File size: 94,606 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 | {
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:27:14.559796Z"
},
"title": "\"I'll be there for you\": The One with Understanding Indirect Answers",
"authors": [
{
"first": "Cathrine",
"middle": [],
"last": "Damgaard",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Copenhagen",
"location": {}
},
"email": ""
},
{
"first": "Paulina",
"middle": [],
"last": "Toborek",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Copenhagen",
"location": {}
},
"email": ""
},
{
"first": "Trine",
"middle": [],
"last": "Eriksen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Copenhagen",
"location": {}
},
"email": ""
},
{
"first": "Barbara",
"middle": [],
"last": "Plank",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Copenhagen",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Indirect answers are replies to polar questions without the direct use of word cues such as 'yes' and 'no'. Humans are very good at understanding indirect answers, such as 'I gotta go home sometime', when asked 'You wanna crash on the couch?'. Understanding indirect answers is a challenging problem for dialogue systems. In this paper, we introduce a new English corpus to study the problem of understanding indirect answers. Instead of crowdsourcing both polar questions and answers, we collect questions and indirect answers from transcripts of a prominent TV series and manually annotate them for answer type. The resulting dataset contains 5,930 question-answer pairs. We release both aggregated and raw human annotations. We present a set of experiments in which we evaluate Convolutional Neural Networks (CNNs) for this task, including a cross-dataset evaluation and experiments with learning from disagreements in annotation. Our results show that the task of interpreting indirect answers remains challenging, yet we obtain encouraging improvements when explicitly modeling human disagreement.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Indirect answers are replies to polar questions without the direct use of word cues such as 'yes' and 'no'. Humans are very good at understanding indirect answers, such as 'I gotta go home sometime', when asked 'You wanna crash on the couch?'. Understanding indirect answers is a challenging problem for dialogue systems. In this paper, we introduce a new English corpus to study the problem of understanding indirect answers. Instead of crowdsourcing both polar questions and answers, we collect questions and indirect answers from transcripts of a prominent TV series and manually annotate them for answer type. The resulting dataset contains 5,930 question-answer pairs. We release both aggregated and raw human annotations. We present a set of experiments in which we evaluate Convolutional Neural Networks (CNNs) for this task, including a cross-dataset evaluation and experiments with learning from disagreements in annotation. Our results show that the task of interpreting indirect answers remains challenging, yet we obtain encouraging improvements when explicitly modeling human disagreement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Humans are very good at interpreting indirect answers to polar questions. In conversations, even if direct answers are possible, humans often prefer indirect answers due to cooperativeness and to advance the dialogue (Stenstr\u00f6m, 1984) . For dialogue systems and Natural Language Processing (NLP) more generally, however, interpreting indirect answers remains a challenge (Clark et al., 2019) . Recent seminal work introduced CIRCA, a new large-scale dataset containing pairs of polar questions and indirect answers in English (Louis et al., 2020) . This allows for data-driven experiments in this question-answering domain. Table 1 : Examples from the dataset with polar question (Q), indirect answer (A) and annotator labels (L).",
"cite_spans": [
{
"start": 217,
"end": 234,
"text": "(Stenstr\u00f6m, 1984)",
"ref_id": "BIBREF17"
},
{
"start": 371,
"end": 391,
"text": "(Clark et al., 2019)",
"ref_id": "BIBREF2"
},
{
"start": 526,
"end": 546,
"text": "(Louis et al., 2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 624,
"end": 631,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Understanding indirect answers is a pragmatic problem, and even though humans typically have little difficulty in interpreting indirect answers, they may not all agree on a possible interpretation. Work on learning from human disagreement has shown that incorporating disagreement from human annotation is not only noise, but can provide valuable information (Plank et al., 2014; Aroyo and Welty, 2015; Rodrigues and Pereira, 2018) .",
"cite_spans": [
{
"start": 359,
"end": 379,
"text": "(Plank et al., 2014;",
"ref_id": "BIBREF14"
},
{
"start": 380,
"end": 402,
"text": "Aroyo and Welty, 2015;",
"ref_id": "BIBREF0"
},
{
"start": 403,
"end": 431,
"text": "Rodrigues and Pereira, 2018)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Motivated by these two lines of research, we propose a new dataset for studying indirect answers. We provide both aggregated (ground truth/gold) annotations and the raw annotations. This allows us to study the effect of learning to integrate human disagreement into understanding indirect answers. Our dataset, called FRIENDS-QIA, was created by collecting question and answer pairs from transcripts of a popular TV series. The data collection differs in comparison to the recently introduced CIRCA corpus. In their study, a set of 10 dialogue prompts were defined, and both questions and answers were collected by crowdsourcing (Louis et al., 2020) . Following this step, the annotation was crowdsourced again, resulting in a set of labels which was later conflated into a relaxed set of six classes. Instead, we opted to collect the data from transcripts and manually annotate the question-answer pairs in house (by three of the authors of this paper, all highly proficient in English). Examples from our dataset are provided in Table 1 . A data statement is provided in Appendix A.",
"cite_spans": [
{
"start": 629,
"end": 649,
"text": "(Louis et al., 2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 1031,
