File size: 82,073 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 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 |
{
"paper_id": "2022",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:13:24.225358Z"
},
"title": "What GPT Knows About Who is Who",
"authors": [
{
"first": "Xiaohan",
"middle": [],
"last": "Yang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harvard University",
"location": {}
},
"email": "yang@g.harvard.edu"
},
{
"first": "Eduardo",
"middle": [],
"last": "Peynetti",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harvard University",
"location": {}
},
"email": "eduardo.peynetti@gmail.com"
},
{
"first": "Vasco",
"middle": [],
"last": "Meerman",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harvard University",
"location": {}
},
"email": "vmeerman@g.harvard.edu"
},
{
"first": "Chris",
"middle": [],
"last": "Tanner",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harvard University",
"location": {}
},
"email": "christanner@g.harvard.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Coreference resolution-which is a crucial task for understanding discourse and language at large-has yet to witness widespread benefits from large language models (LLMs). Moreover, coreference resolution systems largely rely on supervised labels, which are highly expensive and difficult to annotate, thus making it ripe for prompt engineering. In this paper, we introduce a QA-based prompt-engineering method and discern generative, pre-trained LLMs' abilities and limitations toward the task of coreference resolution. Our experiments show that GPT-2 and GPT-Neo can return valid answers, but that their capabilities to identify coreferent mentions are limited and promptsensitive, leading to inconsistent results.",
"pdf_parse": {
"paper_id": "2022",
"_pdf_hash": "",
"abstract": [
{
"text": "Coreference resolution-which is a crucial task for understanding discourse and language at large-has yet to witness widespread benefits from large language models (LLMs). Moreover, coreference resolution systems largely rely on supervised labels, which are highly expensive and difficult to annotate, thus making it ripe for prompt engineering. In this paper, we introduce a QA-based prompt-engineering method and discern generative, pre-trained LLMs' abilities and limitations toward the task of coreference resolution. Our experiments show that GPT-2 and GPT-Neo can return valid answers, but that their capabilities to identify coreferent mentions are limited and promptsensitive, leading to inconsistent results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Coreference resolution (CR) aims to identify and cluster all words (i.e., mentions) that refer to the same entity or event. Solving this task is essential for natural language understanding, as mismatched references will lead to bias. Recent improvements in CR have been incremental (Lee et al., 2017; Cattan et al., 2020) , compared to other NLP tasks that have demonstrated more real-world impact. One reason is the limited training corpora. For example, one of the primary datasets, ECB+ (Cybulska and Vossen, 2014) , contains only 984 documents, including 6,833 mentions and 2,741 clusters. Moreover, this dataset was built around 43 news topics ten years ago, potentially leading to generalization problems for the state-of-the-art (SOTA) models.",
"cite_spans": [
{
"start": 283,
"end": 301,
"text": "(Lee et al., 2017;",
"ref_id": "BIBREF10"
},
{
"start": 302,
"end": 322,
"text": "Cattan et al., 2020)",
"ref_id": "BIBREF2"
},
{
"start": 491,
"end": 518,
"text": "(Cybulska and Vossen, 2014)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "When dealing with low-resource tasks, there is an emerging trend to perform prompt engineering with pre-trained LMs. Unlike fine-tuning (Brown et al., 2020; Wei et al., 2021) , prompt engineering does not update the pre-trained model's weights when completing the downstream task. Instead, one transforms the downstream task to match the original task of the pre-trained model (Liu et al., 2021) . For example, for machine translation, one can create prompts such as \"English: I love bread. French:\" and input them to a generative LM (e.g., . If the pre-trained model encountered similar patterns during training, it should be able to generate the translated French sentence. Nevertheless, to the best of our knowledge, there is scarce research on applying this approach to coreference resolution (Sanh et al., 2021) .",
"cite_spans": [
{
"start": 136,
"end": 156,
"text": "(Brown et al., 2020;",
"ref_id": null
},
{
"start": 157,
"end": 174,
"text": "Wei et al., 2021)",
"ref_id": "BIBREF22"
},
{
"start": 377,
"end": 395,
"text": "(Liu et al., 2021)",
"ref_id": "BIBREF12"
},
{
"start": 797,
"end": 816,
