File size: 87,403 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 1787 1788 1789 1790 1791 1792 1793 1794 1795 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T01:08:45.685277Z"
},
"title": "Training Dynamic based data filtering may not work for NLP datasets",
"authors": [
{
"first": "Arka",
"middle": [],
"last": "Talukdar",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Monika",
"middle": [],
"last": "Dagar",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Prachi",
"middle": [],
"last": "Gupta",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Varun",
"middle": [],
"last": "Menon",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The recent increase in dataset size has brought about significant advances in natural language understanding. These large datasets are usually collected through automation (search engines or web crawlers) or crowdsourcing which inherently introduces incorrectly labeled data. Training on these datasets leads to memorization and poor generalization. Thus, it is pertinent to develop techniques that help in the identification and isolation of mislabelled data. In this paper, we study the applicability of the Area Under the Margin (AUM) metric to identify and remove/rectify mislabelled examples in NLP datasets. We find that mislabelled samples can be filtered using the AUM metric in NLP datasets but it also removes a significant number of correctly labeled points and leads to the loss of a large amount of relevant language information. We show that models rely on the distributional information instead of relying on syntactic and semantic representations.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "The recent increase in dataset size has brought about significant advances in natural language understanding. These large datasets are usually collected through automation (search engines or web crawlers) or crowdsourcing which inherently introduces incorrectly labeled data. Training on these datasets leads to memorization and poor generalization. Thus, it is pertinent to develop techniques that help in the identification and isolation of mislabelled data. In this paper, we study the applicability of the Area Under the Margin (AUM) metric to identify and remove/rectify mislabelled examples in NLP datasets. We find that mislabelled samples can be filtered using the AUM metric in NLP datasets but it also removes a significant number of correctly labeled points and leads to the loss of a large amount of relevant language information. We show that models rely on the distributional information instead of relying on syntactic and semantic representations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Modern deep learning networks are becoming deeper and powerful, and have led to significant advances in Natural Language Processing (NLP) (Devlin et al., 2019) , Computer Vision (He et al., 2015) , and Speech Processing (Graves et al., 2013) . However, these networks rely on large labeled datasets to be effective.",
"cite_spans": [
{
"start": 138,
"end": 159,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF5"
},
{
"start": 178,
"end": 195,
"text": "(He et al., 2015)",
"ref_id": "BIBREF7"
},
{
"start": 220,
"end": 241,
"text": "(Graves et al., 2013)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The creation of large labeled datasets has fueled the advances in NLP (Rajpurkar et al., 2016; Bowman et al., 2015) . Abundant labeled data increases the likelihood of learning diverse phenomena, which in turn leads to models that generalize well (Linzen, 2020) .",
"cite_spans": [
{
"start": 70,
"end": 94,
"text": "(Rajpurkar et al., 2016;",
"ref_id": "BIBREF19"
},
{
"start": 95,
"end": 115,
"text": "Bowman et al., 2015)",
"ref_id": "BIBREF3"
},
{
"start": 247,
"end": 261,
"text": "(Linzen, 2020)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Curating expert annotated datasets is very timeconsuming and costly (Malik and Bhardwaj, 2011) * Equal contribution therefore large language datasets are usually collected through crowd-sourcing, by hiring human annotators (Wang et al., 2019) or by crawling the web. Such methods inherently introduce label noise in the resulting data. Mislabelled training data is particularly problematic for deep neural networks with billions of parameters because they can overfit on the mislabelled data and achieve zero training error even on randomly assigned labels . Training models with noisy labels can also lead to misclassification on easy examples during test-time (Beigman and Beigman Klebanov, 2009) .",
"cite_spans": [
{
"start": 223,
"end": 242,
"text": "(Wang et al., 2019)",
"ref_id": "BIBREF25"
},
{
"start": 662,
"end": 698,
"text": "(Beigman and Beigman Klebanov, 2009)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "It is prohibitively costly to manually remove mislabeled samples from large datasets. Hence, the need arises to create an automated pipeline to analyze and clean datasets. Area Under the Margin (AUM) metric was designed to identify and eliminate noisy data. It can be used as a plug-and-play method within the training pipeline of any classification network with minimal overhead (Pleiss et al., 2020) . AUM shows promising results in identifying mislabelled samples in image classification datasets.",
"cite_spans": [
{
"start": 380,
"end": 401,
"text": "(Pleiss et al., 2020)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Thus, in this paper, we investigate the applicability of the AUM metric on text classification datasets. We make the following contributions: (i) We show that the AUM metric has lower efficacy for filtering mislabelled data in NLP datasets than image datasets. (ii) We hypothesize that AUM does not work as expected in NLP datasets as it did in image datasets because of the intrinsic nature of the data samples. They have high intra-class and low inter-class feature similarity (Ho et al., 2021) , which is usually not the case in NLP datasets. We show samples from NLP datasets to corroborate our hypothesis.",
