File size: 118,321 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 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 | {
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:27:15.586832Z"
},
"title": "Comparison of methods for explicit discourse connective identification across various domains",
"authors": [
{
"first": "Merel",
"middle": [
"C J"
],
"last": "Scholman",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Language Science and Technology",
"location": {}
},
"email": "m.c.j.scholman@coli.uni-saarland.de"
},
{
"first": "Tianai",
"middle": [],
"last": "Dong",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Language Science and Technology",
"location": {}
},
"email": "tdong@coli.uni-saarland.de"
},
{
"first": "Frances",
"middle": [],
"last": "Yung",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Language Science and Technology",
"location": {}
},
"email": "frances@coli.uni-saarland.de"
},
{
"first": "Vera",
"middle": [],
"last": "Demberg",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Language Science and Technology",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Existing parse methods use varying approaches to identify explicit discourse connectives, but their performance has not been consistently evaluated in comparison to each other, nor have they been evaluated consistently on text other than newspaper articles. We here assess the performance on explicit connective identification of four parse methods (",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Existing parse methods use varying approaches to identify explicit discourse connectives, but their performance has not been consistently evaluated in comparison to each other, nor have they been evaluated consistently on text other than newspaper articles. We here assess the performance on explicit connective identification of four parse methods (",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Understanding the discourse relations that hold between segments in natural language text is crucial to many NLP applications, such as text generation, dialogue understanding, and question-answering systems. Shallow discourse parsers are used to uncover such relations by identifying connectives, extracting their arguments, and predicting the sense of the discourse relation. The current contribution focuses on the first step in the pipeline: finding explicit connectives in natural language. This step is not only interesting from the perspective of discourse relation classification, it can also be valuable for downstream applications, as accurate connective identification is crucial to mitigate the effect of cascaded errors downstream (Lin et al., 2014) .",
"cite_spans": [
{
"start": 743,
"end": 761,
"text": "(Lin et al., 2014)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Connective identification is not a trivial task, as some connectives are ambiguous and may not consistently function as discourse connectives. A simple dictionary lookup would therefore not be able to distinguish between the discourse connectiveusage of yet in Example (1), compared to the nonconnective usage in (2) and (3).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(1) Julie wants to buy a house. Yet she has not found the right one. (2) Julie wants to buy a house. She has yet to find the right one. (3) Julie wants to buy a luxurious, yet affordable house.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In 1, yet expresses an adversative relation (equivalent to nevertheless). In (2), it functions as an adverb expressing a temporal meaning (i.e. up until this time). Finally, in (3), yet does not function in a relation with two complete discourse arguments, and would therefore not be annotated as a connective. An accurate parser would need to be able to distinguish between these fine-grained differences in the usage of connective candidates.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Further, connective usage diverges between domains. For example, but and so are often used as discourse structuring markers in spoken language, rather than discourse connectives. Existing parsers, however, have mainly been evaluated on newspaper text, since the largest discourse-annotated corpus available comes from this domain (PDTB, Prasad et al., 2008) . Performance of existing discourse connective identification parsers on domains other than the written one is currently not well known.",
"cite_spans": [
{
"start": 330,
"end": 357,