"text": "(Sanh et al., 2021)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To better understand if pre-trained LMs can help resolve coreferences, we construct a QA-based prompting method and experiment with both GPT-2 (Radford et al., 2019) and GPT-Neo (Gao et al., 2020) . By using this prompting methodology, we measure if these models can predict whether two mentions are coreferent. For evaluation, we use the ECB+ dataset, which provides gold mentions and clustering labels. We compare the results with unsupervised and supervised coreference resolution models, including a classic rule-based system (Lee et al., 2011) , the seminal end-to-end neural model (Lee et al., 2017) , and a recent SOTA model (Cattan et al., 2020) .",
"cite_spans": [
{
"start": 143,
"end": 165,
"text": "(Radford et al., 2019)",
"ref_id": "BIBREF16"
},
{
"start": 170,
"end": 196,
"text": "GPT-Neo (Gao et al., 2020)",
"ref_id": null
},
{
"start": 530,
"end": 548,
"text": "(Lee et al., 2011)",
"ref_id": "BIBREF9"
},
{
"start": 587,
"end": 605,
"text": "(Lee et al., 2017)",
"ref_id": "BIBREF10"
},
{
"start": 632,
"end": 653,
"text": "(Cattan et al., 2020)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Prompt-based learning Prompt-based learning is a fast-growing area in NLP, as it can reduce the need to fine-tune models and rely on supervised labels. According to the survey by Liu et al., over 120 papers have been published since 2019, which collectively demonstrates effectiveness toward many different tasks: text classification (Tam et al., 2021; Holtzman et al., 2021) , factual probing (Perez et al., 2021) , question-answering (Tsimpoukelli et al., 2021) , and more. Nevertheless, to the best of our knowledge, only one prompt-based learning paper concerned CR. Specifically, Sanh et al. introduces T0, a zero-shot generalization of T5 (Raffel et al., 2019) . The authors convert various supervised datasets into task-specific prompts, Figure 1 : An example of prompt-based learning for CR. The green block represents the prefix, which serves as the description of the CR task and remains unchanged throughout an experiment for all inputs x. The purple block is the unfilled prompt, which changes for each input x and serves as the prediction. Moreover, in each block, the yellow part is the prompting function while the blue and red parts are the original data x and y, respectively. including CR. Using the WSC dataset (Levesque et al., 2012) , they achieve over 60% accuracy. Although this result is not comparable with supervised state-of-the-art (SOTA) models, it still offers compelling results and suggests the model might contain CR knowledge without requiring supervised training on the task. However, since the WSC dataset only focuses on highly ambiguous pronouns, it is not as complete as the standard CR task that involves named and nominal mentions.",
"cite_spans": [
{
"start": 334,
"end": 352,
"text": "(Tam et al., 2021;",
"ref_id": null
},
{
"start": 353,
"end": 375,
"text": "Holtzman et al., 2021)",
"ref_id": "BIBREF7"
},
{
"start": 394,
"end": 414,
"text": "(Perez et al., 2021)",
"ref_id": "BIBREF15"
},
{
"start": 436,
"end": 463,
"text": "(Tsimpoukelli et al., 2021)",
"ref_id": "BIBREF21"
},
{
"start": 645,
"end": 666,
"text": "(Raffel et al., 2019)",
"ref_id": "BIBREF17"
},
{
"start": 1230,
"end": 1253,
"text": "(Levesque et al., 2012)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 745,
"end": 753,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Traditional CR Models Similar to other NLP tasks, most CR models can be categorized as being either unsupervised or supervised. A commonly used unsupervised model is the Multi-Pass Sieve model (Lee et al., 2011) . This rule-based system extracts entity mentions and clusters them by applying 13 \"filters\" in successive manner. Amongst supervised models, e2e-coref (Lee et al., 2017) is the seminal end-to-end neural model. This model performs within-document CR and was trained on the OntoNotes (CoNLL-2012) dataset. Building on this architecture, Cattan et al. (2020) performs cross-document CR for entities and events by training on the ECB+ dataset and using RoBERTa as an encoder. Although supervised models offer significant improvements over unsupervised models, they are expensive to train; most SOTA models have O(n 4 ) complexity, where n is the length of each document.",
"cite_spans": [
{
"start": 193,
"end": 211,
"text": "(Lee et al., 2011)",
"ref_id": "BIBREF9"
},
{
"start": 364,
"end": 382,
"text": "(Lee et al., 2017)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "This section introduces our prompt-based learning method for CR. Typically, CR models can be broken down into three sub-tasks: (1) detecting mentions; (2) predicting whether two given mentions are coreferent or not; (3) and clustering mentions accordingly. The crux of CR research centers around the second part, which is also our focus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Building on the approach introduced by Sanh et al. 2021, we define our input x as [text, m 1 , m 2 ] and output y as a binary label. Specifically, m 1 and m 2 are a pair of gold mentions in a document, and the text are the sentences containing those mentions. For example, in Figure 1, within each green box, the successive blue parts are text, m 1 , m 2 , respectively. We define a prompting function f , which takes x as input and produces a question prompt q x (Equation 1). Further details about f are in Appendix A.",
"cite_spans": [],