"end": 1038,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Contributions In this paper, we a) introduce a new dataset, FRIENDS-QIA, with 5,930 polar question-indirect answer pairs in English; 1 b) study the effectiveness of neural classifiers based on Convolutional Neural Networks, both with traditional pre-trained word embeddings and contextualized BERT embeddings; c) provide results on crossdataset evaluation, for which we train a model on both CIRCA and FRIENDS-QIA; and d) show that modeling human disagreement via deep learning from crowds is beneficial for this task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The task of understanding indirect answers in Natural Language Processing is relatively new and has not been attempted by many yet. To enable progress on the task, Louis et al. (2020) created and released the first large-scale English language corpus, CIRCA, consisting of 34,268 polar questions and their corresponding indirect answers. The difficulty of interpreting indirect answers is, however, widely studied in some early papers, (e.g. Carberry, 1992, 1999) . Following this work, de Marneffe et al. (2009) propose a logical inference model with probabilistic methods. They realize the influence of discourse conditions as well as the difference between the literal meaning of the answer and the interpretation by the two speakers. Hockey et al. (1997) further underline the complexity of interpreting indirect answers to polar questions. They explore the existing Edinburgh map task corpus (Thompson et al., 1993) which consists of two-person dialogues already coded for dialogue structure. Clark et al. (2019) recently also explore the understanding of indirect answers and report on the difficulty of the task. They create a new dataset, BOOLQ, by combining search queries from the Google search engine as questions and passages on Wikipedia pages as answers. They attempt to classify such indirect answers by training BERT-based neural models. Louis et al. (2020) further experiment with training BERT models from scratch and by transfer learning from BOOLQ, building on top of Clark et al. (2019) . Their newly created dataset, CIRCA, includes 10 question prompt types, spanning a wide variety of communication situations. Their study shows promising results, and inspired our work.",
"cite_spans": [
{
"start": 164,
"end": 183,
"text": "Louis et al. (2020)",
"ref_id": "BIBREF10"
},
{
"start": 442,
"end": 463,
"text": "Carberry, 1992, 1999)",
"ref_id": null
},
{
"start": 490,
"end": 512,
"text": "Marneffe et al. (2009)",
"ref_id": "BIBREF3"
},
{
"start": 738,
"end": 758,
"text": "Hockey et al. (1997)",
"ref_id": "BIBREF7"
},
{
"start": 897,
"end": 920,
"text": "(Thompson et al., 1993)",
"ref_id": "BIBREF18"
},
{
"start": 998,
"end": 1017,
"text": "Clark et al. (2019)",
"ref_id": "BIBREF2"
},
{
"start": 1354,
"end": 1373,
"text": "Louis et al. (2020)",
"ref_id": "BIBREF10"
},
{
"start": 1488,
"end": 1507,
"text": "Clark et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Preparing a dataset for classification tasks often requires collecting labels from multiple annotators. A unanimous gold standard label cannot always be clearly achieved (e.g. Aroyo and Welty, 2015; Rodrigues et al., 2013; Palomaki et al., 2018; Pavlick and Kwiatkowski, 2019) . The problem of learning from multiple annotators has become more important and several attempts have been made to deal with biases present in such data. Yan et al. (2014) look into different levels of expertise among annotators and how a model can learn from this, taking into consideration the biases present while labeling the dataset. By measuring the inter-annotator agreement and incorporating the annotator uncertainty in model training, Plank et al. (2014) show that modelling annotator disagreement can be useful even in cases of seemingly more objective annotation tasks, like part-of-speech tagging. An interesting and similar approach was suggested by Rodrigues and Pereira (2018) . Their method, deep learning from crowds (DLFC), adds a crowd layer on top of a neural network which takes advantage of the reliability and bias from different annotators. It assumes access to the raw annotations from the training data. This is in contrast to the weighting approach proposed by Plank et al. (2014) , who integrate aggregated disagreement from a sample. As we have the full data available with multiple annotations in FRIENDS-QIA, we here experiment with the deep learning from crowds (DLFC) approach proposed by Rodrigues and Pereira (2018) . The key idea is to train a model directly from the noisy annotator-specific annotations modeled as additional per-annotator auxiliary tasks on top of the gold distribution. By using backpropagation and creating annotator-specific weights in the crowd layer, it allows us to include the label bias in the model's learning. There is emerging interest in learning from disagreement, and we refer the reader to a recent survey on this topic (Uma et al., 2021) .",
"cite_spans": [
{
"start": 176,
"end": 198,
"text": "Aroyo and Welty, 2015;",
"ref_id": "BIBREF0"
},
{
"start": 199,
"end": 222,
"text": "Rodrigues et al., 2013;",
"ref_id": "BIBREF15"
},
{
"start": 223,
"end": 245,
"text": "Palomaki et al., 2018;",
"ref_id": "BIBREF11"
},
{
"start": 246,
"end": 276,
"text": "Pavlick and Kwiatkowski, 2019)",
"ref_id": "BIBREF12"
},
{
"start": 432,
"end": 449,
"text": "Yan et al. (2014)",
"ref_id": "BIBREF20"
},
{
"start": 723,
"end": 742,
"text": "Plank et al. (2014)",
"ref_id": "BIBREF14"
},
{
"start": 942,
"end": 970,
"text": "Rodrigues and Pereira (2018)",
"ref_id": "BIBREF16"
},
{
"start": 1267,
"end": 1286,
"text": "Plank et al. (2014)",
"ref_id": "BIBREF14"
},
{
"start": 1501,
"end": 1529,
"text": "Rodrigues and Pereira (2018)",
"ref_id": "BIBREF16"
},
{
"start": 1969,
"end": 1987,
"text": "(Uma et al., 2021)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "We introduce a new corpus called FRIENDS-QIA for studying indirect answers. The three steps involved in corpus construction are: data collection, data preprocessing and data annotation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "FRIENDS-QIA Corpus",
"sec_num": "3"
},
{
"text": "Instead of crowdsourcing the data collection as done by Louis et al. (2020) , we attempt to exploit already existing data by simply looking for useful question-answer (QA) pairs in dialogues. We use the existing transcriptions of the American televi- Monica: I need more swordfish. Can you get me some more swordfish? Kitchen Worker: I don't speak English.",
"cite_spans": [
{
"start": 56,
"end": 75,
"text": "Louis et al. (2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "FRIENDS-QIA Corpus",
"sec_num": "3"
},
{
"text": "Joey: You know more than one Fun Bobby? Chandler: I happen to know a Fun Bob. Table 2 : Examples of QA pairs from FRIENDS-QIA for each of the 6 labels. sion sitcom \"Friends\". 2 This data provides rich dialogues from conversations covering regular, everyday topics. FRIENDS-QIA includes 10 seasons consisting of 17-25 episodes each. This sums to a total of 228 episodes of the TV series. The full dataset, including metadata, is described in Appendix C.",
"cite_spans": [],
"ref_spans": [
{
"start": 78,
"end": 85,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "FRIENDS-QIA Corpus",
"sec_num": "3"
},
{
"text": "To collect the QA pairs we manually scan all episodes of \"Friends\" for polar questions and retrieve both the question and the answer as well as which character said each of the two. A few examples are listed below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Collection",
"sec_num": "3.1"
},
{
"text": "Example 3.1 Joey: Hey Pheebs, you wanna help? Phoebe: Oh, I wish I could, but I don't want to.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Collection",
"sec_num": "3.1"
},
{
