File size: 112,174 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 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:27:35.401715Z"
},
"title": "Anaphora Resolution in Dialogue: Description of the DFKI-TalkingRobots System for the CODI-CRAC 2021 Shared-Task",
"authors": [
{
"first": "Tatiana",
"middle": [],
"last": "Anikina",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Cennet",
"middle": [],
"last": "Oguz",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Natalia",
"middle": [],
"last": "Skachkova",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Siyu",
"middle": [],
"last": "Tao",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Sharmila",
"middle": [],
"last": "Upadhyaya",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Ivana",
"middle": [],
"last": "Kruijff-Korbayov\u00e1",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We describe the system developed by the DFKI-TalkingRobots Team for the CODI-CRAC 2021 Shared-Task on anaphora resolution in dialogue. Our system consists of three subsystems: (1) the Workspace Coreference System (WCS) incrementally clusters mentions using semantic similarity based on embeddings combined with lexical feature heuristics; (2) the Mention-to-Mention (M2M) coreference resolution system pairs same entity mentions; (3) the Discourse Deixis Resolution (DDR) system employs a Siamese Network to detect discourse anaphor-antecedent pairs. WCS achieved F1-score of 55.6% averaged across the evaluation test sets, M2M 57.2% and DDR 21.5%, all on predicted mentions.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "We describe the system developed by the DFKI-TalkingRobots Team for the CODI-CRAC 2021 Shared-Task on anaphora resolution in dialogue. Our system consists of three subsystems: (1) the Workspace Coreference System (WCS) incrementally clusters mentions using semantic similarity based on embeddings combined with lexical feature heuristics; (2) the Mention-to-Mention (M2M) coreference resolution system pairs same entity mentions; (3) the Discourse Deixis Resolution (DDR) system employs a Siamese Network to detect discourse anaphor-antecedent pairs. WCS achieved F1-score of 55.6% averaged across the evaluation test sets, M2M 57.2% and DDR 21.5%, all on predicted mentions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "This paper describes the anaphora resolution system developed by the DFKI-TalkingRobots Team for the CODI-CRAC 2021 Shared-Task (CCST) 1 (Khosla et al., 2021) . Anaphora resolution (AR) is important for many natural language processing tasks, such as information extraction, summarization, question answering, etc. Most existing systems have been developed for AR in text and focus on coreference resolution. CCST was set up to address AR in dialogue, including not only coreference but also bridging and discourse deixis.",
"cite_spans": [
{
"start": 137,
"end": 158,
"text": "(Khosla et al., 2021)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We were inspired to join CCST, although we did not have an existing AR system to start with. Initial small tests of publicly available pre-trained AR models from the AllenNLP and HuggingFace libraries on our dialogue data in the domain of disaster response team communication (Skachkova and Kruijff-Korbayov\u00e1, 2020) yielded discouraging results. Participating in CCST was a way to kick start the development of an AR system of our own. Our long-term aim is to develop a domain-independent (Willms et al., 2019) .",
"cite_spans": [
{
"start": 276,
"end": 315,
"text": "(Skachkova and Kruijff-Korbayov\u00e1, 2020)",
"ref_id": "BIBREF21"
},
{
"start": 489,
"end": 510,
"text": "(Willms et al., 2019)",
"ref_id": "BIBREF25"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our system for the CCST is the first step in this direction, and should be considered work in progress. It consists of two independent AR subsystems: Workspace Coreference Resolution System (WCS, Sec. 2.1); Mention-to-Mention Coreference Resolution System (M2M, Sec. 2.2); and the Discourse Deixis Resolution System (DDR, Sec. 2.4).The subsystems were developed using the data provided for CCST and evaluated in the CCST tracks Eval-AR and Eval-DD (Pred), respectively. See Tab. 1 for a summary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Sec. 2 describes the subsystems one by one, including the respective approach, mention detection method and how we used the CCST data. We also highlight the differences between our approaches and recent related work in the respective subsections. In Sec. 3 we provide the results and error analysis for each subsystem. We conclude and point out future work in Sec. 4.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We designed and implemented a system for identity anaphora resolution based on the following principles. First, our system processes the input incrementally to make the learning process resemble the way humans read and understand text and resolve coreference. Second, we restrict the number of possible coreference clusters considered at each step to only the salient candidates (i.e., those clusters that have been referred to recently in the conversation). Inspired by the clustering ideas outlined in Rahman and Ng (2011) our system creates clusters on-the-fly by comparing the current mention to the most recent members from each of the clusters available in the workspace.",
"cite_spans": [
{
"start": 504,
"end": 524,
"text": "Rahman and Ng (2011)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "System description 2.1 Workspace Coreference System (WCS)",
"sec_num": "2"
},
{
"text": "We combined workspace-based clustering with a binary classification of referring expressions. Fig. 1 shows an overview of the whole system that considers one mention at each step and computes the probabilities for clustering this mention with other mentions available in the workspace. Our implementation differs from other submissions to the Shared Task. Both winning models (lxucs and UTD_NLP) are based on the pairwise scoring of mentions and extend the coreference model from (Xu and Choi, 2020) . The model submitted by KU_NLP is based on pointer networks whereas our system is more similar in spirit to the incremental approach introduced in (Xia et al., 2020) . Similarly to their work we are building clusters of mentions incrementally and remove old mentions from the workspace if the corresponding clusters have not been updated for a long time. 3 However, there are some important differences between our model and the work by Xia et al. (2020) . For example, we do not use separate representations for clusters, but rather take the latest 5 mentions from each cluster and compute an average probability of assigning a new mention to the same cluster. Moreover, Xia et al. (2020) evict clusters based on their size and distance while our decision to remove a cluster from the workspace is based on the number of singletons that are allowed in the workspace as well as recency of the cluster updates. We also have substantial differences in the training procedure, e.g., rather than considering all correct antecedents, Xia et al. (2020) focus only on the most recent one while our system checks all antecedents, i.e., compares how many mentions in the assigned cluster are actually coreferent with the mention. We also update the parameters after each clustering decision and not once per document. Moreover, our model does not use SpanBERT embeddings or any other components pre-trained on coreference data 4 and is more focused on exploring different kinds of embeddings and how their combinations contribute to the coreference resolution.",
"cite_spans": [
{
"start": 481,
"end": 500,
"text": "(Xu and Choi, 2020)",
"ref_id": "BIBREF27"
},
{
"start": 649,
"end": 667,
"text": "(Xia et al., 2020)",
"ref_id": "BIBREF26"
},
{
"start": 857,
"end": 858,
"text": "3",
"ref_id": null
},
{
"start": 939,
"end": 956,
"text": "Xia et al. (2020)",
"ref_id": "BIBREF26"
},
{
"start": 1174,
"end": 1191,
"text": "Xia et al. (2020)",
"ref_id": "BIBREF26"
},
{
"start": 1531,
"end": 1548,
"text": "Xia et al. (2020)",
"ref_id": "BIBREF26"
}
],
"ref_spans": [
{
"start": 94,
"end": 101,
"text": "Fig. 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "System description 2.1 Workspace Coreference System (WCS)",
"sec_num": "2"
},
{
"text": "We used the data provided by the CCST organizers and split it into train and development set for WCS as follows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "2.1.1"
},
{