"ref_spans": [
{
"start": 276,
"end": 282,
"text": "Figure",
"ref_id": null
}
],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "q x = f (x) (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Moreover, to allow the model to understand the task, we use few-shot learning (Triantafillou et al., 2017) by constructing a filled prefix. In particular, we select k examples, A, from the training dataset and feed these examples into the same prompting function f . Then, we append the true label ('Yes' or 'No') to the outputs, yielding the filled prefix q A (Equation 2). To be clear, each individual prefix q i\u2208k constitutes a single green box in Figure 1 .",
"cite_spans": [
{
"start": 78,
"end": 106,
"text": "(Triantafillou et al., 2017)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [
{
"start": 451,
"end": 459,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "q A = f (A) (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Last, adding the unfilled prompt q x to the filled prefix q A will give us the full prompt for data point x. This allows us to get a prediction z without updating any parameters \u03b8 in the pre-trained LM.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "z = P (q A + q x ; \u03b8)",
"eq_num": "(3)"
}
],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Since we use pre-trained LMs directly, without fine-tuning, we do not have control over its output; the model can generate invalid answers beyond our desired outputs, 'Yes' or 'No'. Therefore, we repeat the process m times to get a more robust predictionz. To mitigate the bias of one specific f , we average the output of n different prompt formulas to get the final prediction (Equation 4).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "y = n i=1z i n (4) 4 Experimental Setup",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Datasets We use the ECB+ dataset (Cybulska and Vossen, 2014) as our input source, which contains both within-and cross-document coreference information for both event and entity mentions. This dataset consists of 984 documents around 43 news topics, among which 196 documents are in the development set. After preprocessing the data, as described in Appendix B, our development set consists of 172 documents.",
"cite_spans": [
{
"start": 33,
"end": 60,
"text": "(Cybulska and Vossen, 2014)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "To generate a prefix x 0 , we experiment with three data sources: the training sets of WSC (Levesque et al., 2012) and ECB+ (Cybulska and Vossen, 2014) , and a simple dataset that we manually generated. The WSC dataset was used in the research most similar to ours, T0 (Sanh et al., 2021 ), which we compare against while using much smaller pretrained LMs (i.e., GPT-2 and GPT-Neo). As mentioned, ECB+ provides more natural and comprehensive references than WSC. Our manually generated dataset uses 10 very simple examples -allowing one to discern the impact on performance.",
"cite_spans": [
{
"start": 91,
"end": 114,
"text": "(Levesque et al., 2012)",
"ref_id": "BIBREF11"
},
{
"start": 124,
"end": 151,
"text": "(Cybulska and Vossen, 2014)",
"ref_id": "BIBREF3"
},
{
"start": 269,
"end": 287,
"text": "(Sanh et al., 2021",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "When using the ECB+ dataset, we only considered pairs of mentions that are within the same or successive sentences. When evaluating our model, we considered all mention-pair combinations, [m 1 , m 2 ], within said sentences. Relying on the gold mentions, we obtain a dataset with 17832 candidate mention pairs, among which 7.86% are positive samples. Finally, we apply 5 prompt functions from Sanh et al. to generate the full prompts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Models We used three traditional CR models as baselines: Multi-Pass Sieve (Lee et al., 2011) , the seminal end-to-end neural model (e2e-coref ) (Lee et al., 2017) , and a SOTA extension (the Streamlining model) (Cattan et al., 2020) . Respectively, these models represent three categories: a rulesbased model, a supervised model trained on a different dataset, and a supervised model trained on the same dataset. In terms of implementations, we use the CoreNLP toolkit for the Multi-Pass Sieve model (Manning et al., 2014) and AllenNLP (Gardner et al., 2018) for e2e-coref. Since there is no publicly available pre-trained Streamlining model (Cattan et al., 2020) , we fully train the model from scratch using a V100 GPU on Google Colab. To fairly compare with other models, we set a 0.5 threshold for the pairwise scorer in the Streamlining model. We evaluate all models by their mention pairwise scorers, not their clustering decisions.",
"cite_spans": [
{
"start": 74,
"end": 92,
"text": "(Lee et al., 2011)",
"ref_id": "BIBREF9"
},
{
"start": 144,
"end": 162,
"text": "(Lee et al., 2017)",
"ref_id": "BIBREF10"
},
{
"start": 211,
"end": 232,
"text": "(Cattan et al., 2020)",
"ref_id": "BIBREF2"
},
{
"start": 500,
"end": 522,
"text": "(Manning et al., 2014)",
"ref_id": "BIBREF14"
},
{
"start": 642,
"end": 663,
