File size: 97,014 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:24:28.317917Z"
},
"title": "\"Politeness, you simpleton!\" retorted [MASK]: Masked prediction of literary characters",
"authors": [
{
"first": "Eric",
"middle": [],
"last": "Holgate",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The University of Texas at Austin",
"location": {}
},
"email": "holgate@utexas.edu"
},
{
"first": "Katrin",
"middle": [],
"last": "Erk",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The University of Texas at Austin",
"location": {}
},
"email": "katrin.erk@utexas.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "What is the best way to learn embeddings for entities, and what can be learned from them? We consider this question for the case of literary characters. We address the highly challenging task of guessing, from a sentence in the novel, which character is being talked about, and we probe the embeddings to see what information they encode about their literary characters. We find that when continuously trained, entity embeddings do well at the masked entity prediction task, and that they encode considerable information about the traits and characteristics of the entities.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "What is the best way to learn embeddings for entities, and what can be learned from them? We consider this question for the case of literary characters. We address the highly challenging task of guessing, from a sentence in the novel, which character is being talked about, and we probe the embeddings to see what information they encode about their literary characters. We find that when continuously trained, entity embeddings do well at the masked entity prediction task, and that they encode considerable information about the traits and characteristics of the entities.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Neural language models have led to huge improvements across many tasks in the last few years (Peters et al., 2018; Devlin et al., 2019; Radford et al., 2019) .",
"cite_spans": [
{
"start": 93,
"end": 114,
"text": "(Peters et al., 2018;",
"ref_id": "BIBREF14"
},
{
"start": 115,
"end": 135,
"text": "Devlin et al., 2019;",
"ref_id": "BIBREF4"
},
{
"start": 136,
"end": 157,
"text": "Radford et al., 2019)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "1 They compute embeddings for words and word pieces. But when we describe the semantics of a sentence, we talk about entities and events and their relations, not words. And it is to be expected that more complex reasoning tasks would eventually require representations at the semantic level rather than the word level. Entities differ from words in that they are persistent, localized, and variable (within a given range). So, would it be beneficial to compute embeddings of entities in addition to embeddings of words for downstream inference? And how should entity embeddings be computed? There has been a steady rise in work on entity representations and how they can be combined with language models, for example Li et al. (2016) ; ; Rashkin et al. (2018) ; Louis and Sutton (2018) .In this paper, we add to the growing literature on neural representations of entities by considering a particularly challenging case: the representations of entities in very long texts, in particular in novels. Intriguingly, Bruera (2019) recently tested whether literary characters, when represented through distributional vectors trained on the first half of a novel, can be recognized in the second half, and found the task to be near impossible. We take up that same task, but train character embeddings in a masked character prediction task. We ask the following questions. (a) Is it possible to use literary character embeddings to do masked character prediction, that is, to guess from a sentence in a novel which character it mentions? (b) If this task is doable, is it doable only locally, or can we train on the first third of a novel and then guess characters towards the end of the novel? (c) What do the resulting embeddings tell us about the literary characters when we probe them? (d) Can the embeddings identify a literary character from a short description of their personality?",
"cite_spans": [
{
"start": 717,
"end": 733,
"text": "Li et al. (2016)",
"ref_id": null
},
{
"start": 738,
"end": 759,
"text": "Rashkin et al. (2018)",
"ref_id": "BIBREF17"
},
{
"start": 762,
"end": 785,
"text": "Louis and Sutton (2018)",
"ref_id": "BIBREF11"
},
{
"start": 1012,
"end": 1025,
"text": "Bruera (2019)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We find that when continuously trained, entity embeddings do well at the masked entity prediction task, and that they encode considerable information about the traits and characteristics of the entities. Modeling semantics for natural language understanding is about modeling entities and events, not words. So we view this work as an initial step in the direction of entity modeling over time.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Entities have been increasingly common subjects within NLP research. There has been recent work aimed at inducing both characteristics of entities, such as personalities, physical and mental states, and character traits, as well as distributed entity representations, similar to lexical embeddings.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Psychologists have studied the relationship between personality traits and human behavior. Within NLP, there have been recent attempts to model this link computationally. Bamman et al. (2013) explored entity modeling by using Bayesian methods to induce moderately fine-grained character archetypes/stereotypes from film plot summaries. The authors utilized dependency relations to identify, for each entity, the verbs for which they were the agent, the verbs for which they were the patient, and any modifiers attributed to them. Bamman et al. successfully induced clusters that could be manually aligned with tropes like the jerk jock, the nerdy klutz, the villain, etc.",
"cite_spans": [
{
"start": 171,
"end": 191,
"text": "Bamman et al. (2013)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modeling Personalities and Characteristics",
"sec_num": "2.1"
},
{
"text": "Plank and Hovy (2015) recently appealed to psychological personality dimensions in relation to linguistic behavior. They constructed a dataset by crawling twitter for mentions of any of the 16 Myers-Briggs Type Indicators comprising four personality dimensions (MBTI; Myers and Myers 2010), labeling tweets with author gender identity. Plank and Hovy then train logistic regression models to predict each of the four dimensions from user tweet data using tweet context features and other features that are traditional for Twitter data (e.g., counts of tweets, followers, favorites, etc.). In all four dimensions, logistic regression classifiers outperform majority baselines, supporting the notion that linguistic behavior correlates with MBTI designations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Modeling Personalities and Characteristics",
"sec_num": "2.1"
},
{
