"middle": [],
"last": "Vig",
"suffix": ""
},
{
"first": "Kalai",
"middle": [],
"last": "Ramea",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the Workshop on Dialog System Technology Challenges",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jesse Vig and Kalai Ramea. 2019. Comparison of Transfer-Learning Approaches for Response Selec- tion in Multi-Turn Conversations. In Proceedings of the Workshop on Dialog System Technology Chal- lenges, page 7, Honolulu, Hawaii.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Transformers: State-of-the-Art Natural Language Processing",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Wolf",
"suffix": ""
},
{
"first": "Lysandre",
"middle": [],
"last": "Debut",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Sanh",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Chaumond",
"suffix": ""
},
{
"first": "Clement",
"middle": [],
"last": "Delangue",
"suffix": ""
},
{
"first": "Anthony",
"middle": [],
"last": "Moi",
"suffix": ""
},
{
"first": "Pierric",
"middle": [],
"last": "Cistac",
"suffix": ""
},
{
"first": "Tim",
"middle": [],
"last": "Rault",
"suffix": ""
},
{
"first": "Remi",
"middle": [],
"last": "Louf",
"suffix": ""
},
{
"first": "Morgan",
"middle": [],
"last": "Funtowicz",
"suffix": ""
},
{
"first": "Joe",
"middle": [],
"last": "Davison",
"suffix": ""
},
{
"first": "Sam",
"middle": [],
"last": "Shleifer",
"suffix": ""
},
{
"first": "Clara",
"middle": [],
"last": "Patrick Von Platen",
"suffix": ""
},
{
"first": "Yacine",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Jernite",
"suffix": ""
},
{
"first": "Canwen",
"middle": [],
"last": "Plu",
"suffix": ""
},
{
"first": "Teven",
"middle": [
"Le"
],
"last": "Xu",
"suffix": ""
},
{
"first": "Sylvain",
"middle": [],
"last": "Scao",
"suffix": ""
},
{
"first": "Mariama",
"middle": [],
"last": "Gugger",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Drame",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
"volume": "",
"issue": "",
"pages": "38--45",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-demos.6"
]
},
"num": null,
"urls": [],
"raw_text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier- ric Cistac, Tim Rault, Remi Louf, Morgan Funtow- icz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Trans- formers: State-of-the-Art Natural Language Process- ing. In Proceedings of the 2020 Conference on Em- pirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "A unified neural architecture for joint dialog act segmentation and recognition in spoken dialog system",
"authors": [
{
"first": "Tianyu",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Tatsuya",
"middle": [],
"last": "Kawahara",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue",
"volume": "",
"issue": "",
"pages": "201--208",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tianyu Zhao and Tatsuya Kawahara. 2018. A unified neural architecture for joint dialog act segmentation and recognition in spoken dialog system. In Pro- ceedings of the 19th Annual SIGdial Meeting on Dis- course and Dialogue, pages 201-208.",
"links": null
}
},
"ref_entries": {
"TABREF1": {
"html": null,
"content": "<table/>",
"num": null,
"type_str": "table",
"text": ""
},
"TABREF3": {
"html": null,
"content": "<table><tr><td>: Comparison between Switchboard and the</td></tr><tr><td>AMI Meeting Corpus</td></tr><tr><td>model's encoder layers are optimized during train-</td></tr><tr><td>ing for the target task.</td></tr></table>",
"num": null,
"type_str": "table",
"text": ""
},
"TABREF4": {
"html": null,
"content": "<table><tr><td/><td colspan=\"2\">SWDA</td><td colspan=\"2\">AMI-DA</td></tr><tr><td/><td>F1</td><td>acc.</td><td>F1</td><td>acc.</td></tr><tr><td>BERT-FZ</td><td colspan=\"4\">7.75 55.61 14.86 48.34</td></tr><tr><td>BERT+ID-FZ</td><td colspan=\"4\">6.46 52.30 14.48 48.18</td></tr><tr><td>BERT+CC-FZ</td><td colspan=\"4\">5.76 51.14 11.34 40.48</td></tr><tr><td>BERT+OS-FZ</td><td colspan=\"4\">9.60 57.67 17.03 51.03</td></tr><tr><td>BERT-RI</td><td colspan=\"4\">32.18 73.80 34.88 60.89</td></tr><tr><td>Majority class</td><td colspan=\"2\">0.78 33.56</td><td colspan=\"2\">1.88 28.27</td></tr><tr><td>SotA</td><td/><td>-83.1 6</td><td>-</td><td>-</td></tr></table>",
"num": null,
"type_str": "table",
"text": "BERT-FT 36.75 76.60 43.42 64.93 BERT+ID-FT 43.63 77.01 46.70 68.88 BERT+CC-FT 47.78 77.35 48.86 68.79 BERT+OS-FT 41.42 76.95 48.65 68.07"
}
}
}
} |