"text": "(Cattan et al., 2020)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Limited by our computational resources, we choose GPT-2 and GPT-Neo-125M as our pretrained LMs 1 . During inference, the output token length is set to 1, since our expected output is one word (i.e., 'Yes' or 'No') . To generate more robust results, the repetition parameter m is set to 5. We ran our text generative models with multiple temperature settings ranging from 0 to 1, none of which produced significant changes. We settled on using a value of 0.7, to limit the greediness of the generated responses. In terms of few-shot learning, we experimented with k \u2208 {0, 2, 4, 10} and display the results from the 4-shot setting since it produces the best accuracy. To reduce bias introduced by prefixes, we ensure each prefix has equally-balanced samples. For example, for the 4-shot setting, the filled prefix will have 2 positive examples and 2 negative examples.",
"cite_spans": [
{
"start": 199,
"end": 213,
"text": "'Yes' or 'No')",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Yes/No Predictions 0-shot 5% 2-shot 93.7% 4-shot 96.2% 10-shot 98% We first question if GPT-based models can produce valid answers. In Figure 1 , we observe that GPT-2 predicts 'Yes' or 'No' for over 93.7% samples when at least 2 filled prefixes are provided.",
"cite_spans": [],
"ref_spans": [
{
"start": 135,
"end": 143,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5"
},
{
"text": "However, although the answers are valid, they are inaccurate. In Figure 2 , we plot the distribution of predicted labels for each model, where the red bars denote the distribution of positive examples (ground truth), and the blue bars denote negative ones (ground truth). Traditional CR models generally predict low values for negative examples, indicated by blue bars being concentrated at 0. As for positive examples, e2e-coref shows better precision since more positive examples are classified correctly at 1. Yet, GPT-2 seems to be both sensitive to prompts and unstable over the repetitions of each prompt. Furthermore, GPT-Neo's predictions are inaccurate and no better than random, even though it predicts consistent results for multiple runs with the same prompt.",
"cite_spans": [],
"ref_spans": [
{
"start": 65,
"end": 73,
"text": "Figure 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5"
},
{
"text": "Similar conclusions can be drawn from Table 2 , where GPT-based models have the lowest AUC and F1 scores. Specifically, the extremely low precision causes the bad results. Since the ECB+ dataset is highly imbalanced, random predictions from GPTbased models will lead to a low precision, reflecting the proportion of positive samples. For completeness, we also perform an experiment on the WSC dataset (see GPT-2 wsc ), which is a test dataset used by Sanh et al. (2021) . GPT-2 also fails on this task, as its mean prediction averaged across different prompts is always \"Yes\" . POS and Entity Types While the overall performance indicates that GPT models are comparable to a random model, we hypothesize that for some subset of mention pairs, GPT might perform better.",
"cite_spans": [
{
"start": 451,
"end": 469,
"text": "Sanh et al. (2021)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 38,
"end": 45,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5"
},
{
"text": "To investigate, we conducted experiments based on part-of-speech (POS) tags and named-entity types. Figure 3 shows that both GPT-2 and GPT-Neo can capture coreferent relationships relatively better when the second mention is a pronoun. Moreover, this trend is stronger when the first mention is a pronoun or a proper noun. Nonetheless, e2e-coref performs better than both GPT models across all POS tags, and the gap is widest when the second mention is a nominal noun phrase. As for named entities, Figure 4 shows that both GPT-2 and GPT-Neo perform better in precision when one mention is of type PERSON. Moreover, GPT-Neo can identify coreferent relationships more precisely if the second mention is Non-GPE locations (i.e., LOC). However, their precision scores are far lower than the scores from classical CR models. In particular, both the multi-pass sieve model and e2e-coref model reach the highest precision when a mention is a PRODUCT object (e.g., vehicle, food) or a NORP object (e.g., nationality, religious or political group). Mention Similarity In addition to inspecting how performance varies with mention types, we also considered how performance is affected by mentions' similarity. Using pre-trained BERT (Devlin et al., 2018) , we encode each mention into span representations by averaging its tokens' last hidden states. Then, we measure cosine similarity between mention pairs. Figure 5 shows that F1 scores generally improve as the semantic similarity increases. Although, the multi-pass sieve model maintains a low F1 because it is a rule-based model that tends to predict False for most samples -which yields a high accuracy for unbalanced datasets. The e2e-coref model performs well on mentions that are not so similar, while the performance of Streamlining model improves drastically as similarity is greater than 50%. However, both GPT-2 and GPT-NEO yield low F1 (approximately 0.2) for mention pairs with less than 70% similarity. When considering mentions of higher similarity, GPT-based models can achieve over 0.4 F1 score.",