"text": "Flekova and Gurevych (2015) similarly explored personality traits, though they utilized the Five-Factor Model of personality instead of MBTIs (John et al., 1999) . Here, authors collected extraversion/intraversion ratings for approximately 300 literary characters, and explore three sources of signal to predict the extraversion scores. The first system aligns most closely with Plank and Hovy's work as it considers only character speech (both style and content). Flekova and Gurevych go slightly farther, however, as they also show that character actions and behaviors as well as the descriptions of characters given in narration carry useful signal for extraversion prediction. Rashkin et al. (2018) modeled the mental state of characters in short stories, including motivations for behaviors and emotional reactions to events. The authors noted a substantial increase in performance in mental state classification when entity-specific contextual information was presented to the classifier, suggesting that entity-specific context may be useful to a wide array of downstream tasks. Louis and Sutton (2018) further explored the relation between character properties and actions taken in online role-playing game data. In Dungeons and Dragons, a giant is more likely than a fairy to wield a giant axe, but a fairy is more likely to be agile or cast spells. Louis and Sutton show that computational models can capture this interaction by using character description information in conjunction with action descriptions to train action and character language models. When a formal representation of a given character is included, performance improves. demonstrated dynamically tracked cooking ingredients, identifying which ingredient entity was selected in any given recipe step, and recognizing what changes in state they underwent as a result of the action described in the step. For example, these dynamic entity representations enabled the model to determine that an ingredient was clean after having been washed.",
"cite_spans": [
{
"start": 142,
"end": 161,
"text": "(John et al., 1999)",
"ref_id": "BIBREF8"
},
{
"start": 681,
"end": 702,
"text": "Rashkin et al. (2018)",
"ref_id": "BIBREF17"
},
{
"start": 1086,
"end": 1109,
"text": "Louis and Sutton (2018)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modeling Personalities and Characteristics",
"sec_num": "2.1"
},
{
"text": "Recently, major work in NLP has begun to explicitly model entities for use in downstream tasks. While still new (and limited in scope), much of this work has relied upon the notion of an Entity Library, a vocabulary of individuals which utilizes consecutive mentions to construct distributed vector representations, though methods of learning these representations have varied. Entity representations have been shown to improve the quality of generated text. In Ji et al. (2017) , researchers build a generative language model (an RNN) which has access to an entity library which contains continuous, dynamic representations of each entity mentioned in the text. The result is that the library explicitly groups coreferential mentions, and each generated mention affects the subsequently generated text.",
"cite_spans": [
{
"start": 462,
"end": 478,
"text": "Ji et al. (2017)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Entity Representations and Entity Libraries",
"sec_num": "2.2"
},
{
"text": "Tracking entity information has also been shown to be useful for increasing the consistency of responses in dialogue agents (Li et al., 2016) . Researchers introduce a conversation model which maintains a persona, defined as the character that the artificial agent performs during conversational interactions. The persona maintains elements of identity such as background facts, linguistic behav-ior (dialect), and interaction style (personality) in continuously updated distributed representations. The model maintains the capability for the persona to be adaptive, as the agent may need to present different characteristics to different interlocutors as interactions take place, but reduces the likelihood of the model providing contradictory information (i.e., maintaining these distributed representations prevents the model from claiming to live in both Los Angeles, Madrid, and England in consecutive queries). Crucially, this desirable change is achieved without the need for a structured ontology of properties, but instead through persona embeddings that are learned jointly with word representations. Fevry et al. (2020) demonstrates that entity representations trained only from text can capture more declarative knowledge about those entities than a similarly sized BERT. Researchers showed that these representations are useful for a variety of downstream tasks, including open domain question answering, relation extraction, entity typing, and generalized knowledge tasks.",
"cite_spans": [
{
"start": 124,
"end": 141,
"text": "(Li et al., 2016)",
"ref_id": null
},
{
"start": 1111,
"end": 1130,
"text": "Fevry et al. (2020)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Entity Representations and Entity Libraries",
"sec_num": "2.2"
},
{
"text": "Yamada et al. (2020) explore another entity masking task in the context of transformer pretraining. They train a large transformer on both masked words and masked entities in Wikipedia text. Here, however, each entity-in-context exists as its own token, rather than a representation that is aggregated over a sequence of mentions. Yamada et al. test on entity typing, relation classification, and named entity recognition.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Entity Representations and Entity Libraries",
"sec_num": "2.2"
},
{
"text": "Finally, Bruera (2019) introduces the data that we will use to build our model (described in detail below), and compares the ability to construct computational embeddings for proper names with that of common nouns. Researchers trained a distributional semantics model to create and store two different representations for literary characters in novels, each from a separate section of text from the novel. The model is then asked to match the characters' representations from one portion of text to the representations computed from the other portion of text, which the authors term the Doppelg\u00e4nger Task. Importantly, their results showed that the ability to match these representations is much reduced in the case of proper names when compared to common nouns. This insight serves as a major motivation for the current work, where we follow the hypothesis that entities can be represented in a distributional fashion after all, though not with the same training as with common nouns. 2 We assume that entity representations must be persistent, continuously available, and dynamic.",
"cite_spans": [
{
"start": 9,
"end": 22,
"text": "Bruera (2019)",
"ref_id": "BIBREF3"
},
{
"start": 986,
"end": 987,
"text": "2",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Entity Representations and Entity Libraries",
"sec_num": "2.2"
},
{
"text": "3 Data",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Entity Representations and Entity Libraries",
"sec_num": "2.2"
},
{