"text": "First, we extracted individual documents from each of the files (RST_DTreeBank, Gnome_Subset2, Pear_Stories, Trains_91, Trains_93, AMI_dev, LDC2021E05_Switchboard_3_dev, light_dev and Persuasion_dev). Then we randomly shuffled these documents to create a training set with 550 documents and a development set with 61 documents. We did not use any additional data for training/development. The system was tested on the CCST test set in the Eval-AR track.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "2.1.1"
},
{
"text": "We use SpaCy 5 to extract mention candidates (markables) for WCS. The candidates include all noun chunks detected by SpaCy (based on the en_core_web_trf model). We add all pronouns and adverbs like 'here', 'there', 'now' and 'then' because they are not recognized by SpaCy as noun phrases. We also check for each noun phrase whether it has a prepositional phrase or clause (based on the dependency relations 'prep' and 'relcl' in SpaCy) and extend the span if necessary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SpaCy-based Mention Detection",
"sec_num": "2.1.2"
},
{
"text": "After the release of the CCST annotated test set we checked the performance of our SpaCy-based approach by comparing the gold spans to those extracted by our system. We got the following micro-F1 scores: 72.4% on AMI, 85.6% on Light, 81.6% on Persuasion and 81.5% on Switchboard. These results show that the quality of the mention extraction significantly impacts the overall performance of the coreference resolution system. E.g., WCS had the lowest F1 score on AMI (43.93%) and the highest score on Light (67.28%), which is in line with the mention extraction performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SpaCy-based Mention Detection",
"sec_num": "2.1.2"
},
{
"text": "After extracting mentions we pre-process all sentences in the documents using BERT (pre-trained cased version) and extract various features for each mention, including the mention position within the document, the head lemma, head number, 6 animacy classification, 7 speaker embedding, BERT embedding (Devlin et al., 2018) of the mention's head, averaged BERT embedding for the span of the mention, Numberbatch concept embedding (Speer et al., 2017) for the head lemma, GloVe embedding (Pennington et al., 2014) for the head lemma, averaged GloVe span embedding and the masked language model (MLM) embedding for the mention. For the MLM embedding we use BERT (pre-trained cased base model) and take the current and the previous dialogue turns, mask the mention's head in the turn and ask the model to predict the top candidate for the masked token. We take the GloVe embedding of this candidate as the MLM embedding.",
"cite_spans": [
{
"start": 301,
"end": 322,
"text": "(Devlin et al., 2018)",
"ref_id": "BIBREF2"
},
{
"start": 429,
"end": 449,
"text": "(Speer et al., 2017)",
"ref_id": "BIBREF22"
},
{
"start": 486,
"end": 511,
"text": "(Pennington et al., 2014)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "WCS Input and Architecture",
"sec_num": "2.1.3"
},
{
"text": "When processing a new mention we compare its representation to each of the workspace clusters. Note that the very first position in the workspace is always reserved for the singleton case. All embeddings for the singleton case have zero values but keep the same dimensions as the real cluster members. In order to compare a mention to the workspace clusters we took up to 5 latest members from each cluster and computed the probability for each member-mention pair. The individual pairbased probabilities were averaged to obtain the cluster-level probabilities as shown in Fig. 2 . When training the model we used the extracted features of different types and concatenated each mention feature with the corresponding member feature (e.g., mention speaker with the member speaker embedding) before passing them through the linear layers and activation functions to obtain the final embedding. A separate list of layers (a.k.a. channels) was used for each embedding type. We concatenated the output of different channels and passed it to the feed-forward network. We used ReLU as an activation function for most of the channels. ReLU was also used for the final decision (when combining the outputs of different channels). In parallel, we performed a binary referring expression classification using a network with two linear layers. This network helped the model to decide whether an input mention is referring or not. If it was classified as non-referring it was not clustered with any of the workspace clusters.",
"cite_spans": [],
"ref_spans": [
{
"start": 573,
"end": 579,
"text": "Fig. 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "WCS Input and Architecture",
"sec_num": "2.1.3"
},
{
"text": "After obtaining a probability for each of the workspace clusters as shown in Fig. 2 we assign the mention to the most likely cluster. Then we add the new mention to the workspace clusters, either as a new member of some existing cluster or as a new singleton cluster. In each iteration the clusters in the workspace have to be re-arranged. If some clusters have not been updated for more than 100 steps they are removed from the workspace and stored in the history. Additionally, we limit the maximum number of singleton clusters in the workspace to 20. When deciding whether to put the cluster in the history or keep it in the workspace we sort all candidates by the recency of the last update.",
"cite_spans": [],
"ref_spans": [
{
"start": 77,
"end": 83,
"text": "Fig. 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "WCS Input and Architecture",
"sec_num": "2.1.3"
},
{
"text": "We experimented with different kinds of embeddings and their combinations on a subset of the data. The details are presented in A.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WCS Input and Architecture",
"sec_num": "2.1.3"
},
{
"text": "In order to provide a learning signal for the network we combined three different loss functions. The first loss compares the gold clusters to the ones generated by the model (clustering loss). The second loss checks how heterogeneous are the mentions that appear in the same cluster (cluster coherence loss) and the third loss is a simple cross-entropy loss for the binary classification of referring/nonreferring mentions (referring loss).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WCS Training",
"sec_num": "2.1.4"
},
{
"text": "While training our model we used a teacher forcing ratio of 0.3 to enforce the correct clustering in 30% of all cases and make the network converge faster. We also checked the compatibility between the clusters based on the features typically used in coreference systems. These features include number and animacy compatibility and nesting. E.g., number compatibility guarantees that mentions in plural are not clustered together with the ones that are in singular and animacy compatibility prevents clustering animate and inanimate objects together. Nesting simply means that if some mention is within the NP represented by another mention these two should not be clustered together (e.g., 'a good student with a book' and 'a book'). These conditions were added in order to make the learning process easier by filtering out incompatible clusters at an early stage. However, since our animacy classification and number annotations were not perfect, the filtering stage could cause some mistakes in the later clustering. Additionally, we checked first person pronouns and speaker consistency. E.g., if the current cluster has a speaker A with the mention 'I', another mention 'I' with a different speaker B cannot be put in the same cluster.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WCS Training",
"sec_num": "2.1.4"
},
{
"text": "At test time we used an additional constraint that all referential mentions that are third person pronouns such as 'it' and 'they' cannot form a new singleton cluster and must be clustered with one of the preceding workspace clusters. In order to score the workspace clusters for this task we ordered them by recency starting with the most recent ones and checked the mention-cluster compatibility as explained above (using animacy, number and nesting constraints). Among the compatible clusters we selected the one that had the highest probability according to the masked language model. For this we masked the current mention in the original sentence and checked one by one the likelihoods of mentions from different clusters to fill the position. These likelihoods were averaged for each cluster based on the 5 most recent members and finally the cluster with the highest likelihood was selected. We trained WCS for 5 epochs with a learning rate of 0.0001 and Adam as an optimizer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WCS Post-processing",
"sec_num": "2.1.5"
},
{