"cite_spans": [
{
"start": 479,
"end": 496,
"text": "(Ho et al., 2021)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Detecting mislabelled Instances. Pleiss et al. (2020) and Swayamdipta et al. (2020) , both use model behavior on each sample over the training process, also known as training dynamics, to identify mislabelled instances in classification datasets but using different metrics. For each sample, Pleiss et al. (2020) finds the difference in logit value of the assigned class (gold label) and the highest other logit value among the non-assigned classes averaged over training epochs called the area under the margin (AUM) metric. They also introduce a fake class with samples having only mislabelled instances by definition to find a threshold AUM value. Samples with low AUM scores are likely to be mislabelled and the threshold AUM value is used to filter out such mislabelled instances. Swayamdipta et al. (2020) uses mean and standard deviation of the gold label probabilities over the training epochs, called confidence and variability scores, respectively, for each sample. They classify samples with low confidence and low variability as either mislabelled or \"hard-to-learn\" for the model. Bhardwaj et al. (2010) uses statistical methods to find annotators whose annotations differ considerably from the remaining annotators and use manual inspection to decide the verdict for samples annotated by these annotators. M\u00fcller and Markert (2019) classifies training samples with the lowest gold label probabilities on a robust classifier as potentially mislabelled followed by their manual review for final decision. Zhang and Sugiyama (2021) detects samples with erroneous labels using an instance-dependent noise model along with instance-based embedding to capture instance-specific label corruption.",
"cite_spans": [
{
"start": 33,
"end": 53,
"text": "Pleiss et al. (2020)",
"ref_id": "BIBREF18"
},
{
"start": 58,
"end": 83,
"text": "Swayamdipta et al. (2020)",
"ref_id": "BIBREF24"
},
{
"start": 292,
"end": 312,
"text": "Pleiss et al. (2020)",
"ref_id": "BIBREF18"
},
{
"start": 786,
"end": 811,
"text": "Swayamdipta et al. (2020)",
"ref_id": "BIBREF24"
},
{
"start": 1094,
"end": 1116,
"text": "Bhardwaj et al. (2010)",
"ref_id": "BIBREF2"
},
{
"start": 1320,
"end": 1345,
"text": "M\u00fcller and Markert (2019)",
"ref_id": "BIBREF16"
},
{
"start": 1517,
"end": 1542,
"text": "Zhang and Sugiyama (2021)",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Learning in the Presence of Noisy Labels. Several efforts have been made to account for noise in the data and prevent the model from memorizing wrong examples without actually identifying and removing such examples from the training set. Li et al. (2020) replaces the last layer of models trained on noisy data with a linear layer trained on a small set of clean data, Jindal et al. (2019) adds a non-linear noise-modeling layer on top of the target text-classification model. Kang and Hashimoto (2020) Other types of Noise in text data. Depending on the type of supervisory signal and data acquisition method, language datasets can have noise in the form other than labeling errors like spelling errors, grammatical errors (Subramaniam et al., 2009; Malik and Bhardwaj, 2011) . Caswell et al. (2021) provides large-scale systematic quality analysis of various web-crawled multilingual datasets and found large amounts of samples with inconsistencies in language codes and mistranslations. Robust-to-noise word embeddings (Malykh, 2019) , noisy data classifiers trained on clean data, and synthetically generated noisy data (Xu and Koehn, 2017) are some efforts to deal with non-label noise in language data. In this work, however, we only study label noise.",
"cite_spans": [
{
"start": 238,
"end": 254,
"text": "Li et al. (2020)",
"ref_id": "BIBREF11"
},
{
"start": 477,
"end": 502,
"text": "Kang and Hashimoto (2020)",
"ref_id": "BIBREF10"
},
{
"start": 724,
"end": 750,
"text": "(Subramaniam et al., 2009;",
"ref_id": "BIBREF23"
},
{
"start": 751,
"end": 776,
"text": "Malik and Bhardwaj, 2011)",
"ref_id": "BIBREF14"
},
{
"start": 779,
"end": 800,
"text": "Caswell et al. (2021)",
"ref_id": null
},
{
"start": 1022,
"end": 1036,
"text": "(Malykh, 2019)",
"ref_id": "BIBREF15"
},
{
"start": 1124,
"end": 1144,
"text": "(Xu and Koehn, 2017)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Relation to annotator disagreement. The work on dealing with noisy labels in classification datasets can also be related to the work on studying annotator disagreements. Previous work (Beigman and Beigman Klebanov, 2009; Beigman Klebanov et al., 2008; Pavlick and Kwiatkowski, 2019) shows that there can be two reasons for disagreements in annotator labels in crowdsourced datasets: difference of opinion and attention slip. Former generally occurs when different groups of people agree with a different assigned label for a sample based on their understanding of the text. Latter generally occurs due to attention slip or genuine mistake during annotation. As a direction of future work, comparing samples identified as mislabelled using the AUM method with samples that get relatively low agreement among crowd worker annotators can provide meaningful insights.",
"cite_spans": [
{
"start": 193,
"end": 220,
"text": "and Beigman Klebanov, 2009;",
"ref_id": "BIBREF0"
},
{
"start": 221,
"end": 251,
"text": "Beigman Klebanov et al., 2008;",