"text": "Example 3.2 Rachel: Hey! So, did you quit? Chandler: No, I almost did, couldn't leave Ross there without a spotter! Example 3.3 Joey: (intrigued) Really? Mr. Treeger: Yeah, you could dance real good with her, she's the same size as me.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Collection",
"sec_num": "3.1"
},
{
"text": "As the examples show, some answers directly give away cues (e.g., Example 3.2 starts with a \"No,\"). We preprocess the data manually, to obfuscate direct cues and enrich short questions with information from the context, to keep the setup similar to prior work and model single questionanswer pairs. The data preprocessing motivation and details are outlined next.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Collection",
"sec_num": "3.1"
},
{
"text": "To prepare the QA pairs in a useful format for the task at hand, we assign each collected pair to one of four categories. The categories are included in the final FRIENDS-QIA dataset and describe if, and in what way, the question or answer had to be modified. Modifications include removing direct cues to the answers 'yes' or 'no', such as 'yeah' or 'nope' (a full list is provided in Appendix B), adding context to the question in case we deem it insufficient, as well as removing metadata and irrelevant parts of the question or answer. The four categories are listed below along with their respective description.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Preprocessing",
"sec_num": "3.2"
},
{
"text": "1. Able to use question and answer, exactly as they are 2. Have to only remove yes/no/nope/yeah (etc.) in the answer 3. Clarifying/adding context to the question (or rarely the answer)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Preprocessing",
"sec_num": "3.2"
},
{
"text": "Questions with an answer only containing a yes/no/nope/yeah (etc.)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "Example 3.1 is a QA pair which is useful for our task exactly as it is. It has a polar question and a perfectly indirect answer with no direct mappers or cues, so we assign this QA pair to category 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "However, in Example 3.2 we remove a bit of irrelevant context, namely the \"Hey!\" in the question. Furthermore, the answer contains an actual \"No\", which should not be present in an indirect answer, so this is also removed. These modifications cause the QA pair to be assigned to category 2. The modified example is shown below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "Example 3.4 Rachel: So, did you quit?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "Chandler: I almost did, couldn't leave Ross there without a spotter! Example 3.3 is from category 3, where we need to do a larger modification to make the QA pair useful. This modification could for example be to add context, which is usually found in the transcript lines immediately above the question. The modified example is shown below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "Example 3.5 Joey: Really? Marge has a girlfriend? Mr. Treeger: You could dance real good with her, she's the same size as me.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "Category 4 contains QA pairs where the answer consists of only a direct mapper or cue alone, so it cannot easily be modified into an indirect answer. Therefore, this category is ultimately excluded from the final FRIENDS-QIA dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "We adopt the RELAXED label set as introduced by Louis et The label distribution of FRIENDS-QIA is shown in Figure 1 .",
"cite_spans": [
{
"start": 48,
"end": 56,
"text": "Louis et",
"ref_id": null
}
],
"ref_spans": [
{
"start": 107,
"end": 115,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data Annotation",
"sec_num": "3.3"
},
{
"text": "We measure the agreement and thereby the reliability of the annotations in FRIENDS-QIA. The three annotators independently labeled all QA pairs in the dataset. The raw agreement distribution of the annotations is listed in Table 3 , which includes the agreement for each of the final three categories We use majority voting to aggregate the gold standard from the three annotations. In few cases (1.59%), all three annotators disagreed on the label, and we map these to N/A. Examples from FRIENDS-QIA are provided in Table 2 , where the QA pairs with label 1-5 were given the same label by at least two annotators, whereas the example with label 6 is a case of full annotator disagreement.",
"cite_spans": [],
"ref_spans": [
{
"start": 223,
"end": 230,
"text": "Table 3",
"ref_id": "TABREF3"
},
{
"start": 517,
"end": 524,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Annotator Agreement and Aggregation",
"sec_num": "3.3.1"
},
{
"text": "For compatibility with CIRCA, we exclude OTHER (as we found it annotated when the question was not polar in CIRCA) and N/A, which is very infrequent in FRIENDS-QIA (only 1.59% of the labels). We note here that OTHER is used in FRIENDS-QIA in cases where the answer is not related to the question -which is possible due to the dialogue context on which FRIENDS-QIA is based.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotator Agreement and Aggregation",
"sec_num": "3.3.1"
},
{
"text": "Dataset FRIENDS-QIA CIRCA The resulting dataset used in the experiments contains 5,930 question-answer pairs over four labels. The dataset sizes after splitting it are listed in Table 4 , including a comparison to the CIRCA dataset (Louis et al., 2020) . The splits for both datasets are stratified to ensure similar label distribution across all three sets. Table 5 provides additional insights into the two datasets, by providing measures on vocabulary size, length of the instance as well as type-token ratio (TTR). We observe that FRIENDS-QIA is richer in terms of vocabulary, has higher lexical variety and contains on average longer utterances, despite the smaller overall size compared to CIRCA. ",
"cite_spans": [
{
"start": 232,
"end": 252,
"text": "(Louis et al., 2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 178,
"end": 185,
"text": "Table 4",
"ref_id": "TABREF5"
},
{
"start": 359,
"end": 366,
"text": "Table 5",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Annotator Agreement and Aggregation",
"sec_num": "3.3.1"
},
{
"text": "We experiment with different variants of a Convolutional Neural Network (CNN). In particular, we explore the use of BERT embeddings, a crowd layer and combining datasets to obtain higher performance on FRIENDS-QIA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Base CNN Our CNNs are implemented with inspiration from Kim (2014) , who uses 1dimensional convolutions in parallel. As shown in Figure 2 , our implementation of the parallel convolutions consists of a convolutional 1d, a max pooling, a flatten and a dropout layer, after which they are concatenated and fed to the final output layer which uses softmax as activation function. The convolutional layer applies the ReLU activation function after which the maximum of each filter is selected. The dropout layer applies a dropout rate of 0.5 which is the same rate we use on the final layer when we apply regularization with a constraint of the L2-norm on the weights. The models either take English BERT embeddings (Devlin et al., 2019) as input or use GloVe embeddings (Pennington et al., 2014) . The models are optimized using the Adam optimizer (Kingma and Ba, 2015). This general architecture is what all of our CNN variants are built upon. In preliminary experiments we took a portion of the training data as hyperparameter tuning set. We performed a grid search on the base CNN to find the optimal hyperparameters. For the CNN with BERT no further hyperparameter tuning was done. Crowd Layer Figure 3 illustrates the key idea of the DLFC approach we adopt on top of our CNN.",