"text": "Coreference resolution systems aim to find the mentions of the same entities which might be referred to by different expressions (e.g., synonyms, pronouns, etc.) or the same expressions according to their presented context in the document. This is challenging because changes in the context and mentions of the entities strongly affect the coreference resolution systems. Additionally, the same expressions might also refer to different entities in the real world. These noted points cause three different challenges that the M2M model aims to overcome. The first is pairing the same mentions of different entities because of the same context. The second is the problem with the different contexts of the same entities when they are referred to by the same expressions. The third one is the different mentions of the same entities in different contexts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mention-to-Mention System (M2M)",
"sec_num": "2.2"
},
{
"text": "Most coreference resolution models have been adapted to rank the antecedents based on average embeddings (Lee et al., 2017 and contextualized embeddings (Joshi et al., 2019 (Joshi et al., , 2020 . Furthermore, Xu and Choi (2020) has adapted the model implementation from and used the contextualized span embeddings (Joshi et al., 2020) to obtain the mention candidates through span enumeration and aggressive pruning. The M2M system combines the average and contextualized embeddings instead of using them alone, to avoid context variation of the same mentions and capture lexical similarities of different mentions. Therefore, the M2M system is designed with the aim of suppressing the context variations of different or the same entities. This simply means that the different mentions might be paired because their contexts are similar, and on the contrary, the same expressions might not be paired together because their context or tokens are different. Therefore, we aim first to investigate the effect of mention-mention matching, and then pooling them together with the paired entities into the same clusters. Therefore, the M2M system is designed to cluster the mentions which are paired with the common candidates.",
"cite_spans": [
{
"start": 105,
"end": 122,
"text": "(Lee et al., 2017",
"ref_id": "BIBREF12"
},
{
"start": 153,
"end": 172,
"text": "(Joshi et al., 2019",
"ref_id": "BIBREF7"
},
{
"start": 173,
"end": 194,
"text": "(Joshi et al., , 2020",
"ref_id": "BIBREF6"
},
{
"start": 210,
"end": 228,
"text": "Xu and Choi (2020)",
"ref_id": "BIBREF27"
},
{
"start": 315,
"end": 335,
"text": "(Joshi et al., 2020)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Mention-to-Mention System (M2M)",
"sec_num": "2.2"
},
{
"text": "Whereas the M2M implementation differs from the system of pointer network by KU_NLP, it is similar to the winning systems lxucs and UTD_NLP because it is designed to match the most similar entities with a pairwise comparison. M2M, however, forms clusters based on transitive closure over the mention-mention pairs, i.e., pairs which have a common mention are clustered together. Additionally, our mention representation method differs by the included distance vector.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mention-to-Mention System (M2M)",
"sec_num": "2.2"
},
{
"text": "To train the M2M model we use the Light, Persuasion, and Switchboard data provided by the CCST organizers. All three datasets are used to train the Pronoun and Noun Models (see below). For training the Self Model (see below) the Switchboard data is not used, because the speaker of the corresponding utterances is not given in the training data, only in test data. Therefore, only the Light and Persuasion data is used to train the Self Model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "2.2.1"
},
{
"text": "We treat mention detection as a sequence tagging task, analogous to the common approach to Named Entity Recognition. Given an input sequence of tokens, our mention detection model extracts markable chunks by tagging the tokens in the IOB2 format, i.e., with the beginning of each chunk tagged B, any subsequent tokens in the chunk tagged I, and tokens not in any chunk tagged O.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mention Detection with BiLSTM-CRF",
"sec_num": "2.2.2"
},
{
"text": "We implemented a Bidirectional LSTM -Conditional Random Field (BiLSTM-CRF) Model, shown to be effective in sequence tagging (Huang et al., 2015) . As the standard IOB2 format only annotates flat chunks, our model only predicts the markables with the widest scope in cases of nesting. We flattened the training data accordingly when converting it to the IOB2 format.",
"cite_spans": [
{
"start": 124,
"end": 144,
"text": "(Huang et al., 2015)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Mention Detection with BiLSTM-CRF",
"sec_num": "2.2.2"
},
{
"text": "The architecture of the model is as follows: the input data (in IOB2 format) is first passed through an ELMo layer (Peters et al., 2018) , where contextualized word representations (embeddings) were extracted for each word. The ELMo embeddings are then passed to the BiLSTM-CRF model. Finally, the CRF model predicts the tags for the input tokens. In a post-processing step, we used SpaCy (en_core_web_trf ) to extract any nested noun chunks from the markables predicted by the model. Maximum sentence length was set to 50, and we used the Adam optimizer with a learning rate of 0.0005 and trained the model for 4 epochs.",
"cite_spans": [
{
"start": 115,
"end": 136,
"text": "(Peters et al., 2018)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Mention Detection with BiLSTM-CRF",
"sec_num": "2.2.2"
},
{
"text": "The training and development sets are created as an 8:2 split of the documents in each CCST dataset. This was done to include sentences from all datasets and prevent bias towards one dataset. The label distribution is shown in Tab. 2. While our model achieves an overall F1 score of 0.98 on the development set, a breakdown of the score by label (Tab. 2) shows that the F1 score for B and I tags are significantly lower, at 0.92 and 0.90, respectively. This means that our current model is overly conservative in predicting the markable chunks, indicated by a large number of false negatives for the B and I tags. The false negatives result in incomplete markables being extracted and impact the performance of the downstream task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mention Detection with BiLSTM-CRF",
"sec_num": "2.2.2"
},
{
"text": "We use contextualized word embeddings BERT and ELMo, and average word embeddings GloVe and Fasttext (Joulin et al., 2016) . The individual utterances are considered as the input for the embedding systems to extract the vectors for the head and span of the mentions. In order to obtain the rich feature representation of the mentions, word embedding models are used under four different contributions: (1) the 10-th layer of BERT is used to obtain a significant local representation without global context of utterances; (2) ELMo is employed for positional encoding of full context of the utterances;",
"cite_spans": [
{
"start": 100,
"end": 121,
"text": "(Joulin et al., 2016)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Training",
"sec_num": "2.3"
},
{
"text": "(3) GloVe is applied to represent the external global context of each token; (4) Fasttext is utilized to extract the n-gram features of words. Because a mention can be represented with different language forms such as a span of multiple words, one word entity, pronoun, etc., we used different embeddings to obtain the vectors: only BERT is used for span representation, while a combination of BERT, ELMo, GloVe, and Fasttext is used for head representation of each mention.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Training",
"sec_num": "2.3"
},
{
"text": "Each mention is represented by using the word embedding methods with four different feature sets:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Training",
"sec_num": "2.3"
},
{
"text": "(1) head embedding to represent similarities between the different mentions of the same entities;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Training",
"sec_num": "2.3"
},
{
"text": "(2) span embeddings to represent the differences between similar mentions of the different entities;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Training",
"sec_num": "2.3"
},
{
"text": "(3) the distance feature vector is obtained by using distance functions between the head embeddings of mentions to represent the similarities via a distance 7-d vector; (4) speaker embeddings by using the average embeddings.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Training",
"sec_num": "2.3"
},
{