"cite_spans": [
{
"start": 1224,
"end": 1245,
"text": "(Devlin et al., 2018)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 100,
"end": 108,
"text": "Figure 3",
"ref_id": "FIGREF1"
},
{
"start": 499,
"end": 507,
"text": "Figure 4",
"ref_id": "FIGREF2"
},
{
"start": 1400,
"end": 1408,
"text": "Figure 5",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Acc Prec",
"sec_num": null
},
{
"text": "In this paper, we rely on prompt-based learning to analyze how much GPT-like models know about coreference resolution. Despite the popularity of prompting in recent NLP research, we find that LLMs perform poorly on this task without finetuning. Nonetheless, these models achieve relatively better performance for specific types of mentions, including pronouns and person objects, and mention pairs with high similarity.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "A Prompt formulas Figure 6 : Prompt Formulas. We experiment with these 5 prompt formulas mentioned in Sanh et al. (2021) . Here, each block is one formula and the parts highlighted in blue are [text, m 1 , m 2 ] respectively.",
"cite_spans": [
{
"start": 102,
"end": 120,
"text": "Sanh et al. (2021)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 18,
"end": 26,
"text": "Figure 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "The original ECB+ dataset is in XML format, where everything is tokenized. Moreover, the information about gold mentions and gold clusters is related to token ids. However, we cannot easily get the plain text by joining tokens with a space character. If we do so, we will get strange looking text as shown below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B Data Preprocessing",
"sec_num": null
},
{
"text": "http : / / www . accesshollywood . com / lindsaylohan -leaves -betty -ford -checks -into -maliburehab article 80744 Lindsay Lohan Leaves Betty Ford , Checks Into Malibu Rehab First Published : June 13 , 2013 4 : 59 PM EDT Lindsay Lohan has left the Betty Ford Center and is moving to a rehab facility in Malibu , Calif . , Access Hollywood has confirmed .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B Data Preprocessing",
"sec_num": null
},
{
"text": "In this example, we can see objects like urls, datetime and punctuation are not in the right format. Since we are using the text as an input to the prompt function, we need to properly format them to align with normal text that GPTs are trained on. Moreover, as gold mention and gold clusters are based on original token ids in ECB+, when we parsed and re-formatted the data, we could match these ids again. Continuing with the previous example, our parsing algorithm cleans up the previous text to be something as follows. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B Data Preprocessing",
"sec_num": null
},
{
"text": "Here are additional results for our experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "C Additional Results",
"sec_num": null
},
{
"text": "Experiments on Prefix The aggregate results from few shot learning are displayed in Table 3 . Our results show that 4-shots learning performs the best for both GPT-2 and GPT-NEO in terms of accuracy. Unexpectedly, as we increase the size of examples, the result does not improve accordingly. Given 10 examples in prefix, the model tend to predict \"yes\" more easily. One possible explanation might be that we have balanced examples in prefix while the actual querying data only have around 8% positive samples.",
"cite_spans": [],
"ref_spans": [
{
"start": 84,
"end": 91,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "C Additional Results",
"sec_num": null
},
{
"text": "F1 AUC 2-shots 0.39 0.08 0.64 0.14 0.50 4-shots 0.51 0.08 0.51 0.14 0.51 10-shots 0.19 0.08 0.90 0.15 0.51 Moreover, we experiment with various datasets for prefix as discussed in section 4. The results in Table 4 shows that prefix does have an impact on the results. The prefix generated from ECB+ dataset performs slightly better than others regarding to AUC. This is understandable because we evaluate on the ECB+ development set. Beyond our expectation, WSC-prefix result in a perfect recall and a super bad accuracy, which means that this prefix lead models to generate \"yes\" regardless of the context. This result further proves that GPT-2 is very sensitive to prompts.",
"cite_spans": [],
"ref_spans": [
{
"start": 206,
"end": 213,
"text": "Table 4",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Acc Prec Recall",
"sec_num": null
},
{
"text": "Our code can be found at https://github.com/ AwesomeCoref/prompt-coref",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF2": {
"ref_id": "b2",
"title": "Streamlining crossdocument coreference resolution: Evaluation and modeling",
"authors": [
{
"first": "Arie",
"middle": [],
"last": "Cattan",
"suffix": ""
},
{
"first": "Alon",
"middle": [],
"last": "Eirew",
"suffix": ""
},
{
"first": "Gabriel",