"ref_id": "BIBREF1"
},
{
"start": 252,
"end": 282,
"text": "Pavlick and Kwiatkowski, 2019)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "3 Implementation Details",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "We use the AUM metric and methodology from Pleiss et al. (2020) to identify training samples with AUM values below a threshold value as mislabelled.",
"cite_spans": [
{
"start": 43,
"end": 63,
"text": "Pleiss et al. (2020)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Filtering data with AUM",
"sec_num": "3.1"
},
{
"text": "To calculate this threshold, original training data is distributed to make a fake class with equal samples from all the original classes. The classification model is trained on this new dataset configuration to generate the AUM values for all the data points. Samples in fake class are by definition mislabelled hence AUM values of fake class samples can be used as a threshold for the samples in the original classes. This method is repeated to find the mislabelled samples among the samples which were in fake class initially. In the second run, a fake class is created such that it does not have any samples which were in the fake class in the first run. As discussed in Section 5, we observed that the heuristic-based thresholding technique suggested in Pleiss et al. (2020) , wherein they used the AUM value of the 99th percentile threshold sample as the threshold to filter the data, does not show major improvement in NLP datasets. We thus consider the AUM threshold value as a hyperparameter and fine-tune it. We also propose a method to rectify the labels and reuse the data for training (discussed in Section 4).",
"cite_spans": [
{
"start": 758,
"end": 778,
"text": "Pleiss et al. (2020)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Filtering data with AUM",
"sec_num": "3.1"
},
{
"text": "We finetuned a distillBERT-base model on SST-2 (Socher et al., 2013) and CoLA (Warstadt et al., 2019) , pretrained using a masked language modeling (MLM) objective (Sanh et al., 2019 ) with a default AdamW optimizer (Loshchilov and Hutter, 2017) . We selected distillBERT for our experimentation because it is small and fast while preserving over 95% of BERT's performance measured on GLUE benchmark (Sanh et al., 2019) . Zhang and Sugiyama (2021) show the mislabelled samples present in CoLA and SST-2 datasets, and we use these results to validate the efficacy of AUM to identify mislabelled data.",
"cite_spans": [
{
"start": 47,
"end": 68,
"text": "(Socher et al., 2013)",
"ref_id": "BIBREF22"
},
{
"start": 78,
"end": 101,
"text": "(Warstadt et al., 2019)",
"ref_id": "BIBREF26"
},
{
"start": 164,
"end": 182,
"text": "(Sanh et al., 2019",
"ref_id": "BIBREF20"
},
{
"start": 216,
"end": 245,
"text": "(Loshchilov and Hutter, 2017)",
"ref_id": "BIBREF13"
},
{
"start": 400,
"end": 419,
"text": "(Sanh et al., 2019)",
"ref_id": "BIBREF20"
},
{
"start": 422,
"end": 447,
"text": "Zhang and Sugiyama (2021)",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "3.2"
},
{
"text": "Following the recommendations from Pleiss et al. (2020), we test the efficacy of AUM on syntheticnoisy and real-world NLP datasets to identify mislabelled samples. To create synthetic-noisy datasets, we injected noise in the real-world datasets by uniformly sampling data points and flipping their labels. We run two experiments on both types of datasets. First, we discard the samples classified by AUM as mislabelled; we will refer to this process as Sieving. Second, instead of discarding samples, we rectify the label and reuse them for training; we will refer to this process as Flipping. Since we train on binary-classification tasks, we flip the label of the samples which are classified as mislabelled.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "As noted in Pleiss et al. (2020) , the AUM threshold for filtration is dataset dependent. The authors provide a simple heuristic for classifying samples as mislabelled; samples with AUM lower than the Table 4 : Cluster of data points in CoLA with high inter-class similarity; Dominant class -class with samples that have high structural and vocab similarity (This similarity is not quantified numerically but was observed during manual inspection of the data); Class 0 is the major class in this example but the high structural and vocab similarity within class 1 reinforces the modeling process to present it as the dominant class 99th percentile threshold sample's AUM will be classified as mislabelled. Further, they also note that the filtration performance is robust to this hyperparameter (percentile threshold). In our experiments with CoLA and SST-2, we observed that this does not translate well to NLP datasets. Indiscriminately removing samples with AUM less than the 99th percentile threshold consistently had poor performance compared to the unfiltered dataset. As the percentile threshold was reduced, the validation accuracy increases as seen in Figure 2 for CoLA and Table 1 for SST-2. It is because that samples with AUM lower than a high percentile threshold could be a hard-to-learn sample which helps generalization (Swayamdipta et al., 2020) . To support this hypothesis, we classified data points using the method (refer to Section 2 for details) suggested by Swayamdipta et al. (2020) . We observed significant overlap in samples identified as mislabelled using AUM and samples identified as hard-to-learn using Swayamdipta et al. (2020) which can be seen in Figure 1 . As we reduce the percentile threshold, we proportionately filter a larger fraction of truly mislabelled data than hard-to-learn samples.",
"cite_spans": [
{
"start": 12,
"end": 32,
"text": "Pleiss et al. (2020)",
"ref_id": "BIBREF18"
},