"text": "Three different models are designed as part of the M2M system. The Self Model is built for personal pronouns with average embeddings of the speaker and head features of the mentions. The Pronoun Model is designed to pair the third person and place pronouns to the corresponding candidate by exploiting the head average embeddings of mentions. The Noun Model attempts to find the pairs of the mentions by using head and span embeddings with contextualized and average embeddings when they are noun phrases. Fig. 3 shows the Noun Model.",
"cite_spans": [],
"ref_spans": [
{
"start": 506,
"end": 512,
"text": "Fig. 3",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "M2M Training",
"sec_num": "2.3"
},
{
"text": "The models are designed with linear layers with a ReLU activation function for the head, speaker and span features. Then the extracted vectors from embeddings are concatenated with a distance vector to train a linear layer with a ReLU activation function. Then a linear layer with a Sigmoid activation function is used for output probabilities.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Architecture and Training",
"sec_num": "2.3.1"
},
{
"text": "The models are trained with the same hyperparameters. The learning rate is defined as 0.0001 and the optimizer is Adam. For computing the loss, we specify BCELoss which is a criterion that measures the Binary Cross Entropy between the target and the output. Training is done for 30 epochs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "M2M Architecture and Training",
"sec_num": "2.3.1"
},
{
"text": "Webber (1991) defines discourse deixis as reference to a discourse entity such as a proposition, description, event, speech act, etc. The discourse entity is a part of the discourse model of the dialogue participant and is realized in a certain discourse segment, i.e. a chunk of a linguistic text, e.g., a sequence of clauses or utterances. The exact boundaries of discourse segments are often hard to define even for humans (Webber, 1991; Artstein and Poesio, 2006) . In addition, the resolution of discourse deixis mentions requires cross-sentence information about rhetorical relations and focus/attention which is often not annotated. All this makes discourse deixis resolution a very challenging task.",
"cite_spans": [
{
"start": 426,
"end": 440,
"text": "(Webber, 1991;",
"ref_id": "BIBREF24"
},
{
"start": 441,
"end": 467,
"text": "Artstein and Poesio, 2006)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Resolution(DDR)",
"sec_num": "2.4"
},
{
"text": "There exist only a few frameworks focusing on the resolution of discourse deixis mentions. Most early works that we know of (e.g., Eckert and Strube (2000) , Navarretta (2000) , or Byron (2002)) are rule-based. They use hand-crafted rules to pick out anaphors, which are usually demonstrative pronouns. The right antecedent for each anaphor is found relying on candidates' dialogue act types. Early machine learning approaches are presented, e.g., by Strube and M\u00fcller (2003) and M\u00fcller (2008) , who use many elaborate manually created feature vectors to represent markables, and train classifiers to predict whether an antecedentanaphor pair is valid or not. Marasovi\u0107 et al. (2017) introduce a model that relies on a neural network (a Siamese Net) to rank potential antecedent candidates given an anaphor. The input for the network, i.e. sentences containing anaphors and antecedent candidates, is represented as feature vectors (embeddings) using bi-LSTM.",
"cite_spans": [
{
"start": 131,
"end": 155,
"text": "Eckert and Strube (2000)",
"ref_id": "BIBREF3"
},
{
"start": 158,
"end": 175,
"text": "Navarretta (2000)",
"ref_id": "BIBREF16"
},
{
"start": 451,
"end": 475,
"text": "Strube and M\u00fcller (2003)",
"ref_id": "BIBREF23"
},
{
"start": 480,
"end": 493,
"text": "M\u00fcller (2008)",
"ref_id": "BIBREF15"
},
{
"start": 660,
"end": 683,
"text": "Marasovi\u0107 et al. (2017)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Resolution(DDR)",
"sec_num": "2.4"
},
{
"text": "Works focusing on automatic identification of discourse deixis anaphors are also not many. An algorithm based on hand-crafted features to recognize discourse deictic it, this and that is suggested, e.g., by M\u00fcller (2008) . Other works, e.g., by Eckert and Strube (2000) , Kolhatkar et al. (2018) , or present analyses of anaphors with non-nominal antecedents and their properties, but no actual implementations.",
"cite_spans": [
{
"start": 207,
"end": 220,
"text": "M\u00fcller (2008)",
"ref_id": "BIBREF15"
},
{
"start": 256,
"end": 269,
"text": "Strube (2000)",
"ref_id": "BIBREF3"
},
{
"start": 272,
"end": 295,
"text": "Kolhatkar et al. (2018)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Resolution(DDR)",
"sec_num": "2.4"
},
{
"text": "Our approach combines machine learning and hand-crafted rules. The machine learning part is inspired by the mention-ranking model for abstract anaphora resolution by Marasovi\u0107 et al. (2017) . Given a discourse deixis mention that needs to be resolved, the task is to choose the most suitable antecedent among several candidates. To do so, we train a scorer implemented as a Siamese Net with the goal to assign a higher score to true anaphorantecedent pairs and lower scores to false ones. Our implementation of the Siamese Net is much simpler than that of Marasovi\u0107 et al. (2017) -instead of bi-LSTM we use BERT's [CLS] token to extract utterance embeddings, and our antecedent-anaphor score is defined as simple Euclidean distance between the two vectors produced by the Siamese Net. Identification of anaphors and the antecedent candidates is rule-based. We explain the details of our system below.",
"cite_spans": [
{
"start": 166,
"end": 189,
"text": "Marasovi\u0107 et al. (2017)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Resolution(DDR)",
"sec_num": "2.4"
},
{
"text": "The scorer is trained on all the CCST train and development data. Before that the light_dev file was used to work on the heuristic rules. We built our dataset as follows. For each utterance containing a discourse deixis anaphor (anchor utterance) we create a pair of examples (see Fig. 4 ) by replacing the anaphor with its true antecedent (a positive example) and with a false candidate (a negative example), which is usually the utterance before the true antecedent. The idea is to make the Siamese Net learn that the true antecedent fits semantically into the utterance containing the anaphor. In total Anchor: That certainly scared the eternal daylights out of me ! Positive example: I myself happened to look out from the kitchen and see his dangling legs in front of the window . certainly scared the eternal daylights out of me ! Negative example: He never experimented with flying contraptions again . certainly scared the eternal daylights out of me ! ",
"cite_spans": [],
"ref_spans": [
{
"start": 281,
"end": 287,
"text": "Fig. 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Data",
"sec_num": "2.4.1"
},
{
"text": "We do not assume that markables are given during testing, so it is our task to recognize and cluster them. For this we employ heuristic rules that are based on observations in the light_dev data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Mention Detection",
"sec_num": "2.4.2"
},
{
"text": "First, we want to find discourse deixis mentions that need to be resolved. For now we focus only on demonstrative pronouns, which we detect using the SpaCy NLP pipeline (Honnibal et al., 2020). SpaCy does not distinguish between different types of non-personal pronouns, and labels them all with a DET tag. So, we consider all subtrees with a DET head linked to a parent node via 'dobj', 'pobj', 'nsubj' or 'attr' relations as discourse deixis markables that need to be resolved.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Mention Detection",
"sec_num": "2.4.2"
},
{
"text": "Next, we need to find the antecedents of the detected discourse deixis mentions and cluster the right mentions together. It should be noticed that, in most cases these clusters only include 2-3 mentions, and the antecedents can be found within the local context of several previous utterances. We assume that the antecedents are verbal phrases, clauses, or sequences of clauses (split antecedent cases).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Mention Detection",
"sec_num": "2.4.2"
},
{