"cite_spans": [
{
"start": 56,
"end": 66,
"text": "Kim (2014)",
"ref_id": "BIBREF8"
},
{
"start": 712,
"end": 733,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF4"
},
{
"start": 767,
"end": 792,
"text": "(Pennington et al., 2014)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [
{
"start": 129,
"end": 137,
"text": "Figure 2",
"ref_id": "FIGREF2"
},
{
"start": 1195,
"end": 1203,
"text": "Figure 3",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Training the models with a crowd layer is closely based on the paper and code by Rodrigues and Pereira (2018) . Following their implementation, our crowd layer is applied on top of the existing network. It takes as input the output of the dense layer and uses the annotator-specific labels, each modeled as a separate task, to propagate errors through the network and adjust the gradients. The layer is applied on the already trained and saved base model with an annotator-specific weight matrix (W r ), and further training is performed with the crowd layer. We found this setup to perform best. For the CNN+crowd layer model we use the same parameters as when training the base model. After training, the crowd layer is removed to allow for the final classification using the original dense layer with a softmax activation function.",
"cite_spans": [
{
"start": 81,
"end": 109,
"text": "Rodrigues and Pereira (2018)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Evaluation Accuracy and macro-average F1score are the evaluation metrics. We first report results on the development set, and report the performance of the best models on the test portion. As baselines, we provide results of a majority baseline and a Naive Bayes model with word trigrams.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "The main results on the development set are given in Table 6 . It provides results of a majority baseline, the CNN with GloVe and BERT embeddings, and the CNN trained with the crowd layer. We first discuss results for models trained on FRIENDS-QIA, followed by results on training with FRIENDS-QIA and CIRCA. Take-aways There are four take-aways. First, we observe the difficulty of the task. This can be seen from the low baseline results: The majority baseline reaches an accuracy of 49.07 and F1-score of 16.46. The Naive Bayes model reaches an accuracy of 52.45 and F1-score of 37.08. Second, we test several variants of the base CNN with GloVe embeddings: one which takes the concatenation of question and answer as input (CNN), one that models question and answer separately (CNN, multi-input), and a CNN using only the answer or the question. The latter provides information on how much signal is represented in the answer (or question) alone. As the results in Table 6 show, multi-input modeling is not consistently the best model (on both metrics). Training the CNN on answers alone is highly predictive and performs substantially better than the question alone. This corroborates findings by Louis et al. (2020) : the answer alone is highly predictive for the task (yet not sufficient). Overall, the best models are obtained when considering both question and answer, reaching an accuracy of 54.86 and F1-score of 33.02 on the development set with base CNN.",
"cite_spans": [
{
"start": 1202,
"end": 1221,
"text": "Louis et al. (2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 53,
"end": 60,
"text": "Table 6",
"ref_id": "TABREF9"
},
{
"start": 969,
"end": 976,
"text": "Table 6",
"ref_id": "TABREF9"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Third, we observe that modeling the annotator uncertainty is beneficial. When we add the crowd layer to the base CNN (CNN + crowd layer), we observe both improved accuracy and F1-score, reaching 55.71 and 39.38, respectively. This is encouraging, as it shows that the disagreement in human annotations is informative for this task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Finally, we observe that using BERT embeddings as input representations consistently and remarkably improves the performance of the model. The CNN model with BERT reaches an accuracy of 64.08 and F1-score of 49.16, which is an absolute improvement of over 9% in accuracy and over 16% in F1-score. What is, however, striking, is that this trend does not hold for the BERT-based model trained on answer or question alone; we attribute this to the limited context, but this result warrants further investigation. Especially the F1score is not even better than the base CNN model with GloVe embeddings on answers alone. Nevertheless, BERT representations overall improve the best models that take question-answer pairs as input, which are necessary for the overall best performance on the task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "We observe that the crowd layer for the CNN taking the concatenation of question and answer as input (CNN) further improves upon BERT representations, for which we obtain an F1-score of 55 and the overall best performance with the CNN with BERT and crowd layer. The accuracy with the crowd layer does not improve considerably, which shows that the crowd layer is not equally helpful for all classes. We hypothesize that the crowd layer is particularly helpful for infrequent classes, which are the most difficult for this task and which we observed the highest annotator disagreement on. This is in fact the case, as we will discuss in Section 6.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Training on FRIENDS-QIA and CIRCA We investigated whether we could improve the model further by using also the CIRCA data as additional training data. The bottom rows of Table 6 show that this is not the case; taking the union of the two datasets is generally not beneficial. The limited added vocabulary of CIRCA does not yield better generalization within FRIENDS-QIA. Test Set Results Table 7 shows the results of evaluating the best models and baselines on the test set. The results corroborate our findings from the development set. Training with the crowd layer is beneficial, and improves F1-score from 45.65 to 47.89. Similarly as on the development set, overall accuracy slightly drops with the crowd layer (it is within a 1% range). Training on the union of the two datasets does not outperform a model trained on FRIENDS-QIA alone.",