{
"start": 1336,
"end": 1362,
"text": "(Swayamdipta et al., 2020)",
"ref_id": "BIBREF24"
},
{
"start": 1482,
"end": 1507,
"text": "Swayamdipta et al. (2020)",
"ref_id": "BIBREF24"
},
{
"start": 1635,
"end": 1660,
"text": "Swayamdipta et al. (2020)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [
{
"start": 201,
"end": 208,
"text": "Table 4",
"ref_id": null
},
{
"start": 1161,
"end": 1169,
"text": "Figure 2",
"ref_id": null
},
{
"start": 1183,
"end": 1190,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 1682,
"end": 1690,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Overfitting is another issue that is facilitated by indiscriminately removing samples with low AUM as the proportion of easy-to-learn samples is increased in the filtered dataset (Swayamdipta et al., 2020) . Table 1 shows the results for Sieving on the realworld dataset (SST-2). This experiment also shows how increasing the percentile threshold decreases the increase in performance, hinting at the fact that large amounts of relevant language information might be getting filtered. Table 5 shows the result for sieving and flipping on synthetic mislabelled samples (SST-2). We expected to observe a drastic dip in performance with noise injection and a relatively large gain once filtered, but we only observed a marginal dip after injecting noise and a marginal increase after filtering in performance. For the Flipping experiment, we only saw~1% increase after flipping 68 samples (<0.1 percentile threshold) with the lowest AUM. We considered such a low threshold in an attempt to flip only the truly mislabelled data. Investigating further, we saw that about 60-65% of the noise samples were filtered from our experiments. Figure 3 shows the distribution of AUM values of the synthetic noise and clean samples. The graphs clearly show that AUM does help in identifying the mislabelled samples to some extent (Table 2 and Table 3 show the mislabelled samples we detected in SST-2 and CoLA with low AUM values) but a lot of correctly labeled samples also get filtered depending on how noisy the dataset is. Although there is a high correlation between noise and correctly labeled samples being filtered, the amount of noise alone does not explain this behavior. This leads us to question the efficacy of the AUM metric in NLP datasets. On manual inspection of the CoLA dataset, we found multiple clusters with high feature similarities. Table 4 shows an example of such clusters. We observed that the model is relying on superficial features like word co-occurrence statistics (Sinha et al., 2021) , within these clusters and builds a bias for the dominant class label in a particular cluster. Thus the non-dominant class samples (which usually are correctly labeled) get low AUM values instead of the synthetic noise samples. This does not go hand in hand with our previous observations where Figure 3 shows that synthetic noise samples have low AUM, but it is important to note that synthetic noise samples also happen to be a part of the non-dominant class in most cases (noise in an acceptable dataset is non-dominant). Again, we emphasize correlation does not imply causation.",
"cite_spans": [
{
"start": 179,
"end": 205,
"text": "(Swayamdipta et al., 2020)",
"ref_id": "BIBREF24"
},
{
"start": 1982,
"end": 2002,
"text": "(Sinha et al., 2021)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [
{
"start": 208,
"end": 215,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 485,
"end": 492,
"text": "Table 5",
"ref_id": "TABREF7"
},
{
"start": 1130,
"end": 1138,
"text": "Figure 3",
"ref_id": null
},
{
"start": 1315,
"end": 1335,
"text": "(Table 2 and Table 3",
"ref_id": "TABREF3"
},
{
"start": 1842,
"end": 1849,
"text": "Table 4",
"ref_id": null
},
{
"start": 2299,
"end": 2307,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "In Table 4 , the synthetically introduced noise (marked in red) and members of class 0 (marked in yellow) are both parts of the non-dominant class which gives these samples a negative AUM. While the red labeled samples are legitimate candidates for removal, the removal of yellow samples causes loss of correctly labeled data points. We observed the same pattern through all clusters.",
"cite_spans": [],
"ref_spans": [
{
"start": 3,
"end": 10,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results & Analysis",
"sec_num": "5"
},
{
"text": "We report on the applicability of AUM on NLP datasets. AUM does help in identifying mislabelled samples available to some extent, but sieving these samples indiscriminately removes large amounts of relevant language information. We hypothesize that the reason AUM works well in image datasets is because of the intrinsic nature of the data samples, i.e., data samples in image datasets have high intra-class and less inter-class feature similarity whereas in NLP datasets, data samples have high inter-class feature similarity as seen in Table 4 and this coupled with the model dependency on superficial features results in low AUM values for the non-dominant class samples instead of the mislabelled class samples consequently reducing the efficacy of the AUM metric in NLP datasets.",
"cite_spans": [],
"ref_spans": [
{
"start": 538,
"end": 545,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results & Analysis",
"sec_num": "5"
}
],
"back_matter": [
{
"text": "We would like to thank Prof. Samuel R. Bowman for his guidance and support. We greatly benefited from the discussions we had with him.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": "7"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Learning with annotation noise",
"authors": [
{
"first": "Eyal",
"middle": [],
"last": "Beigman",
"suffix": ""
},
{
"first": "Beata",