"text": "For extracting the antecedent candidates we consider the current utterance containing the mention, and 3 previous ones. Each previous utterance, as well as their continuous sequences are viewed as possible (split) antecedents. Besides, we find all subtrees with VERB or AUX heads linked to the ROOT node via 'relcl', 'conj', 'ccomp', or 'advcl' relations. We remove all overlapping candidates and candidates that occur after the mention to be resolved. In case of nested candidates, the longest one is chosen. We also apply the following simple heuristics to exclude unlikely candidates and improve the results. First, we limit the size of the clusters by two elements. Here we also consider cases of anaphors referring to split antecedents, which form their own singleton clusters. Second, we disregard all candidates that do not have verbal roots.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixis Mention Detection",
"sec_num": "2.4.2"
},
{
"text": "The discourse deixis resolution system ranks the antecedent candidates using a scorer, sorts the scores in descending order and selects the top candidate. At this step we limit the output of the scorer to 5 candidates and re-sort them giving preference to the candidates closest to the anaphor, as usually the anaphor immediately follows its antecedent.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "DDR Architecture and Training",
"sec_num": "2.4.3"
},
{
"text": "Being a Siamese Net, the scorer consists of two independent pipelines that have identical architectures. One pipeline is designed for the anchor, the other for an antecedent candidate. Each pipeline accepts input in the form of BERT input IDs, which are then fed into the BERT BASE model. We use the models' pooler output to represent each sequence, which is next processed by a simple multilayer perceptron (MLP) network containing two linear layers with a batch normalization and a ReLU activation function. Finally, the Euclidean distance between the feature representations produced by both MLP networks is calculated, and the pair with the smallest distance gets the biggest score. The scorer's architecture is shown in Fig. 5 .",
"cite_spans": [],
"ref_spans": [
{
"start": 725,
"end": 731,
"text": "Fig. 5",
"ref_id": "FIGREF4"
}
],
"eq_spans": [],
"section": "DDR Architecture and Training",
"sec_num": "2.4.3"
},
{
"text": "The scorer is trained for 12 epochs. We use adaptive learning rate starting with 3e-5, the AdamW optimizer and the TripletMarginLoss loss function.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "DDR Architecture and Training",
"sec_num": "2.4.3"
},
{
"text": "WCS achieved an average F1-score of 64.99% on Light, 43.93% on AMI, 59.93% on Persuasion and 53.55% on Switchboard. Although the results are lower than those of the top-performing systems, we believe that our approach presents a new way to treat coreference as an incremental clustering task and can be further improved by experimenting with different configurations of input and hyperparameters as well as more robust feature extraction.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Workspace Coreference System (WCS)",
"sec_num": "3.1"
},
{
"text": "We believe that our system should benefit from better animacy and number annotations because these are crucial for the decision on the mentioncluster compatibility as explained in Section 2.1. Using a better mention extraction model should further improve the scores. We also evaluated WCS using the gold mentions as input and obtained the following results: 73.24% on Light, 56.48% on AMI, 72.06% on Persuasion and 62.85% on Switchboard. The difference in scores shows that WCS gains up to 12% F-score on AMI and Persuasion, up to 9% on Switchboard and 6% on Light when it uses the correct mentions. We noticed that quite often markables extracted by SpaCy were incomplete and sometimes even missing. E.g., NPs such as 'my two year old' and 'the whole time' were not recognized as NPs by SpaCy and some parts of complex NPs such as 'the kind of things you're supposed to look for' were missing. Nested NPs present another big challenge for SpaCy, e.g., 'case' in 'the case material' was not recognized as a markable. Our feature extraction process heavily relies on SpaCy because we use it together with some heuristics to find a head in each mention span and perform the lemmatization. For instance, in one case 'someones' was selected as a head in 'someones free will', hence the incorrect embedding for 'someone' instead of 'will' was used for the clustering decision.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Workspace Coreference System (WCS)",
"sec_num": "3.1"
},
{
"text": "Some of the rules and heuristics we implemented for WCS do not apply in some cases. E.g., the AMI dataset proved to be especially challenging because the dialogues in AMI have several participants and there are references to objects that were not previously mentioned in the text. For instance, a third person pronoun 'it' in 'do you wanna hold it' does not have an antecedent in the preceding text because the object can be inferred from the physical context. But our algorithm requires to find an antecedent for each mention of 'it' that is a referring expression, hence 'it' ends up in the wrong cluster.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Workspace Coreference System (WCS)",
"sec_num": "3.1"
},
{
"text": "We plan to continue our experiments with WCS by improving the feature extraction process and applying WCS to different domains and genres.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Workspace Coreference System (WCS)",
"sec_num": "3.1"
},
{
"text": "The M2M system achieved F1-score of 61.26% on Light, 59.20% on Persuasion and 51.24% on Switchboard. Whereas Self Model shows high pair-ing performance, we observe that Pronoun Model struggles with lexical variation between pronouns and the corresponding nouns. We assume that the high performance of Self Model is based on the speaker invariant, i.e. two speakers occur in the data. For example, the model pairs the mentions 'I' and 'my' of a speaker always with the corresponding speaker. On the other hand, the delicate performance with the Pronoun Model is observed based on the high lexical variation: 'He' needs to be paired with the 'father' as well as 'his'. However, the design of M2M is not able to handle more than one possible pair even though the correct pair count is more. The Noun Model is also affected by this problem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mention-to-Mention System (M2M)",
"sec_num": "3.2"
},
{
"text": "To conclude, M2M performs effectively with the similar lexical features of mention-candidate pairs where the candidate variation is lower, e.g., personal pronouns of speakers. However, it struggles to pair the mentions when they are represented with different lexical features. At the same time, It suffers from the same mentions of different entities. In order to overcome the deficiencies of the M2M system, we expect that the clustering mechanism solves the lexical variation problem between mentions by obtaining one possible cluster instead of many possible candidate mentions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mention-to-Mention System (M2M)",
"sec_num": "3.2"
},
{
"text": "The DDR system achieved F1-score 20.97% on the Light dataset, 17.43% on AMI, 23.76% on Persuation and 23.86% on Switchboard in the Eval-DD (Pred) track. The submitted system combines the scorer and four heuristic rules partially described in Sec. 2.4, namely limiting cluster size to two mentions, removing all non-verbal antecedent candidates, restricting the scorer's output to five candidates and re-scoring them by distance to the anaphor. For the error analysis we use fully annotated test data released after the competition ended. We perform discourse deixis mention detection as presented in Sec. 2.4.2, and check the performance of the system adding rules incrementally. These results are shown in Tab. 3. The first row of shows the F1 scores achieved by the scorer on the three fully annotated test sets. The next two rows present the scores for cases when the scorer is not forced to search for clusters larger than two mentions, and is applied after the removal of non-verbal candidates. The last four rows show the results after re-sorting the scorer's output by distance to the anaphor. N stands for the number of candidates ranked by the scorer that we re-sort.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixic Resolution (DDR)",
"sec_num": "3.3"
},
{
"text": "Without additional heuristic rules the performance of the scorer is really poor. We assume that the main reasons for this are as follows. First, our model does not possess any information about the semantics of separate tokens and utterances. The example in Fig. 4 illustrates the problem. Using only local context, it is impossible for the model to understand why 'seeing his dangling legs in front of the window' is a better candidate for being scary than 'never experimenting with flying contraptions'. BERT, which we use to get utterance representations, is not capable of providing such deep semantics. In the result, both positive and negative examples look legitimate for the model. Second, the model also lacks information about the discourse structure and how different utterances relate to each other, and thus has difficulties recognizing split antecedents.",