"cite_spans": [],
"ref_spans": [
{
"start": 170,
"end": 177,
"text": "Table 6",
"ref_id": "TABREF9"
},
{
"start": 388,
"end": 395,
"text": "Table 7",
"ref_id": "TABREF11"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "In the following section, we provide additional insights on the task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "We observe a consistent improvement of overall macro F1-score, at a cost of a slight drop in accuracy. Therefore we analyze the per-class F1-score for the BERT-based model trained with and without the crowd layer. Moreover, we analyze the resulting estimated annotator bias matrices obtained from the crowd layer. The results are presented in Table 8 . The results confirm that the crowd layer is particularly useful for more easily confused (in this task low frequency) classes, namely 3 (YES, SUBJECT TO SOME CONDITIONS). It helps on 2 (NO) as well. The crowd layer hurts the most frequent class, which aligns well with the overall gain in macro F1-score at a slight cost in accuracy. Figure 4 shows the three annotator-specific weight matrices extracted from the crowd layer. Overall, the weight matrices show several patterns. The dark diagonals mean that the estimated weights are high and the annotators agree with the gold standard. This is also reflected in the raw agreement as shown in Table 3 . The crowd layer generally models each annotator similarly with a few exceptions. One of them is a clear uncertainty of label 3 (YES, SUBJECT TO SOME CONDITIONS). This fits exactly with the label distribution in the FRIENDS-QIA, where class 3 is heavily underrepresented. For annotator 2, the estimated weights are visibly lower for the gold standard 3 (YES, SUBJECT TO SOME CONDITIONS). That means that this annotator tends to give other labels, mostly 1 (YES) and 4 (NEITHER YES NOR NO), when the actual gold standard is 3 (YES, SUBJECT TO SOME CON-DITIONS). The annotators 1 and 2 assign label 4 (NEITHER YES NOR NO) more often, in cases where the gold standard was actually one of the other labels. This pattern is not as apparent for annotator 3, which means that this annotator agreed slightly more with the gold labels than the other two annotators. Figure 5 shows the annotator agreement for correct and incorrect predictions of CNN with BERT (using FRIENDS-QIA on the test set). It is clear that the model is more likely to predict the wrong label, when the annotators are also in disagreement with each other. Furthermore, we show the amount of correct and incorrect predictions of CNN with BERT for each category (which were described in Section 3.2) in Figure 6 . The amounts are generally similar, but category 3, which required the largest modifications such as adding context to the question, is slightly harder to classify than category 1 and 2, which required either no modification or removing direct cues in the answer, respectively. FRIENDS-QIA versus CIRCA In general, we see a lower performance when we train and evaluate on FRIENDS-QIA, compared to using the CIRCA data for training and evaluation (we also trained models on CIRCA alone and evaluate on FRIENDS-QIA but leave them out for space reasons). The difference in performance is largely due to the general differences between the two datasets, and we describe a few of those differences as well as the reasons for them in the following paragraphs.",
"cite_spans": [],
"ref_spans": [
{
"start": 343,
"end": 350,
"text": "Table 8",
"ref_id": "TABREF13"
},
{
"start": 687,
"end": 695,
"text": "Figure 4",
"ref_id": "FIGREF4"
},
{
"start": 996,
"end": 1003,
"text": "Table 3",
"ref_id": "TABREF3"
},
{
"start": 1862,
"end": 1870,
"text": "Figure 5",
"ref_id": "FIGREF5"
},
{
"start": 2270,
"end": 2278,
"text": "Figure 6",
"ref_id": "FIGREF6"
}
],
"eq_spans": [],
"section": "Crowd Layer Analysis",
"sec_num": null
},
{
"text": "First of all, the data is collected in a different way, which greatly affects the content of the questions and answers. CIRCA is created specifically for the task at hand by Louis et al. (2020) , namely understanding indirect answers, and was created under specific, topic-restricted settings, consisting of 10 different dialogue prompts. FRIENDS-QIA is obtained from a more open domain (yet, confined to typical TV series dialogues) and includes a broader context (also reflected in the vocabulary size differences). For example, it includes cases of both sarcasm and irony as well as additional information in the utterances, which might not be related to the actual question or answer.",
"cite_spans": [
{
"start": 174,
"end": 193,
"text": "Louis et al. (2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Incorrect Predictions",
"sec_num": null
},
{
"text": "Secondly, FRIENDS-QIA includes multisentence responses. Given the dialogue context, we observe cases where the speaker might change their mind in-between the sentences, further complicating the task of interpreting the questions and answers. This is very different from CIRCA due to the restricted and written setup, resulting in the CIRCA data being much more concise in meaning and structure than FRIENDS-QIA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Incorrect Predictions",
"sec_num": null
},
{
"text": "Third of all, a major difference between CIRCA and FRIENDS-QIA is the data size. FRIENDS-QIA contains much fewer QA pairs than CIRCA, which makes it considerably more difficult for the CNNs to learn well from. Additionally, FRIENDS-QIA has fewer annotations contributing to the gold standard (three versus five annotations in CIRCA), yet each of the three corresponds to a single annotator for FRIENDS-QIA, which is not the case in CIRCA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Incorrect Predictions",
"sec_num": null
},
{
"text": "In this paper, we present FRIENDS-QIA, a corpus for studying indirect answers in English dialogues. We propose to mine TV series transcripts of a wellknown TV series. Recent work proposed this challenging answer-understanding task, and collected CIRCA, a dataset on question-answer pairs from 10 dialogue prompts constructed instead using crowdsourcing (Louis et al., 2020) . Motivated by their work, we propose FRIENDS-QIA. It contains a total of 5,930 question-answer pairs, and is released both with a majority label and the raw annotations.",
"cite_spans": [
{
"start": 353,
"end": 373,