"text": "In the current paper, we present a model that is able to construct entity representations for characters in classic literary novels. Novels are a compelling environment for this exploration as they feature a relatively small number of entities that appear frequently over a long document. To this end, we turn to the Novel Aficionados dataset introduced by Bruera (2019). The dataset comprises 62 pieces of classic literature, represented as both their original texts (deemed the OriginalNovels dataset; these texts are distributed by Project Gutenberg, which maintains a repository of free eBooks of works no longer protected by copyright), and their English Wikipedia summaries (the WikiNovels dataset). In order to have sufficient description of as many characters as possible, we only utilize the corpus of original novels in training our representations, as this corpus yields significantly more mentions per character. We utilize the Wikipedia summaries as a test set to determine how well our entity representations work outside the domain of the novels themselves.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Entity Representations and Entity Libraries",
"sec_num": "2.2"
},
{
"text": "The novels are distributed within the dataset in both their original form and having been preprocessed with BookNLP (Bamman et al., 2014) . BookNLP is a natural language processing pipeline that extends from Stanford CoreNLP (Manning et al., 2014) and is specifically aimed at scaling to books and other long documents. BookNLP includes part of speech tagging, dependency parsing, NER, and supersense tagging. Most critical to our application, BookNLP provides quotation speaker identification, pronominal coreference resolution, and character name clustering. This means that, in addition to standard anaphoric coreference resolution, BookNLP can identify different proper names as character aliases (i.e., Jane Fairfax, a character from Jane Austen's Emma, is referenced throughout the text not only by her full name, but also Jane, Miss Jane Fairfax, and Miss Fairfax; BookNLP is able to recognize this and map all of these aliases to a single, unique character ID). Concerning the quality of the coreference resolution, Bamman et al. report average accuracy of 82.7% in a 10-fold cross-validation experiment on predicting the nearest antecedent for a pronominal anaphor. While the accuracy of character clustering was not evaluated, manual inspection of the data revealed it to be very reliable.",
"cite_spans": [
{
"start": 116,
"end": 137,
"text": "(Bamman et al., 2014)",
"ref_id": "BIBREF1"
},
{
"start": 225,
"end": 247,
"text": "(Manning et al., 2014)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Entity Representations and Entity Libraries",
"sec_num": "2.2"
},
{
"text": "Our hypothesis is that it is possible to represent characters in a novel through an embedding in such a way that it is possible for a model to recognize who is who, or, as we call the task here, Is this Me?. Bruera (2019) found that with an approach that treated characters like common nouns, the related Doppelg\u00e4nger Task was not feasible. 4 We hypothesize that if embeddings are learned to best facilitate Is this Me? prediction, the task will be feasible. We further hypothesize that the resulting embeddings can be found to contain information about the characters. In a way, our approach is similar to recent contextualized language models like BERT (Devlin et al., 2019 ) in that we, too, train on a masked prediction task, and we, too, hope to find the resulting embeddings to be useful beyond the prediction task itself.",
"cite_spans": [
{
"start": 208,
"end": 221,
"text": "Bruera (2019)",
"ref_id": "BIBREF3"
},
{
"start": 341,
"end": 342,
"text": "4",
"ref_id": null
},
{
"start": 655,
"end": 675,
"text": "(Devlin et al., 2019",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modeling",
"sec_num": "4"
},
{
"text": "4.1 A model for the \"Is this Me?\" task",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Modeling",
"sec_num": "4"
},
{
"text": "Our model keeps track of characters as they appear in a novel, and trains an embedding for each character through the Is this Me? task, a masked prediction task: Given a sentence of the novel with a masked character mention, and given the current embedding for character c, a classifier decides whether this is a mention of c or not. This is a binary task. The embedding for each character is updated incrementally as the novel is read, and as such, the entity embeddings are learned directly by the model. 4 Although related, the Is this Me? and Doppelg\u00e4nger tasks are truly different in nature. As such, we cannot compare results on the Is this Me? task to results on the Doppelg\u00e4nger Task directly. from the data. The classifier weights are updated alongside the character embeddings.",
"cite_spans": [
{
"start": 507,
"end": 508,
"text": "4",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modeling",
"sec_num": "4"
},
{
"text": "Because the classifier weights are learned as the model reads the novels, we read all novels in parallel. The classifier is trained on a binary masked prediction task, where negative examples are drawn from the same novel. (That is, a negative example for Emma in the novel Emma might be Harriet, but it would never be Heathcliff.) A sketch of the model is shown in Figure 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 366,
"end": 374,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Modeling",
"sec_num": "4"
},
{
"text": "Entity Library. The entity library, shown in blue in Figure 1 is a collection of embeddings of literary characters, each represented by a 300 dimensional embedding learned incrementally throughout the novel. Entity embeddings are randomly initialized and passed through a projection layer (green in the figure) before being received by the classifier.",
"cite_spans": [],
"ref_spans": [
{
"start": 53,
"end": 61,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Modeling",
"sec_num": "4"
},
{
"text": "Contextual Sentence and Target Mention Representations. We utilize the base, uncased distribution of BERT to compute contextualized sentence representations of each target sentence, shown in orange in Figure 1 . Contextualized sentence representations are truncated to a maximum of 150 subword tokens. 5 We do not fine tune BERT on our data. All character mentions in a sentence are masked. The input to the classifier is a target representation of one of the masked entity mentions, using mix representations introduced in (Tenney et al., 2019). The target mention representation is computed directly from the contextualized sentence representations obtained from BERT and is a scalar mix of the layer activations using learned scalars.",
"cite_spans": [
{
"start": 302,
"end": 303,
"text": "5",
"ref_id": null
}
],
"ref_spans": [
{
"start": 201,
"end": 209,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Modeling",
"sec_num": "4"
},
{
"text": "Is this Me? Binary Classifier. The classifier for the binary Is this Me? task takes as input an entity embedding, transformed through the projection layer, along with a target mention embedding from BERT as described above. The classifier consists of a single, ReLU activation layer. We keep the classifier this simple intentionally, as, to be successful, we want the entity representations to do the heavy lifting.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Modeling",
"sec_num": "4"
},
{
"text": "Training Data. We restrict our modeling to characters that appear at least 10 times to ensure that there is enough information to train a representation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model details",
"sec_num": "4.2"
},
{