"middle": [],
"last": "Stanovsky",
"suffix": ""
},
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Ido",
"middle": [],
"last": "Dagan",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2009.11032"
]
},
"num": null,
"urls": [],
"raw_text": "Arie Cattan, Alon Eirew, Gabriel Stanovsky, Mandar Joshi, and Ido Dagan. 2020. Streamlining cross- document coreference resolution: Evaluation and modeling. arXiv preprint arXiv:2009.11032.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Using a sledgehammer to crack a nut? lexical diversity and event coreference resolution",
"authors": [
{
"first": "Agata",
"middle": [],
"last": "Cybulska",
"suffix": ""
},
{
"first": "Piek",
"middle": [],
"last": "Vossen",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14)",
"volume": "",
"issue": "",
"pages": "4545--4552",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Agata Cybulska and Piek Vossen. 2014. Using a sledge- hammer to crack a nut? lexical diversity and event coreference resolution. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 4545-4552, Reyk- javik, Iceland. European Language Resources Asso- ciation (ELRA).",
"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": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1810.04805"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "The pile: An 800gb dataset of diverse text for language modeling",
"authors": [
{
"first": "Leo",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Stella",
"middle": [],
"last": "Biderman",
"suffix": ""
},
{
"first": "Sid",
"middle": [],
"last": "Black",
"suffix": ""
},
{
"first": "Laurence",
"middle": [],
"last": "Golding",
"suffix": ""
},
{
"first": "Travis",
"middle": [],
"last": "Hoppe",
"suffix": ""
},
{
"first": "Charles",
"middle": [],
"last": "Foster",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Phang",
"suffix": ""
},
{
"first": "Horace",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Anish",
"middle": [],
"last": "Thite",
"suffix": ""
},
{
"first": "Noa",
"middle": [],
"last": "Nabeshima",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2101.00027"
]
},
"num": null,
"urls": [],
"raw_text": "Leo Gao, Stella Biderman, Sid Black, Laurence Gold- ing, Travis Hoppe, Charles Foster, Jason Phang, Ho- race He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for lan- guage modeling. arXiv preprint arXiv:2101.00027.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Allennlp: A deep semantic natural language processing platform",
"authors": [
{
"first": "Matt",
"middle": [],
"last": "Gardner",
"suffix": ""
},
{
"first": "Joel",
"middle": [],
"last": "Grus",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Neumann",
"suffix": ""
},
{
"first": "Oyvind",
"middle": [],
"last": "Tafjord",
"suffix": ""
},
{
"first": "Pradeep",
"middle": [],
"last": "Dasigi",
"suffix": ""
},
{
"first": "Nelson",
"middle": [
"F"
],
"last": "Liu",
"suffix": ""
},
{
"first": "Matthew",
"middle": [
"E"
],
"last": "Peters",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Schmitz",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew E. Peters, Michael Schmitz, and Luke Zettlemoyer. 2018. Allennlp: A deep semantic natural language processing platform. CoRR, abs/1803.07640.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Surface form competition: Why the highest probability answer isn't always right",
"authors": [
{
"first": "Ari",
"middle": [],
"last": "Holtzman",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "West",
"suffix": ""
},
{
"first": "Vered",
"middle": [],
"last": "Schwartz",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2104.08315"
]
},
"num": null,
"urls": [],
"raw_text": "Ari Holtzman, Peter West, Vered Schwartz, Yejin Choi, and Luke Zettlemoyer. 2021. Surface form competi- tion: Why the highest probability answer isn't always right. arXiv preprint arXiv:2104.08315.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Span-BERT: Improving Pre-training by Representing and Predicting Spans",
"authors": [
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Danqi",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Yinhan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Daniel",
"middle": [
"S"
],
"last": "Weld",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
}
],
"year": 2020,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "8",
"issue": "",
"pages": "64--77",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. Span- BERT: Improving Pre-training by Representing and Predicting Spans. Transactions of the Association for Computational Linguistics, 8:64-77.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Stanford's multi-pass sieve coreference resolution system at the conll-2011 shared task",
"authors": [
{
"first": "Heeyoung",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Yves",
"middle": [],
"last": "Peirsman",
"suffix": ""
},
{
"first": "Angel",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Nathanael",
"middle": [],
"last": "Chambers",
"suffix": ""
},
{
"first": "Mihai",
"middle": [],
"last": "Surdeanu",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the 15th conference on computational natural language learning: Shared task",