"cite_spans": [],
"ref_spans": [
{
"start": 258,
"end": 264,
"text": "Fig. 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Discourse Deixic Resolution (DDR)",
"sec_num": "3.3"
},
{
"text": "Next, our simple approach has a problem with both recall and precision. Recall is bad because mentions with determiner heads make up only 13%-35% (depending on the sub-corpus) of all discourse deixis mentions in the data. Precision is negatively affected by the fact that the DET tag in SpaCy is very general. As a result, we extract many false discourse deixis mentions, like relative pronoun 'that' in \"I use it to clean the tower that I work in .\" or a determiner phrase 'some of these beautiful flowers' which is a case of 'standard' anaphora.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixic Resolution (DDR)",
"sec_num": "3.3"
},
{
"text": "Using heuristic rules mostly has a positive effect up to 5% on the test sets. However, not all of these rules work well with different dialogue data. Some sub-corpora contain a lot of fragments, many of which are rather short. As a result, the true antecedent may lie more than three utterances back. Some utterances are ungrammatical, and SpaCy makes mistakes parsing them, which leads to wrong candidate boundaries. Searching for antecedent candidates, we try to split complex and compound utterances. This often leads to rather cumbersome SpaCy-based rules that easily fail, because dialogue speech patterns are very diverse, so it is impossible to create an exhaustive set of rules to capture them. Our focus on verbal (split) antecedents may also lead to errors -we work with transcribed dialogue data, and antecedents are not necessarily verbal, they can also be represented by noun phrases like 'the inscriptions on the side of the fountain', prepositional phrases 'to Bath', adverbial phrases 'u : and back to Avon', etc. Finally, re-sorting the scorer's output by distance does not work for all types of dialogue data (see Tab. 3).",
"cite_spans": [],
"ref_spans": [
{
"start": 1132,
"end": 1136,
"text": "Tab.",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discourse Deixic Resolution (DDR)",
"sec_num": "3.3"
},
{
"text": "We plan to continue working on discourse deixis resolution, as our naive system offers some space for improvement. Recall and precision of discourse deixis mentions are first obvious aspects to work on. For this purpose current approaches to shell nouns resolution and event detection can be considered.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse Deixic Resolution (DDR)",
"sec_num": "3.3"
},
{
"text": "We described the three anaphora resolution subsystems we developed from scratch for CCST using only the datasets provided by the CCST organizers. WCS is a coreference resolution approach that incrementally clusters mentions using semantic similarity based on embeddings combined with lexical feature heuristics. It employs SpaCy for markable detection. M2M is another coreference resolution approach which pairs the mentions for the same entities. For markable detection it uses our own BiLSTM-CRF model trained on the CCST data. WCS achieved F1-score of 55.6% averaged across the CCST evaluation test sets, using predicted markables; the M2M system achieved 57.2%. Finally, DDR employs a Siamese Network to distinguish true discourse anaphor-antecedent pairs. Its markable detection is also based on SpaCy. It has average F1-score 21.5% using predicted mentions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Outlook",
"sec_num": "4"
},
{
"text": "Our models employ novel ideas, but are still in the early stages of development. CCST gave us a boost and we see many options for future work regarding both markable detection as well as the AR resolution models themselves. For example, we would like to improve the feature extraction for WCS; enhance M2M by clustering; combine WCS and M2M; and improve the detection of discourse deixis mentions that need to be resolved. We also plan to apply the systems on other datasets, which we are currently preparing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Outlook",
"sec_num": "4"
},
{
"text": "We experimented with using different kinds of embeddings and their combinations for the WCS subsystem on a subset of the data. We used RST_DTreeBank_dev, Switchboard_3_dev and Persuasion_dev files for training and evaluated performance on the light_dev data. In total, we used 50 documents for training and 20 for testing. Unlike other participants (lxucs and UTD_NLP) we did not use SpanBERT embeddings, but we explored GloVe, BERT, Numberbatch, masked LM embeddings as well as separate embeddings for speaker, position and distance between mentions. Table 4 shows some of the results. GloVe head embedding works better than GloVe span embedding (experiments 1 and 2); the difference in CoNLL score is 2.3% (65.09% vs. 62.79%).",
"cite_spans": [],
"ref_spans": [
{
"start": 552,
"end": 559,
"text": "Table 4",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "A Appendix: WCS Embeddings",
"sec_num": null
},
{
"text": "Combined head and span embeddings (3) further improve the performance by 1.2%. Interestingly, the trend is different for BERT: BERT head (4) achieves 64.82% and BERT span (5) 66.43%. Note that in our case BERT span refers to an average over all BERT embeddings of the tokens within the mention span (this is different from SpanBERT). We observe a very small increase in performance when BERT span is combined with BERT head (6), 66.59% CoNLL score; BERT and Glove combined together (7) achieve 66.96%. We also found that WCS performs better when we add Numberbatch embeddings, e.g., BERT with Numberbatch (8) has 68.08% CoNLL score. We found that the combination of GloVe, BERT, Numberbatch, Masked LM, distance and speaker embeddings (9) gives us 68.13% on the test data. Furthermore, when the position embedding is added (10) the score increases to 70.09 that indicates that the positions of mentions within the document 8 play a very important role in resolving coreference.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A Appendix: WCS Embeddings",
"sec_num": null
},
{
"text": "Although we tested various combinations of embeddings, our experiments were limited to a subset of the data (we used only 50 documents for training). Another limitation of our approach is that we defined the dimensionality of each embedding separately, i.e., when passing an embedding through several linear layers we defined the output dimensionality independently for each embedding type. It would be interesting to see how interactions between different embeddings affect the dimensionality. For example, it might be that 100 dimensions work well for GloVe when it is used alone but when it is combined with BERT the dimensionality should be increased (or decreased) to achieve better performance. We would also like to test whether ablation studies show similar trends with different types of data (e.g., dialogue vs. non-dialogue, different genres) and perform a proper statistical analysis of the results. We see our present results as a starting point of an investigation of interactions between the embedding types in dialogue, to determine what embeddings work best in which case and what dimensionality they should have.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A Appendix: WCS Embeddings",
"sec_num": null
},
{
"text": "URL: dfki.de/mlt",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The limit was set to 100 steps without updates in our experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Xia et al. (2020) used encoder weights that were already fine-tuned on the OntoNotes dataset and adopted both mention and pairwise scorers from(Joshi et al., 2020).5 https://spacy.io",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We use SpaCy Lemmatization and number annotation.7 We trained a separate classifier for this using the CCST entity type annotations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Position embedding for each mention is computed as a start position divided by the total number of tokens in the document. When computing a score for clustering we concatenate both referent and antecedent positions and then pass them through a linear layer that increases the dimensionality of embedding to 200.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Identifying reference to abstract objects in dialogue",
"authors": [
{
"first": "Ron",
"middle": [],
"last": "Artstein",