"text": "(PDTB, Prasad et al., 2008)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Finally, we note that there is a lack of information on which connectives remain difficult to identify, even in state-of-the-art parsers. This is due to the tendency of studies reporting only the general accuracy, without providing detail on the accuracy on specific connectives. However, such information can provide the field with more information on what to focus on.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In sum, performance of different parsing methods have not been consistently evaluated in comparison to each other, nor have they been evaluated consistently on text other than newspaper articles. In addition, previous work tends to report only overall accuracy, whereas a more fine-grained analysis of connectives might prove to be valuable for future efforts. The current contribution fills these gaps in the literature.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Specifically, we evaluate the performance on explicit connective identification of four parse methods (PDTB e2e, Lin et al., 2014 ; the winner of CONLL2015, Wang and Lan, 2015; DisSent, Nie et al., 2019; and Discopy, Knaebel and Stede, 2020) , along with a simple heuristic as a baseline. The heuristic will identify all connectives in the data, without being able to distinguish discourse versus non-discourse usages of connectives. We include this \"parse method\" to provide insight into how discourse usage of connectives affects the performance of such a heuristic.",
"cite_spans": [
{
"start": 113,
"end": 129,
"text": "Lin et al., 2014",
"ref_id": "BIBREF14"
},
{
"start": 157,
"end": 176,
"text": "Wang and Lan, 2015;",
"ref_id": "BIBREF24"
},
{
"start": 177,
"end": 203,
"text": "DisSent, Nie et al., 2019;",
"ref_id": null
},
{
"start": 204,
"end": 241,
"text": "and Discopy, Knaebel and Stede, 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We also evaluate how well these systems generalize to different datasets from new domains, namely scientific text (BioDRB, Prasad et al., 2011) , prepared spoken text (TED-MDB, Zeyrek et al., 2019) and spontaneous spoken text (DiscoSPICE, Rehbein et al., 2016) .",
"cite_spans": [
{
"start": 114,
"end": 143,
"text": "(BioDRB, Prasad et al., 2011)",
"ref_id": null
},
{
"start": 177,
"end": 197,
"text": "Zeyrek et al., 2019)",
"ref_id": "BIBREF26"
},
{
"start": 226,
"end": 260,
"text": "(DiscoSPICE, Rehbein et al., 2016)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Connective identification In recent years, shallow discourse parsing has received notable attention in the field. Much of this work has focused on the automatic labeling of implicit connectives, as performance there is significantly lacking (the current state-of-the-art achieves an F1 around 64% on a four-way classification, e.g. Ji et al., 2016; Lan et al., 2017; Shi and Demberg, 2019) . Explicit relation identification has received less attention since prior work has reported high accuracy on explicit connectives in news articles. An open question is how these parsers perform on out-of-domain data. Pitler and Nenkova (2009) 's work on explicit connective identification formed the basis for much subsequent research efforts in this direction. They show that syntactic features provide highly useful information for predicting whether a connective candidate functions as a discourse connective (an accuracy of 94.2% F1 on PDTB2 with a 10-fold cross validation on Sec. 02-22). Lin et al. (2014) built on this work to develop the PDTB end-to-end (PDTB e2e) discourse parser. Their parser per-forms at 95.4% F1 on the PDTB with a 10-fold cross validation on Sec. 02-22. In the context of the CONLL2015 shared task, Wang and Lan (2015) built on both of these approaches and presented the top-ranked system, which achieved an F1 score of 94.2% on connective identification in the PDTB2 section 23 test set and 91.9% in the CONLL 2015 blind test set. All three models rely heavily on various combinations of syntactic and lexical features extracted from texts.",
"cite_spans": [
{
"start": 332,
"end": 348,
"text": "Ji et al., 2016;",
"ref_id": "BIBREF7"
},
{
"start": 349,
"end": 366,
"text": "Lan et al., 2017;",
"ref_id": "BIBREF13"
},
{
"start": 367,
"end": 389,
"text": "Shi and Demberg, 2019)",
"ref_id": "BIBREF23"
},
{
"start": 608,
"end": 633,
"text": "Pitler and Nenkova (2009)",
"ref_id": "BIBREF17"
},
{
"start": 985,
"end": 1002,
"text": "Lin et al. (2014)",
"ref_id": "BIBREF14"
},
{
"start": 1221,
"end": 1240,
"text": "Wang and Lan (2015)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "More recent models have taken different approaches. DisSent uses dependency parsing and sentence embeddings to annotate discourse relations (Nie et al., 2019) . They report an accuracy of 87.9% F1 on PDTB relations in determining whether a connective is present. Notably, they used sentences extracted from books as training data, rather than PDTB. This might make their approach more stable across domains. Finally, Knaebel and Stede (2020) use a neural approach that integrates contextualized word embeddings and predicts whether a connective candidate is part of a discourse relation or not. They achieve an F1 score of 97% on the PDTB2 section 23 test set.",