"middle": [],
"last": "Beigman Klebanov",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP",
"volume": "",
"issue": "",
"pages": "280--287",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eyal Beigman and Beata Beigman Klebanov. 2009. Learning with annotation noise. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 280-287, Suntec, Singapore. Association for Computational Linguistics.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Analyzing disagreements",
"authors": [
{
"first": "Eyal",
"middle": [],
"last": "Beata Beigman Klebanov",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Beigman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Diermeier",
"suffix": ""
}
],
"year": 2008,
"venue": "Coling 2008: Proceedings of the workshop on Human Judgements in Computational Linguistics",
"volume": "",
"issue": "",
"pages": "2--7",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Beata Beigman Klebanov, Eyal Beigman, and Daniel Diermeier. 2008. Analyzing disagreements. In Col- ing 2008: Proceedings of the workshop on Human Judgements in Computational Linguistics, pages 2- 7, Manchester, UK. Coling 2008 Organizing Com- mittee.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Anveshan: A framework for analysis of multiple annotators' labeling behavior",
"authors": [
{
"first": "Vikas",
"middle": [],
"last": "Bhardwaj",
"suffix": ""
},
{
"first": "Rebecca",
"middle": [],
"last": "Passonneau",
"suffix": ""
},
{
"first": "Ansaf",
"middle": [],
"last": "Salleb-Aouissi",
"suffix": ""
},
{
"first": "Nancy",
"middle": [],
"last": "Ide",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the Fourth Linguistic Annotation Workshop",
"volume": "",
"issue": "",
"pages": "47--55",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vikas Bhardwaj, Rebecca Passonneau, Ansaf Salleb- Aouissi, and Nancy Ide. 2010. Anveshan: A frame- work for analysis of multiple annotators' labeling be- havior. In Proceedings of the Fourth Linguistic An- notation Workshop, pages 47-55, Uppsala, Sweden. Association for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A large annotated corpus for learning natural language inference",
"authors": [
{
"first": "R",
"middle": [],
"last": "Samuel",
"suffix": ""
},
{
"first": "Gabor",
"middle": [],
"last": "Bowman",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Angeli",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Potts",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "632--642",
"other_ids": {
"DOI": [
"10.18653/v1/D15-1075"
]
},
"num": null,
"urls": [],
"raw_text": "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large anno- tated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empiri- cal Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Compu- tational Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"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
},
"BIBREF6": {
"ref_id": "b6",
"title": "Speech recognition with deep recurrent neural networks",
"authors": [
{
"first": "Alex",
"middle": [],
"last": "Graves",
"suffix": ""
},
{
"first": "Mohamed",
"middle": [],
"last": "Abdel-Rahman",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [],
"last": "Hinton",
"suffix": ""
}
],
"year": 2013,
"venue": "2013 IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "",
"issue": "",
"pages": "6645--6649",
"other_ids": {
"DOI": [
"10.1109/icassp.2013.6638947"
]
},
"num": null,
"urls": [],
"raw_text": "Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. 2013. Speech recognition with deep recur- rent neural networks. In 2013 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing, pages 6645-6649. IEEE.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Deep residual learning for image recognition",
"authors": [
{
"first": "Kaiming",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Xiangyu",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Shaoqing",
"middle": [],
"last": "Ren",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Sun",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep residual learning for image recog- nition.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Learning embeddings for image clustering: An empirical study of triplet loss approaches",
"authors": [
{
"first": "Kalun",
"middle": [],
"last": "Ho",
"suffix": ""
},
{
"first": "Janis",
"middle": [],
"last": "Keuper",
"suffix": ""
},
{
"first": "Franz-Josef",
"middle": [],
"last": "Pfreundt",
"suffix": ""
},
{
"first": "Margret",
"middle": [],
"last": "Keuper",
"suffix": ""
}
],
"year": 2021,
"venue": "2020 25th International Conference on Pattern Recognition (ICPR)",
"volume": "",
"issue": "",
"pages": "87--94",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kalun Ho, Janis Keuper, Franz-Josef Pfreundt, and Margret Keuper. 2021. Learning embeddings for im- age clustering: An empirical study of triplet loss ap- proaches. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 87-94. IEEE.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "An effective label noise model for DNN text classification",
"authors": [
{
"first": "Ishan",
"middle": [],
"last": "Jindal",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Pressel",
"suffix": ""
},
{
"first": "Brian",
"middle": [],
"last": "Lester",
"suffix": ""
},
{
"first": "Matthew",
"middle": [],
"last": "Nokleby",
"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": "3246--3256",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1328"
]
},
"num": null,
"urls": [],
"raw_text": "Ishan Jindal, Daniel Pressel, Brian Lester, and Matthew Nokleby. 2019. An effective label noise model for DNN text classification. 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 3246-3256, Minneapolis, Minnesota. Association for Computational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Improved natural language generation via loss truncation",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Kang",