"text": "As our intent is to induce entity representations for each character, we must mask each character mention. For each mention of any character predicted by BookNLP in each sentence in a novel, we replace the mention with a single [MASK] token in order to obscure the character's identity from the model. Multiword mentions are reduced to a single [MASK] token in order to prevent the model from being able to detect signal from mention length. Masking is applied to any mention, even for characters that appear fewer than 10 times.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model details",
"sec_num": "4.2"
},
{
"text": "For each sentence in a novel that contains at least one character mention, we produce at least two examples for the model: one positive example from the gold data, and one hallucinated example by randomly selecting a confound character from the same novel. If a character is mentioned more than one time in the same sentence, one mention is randomly selected to be the target mention for that character in that sentence. If a sentence talks about more than one character, a single positive example is generated for each character. Consider this sentence from Jane Austen's Emma:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model details",
"sec_num": "4.2"
},
{
"text": "Whenever [MASK] James goes over to see [MASK] James' daughter, you know, [MASK] Miss Taylor will be hearing of us.",
"cite_spans": [
{
"start": 39,
"end": 45,
"text": "[MASK]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model details",
"sec_num": "4.2"
},
{
"text": "We first have to decide whether to first generate examples for James or for Miss Taylor. We pick one of the two at random, let us assume it is James. We next randomly select one of the two mentions of James to be the target mention. Let us say we pick the first. The input to the model for the posi-tive example is then the Tenney et al. (2019) mix representation of the target mention concatenated with the current entity representation of James. We then construct a negative example by randomly selecting a character other than James to serve as a confound, following standard practice. If, for example, we were to sample Isabella (Emma's sister), the input to the model for the negative example from this mention would be the exact same mix representation of the target mention concatenated with the current entity embedding of the confound character, Isabella. With positive and negative examples constructed for James's mention, we then turn to the remaining character, Miss Taylor, and construct a positive and negative example for her mention.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model details",
"sec_num": "4.2"
},
{
"text": "Note that restricting the possible set of confounds for a given character to characters in the same novel, we have created a more difficult negative example than if we were to sample across all novels. For example, telling the difference between Elizabeth Bennet and Jane Bennet (both from Austen's Pride and Prejudice) is significantly more difficult than telling the difference between Elizabeth Bennet and the Cowardly Lion (from Baum's The Wonderful Wizard of Oz).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model details",
"sec_num": "4.2"
},
{
"text": "Training. All learned weights (the entity embeddings themselves, those in the projection layer, the scalars guiding the target mention representation, and those in the classifier) are updated with respect to cross entropy loss, optimized with Adam (Kingma and Ba, 2015) at a learning rate of 2e-05.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model details",
"sec_num": "4.2"
},
{
"text": "We first address our questions (a) and (b) from above: Is it possible to predict a masked mention of a literary character from an entity embedding, either within the same novel or in a summary of the same novel? And does performance degrade if we \"skip ahead\", using a character embedding trained on the beginning of a novel to predict a mention near the end?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Is this Me?",
"sec_num": "5.1"
},
{
"text": "We begin by allowing the model to train entity representations (and all other learned weights) continuously throughout each novel. This means that we treat each example as a test example, and only allow the model to update based on its performance on a given example after its prediction has been made, as in a standard learning curve. As such, although the model is updated after every example, our performance statistics are computed over its prediction made before the update operation (meaning there is no performance computed over alreadyseen examples). As Table 2 shows, the model does well at this task, with overall accuracy across all characters and all novels of 74.37%. Accuracy was consistent across positive and negative examples. Most learning happened quickly within the first 50,000 examples, though accuracy did continue to increase through the entire run ( Figure 2 ). As should be expected, overall accuracy at the book level in this task is subject to frequency effects. Book-level accuracy exhibits strong positive correlation with the total number of examples per novel (r = 0.584; p \u226a 0.01; Figure 3, left) . Interestingly, however, book-level accuracy also increases with the number of characters modeled per book (r = 0.500; p \u226a 0.01; Figure 3, right) .",
"cite_spans": [],
"ref_spans": [
{
"start": 562,
"end": 569,
"text": "Table 2",
"ref_id": "TABREF3"
},
{
"start": 875,
"end": 883,
"text": "Figure 2",
"ref_id": "FIGREF1"
},
{
"start": 1114,
"end": 1129,
"text": "Figure 3, left)",
"ref_id": "FIGREF2"
},
{
"start": 1260,
"end": 1276,
"text": "Figure 3, right)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Continuous Training",
"sec_num": "5.1.1"
},
{
"text": "To see whether the model is affected by language differences between older and more recent books, we used linear regression to predict book-level accuracy from novel publication date, finding very low correlation (R 2 = 0.008; p = 0.663). At the character level, frequency effects were not nearly as strong, except in cases where characters were mentioned very frequently (defined here as characters with over 300 mentions). Across all characters, testing showed moderate positive correlation with mention frequency (r = 0.174; p \u226a 0.01; Figure 4 , left). Within frequently appearing characters, correlation with mention frequency was much higher (r = 0.633; p \u226a 0.01; Figure 4 , right). ",
"cite_spans": [],
"ref_spans": [
{
"start": 538,
"end": 546,
"text": "Figure 4",
"ref_id": "FIGREF3"
},
{
"start": 669,
"end": 677,
"text": "Figure 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Continuous Training",
"sec_num": "5.1.1"
},
{
"text": "We also explored the extent to which the entity representations after having been trained on the full novel, could identify the same entities in short summaries of the same novel. To that end, we used the the WikiNovel summaries distributed with the Novel Aficionados dataset. The summaries show a strong domain shift compared to the novels. While they frequently do contain proper, if succinct, descriptions of the novel's plot and the involvement of major characters, they also exhibit significantly different patterns of language. Wikipedia entries do not just summarize the novels, they also frequently include metadiscursive language, as in this sentence from the WikiNovels summary of Austen's Emma:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Applicability to Novel Summaries",