"volume": "",
"issue": "",
"pages": "28--34",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Heeyoung Lee, Yves Peirsman, Angel Chang, Nathanael Chambers, Mihai Surdeanu, and Dan Ju- rafsky. 2011. Stanford's multi-pass sieve coreference resolution system at the conll-2011 shared task. In Proceedings of the 15th conference on computational natural language learning: Shared task, pages 28-34. Association for Computational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "End-to-end neural coreference resolution",
"authors": [
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Luheng",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kenton Lee, Luheng He, Mike Lewis, and Luke Zettle- moyer. 2017. End-to-end neural coreference resolu- tion.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "The winograd schema challenge",
"authors": [
{
"first": "Hector",
"middle": [],
"last": "Levesque",
"suffix": ""
},
{
"first": "Ernest",
"middle": [],
"last": "Davis",
"suffix": ""
},
{
"first": "Leora",
"middle": [],
"last": "Morgenstern",
"suffix": ""
}
],
"year": 2012,
"venue": "Thirteenth International Conference on the Principles of Knowledge Representation and Reasoning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thir- teenth International Conference on the Principles of Knowledge Representation and Reasoning.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing",
"authors": [
{
"first": "Pengfei",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Weizhe",
"middle": [],
"last": "Yuan",
"suffix": ""
},
{
"first": "Jinlan",
"middle": [],
"last": "Fu",
"suffix": ""
},
{
"first": "Zhengbao",
"middle": [],
"last": "Jiang",
"suffix": ""
},
{
"first": "Hiroaki",
"middle": [],
"last": "Hayashi",
"suffix": ""
},
{
"first": "Graham",
"middle": [],
"last": "Neubig",
"suffix": ""
}
],
"year": 2021,
"venue": "ArXiv",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2021. Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ArXiv, abs/2107.13586.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Roberta: A robustly optimized bert pretraining approach",
"authors": [
{
"first": "Yinhan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Myle",
"middle": [],
"last": "Ott",
"suffix": ""
},
{
"first": "Naman",
"middle": [],
"last": "Goyal",
"suffix": ""
},
{
"first": "Jingfei",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Danqi",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Veselin",
"middle": [],
"last": "Stoyanov",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1907.11692"
]
},
"num": null,
"urls": [],
"raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining ap- proach. arXiv preprint arXiv:1907.11692.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "The stanford corenlp natural language processing toolkit",
"authors": [
{
"first": "D",
"middle": [],
"last": "Christopher",
"suffix": ""
},
{
"first": "Mihai",
"middle": [],
"last": "Manning",
"suffix": ""
},
{
"first": "John",
"middle": [],
"last": "Surdeanu",
"suffix": ""
},
{
"first": "Jenny",
"middle": [
"Rose"
],
"last": "Bauer",
"suffix": ""
},
{
"first": "Steven",
"middle": [],
"last": "Finkel",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Bethard",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Mc-Closky",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of 52nd annual meeting of the association for computational linguistics: system demonstrations",
"volume": "",
"issue": "",
"pages": "55--60",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christopher D Manning, Mihai Surdeanu, John Bauer, Jenny Rose Finkel, Steven Bethard, and David Mc- Closky. 2014. The stanford corenlp natural language processing toolkit. In Proceedings of 52nd annual meeting of the association for computational linguis- tics: system demonstrations, pages 55-60.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "True few-shot learning with language models",
"authors": [
{
"first": "Ethan",
"middle": [],
"last": "Perez",
"suffix": ""
},
{
"first": "Douwe",
"middle": [],
"last": "Kiela",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2105.11447"
]
},
"num": null,
"urls": [],
"raw_text": "Ethan Perez, Douwe Kiela, and Kyunghyun Cho. 2021. True few-shot learning with language models. arXiv preprint arXiv:2105.11447.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Language models are unsupervised multitask learners",
"authors": [
{
"first": "Alec",
"middle": [],
"last": "Radford",
"suffix": ""
},
{
"first": "Jeffrey",
"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": "OpenAI blog",
"volume": "1",
"issue": "8",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"authors": [
{
"first": "Colin",
"middle": [],
"last": "Raffel",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Roberts",
"suffix": ""
},
{
"first": "Katherine",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Sharan",
"middle": [],
"last": "Narang",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Matena",
"suffix": ""
},
{
"first": "Yanqi",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Peter J",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1910.10683"