"suffix": ""
},
{
"first": "Massimo",
"middle": [],
"last": "Poesio",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ron Artstein and Massimo Poesio. 2006. Identifying reference to abstract objects in dialogue. Univer- sit\u00e4t Potsdam.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Resolving pronominal reference to abstract entities",
"authors": [
{
"first": "Byron",
"middle": [],
"last": "Donna Karen",
"suffix": ""
}
],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Donna Karen Byron. 2002. Resolving pronominal ref- erence to abstract entities. University of Rochester.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1810.04805"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of deep bidirectional transformers for language under- standing. arXiv preprint arXiv:1810.04805.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Dialogue acts, synchronizing units, and anaphora resolution",
"authors": [
{
"first": "Miriam",
"middle": [],
"last": "Eckert",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Strube",
"suffix": ""
}
],
"year": 2000,
"venue": "Journal of Semantics",
"volume": "17",
"issue": "1",
"pages": "51--89",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Miriam Eckert and Michael Strube. 2000. Dialogue acts, synchronizing units, and anaphora resolution. Journal of Semantics, 17(1):51-89.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "SpaCy: Industrial-strength Natural Language Processing in Python",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.5281/zenodo.1212303"
]
},
"num": null,
"urls": [],
"raw_text": "Matthew Honnibal, Ines Montani, Sofie Van Lan- deghem, and Adriane Boyd. 2020. SpaCy: Industrial-strength Natural Language Processing in Python.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Bidirectional LSTM-CRF Models for Sequence Tagging",
"authors": [
{
"first": "Zhiheng",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Kai",
"middle": [],
"last": "Yu",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirec- tional LSTM-CRF Models for Sequence Tagging.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Spanbert: Improving pre-training by representing and predicting spans",
"authors": [
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Danqi",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Yinhan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Daniel",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Weld",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Levy",
"suffix": ""
}
],
"year": 2020,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "8",
"issue": "",
"pages": "64--77",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. 2020. Spanbert: Improving pre-training by representing and predict- ing spans. Transactions of the Association for Com- putational Linguistics, 8:64-77.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "BERT for coreference resolution: Baselines and analysis",
"authors": [
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Weld",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "5803--5808",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1588"
]
},
"num": null,
"urls": [],
"raw_text": "Mandar Joshi, Omer Levy, Luke Zettlemoyer, and Daniel Weld. 2019. BERT for coreference reso- lution: Baselines and analysis. In Proceedings of the 2019 Conference on Empirical Methods in Nat- ural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5803-5808, Hong Kong, China. Association for Computational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "FastText.zip: Compressing text classification models",
"authors": [
{
"first": "Armand",
"middle": [],
"last": "Joulin",
"suffix": ""
},
{
"first": "Edouard",
"middle": [],
"last": "Grave",
"suffix": ""
},
{
"first": "Piotr",
"middle": [],
"last": "Bojanowski",
"suffix": ""
},
{
"first": "Matthijs",
"middle": [],
"last": "Douze",
"suffix": ""
},
{
"first": "H\u00e9rve",
"middle": [],
"last": "J\u00e9gou",
"suffix": ""
},
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1612.03651"
]
},
"num": null,
"urls": [],
"raw_text": "Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, H\u00e9rve J\u00e9gou, and Tomas Mikolov. 2016. FastText.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The CODI-CRAC 2021",
"authors": [
{
"first": "Sopan",
"middle": [],
"last": "Khosla",
"suffix": ""
},
{
"first": "Juntao",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "Ramesh",
"middle": [],
"last": "Manuvinakurike",
"suffix": ""
},
{
"first": "N",
"middle": [
"G"
],
"last": "Vincent",
"suffix": ""
},
{
"first": "Massimo",
"middle": [],
"last": "Poesio",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Strube",
"suffix": ""
},
{
"first": "Carolyn",
"middle": [],
"last": "Ros\u00e9",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sopan Khosla, Juntao Yu, Ramesh Manuvinakurike, Vincent NG, Massimo Poesio, Michael Strube, and Carolyn Ros\u00e9. 2021. The CODI-CRAC 2021",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Shared Task on Anaphora, Bridging, and Discourse Deixis in Dialogue",
"authors": [],
"year": null,
"venue": "Proceedings of the CODI-CRAC 2021 Shared Task on Anaphora, Bridging, and Discourse Deixis in Dialogue",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shared Task on Anaphora, Bridging, and Discourse Deixis in Dialogue. In Proceedings of the CODI- CRAC 2021 Shared Task on Anaphora, Bridging, and Discourse Deixis in Dialogue. Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Anaphora with nonnominal antecedents in computational linguistics: A survey",
"authors": [
{
"first": "Varada",
"middle": [],
"last": "Kolhatkar",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Roussel",
"suffix": ""
},
{
"first": "Stefanie",
"middle": [],
"last": "Dipper",
"suffix": ""
},
{
"first": "Heike",
"middle": [],
"last": "Zinsmeister",
"suffix": ""
}
],
"year": 2018,
"venue": "Computational Linguistics",
"volume": "44",
"issue": "3",
"pages": "547--612",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Varada Kolhatkar, Adam Roussel, Stefanie Dipper, and Heike Zinsmeister. 2018. Anaphora with non- nominal antecedents in computational linguistics: A survey. Computational Linguistics, 44(3):547-612.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "End-to-end neural coreference resolution",
"authors": [
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Luheng",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "188--197",
"other_ids": {
"DOI": [
"10.18653/v1/D17-1018"
]
},
"num": null,
"urls": [],
"raw_text": "Kenton Lee, Luheng He, Mike Lewis, and Luke Zettle- moyer. 2017. End-to-end neural coreference reso- lution. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 188-197, Copenhagen, Denmark. ACL.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Higher-order coreference resolution with coarse-tofine inference",
"authors": [
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Luheng",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "2",
"issue": "",
"pages": "687--692",
"other_ids": {
"DOI": [
"10.18653/v1/N18-2108"
]
},
"num": null,
"urls": [],
"raw_text": "Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-to- fine inference. In Proceedings of the 2018 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies, Volume 2 (Short Papers), pages 687-692, New Orleans, Louisiana. ACL.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A mention-ranking model for abstract anaphora resolution",
"authors": [
{
"first": "Ana",
"middle": [],
"last": "Marasovi\u0107",
"suffix": ""
},
{
"first": "Leo",
"middle": [],
"last": "Born",
"suffix": ""
},
{
"first": "Juri",
"middle": [],
"last": "Opitz",
"suffix": ""
},
{
"first": "Anette",
"middle": [],
"last": "Frank",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1706.02256"
]
},
"num": null,
"urls": [],
"raw_text": "Ana Marasovi\u0107, Leo Born, Juri Opitz, and Anette Frank. 2017. A mention-ranking model for abstract anaphora resolution. arXiv preprint arXiv:1706.02256.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Fully automatic resolution of'it','this', and'that'in unrestricted multi-party dialog",
"authors": [
{
"first": "Mark-Christoph",
"middle": [],
"last": "M\u00fcller",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mark-Christoph M\u00fcller. 2008. Fully automatic resolu- tion of'it','this', and'that'in unrestricted multi-party dialog. Ph.D. thesis, Universit\u00e4t T\u00fcbingen.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Abstract anaphora resolution in danish",
"authors": [
{
"first": "Costanza",
"middle": [],
"last": "Navarretta",
"suffix": ""
}
],
"year": 2000,
"venue": "1st SIGdial Workshop on Discourse and Dialogue",
"volume": "",
"issue": "",