"sec_num": "5.1.2"
},
{
"text": "This point of view appears both as something perceived by [emma woodhouse] an external perspective on events and characters that the reader encounters as and when [emma woodhouse] recognises it and as an independent discourse appearing in the text alongside the discourse of the narrator and characters.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Applicability to Novel Summaries",
"sec_num": "5.1.2"
},
{
"text": "Because of this shift in domain, we see vastly reduced performance in character prediction and a heavy bias towards claiming the target mention is not a given character when using the model trained on the sentences from the novel. This is shown in Table 3 . We evaluated the model in two settings. In the Pos. Only setting, all data points were positives, such that the model would have perfect accuracy by always saying yes. In the Pos. & Neg. setting, we use the same negative example generation technique as used in the model's training. While the model performs slightly better than chance when negative examples are included, it remains clear that future work should explore ways to generalize the entity representations such that they may be more informative across domain boundaries. ",
"cite_spans": [],
"ref_spans": [
{
"start": 248,
"end": 255,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Applicability to Novel Summaries",
"sec_num": "5.1.2"
},
{
"text": "In \u00a72 we noted that identifying masked character mentions is a not trivial due to the nature of narratives themselves. Literary plots are often constructed to force profound change in the behaviors, beliefs, and characteristics of central characters. This may be among the reasons that Bruera (2019) reported such difficulty with the Doppelg\u00e4nger task. To see if distance in the novel affects our representations, we experimented with \"skipping ahead\" in the novel in order to determine the impact on performance when entities are not continuously updated.",
"cite_spans": [
{
"start": 286,
"end": 299,
"text": "Bruera (2019)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Non-Continuous Training",
"sec_num": "5.1.3"
},
{
"text": "Inspired by traditional character arcs, we split each novel into three sections of equal length (determined by number of sentences). The underlying assumption is that, due to the structure of the narrative, each character (especially main characters) will undergo some form of growth or change in between each novel section, suggesting that the learned entity representations should never be static in order to encode the results of that growth. We allowed a new Is this Me? classifier to learn representations for all literary entities using only the first third of the novels as training data, then froze the entity embeddings, and evaluated classifier performance against the middle and final thirds independently. We hypothesized that the model would exhibit a gradual decrease in performance as it moved further away from the point in time at which the entity representations were fixed, with the performance on the middle third better than performance toward the ends of the novels. Instead, we found a fairly rapid decline in performance (Table 4) . Performance stays above chance, however, suggesting there is a kernel of each representation that is informative regardless. While this experiment does not explicitly demonstrate character development/change, the sharp decrease in performance when entity representations are fixed implicitly supports the claim that such change is present. Capturing that development directly, however, is another very difficult task and well-worthy of being the subject of future work. ",
"cite_spans": [],
"ref_spans": [
{
"start": 1045,
"end": 1054,
"text": "(Table 4)",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Non-Continuous Training",
"sec_num": "5.1.3"
},
{
"text": "We have found that, at least when entity embeddings are continuously trained, they can be used to predict a masked mention in a novel with reasonable accuracy. But are the resulting embeddings useful beyond the masked entity prediction task? To find this out, we turn to our questions (c) and (d) from above, and see if we can predict character gender from entity representation, and whether the identity of a character can be predicted from a description.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Probing Entity Embeddings",
"sec_num": "5.2"
},
{
"text": "We used a simple logistic regression model to probe the extent to which gender is encoded in the entity representations obtained from the continuous training in \u00a75.1.1. As we have no gold annotation of literary character gender, we utilize the BookNLP preprocessing to look for gendered pronouns (she/he) for each character as a form of distant supervision. Manual inspection shows this heuristic to be very reliable after omitting characters for which no pronominal coreference link is available and characters who exhibit coreference chains featuring both gendered pronouns. This left a total of 2,195 characters (1,533 male, 662 female) to be considered for this experiment. We learn a single weight for each embedding dimension for a total of 300 weights. In each case, we trained the classifiers on 80% of the characters across all novels (1,756 characters), leaving a test set of 439 characters. Each model was run four times, and we present the mean performance statistics in Table 5 . Results were favorable across all runs, suggesting the learned character representations do encapsulate some knowledge of literary character gender.",
"cite_spans": [],
"ref_spans": [
{
"start": 983,
"end": 990,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Predicting Gender from Literary Character Representation",
"sec_num": "5.2.1"
},
{
"text": "\u00b5 Acc \u00b5 MSE \u00b5 F1 60.15% 0.3984 0.7208 Table 5 : Model performance on predicting character gender from entity embeddings: Accuracy, mean squared error, and F1.",
"cite_spans": [],
"ref_spans": [
{
"start": 38,
"end": 45,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Predicting Gender from Literary Character Representation",
"sec_num": "5.2.1"
},
{
"text": "While the WikiNovels corpus is noisy and cluttered with metadiscursive literary commentary, as noted in \u00a75.1.2, certain Wikipedia novel summaries do contain detailed descriptions of major characters. To better evaluate the ability of our learned entity representations to generalize outside of the domain of the novels on which they where trained, we manually extracted a subset of sentences which more readily pertained to our research question.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Character Descriptions",
"sec_num": "5.2.2"
},
{
"text": "We isolated five novels which featured clean character descriptions within their summaries: Jane Austen's Emma, Charles Dickens's A Tale of Two Cities and Great Expectations, Fyodor Dostoevsky's The Brothers Karamazov, and Charlotte Bront\u00eb's Jane Eyre. From the character descriptions within these summaries we generated a total of 605 Is this Me?-style examples (positive and negative).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Character Descriptions",
"sec_num": "5.2.2"
},
{