"cite_spans": [
{
"start": 140,
"end": 158,
"text": "(Nie et al., 2019)",
"ref_id": "BIBREF16"
},
{
"start": 417,
"end": 441,
"text": "Knaebel and Stede (2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "In the current paper we focus on comparing the parsers' performance on two written corpora with the performance on two spoken corpora. Spoken data differs from written in a number of ways, including shorter sentence length on average and a higher rate of elliptical structures and omissions. Moreover, discourse connectives are used differently across domains. For example, compared to written data, spoken data tends to have a higher rate of explicit connectives, fewer connective types, and more non-discourse connective usage of connectives (see, e.g. Crible and Cuenca, 2017; Rehbein et al., 2016) . It is also not uncommon for relations in spoken data to have an incomplete or even implicit sentence argument, as in Example (4), or to have connectives function as a discourse marker rather than connective, as in Example (5) (examples taken from DiscoSPICE):",
"cite_spans": [
{
"start": 555,
"end": 579,
"text": "Crible and Cuenca, 2017;",
"ref_id": "BIBREF2"
},
{
"start": 580,
"end": 601,
"text": "Rehbein et al., 2016)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Domain differences",
"sec_num": "2.1"
},
{
"text": "(4) And uhm, bring cos uh unfortunately just she's been up in Belfast this week. (5) And his face got really red. So I thought oh God. So yesterday when he he came back he said what were you saying to me about John.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Domain differences",
"sec_num": "2.1"
},
{
"text": "In Example (4), the second argument (Arg2) for the connective and is not fully uttered; instead, the speaker cut off her utterance after the verb bring and switched to a different discourse relation. In Example (5), the speaker uses so in a non-discourse connective usage (the utterance cannot be paraphrased as \"I thought oh God and as a result he asked me...\"). It is unclear how parsers developed for the written domain would handle such cases that are more typical of the spoken domain.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Domain differences",
"sec_num": "2.1"
},
{
"text": "Even within the written domain, differences in connective usage can occur between various types of written text. For example, Roman et al. 2016found a higher rate of discourse connectives used in science textbooks compared to social studies textbooks. Moreover, specific connectives can occur more in one domain than another; for example, in summary occurs more commonly in biomedical abstracts than in the general written domain (Gopalan and Devi, 2016) .",
"cite_spans": [
{
"start": 430,
"end": 454,
"text": "(Gopalan and Devi, 2016)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Domain differences",
"sec_num": "2.1"
},
{
"text": "Heuristic The heuristic is based on the list of 100 connectives from the PDTB2. This method simply extracts all connective candidates from the PDTB connective list, without distinguishing between usages. The heuristic will function as the baseline model. PDTB end-to-end The PDTB end-to-end model is trained using the PDTB dataset sections 02-21, evaluated using sec 22, and tested on sec 23. To distinguish a discourse connective from its nonconnective usage, Lin et al. (2014) extract a set of lexical and syntactic features for a connective and its preceding and following word. They also utilize the syntactic parse path from the connective to the root of the tree model, as well as the compressed path where adjacent identical tags are combined.",
"cite_spans": [
{
"start": 461,
"end": 478,
"text": "Lin et al. (2014)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parse methods",
"sec_num": "3.1"
},
{
"text": "Note that a version of this parser specifically aimed at parsing BioDRB has been made available. Here, we use the general version of the parser to be able to consistently evaluate its performance on out-of-domain text across datasets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parse methods",
"sec_num": "3.1"
},
{
"text": "CONLL2015 Similar to the PDTB e2e parser, the CONLL2015 winning parser is trained on PDTB sec 02-21, evaluated using sec 22, and tested on sec 23. Wang and Lan (2015) reimplemented well-established techniques from Pitler and Nenkova (2009) and Lin et al. (2014) , and added the POS tags of nodes from the connective's parent to capture more syntactic context information from the connective.",
"cite_spans": [
{
"start": 147,
"end": 166,
"text": "Wang and Lan (2015)",