"suffix": ""
},
{
"first": "Tatsunori",
"middle": [
"B"
],
"last": "Hashimoto",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "718--731",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.66"
]
},
"num": null,
"urls": [],
"raw_text": "Daniel Kang and Tatsunori B. Hashimoto. 2020. Im- proved natural language generation via loss trunca- tion. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 718-731, Online. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Noisy labels can induce good representations",
"authors": [
{
"first": "Jingling",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Mozhi",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Keyulu",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "John",
"middle": [
"P"
],
"last": "Dickerson",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Ba",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2012.12896"
]
},
"num": null,
"urls": [],
"raw_text": "Jingling Li, Mozhi Zhang, Keyulu Xu, John P. Dickerson, and Jimmy Ba. 2020. Noisy labels can induce good representations. ArXiv preprint arXiv:2012.12896.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "How can we accelerate progress towards human-like linguistic generalization?",
"authors": [
{
"first": "",
"middle": [],
"last": "Tal Linzen",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "5210--5217",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.465"
]
},
"num": null,
"urls": [],
"raw_text": "Tal Linzen. 2020. How can we accelerate progress to- wards human-like linguistic generalization? In Pro- ceedings of the 58th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 5210- 5217, Online. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Sgdr: Stochastic gradient descent with warm restarts",
"authors": [
{
"first": "Ilya",
"middle": [],
"last": "Loshchilov",
"suffix": ""
},
{
"first": "Frank",
"middle": [],
"last": "Hutter",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ilya Loshchilov and Frank Hutter. 2017. Sgdr: Stochas- tic gradient descent with warm restarts.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Automatic training data cleaning for text classification",
"authors": [
{
"first": "H",
"middle": [],
"last": "Hassan",
"suffix": ""
},
{
"first": "Vikas",
"middle": [
"S"
],
"last": "Malik",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Bhardwaj",
"suffix": ""
}
],
"year": 2011,
"venue": "2011 IEEE 11th International Conference on Data Mining Workshops",
"volume": "",
"issue": "",
"pages": "442--449",
"other_ids": {
"DOI": [
"10.1109/ICDMW.2011.36"
]
},
"num": null,
"urls": [],
"raw_text": "Hassan H. Malik and Vikas S. Bhardwaj. 2011. Au- tomatic training data cleaning for text classifica- tion. In 2011 IEEE 11th International Conference on Data Mining Workshops, pages 442-449.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Robust to noise models in natural language processing tasks",
"authors": [
{
"first": "Valentin",
"middle": [],
"last": "Malykh",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop",
"volume": "",
"issue": "",
"pages": "10--16",
"other_ids": {
"DOI": [
"10.18653/v1/P19-2002"
]
},
"num": null,
"urls": [],
"raw_text": "Valentin Malykh. 2019. Robust to noise models in nat- ural language processing tasks. In Proceedings of the 57th Annual Meeting of the Association for Com- putational Linguistics: Student Research Workshop, pages 10-16, Florence, Italy. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Identifying mislabeled instances in classification datasets",
"authors": [
{
"first": "Nicolas",
"middle": [
"M"
],
"last": "M\u00fcller",
"suffix": ""
},
{
"first": "Karla",
"middle": [],
"last": "Markert",
"suffix": ""
}
],
"year": 2019,
"venue": "2019 International Joint Conference on Neural Networks (IJCNN)",
"volume": "",
"issue": "",
"pages": "1--8",
"other_ids": {
"DOI": [
"10.1109/IJCNN.2019.8851920"
]
},
"num": null,
"urls": [],
"raw_text": "Nicolas M. M\u00fcller and Karla Markert. 2019. Identi- fying mislabeled instances in classification datasets. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Inherent disagreements in human textual inferences",
"authors": [
{
"first": "Ellie",
"middle": [],
"last": "Pavlick",
"suffix": ""
},
{
"first": "Tom",
"middle": [],
"last": "Kwiatkowski",
"suffix": ""
}
],
"year": 2019,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "7",
"issue": "",
"pages": "677--694",
"other_ids": {
"DOI": [
"10.1162/tacl_a_00293"
]
},
"num": null,
"urls": [],
"raw_text": "Ellie Pavlick and Tom Kwiatkowski. 2019. Inherent disagreements in human textual inferences. Transac- tions of the Association for Computational Linguis- tics, 7:677-694.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Identifying mislabeled data using the area under the margin ranking",
"authors": [
{
"first": "Geoff",
"middle": [],
"last": "Pleiss",
"suffix": ""
},
{
"first": "Tianyi",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Ethan",
"middle": [
"R"
],
"last": "Elenberg",
"suffix": ""
},
{
"first": "Kilian",
"middle": [
"Q"
],
"last": "Weinberger",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geoff Pleiss, Tianyi Zhang, Ethan R. Elenberg, and Kilian Q. Weinberger. 2020. Identifying mislabeled data using the area under the margin ranking.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "SQuAD: 100,000+ questions for machine comprehension of text",
"authors": [
{
"first": "Pranav",
"middle": [],
"last": "Rajpurkar",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Konstantin",
"middle": [],
"last": "Lopyrev",
"suffix": ""
},
{
"first": "Percy",