"text": "6 The pre-trained classifier exhibited performance above chance (61.63% accuracy), and a surprising ability to handle challenging out of domain sentences. While the model successfully predicted a high level description of Emma Woodhouse (Table 6 ; Row 1), it struggled with a similar description of Estella Havisham (Row 2). The model was also able to identify a character based on the description of a pivotal plot point (Row 3), but unsurprisingly struggled with more critical descriptions (Row 4).",
"cite_spans": [],
"ref_spans": [
{
"start": 237,
"end": 245,
"text": "(Table 6",
"ref_id": "TABREF9"
}
],
"eq_spans": [],
"section": "Character Descriptions",
"sec_num": "5.2.2"
},
{
"text": "In the ideal case, an entity embedding would constitute a compact representation of a person, their character traits and life story, and would allow for inferences about that person, including story arcs in which that person is likely to occur. What is the best way to learn embeddings for entities, and what can be learned from them? We have considered this question for the case of literary characters. We have trained entity embeddings through a masked prediction task, reading a collection of novels from begininng to end. We found that when trained continuously, the entity embeddings did well at the Is this Me? task: Given a target entity embedding, and given a sentence of the novel with a masked entity mention, is this a mention of the target entity?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "The Is this Me? task becomes much harder when we \"skip ahead\", training only on the first third of a novel and then evaluating on the middle and end. The task also becomes much harder when applied to Wikipedia summaries of novels, which show a marked domain difference from the novels themselves. Probing the entity embeddings that result from the masked prediction task, we find that they encode a good amount of information about the She is the golden thread after whom book the second is named so called because [MASK] holds her father and her family lives together and because of her blond hair like her mother. entities. The gender of the literary character can in many cases be recovered from the embedding, and it is even often possible to identify a person from a Wikipedia description of their characteristic traits.",
"cite_spans": [
{
"start": 515,
"end": 521,
"text": "[MASK]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "Looking ahead, the training regime and trained embeddings allow for many further analyses. We would like to probe further into the \"skipping ahead\" to better understand why it is so difficult. Intuitively, characters that undergo more development across the length of a novel should be more difficult to predict. It is not clear to what extent this is the case with the current model; this needs further study. In addition, we would like to model the change and development of characters more explicitly, for example by representing them as a trajectory over time rather than a single embedding. It would also be beneficial to further explore the ways in which character traits are implicitly present within entity representations learned from the Is this Me? task. While we attempted to probe this superficially via the evaluation on out-of-domain Wikipedia data, this data does not offer the annotation that would be necessary to perform a more in-depth analysis",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "We would also like to extend the model by including additional relevant input. At the moment, we essentially ask the model to bootstrap entity representations from scratch, using only the contextualized sentence representations produced by BERT and the current entity representations as input. Other useful information such as semantic relations (retrievable via dependency parse) may be useful. We may also consider the kind of events and modifiers that a given entity participates in to be able to exploit patterns across character archetypes (similar to Bamman et al. (2014) ). We are also looking to extend the model to directly model relations between characters as relations between entity embeddings, to see whether this would help performance and to see to what extent the interpersonal relations of characters would be encoded in their embeddings.",
"cite_spans": [
{
"start": 557,
"end": 577,
"text": "Bamman et al. (2014)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "Overall, we find the results presented in the current paper to be promising as a first step towards natural language understanding systems that use neural models of entities over time. As we have outlined here, however, there is still much work to be done.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "The[MASK] in the title is actually La Carconte, from the Count of Monte Cristo by Alexandre Dumas.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "2 While our work is inspired byBruera (2019) and conducted on the same data, we introduce a different task that is not directly comparable to the Doppelg\u00e4nger Task.3 Unfortunately, the texts are not distributed with more general coreference resolution (outside of character aliases and pronominal resolution). This means we are unable to include nominal expressions as character mentions to be considered",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "This limit was determined by inspecting the length of each sentence in the corpus in subword tokens and permits nearly all sentences to remain untruncated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "This set of examples may be found at http://www.katrinerk.com/home/software-and-data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This research was supported by the DARPA AIDA program under AFRL grant FA8750-18-2-0017. We acknowledge the Texas Advanced Computing Center for providing grid resources that contributed to these results, and results presented in this paper were obtained using the Chameleon testbed supported by the National Science Foundation. We would like to thank the anonymous reviewers for their valuable feedback, as well as Jessy Li and Pengxiang Cheng.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": "7"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Learning latent personas of film characters",
"authors": [
{
"first": "David",
"middle": [],
"last": "Bamman",
"suffix": ""
},
{
"first": "Brendan",
"middle": [],
"last": "Oconnor",
"suffix": ""
},
{
"first": "Noah A",
"middle": [],
"last": "Smith",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "352--361",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Bamman, Brendan OConnor, and Noah A Smith. 2013. Learning latent personas of film characters. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 352-361.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A Bayesian mixed effects model of literary character",
"authors": [
{
"first": "David",
"middle": [],
"last": "Bamman",
"suffix": ""
},
{
"first": "Ted",
"middle": [],
"last": "Underwood",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "Smith",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Bamman, Ted Underwood, and Noah Smith. 2014. A Bayesian mixed effects model of literary character. In Proceedings of the 52nd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers).",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Simulating action dynamics with neural process networks",
"authors": [
{
"first": "Antoine",
"middle": [],