"ref_id": "BIBREF24"
},
{
"start": 214,
"end": 239,
"text": "Pitler and Nenkova (2009)",
"ref_id": "BIBREF17"
},
{
"start": 244,
"end": 261,
"text": "Lin et al. (2014)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parse methods",
"sec_num": "3.1"
},
{
"text": "As part of the CONLL2015 shared task, competitors had access to the dependency tree, which we do not have access to. Instead, we generated parse trees using Stanford's CoreNLP Natural Language Processing Toolkit (the same parser that was used by the e2e parser). We note that the parse tree result after running Stanford CoreNLP might be different from the one that is given by CONLL2015, which in turn might affect performance of this model. DisSent In this model, connectives are used in the downstream task for learning sentence representation from explicit discourse relations. Nie et al. (2019) used texts from a BookCorpus (Zhu et al., 2015) to train and test their models. They identified common connectives in these texts, choosing those with a frequency greater than 1% in PDTB. As a result, the list of connectives that their parser can identify is much smaller than that used by other parsers (18 connectives compared to approximately 100 used by other parsers). Moreover, the possible dependency patters that are defined for each of these connective types is also restricted, and so this parser runs the risk of missing instances of the 18 connective types that it does aim to identify. We prepared the input for this parser using Stanford's CoreNLP dependency parser. Discopy The final model included is a recent neural model proposed by Knaebel and Stede (2020) , which achieves state-of-the-art results in connective identification. Instead of using various combinations of syntactic or lexical features, the method implements a neural model which relies on pretrained word embeddings. The model concatenates contextualized embeddings of connectives with the embeddings of their contexts. For multi-word connectives, the contextualized embeddings of the single words are averaged. The model is then trained in a multi-task setting, to predict the connective or predict the coherence relation. They used PDTB sec 02-22 for training and sec 23-24 for testing. Here, we assess their best performing bert-based model on connective identification task, with context size of 1.",
"cite_spans": [
{
"start": 582,
"end": 599,
"text": "Nie et al. (2019)",
"ref_id": "BIBREF16"
},
{
"start": 629,
"end": 647,
"text": "(Zhu et al., 2015)",
"ref_id": "BIBREF27"
},
{
"start": 1351,
"end": 1375,
"text": "Knaebel and Stede (2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parse methods",
"sec_num": "3.1"
},
{
"text": "Parser performance was measured on four different datasets in order to determine how well the parsers can identify connectives in various domains.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "3.2"
},
{
"text": "All datasets have already been annotated with discourse relations, and therefore we can use the gold connectives from these annotations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "3.2"
},
{
"text": "We evaluated all parsers on the Penn Discourse Treebank 2.0 (PDTB2, Prasad et al., 2008) . The PDTB consists of discourse annotations on the Wall Street Journal texts. We here evaluate performance on section 23, which is commonly used as test dataset. According to the gold label annotations, the dataset contains 923 explicit connective tokens, with 62 unique types.",
"cite_spans": [
{
"start": 68,
"end": 88,
"text": "Prasad et al., 2008)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "PDTB2 sec 23",
"sec_num": null
},
{
"text": "BioDRB We included text from the Biomedical Discourse Relation Bank (BioDRB; Prasad et al., 2011) . This corpus consists of discourse annotations of 24 biomedical research articles from the GENIA corpus, using an adapted version of the PDTB2 annotation framework. These texts represent the biomedical, scientific text genre.",
"cite_spans": [
{
"start": 77,
"end": 97,
"text": "Prasad et al., 2011)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "PDTB2 sec 23",
"sec_num": null
},
{
"text": "The BioDRB has 2636 gold explicit connective tokens with 180 unique connective types. The higher number of connective types in BioDRB compared to PDTB is mainly due to BioDRB having annotated modified connectives as unique types (e.g. 180 seconds after, due mainly to), and to BioDRB annotating post-modified connectives (because of), which PDTB2 does not annotate. To make the comparison consistent across datasets, we mapped the gold connectives in BioDRB to the corresponding connective heads annotated in PDTB. We removed connectives considered to be alternative lexicalizations in PDTB. The final dataset contains 2574 connective tokens with 134 unique connective types.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PDTB2 sec 23",