"middle": [],
"last": "Liang",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "2383--2392",
"other_ids": {
"DOI": [
"10.18653/v1/D16-1264"
]
},
"num": null,
"urls": [],
"raw_text": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natu- ral Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter",
"authors": [
{
"first": "Victor",
"middle": [],
"last": "Sanh",
"suffix": ""
},
{
"first": "Lysandre",
"middle": [],
"last": "Debut",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Chaumond",
"suffix": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Wolf",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1910.01108"
]
},
"num": null,
"urls": [],
"raw_text": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Masked language modeling and the distributional hypothesis: Order word matters pre-training for little",
"authors": [
{
"first": "Koustuv",
"middle": [],
"last": "Sinha",
"suffix": ""
},
{
"first": "Robin",
"middle": [],
"last": "Jia",
"suffix": ""
},
{
"first": "Dieuwke",
"middle": [],
"last": "Hupkes",
"suffix": ""
},
{
"first": "Joelle",
"middle": [],
"last": "Pineau",
"suffix": ""
},
{
"first": "Adina",
"middle": [],
"last": "Williams",
"suffix": ""
},
{
"first": "Douwe",
"middle": [],
"last": "Kiela",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2104.06644"
]
},
"num": null,
"urls": [],
"raw_text": "Koustuv Sinha, Robin Jia, Dieuwke Hupkes, Joelle Pineau, Adina Williams, and Douwe Kiela. 2021. Masked language modeling and the distributional hypothesis: Order word matters pre-training for lit- tle. arXiv preprint arXiv:2104.06644.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Recursive deep models for semantic compositionality over a sentiment treebank",
"authors": [
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Perelygin",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Jean",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Chuang",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Christopher",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Manning",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Andrew",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Ng",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Potts",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the conference on empirical methods in natural language processing",
"volume": "1631",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment tree- bank. In Proceedings of the conference on empirical methods in natural language processing (EMNLP), volume 1631, page 1642. Citeseer.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "A survey of types of text noise and techniques to handle noisy text",
"authors": [
{
"first": "L",
"middle": [
"Venkata"
],
"last": "Subramaniam",
"suffix": ""
},
{
"first": "Shourya",
"middle": [],
"last": "Roy",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Tanveer",
"suffix": ""
},
{
"first": "Sumit",
"middle": [],
"last": "Faruquie",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Negi",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of The Third Workshop on Analytics for Noisy Unstructured Text Data, AND '09",
"volume": "",
"issue": "",
"pages": "115--122",
"other_ids": {
"DOI": [
"10.1145/1568296.1568315"
]
},
"num": null,
"urls": [],
"raw_text": "L. Venkata Subramaniam, Shourya Roy, Tanveer A. Faruquie, and Sumit Negi. 2009. A survey of types of text noise and techniques to handle noisy text. In Proceedings of The Third Workshop on Analyt- ics for Noisy Unstructured Text Data, AND '09, page 115-122, New York, NY, USA. Association for Computing Machinery.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Dataset cartography: Mapping and diagnosing datasets with training dynamics",
"authors": [
{
"first": "Swabha",
"middle": [],
"last": "Swayamdipta",
"suffix": ""
},
{
"first": "Roy",
"middle": [],
"last": "Schwartz",
"suffix": ""
},
{
"first": "Nicholas",
"middle": [],
"last": "Lourie",
"suffix": ""
},
{
"first": "Yizhong",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Hannaneh",
"middle": [],
"last": "Hajishirzi",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "9275--9293",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.746"
]
},
"num": null,
"urls": [],
"raw_text": "Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A. Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dy- namics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 9275-9293, Online. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Glue: A multi-task benchmark and analysis platform for natural language understanding",
"authors": [
{
"first": "Alex",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Amanpreet",
"middle": [],
"last": "Singh",
"suffix": ""
},
{
"first": "Julian",
"middle": [],
"last": "Michael",
"suffix": ""
},
{
"first": "Felix",
"middle": [],
"last": "Hill",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Samuel",
"middle": [
"R"
],
"last": "Bowman",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. Glue: A multi-task benchmark and analysis platform for natural language understanding.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Neural network acceptability judgments",
"authors": [
{
"first": "Alex",
"middle": [],
"last": "Warstadt",
"suffix": ""
},
{
"first": "Amanpreet",
"middle": [],
"last": "Singh",
"suffix": ""
},
{
"first": "Samuel",
"middle": [
"R"
],
"last": "",
"suffix": ""
}
],
"year": 2019,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "7",
"issue": "",
"pages": "625--641",
"other_ids": {
"DOI": [
"10.1162/tacl_a_00290"
]
},
"num": null,
"urls": [],
"raw_text": "Alex Warstadt, Amanpreet Singh, and Samuel R. Bow- man. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625-641.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Zipporah: a fast and scalable data cleaning system for noisy webcrawled parallel corpora",