"last": "Bosselut",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Ari",
"middle": [],
"last": "Holtzman",
"suffix": ""
},
{
"first": "Corin",
"middle": [],
"last": "Ennis",
"suffix": ""
},
{
"first": "Dieter",
"middle": [],
"last": "Fox",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 6th International Conference for Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Antoine Bosselut, Omer Levy, Ari Holtzman, Corin En- nis, Dieter Fox, and Yejin Choi. 2018. Simulating action dynamics with neural process networks. In Proceedings of the 6th International Conference for Learning Representations (ICLR).",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Modelling the semantic memory of proper names with distributional semantics. Master's thesis",
"authors": [
{
"first": "Andrea",
"middle": [],
"last": "Bruera",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrea Bruera. 2019. Modelling the semantic memory of proper names with distributional semantics. Mas- ter's thesis, Universita degli Studi di Trento.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"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 under- standing. 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
},
"BIBREF5": {
"ref_id": "b5",
"title": "Entities as experts: Sparse memory access with entity supervision",
"authors": [
{
"first": "Thibault",
"middle": [],
"last": "Fevry",
"suffix": ""
},
{
"first": "Baldini",
"middle": [],
"last": "Livio",
"suffix": ""
},
{
"first": "Nicholas",
"middle": [],
"last": "Soares",
"suffix": ""
},
{
"first": "Eunsol",
"middle": [],
"last": "Fitzgerald",
"suffix": ""
},
{
"first": "Tom",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kwiatkowski",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "4937--4951",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.400"
]
},
"num": null,
"urls": [],
"raw_text": "Thibault Fevry, Livio Baldini Soares, Nicholas FitzGer- ald, Eunsol Choi, and Tom Kwiatkowski. 2020. En- tities as experts: Sparse memory access with entity supervision. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 4937-4951. Association for Computational Linguistics.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Personality profiling of fictional characters using sense-level links between lexical resources",
"authors": [
{
"first": "Lucie",
"middle": [],
"last": "Flekova",
"suffix": ""
},
{
"first": "Iryna",
"middle": [],
"last": "Gurevych",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1805--1816",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lucie Flekova and Iryna Gurevych. 2015. Personal- ity profiling of fictional characters using sense-level links between lexical resources. In Proceedings of the 2015 Conference on Empirical Methods in Natu- ral Language Processing, pages 1805-1816.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Dynamic entity representations in neural language models",
"authors": [
{
"first": "Yangfeng",
"middle": [],
"last": "Ji",
"suffix": ""
},
{
"first": "Chenhao",
"middle": [],
"last": "Tan",
"suffix": ""
},
{
"first": "Sebastian",
"middle": [],
"last": "Martschat",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1830--1839",
"other_ids": {
"DOI": [
"10.18653/v1/D17-1195"
]
},
"num": null,
"urls": [],
"raw_text": "Yangfeng Ji, Chenhao Tan, Sebastian Martschat, Yejin Choi, and Noah A. Smith. 2017. Dynamic entity representations in neural language models. In Pro- ceedings of the 2017 Conference on Empirical Meth- ods in Natural Language Processing, pages 1830- 1839, Copenhagen, Denmark. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "The Big Five trait taxonomy: History, measurement, and theoretical perspectives",
"authors": [
{
"first": "Sanjay",
"middle": [],
"last": "Oliver P John",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Srivastava",
"suffix": ""
}
],
"year": 1999,
"venue": "Handbook of personality: Theory and research",
"volume": "2",
"issue": "",
"pages": "102--138",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Oliver P John, Sanjay Srivastava, et al. 1999. The Big Five trait taxonomy: History, measurement, and the- oretical perspectives. Handbook of personality: The- ory and research, 2(1999):102-138.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Adam: A method for stochastic optimization",
"authors": [
{
"first": "P",
"middle": [],
"last": "Diederik",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ba",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. CoRR, abs/1412.6980.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Georgios P Spithourakis, Jianfeng Gao, and Bill Dolan. 2016. A persona-based neural conversation model",
"authors": [
{
"first": "Jiwei",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Michel",
"middle": [],
"last": "Galley",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Brockett",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jiwei Li, Michel Galley, Chris Brockett, Georgios P Sp- ithourakis, Jianfeng Gao, and Bill Dolan. 2016. A persona-based neural conversation model.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Deep Dungeons and Dragons: Learning character-action interactions from role-playing game transcripts",
"authors": [
{
"first": "Annie",
"middle": [],
"last": "Louis",
"suffix": ""
},
{
"first": "Charles",
"middle": [],
"last": "Sutton",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "2",
"issue": "",
"pages": "708--713",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Annie Louis and Charles Sutton. 2018. Deep Dun- geons and Dragons: Learning character-action in- teractions from role-playing game transcripts. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 708-713.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "The Stanford CoreNLP natural language processing toolkit",
"authors": [
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
},
{
"first": "Mihai",
"middle": [],
"last": "Surdeanu",
"suffix": ""
},
{
"first": "John",
"middle": [],
"last": "Bauer",
"suffix": ""
},
{
"first": "Jenny",
"middle": [],
"last": "Finkel",
"suffix": ""
},
{
"first": "Steven",
"middle": [
"J"
],
"last": "Bethard",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Mc-Closky",
"suffix": ""
}
],
"year": 2014,
"venue": "Association for Computational Linguistics (ACL) System Demonstrations",
"volume": "",
"issue": "",
"pages": "55--60",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David Mc- Closky. 2014. The Stanford CoreNLP natural lan- guage processing toolkit. In Association for Compu- tational Linguistics (ACL) System Demonstrations, pages 55-60.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Gifts Differing: Understanding Personality Type",
"authors": [
{
"first": "Isabel",
"middle": [
"Briggs"
],
"last": "Myers",
"suffix": ""
},
{
"first": "Peter B",
"middle": [],
"last": "Myers",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Isabel Briggs Myers and Peter B Myers. 2010. Gifts Differing: Understanding Personality Type. Nicholas Brealey.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Deep contextualized word representations",
"authors": [
{
"first": "Matthew",
"middle": [],
"last": "Peters",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Neumann",
"suffix": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Iyyer",
"suffix": ""
},
{
"first": "Matt",
"middle": [],
"last": "Gardner",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "2227--2237",
"other_ids": {
"DOI": [