"text": "(Louis et al., 2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "Our results with CNNs show that a model trained with BERT embeddings outperforms a CNN trained with GloVe word representations. Most interestingly, this is, however, only the case for a model that considers both question and answer. Training on the answers alone provides reasonable signal for the task, but is not sufficient to resolve the indirect answer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "Understanding indirect answers is a challenging pragmatic task, and even human annotators might not agree on a single gold label. We experiment with a way to leverage disagreement in labeling, which proves encouraging: the macro F1-score of the two best CNN models further improves, when fine-tuning with a crowd layer that encodes individual annotator preferences. This is encouraging, as -to the best of our knowledge -human disagreement has not yet been leveraged in modeling for understanding indirect answers; our dataset can be considered a starting point for further research in learning from disagreement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "Overall, the data in FRIENDS-QIA was originally written in a manuscript for the TV series, then uttered in a spoken dialogue context on the show and ultimately transcribed again. Our results show that understanding indirect answers remains a challenging task. We are missing out on influ-ential factors such as intonation, body language and discourse relations from this dialogue context, when we only process the QA pairs in written form. Modeling such factors is interesting and challenging future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "FRIENDS-QIA is available at: https://github. com/friendsQIA/Friends_QIA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The transcriptions are available at https://fangj. github.io/friends/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We would like to thank the anonymous reviewers for their thorough feedback, as well as ITU's Highperformance Computing Cluster team for providing computational resources used in this work. This research is in part supported by the Independent Research Fund Denmark (DFF) grant 9063-00077B.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
},
{
"text": "Following (Bender and Friedman, 2018) , the following outlines the data statement for FRIENDS-QIA:A. CURATION RATIONALE: Collection of examples of polar questions and curated answers for identification of indirect answer types.B. LANGUAGE VARIETY: US (en-US) mainstream English.C. SPEAKER DEMOGRAPHIC: Characters in a TV series.D. ANNOTATOR DEMOGRAPHIC: All annotators are female, highly proficient in English. Native languages include Danish and Polish. Socioeconomic status: higher-education student.D. SPEECH SITUATION: Both standard and colloquial US English, i.e., spontaneous speech.D. TEXT CHARACTERISTICS: Sentences from transcripts of a TV series.PROVENANCE APPENDIX: The data originates from https://fangj.github.io/ friends/.",
"cite_spans": [
{
"start": 10,
"end": 37,
"text": "(Bender and Friedman, 2018)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "A FRIENDS-QIA Data Statement",
"sec_num": null
},
{
"text": "To avoid having direct mappers or cues in the answers, we remove all occurrences of the words listed below, if they \"stand alone\".\u2022 Yes / yeah / yup / yep / yeah-eah / yah\u2022 No / nope / nah / na / noo(...) / na-ah B.2 Direct mappers which are not removedWe had to draw a line somewhere, between what to keep and what to remove. We ultimately decided to only remove variants of \"yes\" and \"no\" listed in the previous section. This results in keeping terms such as the ones listed below, which otherwise in certain cases might also act as a direct mapper or cue to the answer meaning \"yes\" or \"no\".\u2022 Of course (not) / absolutely (not) / totally (not) / definitely (not) / certainly (not) / obviously (not) / apparently (not) / I guess (not)\u2022 Alright / all right / affirmative / that's right / right / that's correct / sure / exactly / fine / please / okay / OK / 'kay / exclusively\u2022 No way / no can do\u2022 Thank you / thanks / no thanks / I don't know / maybe / kind of / (not) really (not)\u2022 Uhuh / a-huh / mm-hm / mm-mh",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Direct mappers which are removed",
"sec_num": "1"
},
{
"text": "Here we list the variables which FRIENDS-QIA contains as well as a description of them.\u2022 SEASON is the season of the TV series from which the QA pair was extracted.\u2022 EPISODE is the episode (of the season) of the TV series from which the QA pair was extracted.\u2022 CATEGORY tells the category which the QA pair was assigned to during data collection in order to know which preprocessing had to be performed.\u2022 Q_PERSON is the name of the character of \"Friends\" who asked the question.\u2022 A_PERSON is the name of the character of \"Friends\" who said the answer.\u2022 Q_ORIGINAL is the original, non-modified question taken directly from the transcription.\u2022 A_ORIGINAL is the original, non-modified answer taken directly from the transcription.\u2022 Q_MODIFIED is the modified question (might be exactly the same as Q_ORIGINAL, if no modification was needed).\u2022 A_MODIFIED is the modified answer (might be exactly the same as A_ORIGINAL, if no modification was needed).\u2022 ANNOTATION_1 is the annotation given by annotator 1.\u2022 ANNOTATION_2 is the annotation given by annotator 2.\u2022 ANNOTATION_3 is the annotation given by annotator 3.\u2022 GOLDSTANDARD is the aggregated gold label from the three annotators.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "C FRIENDS-QIA Variables",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Truth is a lie: Crowd truth and the seven myths of human annotation",
"authors": [
{
"first": "Lora",
"middle": [],
"last": "Aroyo",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Welty",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "36",
"issue": "",
"pages": "15--24",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lora Aroyo and Chris Welty. 2015. Truth is a lie: Crowd truth and the seven myths of human annota- tion. AI Magazine, 36(1):15-24.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Data statements for natural language processing: Toward mitigating system bias and enabling better science",
"authors": [
{
"first": "Emily",
"middle": [
"M"
],
"last": "Bender",
"suffix": ""
},
{
"first": "Batya",
"middle": [],
"last": "Friedman",
"suffix": ""
}
],
"year": 2018,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "6",
"issue": "",
"pages": "587--604",
"other_ids": {
"DOI": [
"10.1162/tacl_a_00041"
]
},
"num": null,
"urls": [],
"raw_text": "Emily M. Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587-604.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "BoolQ: Exploring the surprising difficulty of natural yes/no questions",
"authors": [
{
"first": "Christopher",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Tom",
"middle": [],
"last": "Kwiatkowski",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Collins",
"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": "2924--2936",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1300"
]
},
"num": null,
"urls": [],
"raw_text": "Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceed- ings of the 2019 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924-2936, Min- neapolis, Minnesota. Association for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Not a simple yes or no: Uncertainty in indirect answers",
"authors": [
{
"first": "Marie-Catherine",
"middle": [],
"last": "De Marneffe",
"suffix": ""
},
{
"first": "Scott",
"middle": [],