]
},
"num": null,
"urls": [],
"raw_text": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text trans- former. arXiv preprint arXiv:1910.10683.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "2021. Improving and simplifying pattern exploiting training 2021",
"authors": [
{
"first": "D",
"middle": [],
"last": "Tam",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Menon",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Bansal",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Srivastava",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2103.11955"
]
},
"num": null,
"urls": [],
"raw_text": "D Tam, RR Menon, M Bansal, S Srivastava, and C Raf- fel. 2021. Improving and simplifying pattern exploit- ing training 2021. arXiv preprint arXiv:2103.11955.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Few-shot learning through an information retrieval lens",
"authors": [
{
"first": "Eleni",
"middle": [],
"last": "Triantafillou",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Zemel",
"suffix": ""
},
{
"first": "Raquel",
"middle": [],
"last": "Urtasun",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eleni Triantafillou, Richard Zemel, and Raquel Urta- sun. 2017. Few-shot learning through an information retrieval lens.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Multimodal few-shot learning with frozen language models",
"authors": [
{
"first": "Maria",
"middle": [],
"last": "Tsimpoukelli",
"suffix": ""
},
{
"first": "Jacob",
"middle": [],
"last": "Menick",
"suffix": ""
},
{
"first": "Serkan",
"middle": [],
"last": "Cabi",
"suffix": ""
},
{
"first": "Ali",
"middle": [],
"last": "Eslami",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "Felix",
"middle": [],
"last": "Hill",
"suffix": ""
}
],
"year": 2021,
"venue": "Thirty-Fifth Conference on Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, SM Ali Eslami, Oriol Vinyals, and Felix Hill. 2021. Multimodal few-shot learning with frozen language models. In Thirty-Fifth Conference on Neural Infor- mation Processing Systems.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Finetuned language models are zero-shot learners",
"authors": [
{
"first": "Jason",
"middle": [],
"last": "Wei",
"suffix": ""
},
{
"first": "Maarten",
"middle": [],
"last": "Bosma",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Vincent",
"suffix": ""
},
{
"first": "Kelvin",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Adams",
"middle": [
"Wei"
],
"last": "Guu",
"suffix": ""
},
{
"first": "Brian",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "Nan",
"middle": [],
"last": "Lester",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Andrew",
"suffix": ""
},
{
"first": "Quoc V",
"middle": [],
"last": "Dai",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Le",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2109.01652"
]
},
"num": null,
"urls": [],
"raw_text": "Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, An- drew M Dai, and Quoc V Le. 2021. Finetuned lan- guage models are zero-shot learners. arXiv preprint arXiv:2109.01652.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"uris": null,
"type_str": "figure",
"text": "Distribution of predicted values"
},
"FIGREF1": {
"num": null,
"uris": null,
"type_str": "figure",
"text": "Model's precision over various types of noun phrases, including pronouns, proper nouns and nominal nouns. Each bar's hue intensity denotes the data density."
},
"FIGREF2": {
"num": null,
"uris": null,
"type_str": "figure",
"text": "GPT-2's performance on different namedentity types. We use colors to denote performance and the text to show data density in each category."
},
"FIGREF3": {
"num": null,
"uris": null,
"type_str": "figure",
"text": "Different models' F1 score over various level of mention similarities based on BERT embedding."
},
"TABREF0": {
"type_str": "table",
"html": null,
"content": "<table/>",
"num": null,
"text": ""
},
"TABREF2": {
"type_str": "table",
"html": null,
"content": "<table/>",
"num": null,
"text": "Performance of different models."
},
"TABREF3": {
"type_str": "table",
"html": null,
"content": "<table/>",
"num": null,
"text": "http://www.accesshollywood.com/lindsaylohan-leaves-betty-ford-checks-into-maliburehab article 80744 [EOS] Lindsay Lohan Leaves Betty Ford, Checks Into Malibu Rehab First Published: June 13, 2013 4: 59 PM EDT [EOS] Lindsay Lohan has left the Betty Ford Center and is moving to a rehab facility in Malibu, Calif., Access Hollywood has confirmed. [EOS]"
},
"TABREF4": {
"type_str": "table",
"html": null,
"content": "<table><tr><td/><td colspan=\"2\">Acc Prec Recall</td><td>F1 AUC</td></tr><tr><td colspan=\"2\">simple 0.61 0.08</td><td colspan=\"2\">0.36 0.13 0.50</td></tr><tr><td>WSC</td><td>0.08 0.08</td><td colspan=\"2\">1.00 0.15 0.50</td></tr><tr><td>ecb+</td><td>0.54 0.08</td><td colspan=\"2\">0.48 0.14 0.51</td></tr></table>",
"num": null,
"text": ": n-shot performance from the text generative models"
},
"TABREF5": {
"type_str": "table",
"html": null,
"content": "<table/>",
"num": null,
"text": "Average results from each dataset that is used for the experiments"
}
}
}
} |