"10.18653/v1/N18-1202"
]
},
"num": null,
"urls": [],
"raw_text": "Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word rep- resentations. In Proceedings of the 2018 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana. Association for Computational Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Personality traits on Twitter -or-how to get 1,500 personality tests in a week",
"authors": [
{
"first": "Barbara",
"middle": [],
"last": "Plank",
"suffix": ""
},
{
"first": "Dirk",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 6th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis",
"volume": "",
"issue": "",
"pages": "92--98",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Barbara Plank and Dirk Hovy. 2015. Personality traits on Twitter -or-how to get 1,500 personality tests in a week. In Proceedings of the 6th Workshop on Computational Approaches to Subjectivity, Sen- timent and Social Media Analysis, pages 92-98.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Language models are unsupervised multitask learners",
"authors": [
{
"first": "Alec",
"middle": [],
"last": "Radford",
"suffix": ""
},
{
"first": "Jeff",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Rewon",
"middle": [],
"last": "Child",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Luan",
"suffix": ""
},
{
"first": "Dario",
"middle": [],
"last": "Amodei",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Modeling naive psychology of characters in simple commonsense stories",
"authors": [
{
"first": "Hannah",
"middle": [],
"last": "Rashkin",
"suffix": ""
},
{
"first": "Antoine",
"middle": [],
"last": "Bosselut",
"suffix": ""
},
{
"first": "Maarten",
"middle": [],
"last": "Sap",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "2289--2299",
"other_ids": {
"DOI": [
"10.18653/v1/P18-1213"
]
},
"num": null,
"urls": [],
"raw_text": "Hannah Rashkin, Antoine Bosselut, Maarten Sap, Kevin Knight, and Yejin Choi. 2018. Modeling naive psychology of characters in simple common- sense stories. In Proceedings of the 56th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 2289- 2299, Melbourne, Australia. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "What do you learn from context? probing for sentence structure in contextualized word representations",
"authors": [
{
"first": "Ian",
"middle": [],
"last": "Tenney",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Xia",
"suffix": ""
},
{
"first": "Berlin",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Poliak",
"suffix": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Mccoy",
"suffix": ""
},
{
"first": "Najoung",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Van Durme",
"suffix": ""
},
{
"first": "Sam",
"middle": [],
"last": "Bowman",
"suffix": ""
},
{
"first": "Dipanjan",
"middle": [],
"last": "Das",
"suffix": ""
},
{
"first": "Ellie",
"middle": [],
"last": "Pavlick",
"suffix": ""
}
],
"year": 2019,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R Thomas McCoy, Najoung Kim, Benjamin Van Durme, Sam Bowman, Dipanjan Das, and Ellie Pavlick. 2019. What do you learn from context? probing for sentence structure in contextu- alized word representations. In International Con- ference on Learning Representations.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "LUKE: Deep contextualized entity representations with entityaware self-attention",
"authors": [
{
"first": "Ikuya",
"middle": [],
"last": "Yamada",
"suffix": ""
},
{
"first": "Akari",
"middle": [],
"last": "Asai",
"suffix": ""
},
{
"first": "Hiroyuki",
"middle": [],
"last": "Shindo",
"suffix": ""
},
{
"first": "Hideaki",
"middle": [],
"last": "Takeda",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "6442--6454",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.523"
]
},
"num": null,
"urls": [],
"raw_text": "Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, and Yuji Matsumoto. 2020. LUKE: Deep contextualized entity representations with entity- aware self-attention. In Proceedings of the 2020 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), pages 6442-6454, On- line. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Sketch of the model. The sentence is from Jane Austen's \"Emma\". All characters are masked. In this case, the first character is Knightley, the second -the target -is Emma. Blue: entity library. Red: Is this Me? classifier."
},
"FIGREF1": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Is this Me? Continuous Training learning curve."
},
"FIGREF2": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Is this Me? Continuous Training -Book-Level Accuracy. Accuracy within book (y-axis) is plotted against the number of examples for that book (xaxis)."
},
"FIGREF3": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Is this Me? Continuous Training -Character-Level Accuracy. Accuracy within character (y-axis) is plotted against the number of examples for that character (x-axis)."
},
"TABREF1": {
"content": "<table><tr><td>Alexei Fyodorovich</td><td/></tr><tr><td>Karamazov</td><td>True/False</td></tr><tr><td>Heathcliff</td><td/></tr><tr><td/><td>Is this Me?</td></tr><tr><td>Emma Woodhouse</td><td/></tr><tr><td>Harriet Smith</td><td>span</td></tr><tr><td/><td>repr.</td></tr><tr><td/><td>{</td></tr><tr><td>BERT</td><td/></tr></table>",
"type_str": "table",
"text": "[MASK], in fact, was one of the few people who could see faults in[MASK].",
"num": null,
"html": null
},
"TABREF3": {
"content": "<table/>",
"type_str": "table",
"text": "Is this Me? accuracy for continuously trained entity representations.",
"num": null,
"html": null
},
"TABREF5": {
"content": "<table/>",
"type_str": "table",
"text": "Is this Me? accuracy for continuously trained entity representations on WikiNovel summaries.",
"num": null,
"html": null
},
"TABREF7": {
"content": "<table/>",
"type_str": "table",
"text": "Is this Me? accuracy on novels split into thirds.",
"num": null,
"html": null
},
"TABREF8": {
"content": "<table><tr><td colspan=\"2\">Novel</td><td>Target</td><td colspan=\"3\">Candidate Result Sentence</td></tr><tr><td colspan=\"6\">Emma [Great Emma Emma + Estella Estella -She hates all men and plots to wreak twisted</td></tr><tr><td colspan=\"2\">Expecta-</td><td/><td/><td/><td>revenge by teaching [MASK] to torment and</td></tr><tr><td>tions</td><td/><td/><td/><td/><td>spurn men, including Pip who loves her.</td></tr><tr><td>A</td><td>Tale</td><td>Miss</td><td>Miss</td><td>+</td><td>[MASK] permanently loses her hearing when</td></tr><tr><td colspan=\"2\">of Two</td><td>Pross</td><td>Pross</td><td/><td>the fatal pistol shot goes off during her climac-</td></tr><tr><td colspan=\"2\">Cities</td><td/><td/><td/><td>tic fight with Madame Defarge.</td></tr><tr><td>A</td><td>Tale</td><td>Lucy</td><td>Lucy</td><td>-</td><td/></tr><tr><td colspan=\"2\">of Two</td><td>Manette</td><td>Manette</td><td/><td/></tr><tr><td colspan=\"2\">Cities</td><td/><td/><td/><td/></tr></table>",
"type_str": "table",
"text": "MASK] the protagonist of the story is beautiful high spirited intelligent and lightly spoiled young woman from the landed gentry.",
"num": null,
"html": null
},
"TABREF9": {
"content": "<table/>",
"type_str": "table",
"text": "Examples of the Is this Me? continuously trained classifier's performance on out-of-domain masked mentions found within the WikiNovels corpus. Non-target mentions have been de-masked for better readability.",
"num": null,
"html": null
}
}
}
} |