"last": "Grimm",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Potts",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the SIGDIAL 2009 Conference",
"volume": "",
"issue": "",
"pages": "136--143",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marie-Catherine de Marneffe, Scott Grimm, and Christopher Potts. 2009. Not a simple yes or no: Un- certainty in indirect answers. In Proceedings of the SIGDIAL 2009 Conference, pages 136-143, London, UK. Association for Computational Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "4171--4186",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1423"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Conversational implicatures in indirect replies",
"authors": [
{
"first": "Nancy",
"middle": [],
"last": "Green",
"suffix": ""
},
{
"first": "Sandra",
"middle": [],
"last": "Carberry",
"suffix": ""
}
],
"year": 1992,
"venue": "30th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "64--71",
"other_ids": {
"DOI": [
"10.3115/981967.981976"
]
},
"num": null,
"urls": [],
"raw_text": "Nancy Green and Sandra Carberry. 1992. Conversa- tional implicatures in indirect replies. In 30th An- nual Meeting of the Association for Computational Linguistics, pages 64-71, Newark, Delaware, USA. Association for Computational Linguistics.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Interpreting and generating indirect answers",
"authors": [
{
"first": "Nancy",
"middle": [],
"last": "Green",
"suffix": ""
},
{
"first": "Sandra",
"middle": [],
"last": "Carberry",
"suffix": ""
}
],
"year": 1999,
"venue": "Computational Linguistics",
"volume": "25",
"issue": "3",
"pages": "389--435",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nancy Green and Sandra Carberry. 1999. Interpret- ing and generating indirect answers. Computational Linguistics, 25(3):389-435.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Can you predict responses to yes/no questions? yes, no, and stuff",
"authors": [
{
"first": "Deborah",
"middle": [],
"last": "Beth Ann Hockey",
"suffix": ""
},
{
"first": "Beverly",
"middle": [],
"last": "Rossen-Knill",
"suffix": ""
},
{
"first": "Matthew",
"middle": [],
"last": "Spejewski",
"suffix": ""
},
{
"first": "Stephen",
"middle": [],
"last": "Stone",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Isard",
"suffix": ""
}
],
"year": 1997,
"venue": "Fifth European Conference on Speech Communication and Technology",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Beth Ann Hockey, Deborah Rossen-Knill, Beverly Spe- jewski, Matthew Stone, and Stephen Isard. 1997. Can you predict responses to yes/no questions? yes, no, and stuff. In Fifth European Conference on Speech Communication and Technology.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Convolutional neural networks for sentence classification",
"authors": [
{
"first": "Yoon",
"middle": [],
"last": "Kim",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "1746--1751",
"other_ids": {
"DOI": [
"10.3115/v1/D14-1181"
]
},
"num": null,
"urls": [],
"raw_text": "Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746-1751, Doha, Qatar. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Adam: A method for stochastic optimization",
"authors": [
{
"first": "P",
"middle": [],
"last": "Diederik",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ba",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. volume abs/1412.6980.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "I'd rather just go to bed\": Understanding indirect answers",
"authors": [
{
"first": "Annie",
"middle": [],
"last": "Louis",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Roth",
"suffix": ""
},
{
"first": "Filip",
"middle": [],
"last": "Radlinski",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "7411--7425",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.601"
]
},
"num": null,
"urls": [],
"raw_text": "Annie Louis, Dan Roth, and Filip Radlinski. 2020. \"I'd rather just go to bed\": Understanding indirect an- swers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 7411-7425, Online. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A case for a range of acceptable annotations",
"authors": [
{
"first": "Jennimaria",
"middle": [],
"last": "Palomaki",
"suffix": ""
},
{
"first": "Olivia",
"middle": [],
"last": "Rhinehart",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Tseng",
"suffix": ""
}
],
"year": 2018,
"venue": "SAD/CrowdBias@ HCOMP",
"volume": "",
"issue": "",
"pages": "19--31",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jennimaria Palomaki, Olivia Rhinehart, and Michael Tseng. 2018. A case for a range of acceptable anno- tations. In SAD/CrowdBias@ HCOMP, pages 19- 31.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"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": {},
"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
},
"BIBREF13": {
"ref_id": "b13",
"title": "Glove: Global vectors for word representation",
"authors": [
{
"first": "Jeffrey",
"middle": [],
"last": "Pennington",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "1532--1543",
"other_ids": {
"DOI": [
"10.3115/v1/D14-1162"
]
},
"num": null,
"urls": [],
"raw_text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word rep- resentation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), page 1532-1543, Doha, Qatar. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Learning part-of-speech taggers with inter-annotator agreement loss",
"authors": [
{
"first": "Barbara",
"middle": [],
"last": "Plank",
"suffix": ""
},
{
"first": "Dirk",
"middle": [],
"last": "Hovy",
"suffix": ""
},
{
"first": "Anders",
"middle": [],
"last": "S\u00f8gaard",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "742--751",
"other_ids": {
"DOI": [
"10.3115/v1/E14-1078"
]
},
"num": null,
"urls": [],
"raw_text": "Barbara Plank, Dirk Hovy, and Anders S\u00f8gaard. 2014. Learning part-of-speech taggers with inter-annotator agreement loss. In Proceedings of the 14th Confer- ence of the European Chapter of the Association for Computational Linguistics, pages 742-751, Gothen- burg, Sweden. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Learning from multiple annotators: Distinguishing good from random labelers",
"authors": [
{
"first": "Filipe",
"middle": [],
"last": "Rodrigues",
"suffix": ""
},
{
"first": "Francisco",
"middle": [],
"last": "Pereira",
"suffix": ""
},
{
"first": "Bernardete",
"middle": [],
"last": "Ribeiro",
"suffix": ""
}
],
"year": 2013,
"venue": "Pattern Recognition Letters",
"volume": "34",
"issue": "12",
"pages": "1428--1436",
"other_ids": {
"DOI": [
"10.1016/j.patrec.2013.05.012"
]
},
"num": null,
"urls": [],
"raw_text": "Filipe Rodrigues, Francisco Pereira, and Bernardete Ribeiro. 2013. Learning from multiple annotators: Distinguishing good from random labelers. Pattern Recognition Letters, 34(12):1428-1436.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Deep learning from crowds",
"authors": [
{
"first": "Filipe",
"middle": [],
"last": "Rodrigues",
"suffix": ""
},
{
"first": "Francisco",
"middle": [
"C"
],
"last": "Pereira",
"suffix": ""
}
],
"year": 2018,