"pages": "56--65",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Costanza Navarretta. 2000. Abstract anaphora resolu- tion in danish. In 1st SIGdial Workshop on Dis- course and Dialogue, pages 56-65.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Glove: Global vectors for word representation",
"authors": [
{
"first": "Jeffrey",
"middle": [],
"last": "Pennington",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
},
{
"first": "Christopher D",
"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": {},
"num": null,
"urls": [],
"raw_text": "Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word rep- resentation. In Proceedings of the 2014 conference on empirical methods in natural language process- ing (EMNLP), pages 1532-1543.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Deep contextualized word representations",
"authors": [
{
"first": "Matthew",
"middle": [
"E"
],
"last": "Peters",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Neumann",
"suffix": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Iyyer",
"suffix": ""
},
{
"first": "Matt",
"middle": [],
"last": "Gardner",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word repre- sentations. CoRR, abs/1802.05365.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Narrowing the modeling gap: A cluster-ranking approach to coreference resolution",
"authors": [
{
"first": "Altaf",
"middle": [],
"last": "Rahman",
"suffix": ""
},
{
"first": "Vincent",
"middle": [],
"last": "Ng",
"suffix": ""
}
],
"year": 2011,
"venue": "J. Artif. Intell. Res",
"volume": "40",
"issue": "",
"pages": "469--521",
"other_ids": {
"DOI": [
"10.1613/jair.3120"
]
},
"num": null,
"urls": [],
"raw_text": "Altaf Rahman and Vincent Ng. 2011. Narrowing the modeling gap: A cluster-ranking approach to coref- erence resolution. J. Artif. Intell. Res., 40:469-521.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Towards the automatic resolution of anaphora with non-nominal antecedents: Insights from annotation",
"authors": [
{
"first": "Adam",
"middle": [],
"last": "Roussel",
"suffix": ""
},
{
"first": "Stefanie",
"middle": [],
"last": "Dipper",
"suffix": ""
},
{
"first": "Sarah",
"middle": [],
"last": "Jablotschkin",
"suffix": ""
},
{
"first": "Heike",
"middle": [],
"last": "Zinsmeister",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Adam Roussel, Stefanie Dipper, Sarah Jablotschkin, and Heike Zinsmeister. 2018. Towards the auto- matic resolution of anaphora with non-nominal an- tecedents: Insights from annotation.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Reference in team communication for robot-assisted disaster response: An initial analysis",
"authors": [
{
"first": "Natalia",
"middle": [],
"last": "Skachkova",
"suffix": ""
},
{
"first": "Ivana",
"middle": [],
"last": "Kruijff-Korbayov\u00e1",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 3rd Workshop on Computational Models of Reference, Anaphora and Coreference. Workshop on Computational Models of Reference, Anaphora and Coreference (CRAC-2020), befindet sich COL-ING'2020",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Natalia Skachkova and Ivana Kruijff-Korbayov\u00e1. 2020. Reference in team communication for robot-assisted disaster response: An initial analysis. In Proceed- ings of the 3rd Workshop on Computational Models of Reference, Anaphora and Coreference. Workshop on Computational Models of Reference, Anaphora and Coreference (CRAC-2020), befindet sich COL- ING'2020, December 12, Virtual, Spain. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "ConceptNet 5.5: An Open Multilingual Graph of General Knowledge",
"authors": [
{
"first": "Robyn",
"middle": [],
"last": "Speer",
"suffix": ""
},
{
"first": "Joshua",
"middle": [],
"last": "Chin",
"suffix": ""
},
{
"first": "Catherine",
"middle": [],
"last": "Havasi",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, AAAI'17",
"volume": "",
"issue": "",
"pages": "4444--4451",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. ConceptNet 5.5: An Open Multilingual Graph of General Knowledge. In Proceedings of the Thirty- First AAAI Conference on Artificial Intelligence, AAAI'17, page 4444-4451. AAAI Press.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "A machine learning approach to pronoun resolution in spoken dialogue",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Strube",
"suffix": ""
},
{
"first": "Christoph",
"middle": [],
"last": "M\u00fcller",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "168--175",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Strube and Christoph M\u00fcller. 2003. A ma- chine learning approach to pronoun resolution in spoken dialogue. In Proceedings of the 41st Annual Meeting of the Association for Computational Lin- guistics, pages 168-175.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Structure and ostension in the interpretation of discourse deixis",
"authors": [
{
"first": "Bonnie",
"middle": [
"Lynn"
],
"last": "Webber",
"suffix": ""
}
],
"year": 1991,
"venue": "Language and Cognitive processes",
"volume": "6",
"issue": "2",
"pages": "107--135",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bonnie Lynn Webber. 1991. Structure and ostension in the interpretation of discourse deixis. Language and Cognitive processes, 6(2):107-135.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Team communication processing and process analytics for supporting robot-assisted emergency response",
"authors": [
{
"first": "Christian",
"middle": [],
"last": "Willms",
"suffix": ""
},
{
"first": "Constantin",
"middle": [],
"last": "Houy",
"suffix": ""
},
{
"first": "Jana-Rebecca",
"middle": [],
"last": "Rehse",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "Fettke",
"suffix": ""
},
{
"first": "Ivana",
"middle": [],
"last": "Kruijff-Korbayov\u00e1",
"suffix": ""
}
],
"year": 2019,
"venue": "IEEE SSRR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christian Willms, Constantin Houy, Jana-Rebecca Rehse, Peter Fettke, and Ivana Kruijff-Korbayov\u00e1. 2019. Team communication processing and process analytics for supporting robot-assisted emergency re- sponse. In IEEE SSRR.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Incremental neural coreference resolution in constant memory",
"authors": [
{
"first": "Patrick",
"middle": [],
"last": "Xia",
"suffix": ""
},
{
"first": "Jo\u00e3o",
"middle": [],
"last": "Sedoc",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Van Durme",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing",
"volume": "2020",
"issue": "",
"pages": "8617--8624",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.695"
]
},
"num": null,
"urls": [],
"raw_text": "Patrick Xia, Jo\u00e3o Sedoc, and Benjamin Van Durme. 2020. Incremental neural coreference resolution in constant memory. In Proceedings of the 2020 Con- ference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 8617-8624.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Revealing the myth of higher-order inference in coreference resolution",
"authors": [
{
"first": "Liyan",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Jinho",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "8527--8533",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Liyan Xu and Jinho D. Choi. 2020. Revealing the myth of higher-order inference in coreference resolution. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8527-8533. ACL.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"text": "Figure 1: Workspace Coreference System Overview",
"uris": null
},
"FIGREF1": {
"num": null,
"type_str": "figure",
"text": "Figure 2: Workspace Clustering Example",
"uris": null
},
"FIGREF2": {
"num": null,
"type_str": "figure",
"text": "M2M Noun Model Architecture and Training",
"uris": null
},
"FIGREF3": {
"num": null,
"type_str": "figure",
"text": "Discourse deixis dataset instance the dataset contains 5,171 examples.",
"uris": null
},
"FIGREF4": {
"num": null,
"type_str": "figure",
"text": "DDR scorer architecture",
"uris": null
},
"TABREF1": {
"html": null,
"content": "<table/>",
"text": "Systems summary AR system to be used in several projects of the Talking Robots Group at DFKI 2 , including for the interpretation of team communication in disaster response",
"num": null,
"type_str": "table"
},
"TABREF3": {
"html": null,
"content": "<table/>",
"text": "",
"num": null,
"type_str": "table"
},
"TABREF5": {
"html": null,
"content": "<table/>",
"text": "DDR F1 scores on fully annotated test sets",
"num": null,
"type_str": "table"
},
"TABREF6": {
"html": null,
"content": "<table/>",
"text": "CoNLL score: 65.09 62.79 66.28 64.82 66.43 66.59 66.96 68.08 68.13 70.09",
"num": null,
"type_str": "table"
},
"TABREF7": {
"html": null,
"content": "<table/>",
"text": "Average CoNLL scores for different combinations of embeddings in WCS",
"num": null,
"type_str": "table"
}
}
}
} |