"authors": [
{
"first": "Hainan",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "2945--2950",
"other_ids": {
"DOI": [
"10.18653/v1/D17-1319"
]
},
"num": null,
"urls": [],
"raw_text": "Hainan Xu and Philipp Koehn. 2017. Zipporah: a fast and scalable data cleaning system for noisy web- crawled parallel corpora. In Proceedings of the 2017 Conference on Empirical Methods in Natural Lan- guage Processing, pages 2945-2950, Copenhagen, Denmark. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Understanding deep learning requires rethinking generalization",
"authors": [
{
"first": "Chiyuan",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Samy",
"middle": [],
"last": "Bengio",
"suffix": ""
},
{
"first": "Moritz",
"middle": [],
"last": "Hardt",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chiyuan Zhang, Samy Bengio, Moritz Hardt, Ben- jamin Recht, and Oriol Vinyals. 2016. Understand- ing deep learning requires rethinking generalization. CoRR, abs/1611.03530.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Approximating instance-dependent noise via instanceconfidence embedding",
"authors": [
{
"first": "Yivan",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Masashi",
"middle": [],
"last": "Sugiyama",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yivan Zhang and Masashi Sugiyama. 2021. Ap- proximating instance-dependent noise via instance- confidence embedding.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Visualizing AUM values of SST-2 samples along with their Data Map plotted as perSwayamdipta et al. (2020) shows considerable overlap in samples with low AUM values and samples identified as hard-tolearn/mislabelled as per Data map (sampled with low confidence and low variability scores).",
"type_str": "figure",
"uris": null,
"num": null
},
"TABREF0": {
"type_str": "table",
"content": "<table><tr><td>Percentile</td></tr></table>",
"text": "improves faithfulness in text generation by adaptively removing high log loss examples during the training process. Acc. on Unfiltered Data Acc. on Filtered Data 1 87.91 \u00b1 0.25 88.02 \u00b1 0.79 10 87.79 \u00b1 0.50 87.79 \u00b1 0.61 50 87.87 \u00b1 0.36 87.72 \u00b1 0.48 90 87.79 \u00b1 0.49 87.72 \u00b1 0.48",
"html": null,
"num": null
},
"TABREF1": {
"type_str": "table",
"content": "<table/>",
"text": "Accuracy on SST-2 dataset at different AUM threshold percentiles; Sieving",
"html": null,
"num": null
},
"TABREF3": {
"type_str": "table",
"content": "<table><tr><td>: Filtered examples from CoLA dataset (1 =</td></tr><tr><td>grammatically acceptable; 0 = grammatically unaccept-</td></tr><tr><td>able)</td></tr><tr><td>Figure 2: CoLA: Percentile threshold vs Validation ac-</td></tr><tr><td>curacy</td></tr></table>",
"text": "",
"html": null,
"num": null
},
"TABREF4": {
"type_str": "table",
"content": "<table><tr><td>Sample Id</td><td>Text</td><td>Label</td></tr><tr><td>41767</td><td>a damn fine and a truly distinctive and a deeply pertinent film</td><td>Negative</td></tr><tr><td>42407</td><td>guts and</td><td>Negative</td></tr><tr><td>62237</td><td colspan=\"2\">as original and Positive</td></tr><tr><td>19159</td><td>he script is n't up to the level of the direction</td><td>Positive</td></tr><tr><td>62178</td><td>though it runs 163 minutes , safe conduct is anything but languorous .</td><td>Positive</td></tr></table>",
"text": "insightful as last week 's episode of behind the music . Negative 6886 loves the members of the upper class almost as much as they love themselves Negative 19153 of how horrible we are to ourselves and each other",
"html": null,
"num": null
},
"TABREF5": {
"type_str": "table",
"content": "<table><tr><td>Sample Id</td><td>Text</td><td>Label</td><td>AUM</td></tr><tr><td>1432</td><td>I disliked the boy's playing the piano loudly.</td><td>0</td><td>-0.501698</td></tr><tr><td>1433</td><td>The boy whose loud playing of the piano I disliked was a student.</td><td>1</td><td>0.163168</td></tr><tr><td>1434</td><td>The piano which I disliked the boy's playing loudly was badly out of tune.</td><td>0</td><td>-0.349795</td></tr><tr><td>1435</td><td>The boy's loud playing of the piano drove everyone crazy.</td><td>1</td><td>0.99208</td></tr><tr><td>1436</td><td>The boy's playing the piano loudly drove everyone crazy.</td><td>1</td><td>0.521766</td></tr><tr><td>1437</td><td>That piano, the boy's loud playing of which drove everyone crazy, was badly out of tune.</td><td>1</td><td>0.49058</td></tr><tr><td>1438</td><td>That piano, the boy's playing which loudly drove everyone crazy, was badly out of tune.</td><td>0</td><td>-0.333451</td></tr><tr><td>1439</td><td>That piano, which the boy's playing loudly drove everyone crazy, was badly out of tune.</td><td>0</td><td>-0.628104</td></tr><tr><td>1440</td><td>Did that he played the piano surprise you?</td><td>0</td><td>-0.290151</td></tr><tr><td>1441</td><td>Would for him to have played the piano have surprised you?</td><td>0</td><td>-0.430419</td></tr><tr><td>1442</td><td>Is whether he played the piano known?</td><td>0</td><td>-0.292625</td></tr><tr><td>1443</td><td>Did his having played the piano surprise you?</td><td>1</td><td>0.332026</td></tr></table>",
"text": "Filtered examples from Stanford Sentiment Treebank (SST2)",
"html": null,
"num": null
},
"TABREF7": {
"type_str": "table",
"content": "<table><tr><td>Synthetic noise)</td></tr></table>",
"text": "Accuracy on Synthetic mislabelled Samples (SST-2); Seed: 100; Threshold Percentile: 90 Figure 3: Histogram of AUM values of synthetic noise and unaltered data. (Blue \u2192 Unaltered data, Orange \u2192",
"html": null,
"num": null
}
}
}
} |