"venue": "Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18). Association for the Advancement of Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Filipe Rodrigues and Francisco C. Pereira. 2018. Deep learning from crowds. In Thirty-Second AAAI Con- ference on Artificial Intelligence (AAAI-18). Associ- ation for the Advancement of Artificial Intelligence.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Questions and responses in English conversation",
"authors": [
{
"first": "Anna-Brita",
"middle": [],
"last": "Stenstr\u00f6m",
"suffix": ""
}
],
"year": 1984,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anna-Brita Stenstr\u00f6m. 1984. Questions and responses in English conversation. Krieger Pub Co.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "The HCRC map task corpus: natural dialogue for speech recognition",
"authors": [
{
"first": "Henry",
"middle": [
"S"
],
"last": "Thompson",
"suffix": ""
},
{
"first": "Anne",
"middle": [
"H"
],
"last": "Anderson",
"suffix": ""
},
{
"first": "Ellen",
"middle": [
"Gurman"
],
"last": "Bard",
"suffix": ""
},
{
"first": "Gwyneth",
"middle": [],
"last": "Doherty-Sneddon",
"suffix": ""
},
{
"first": "Alison",
"middle": [],
"last": "Newlands",
"suffix": ""
},
{
"first": "Cathy",
"middle": [],
"last": "Sotillo",
"suffix": ""
}
],
"year": 1993,
"venue": "Human Language Technology: Proceedings of a Workshop Held at",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Henry S. Thompson, Anne H. Anderson, Ellen Gurman Bard, Gwyneth Doherty-sneddon, Alison Newlands, and Cathy Sotillo. 1993. The HCRC map task cor- pus: natural dialogue for speech recognition. In Hu- man Language Technology: Proceedings of a Work- shop Held at Plainsboro, New Jersey, March 21-24, 1993.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Learning from disagreement: A survey",
"authors": [
{
"first": "Alexandra",
"middle": [
"N"
],
"last": "Uma",
"suffix": ""
},
{
"first": "Tommaso",
"middle": [],
"last": "Fornaciari",
"suffix": ""
},
{
"first": "Dirk",
"middle": [],
"last": "Hovy",
"suffix": ""
},
{
"first": "Silviu",
"middle": [],
"last": "Paun",
"suffix": ""
},
{
"first": "Barbara",
"middle": [],
"last": "Plank",
"suffix": ""
},
{
"first": "Massimo",
"middle": [],
"last": "Poesio",
"suffix": ""
}
],
"year": 2021,
"venue": "The Journal of Artificial Intelligence Research",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexandra N. Uma, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, and Massimo Poesio. 2021. Learning from disagreement: A survey. The Journal of Artificial Intelligence Research, Forth- coming.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Learning from multiple annotators with varying expertise",
"authors": [
{
"first": "Yan",
"middle": [],
"last": "Yan",
"suffix": ""
},
{
"first": "R\u00f3mer",
"middle": [],
"last": "Rosales",
"suffix": ""
},
{
"first": "Glenn",
"middle": [],
"last": "Fung",
"suffix": ""
},
{
"first": "Ramanathan",
"middle": [],
"last": "Subramanian",
"suffix": ""
},
{
"first": "Jennifer",
"middle": [],
"last": "Dy",
"suffix": ""
}
],
"year": 2014,
"venue": "Machine Learning",
"volume": "",
"issue": "",
"pages": "291--327",
"other_ids": {
"DOI": [
"https://link.springer.com/content/pdf/10.1007/s10994-013-5412-1"
]
},
"num": null,
"urls": [],
"raw_text": "Yan Yan, R\u00f3mer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. 2014. Learning from multiple annotators with varying expertise. Machine Learning, page 291-327.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "The authors contributed equally to this work. Q: Hey. Everything ok? A: I'm just mad at my agent. L: NO, NO, YES Q: Are you back from Minsk? A: Well, just for a couple of days. L: YES, NO, YES, SUBJECT TO SOME CONDITIONS",
"type_str": "figure",
"uris": null
},
"FIGREF2": {
"num": null,
"text": "CNN architecture.",
"type_str": "figure",
"uris": null
},
"FIGREF3": {
"num": null,
"text": "Illustration of deep learning from crowds proposed byRodrigues and Pereira (2018).",
"type_str": "figure",
"uris": null
},
"FIGREF4": {
"num": null,
"text": "Normalized weight matrices from the crowd layer for each annotator. 1: YES, 2: NO, 3: YES, SUBJECT TO SOME CONDI-TIONS, 4: NEITHER YES NOR NO.",
"type_str": "figure",
"uris": null
},
"FIGREF5": {
"num": null,
"text": "Correct and incorrect predictions of CNN with BERT vs. annotator agreement.",
"type_str": "figure",
"uris": null
},
"FIGREF6": {
"num": null,
"text": "Correct and incorrect predictions of CNN with BERT vs. category.",
"type_str": "figure",
"uris": null
},
"TABREF0": {
"text": "You still work at the multiplex? Chip: Oh, like I'd give up that job! Free popcorn and candy, anytime I want. I can get you free posters for your room. Monica: He is right, isn't he? Chandler: Y'know what, I think this might be one of the times he's wrong. Okay-ooh, but are you going to have time to read it? Rachel: Oh, I read that in high school.",
"type_str": "table",
"num": null,
"content": "<table><tr><td>1: Yes</td><td>2: No</td><td>3: Yes, subject to some conditions</td></tr><tr><td colspan=\"3\">Monica: Joey: Now if a cow should die of natural</td></tr><tr><td/><td/><td>causes, I can have one of those right?</td></tr><tr><td/><td/><td>Phoebe: Not if I get there first.</td></tr><tr><td>4: Neither yes nor no</td><td>5: Other</td><td>6: N/A</td></tr><tr><td>Phoebe:</td><td/><td/></tr></table>",
"html": null
},
"TABREF1": {
"text": "al. (2020). It consists of six classes: YES, NO, YES SUBJECT TO SOME CONDITIONS, NEITHER YES NOR NO, OTHER and N/A. A description of each class as used in FRIENDS-QIA is",
"type_str": "table",
"num": null,
"content": "<table><tr><td>as follows:</td></tr><tr><td>1. YES: The answer is either a definite yes, a</td></tr><tr><td>probably yes or a sometimes yes</td></tr><tr><td>2. NO: The answer is either a definite no or a</td></tr><tr><td>probably no</td></tr><tr><td>3. YES, SUBJECT TO SOME CONDITIONS: The</td></tr><tr><td>answer is only yes/no, if a certain condition is</td></tr><tr><td>satisfied</td></tr><tr><td>4. NEITHER YES NOR NO: The answer does not</td></tr><tr><td>imply yes nor no</td></tr><tr><td>5. OTHER: The answer is not related to the ques-</td></tr><tr><td>tion</td></tr><tr><td>6. N/A: Lack of majority agreement</td></tr></table>",
"html": null
},
"TABREF3": {
"text": "",
"type_str": "table",
"num": null,
"content": "<table/>",
"html": null
},
"TABREF5": {
"text": "",
"type_str": "table",
"num": null,
"content": "<table/>",
"html": null
},
"TABREF7": {
"text": "",
"type_str": "table",
"num": null,
"content": "<table/>",
"html": null
},
"TABREF9": {
"text": "Results on the FRIENDS-QIA development data.",
"type_str": "table",
"num": null,
"content": "<table/>",
"html": null
},
"TABREF11": {
"text": "Results on the FRIENDS-QIA test data.",
"type_str": "table",
"num": null,
"content": "<table/>",
"html": null
},
"TABREF13": {
"text": "Per-class results on the development set. 1: YES, 2: NO, 3: YES, SUBJECT TO SOME CONDI-TIONS, 4: NEITHER YES NOR NO.",
"type_str": "table",
"num": null,
"content": "<table/>",
"html": null
}
}
}
} |