"sec_num": null
},
{
"text": "TED-MDB TED-MDB (Zeyrek et al., 2019 ) is a resource of TED talk transcripts manually annotated for discourse relations. TED talks are highly structured speeches that are often minutely prepared and are meant to provide targeted information on various topics or ideas. The resource is multilingual, but we focus only on English in the current contribution.",
"cite_spans": [
{
"start": 16,
"end": 36,
"text": "(Zeyrek et al., 2019",
"ref_id": "BIBREF26"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "PDTB2 sec 23",
"sec_num": null
},
{
"text": "TED-MDB currently consists of 6 TED talk transcripts annotated in PDTB3-style, with a total of 304 explicit connective tokens and 35 unique connective types. The reduced number of connective types compared to the PDTB can be attributed to the smaller size of this dataset as well as the genre. However, TED-MDB also includes connective types that are not included in PDTB2's connective list (some of which are part of PDTB3's connective list), such as at, by, in and through.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PDTB2 sec 23",
"sec_num": null
},
{
"text": "Disco-SPICE Disco-SPICE (Rehbein et al., 2016 ) is a corpus of transcribed broadcast interviews and telephone conversations from the SPICE-Ireland corpus (Kallen and Kirk, 2008) , annotated in PDTB3-style. These texts represent a more informal, spontaneous spoken genre than the TED talks. This dataset contains 1163 explicit connective tokens with 50 unique connective types. Again, the reduced number of connective types can be attributed to the domain. Table 1 presents the accuracy of each parser. The results show that Discopy is most accurate in identifying connectives in the PDTB, BioDRB and TED-MDB, but the e2e parser displays a higher F1 score in DiscoSPICE. Moreover, we find that, across the board, performance significantly drops for datasets other than the PDTB. This emphasizes the need for out-of-domain evaluation and development. Chi-squared tests confirm that the difference in performance between the parsers and between the datasets are significant, see Appendix A.",
"cite_spans": [
{
"start": 24,
"end": 45,
"text": "(Rehbein et al., 2016",
"ref_id": "BIBREF21"
},
{
"start": 154,
"end": 177,
"text": "(Kallen and Kirk, 2008)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 456,
"end": 463,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "PDTB2 sec 23",
"sec_num": null
},
{
"text": "Heuristic The heuristic uses PDTB's connective list as input to identify connectives, resulting in the highest recall on all datasets. 1 The heuristic's recall on BioDRB is lower than on other datasets because a portion of the connectives annotated in BioDRB are not included in PDTB's connective lexicon.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "Its F1 score on spoken data is relatively high compared to other parsers, particularly for Dis-coSPICE. This can be attributed to its insensitivity to syntactic requirements for Arg2 that are based on the written domain but are often impractical for the spoken domain. Hence, when syntactic features are too complex/inaccurate for spoken texts, connectives themselves can be used as reliable features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "Nevertheless, of all parse methods included, the heuristic's precision is lowest on all datasets. This reflects its high false positive rate, which is due to the heuristic not being able to distinguish between discourse versus non-discourse usage of connectives. Simply extracting all connectives using a heuristic might therefore be helpful for identifying connectives, but human input or additional computational input would still be necessary to then decide on whether the candidate connective functions as a discourse connective.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "PDTB e2e Performance of the PDTB e2e parser on the PDTB corpus is lower than reported in Lin et al. (2014) (91% versus 95% F1, respectively). This drop in the F1 score may be due to minor differences between our experimental set-up and theirs: we evaluated connective identification from final outputs of the end-to-end parser, while the original paper performed a separate training and testing on connective classifiers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "Applying the parser to datasets other than PDTB leads to a drop in performance. On BioDRB and TED-MDB, the e2e parser shows a high precision score but low recall score, which can be explained in part by the corpora having annotated a larger number of connective types than the PDTB2. Finally, we note low precision and recall scores on DiscoSPICE, with F1 scores on this spoken genre considerably lower than on the PDTB texts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "CONLL2015 Performance of the CONLL2015 winning parser on PDTB is lower than reported in the original paper (87% versus 94% F1, respectively, Wang and Lan, 2015) . This drop in performance is likely caused by the difference in dependency tree parser used to prepare the data.",
"cite_spans": [
{
"start": 141,
"end": 160,
"text": "Wang and Lan, 2015)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "Further, this parser shows a low recall score on BioDRB, which can again be explained by the difference in connective lists. The performance on TED-MDB is in line with PDTB e2e, but performance on DiscoSPICE is relatively low. This might be attributed to the syntactic structure of the texts: Wang and Lan (2015)'s new features can capture more syntactic features from texts, but this might result in lower accuracy when the parsed structure diverges from that of the trained texts. This indicates that complex lexical and syntactic features are too restrictive and therefore not appropriate for spontaneous spoken texts like those in DiscoSPICE.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "DisSent As mentioned in Section 3, DisSent targets only the 18 most frequently occurring connective types. Considering DisSent only takes a small portion of PDTB's connective types into account, it has the potential to uncover a relatively high number of connective tokens: a maximum of 86% of all gold connectives in PDTB, 43% in the BioDRB, 87% in TED-MDB, and 89% in Dis-coSPICE. However, the results show that it missed a significant portion of these connectives: it identified 26% of all gold connectives in PDTB, 11% in the BioDRB, 19% in TED-MDB, and 18% in Dis-coSPICE. Consequently, DisSent shows low recall scores in every corpus. In addition, it shows poor precision in BioDRB and TED-MDB specifically (but relatively high precision in DiscoSPICE compared to other parsers, which might be due to less diverse dependency patterns in DiscoSPICE). We conclude that DisSent can only be competitive if it is extended to include more connective types and syntactic patterns.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "Discopy Discopy outperforms the other parsers on the PDTB, BioDRB, and TED-MDB. This indicates that word embeddings are more flexible than hand-engineered features, and can perform well without domain knowledge. The outlier for this parser is DiscoSPICE: e2e outperforms Discopy on this dataset (47% versus 43% F1). It seems that none of the parsers can parse connectives in this dataset with high accuracy, which could be explained by the features of transcribed spontaneous spoken language that are very different from written language (such as fragments, disfluencies and interjections).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "The performance of the parser on PDTB is slightly lower than reported in the original paper, which could be a result of the implementation of the models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overall accuracy",
"sec_num": "4.1"
},
{
"text": "Detailed results on accuracy per connective for every dataset and parser are provided as an online appendix. 2 We here highlight some observations based on this data. We focus on results from the PDTB e2e, CONLL2015, and Discopy, since these provide the most coverage and behave similarly.",
"cite_spans": [
{
"start": 109,
"end": 110,
"text": "2",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Certain connectives were identified accurately in all corpora by the parsers. These include additionally, although, however, while, instead, meanwhile, nevertheless, therefore, unless, and whereas. Many of these tend to function consistently as discourse connectives (i.e. they do not occur frequently in a non-connective usage), which might explain why they are easier to identify.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "The remainder of this section presents a qualitative analysis, in which we consider in particular connectives that were included in the PDTB connective list, but were not identified correctly. We take the PDTB dataset as our reference model, and compare the other datasets to performance in this dataset. Most findings can be classified as an issue relating to the gold label, an issue relating to the parser, or an issue relating to token frequency. We end with general observations regarding DisSent. Table 4 in Appendix B presents a summary of the findings.",
"cite_spans": [],
"ref_spans": [
{
"start": 503,
"end": 510,
"text": "Table 4",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Performance on the PDTB dataset All three parsers show relatively poor accuracy on or, nor, once and previously. Moreover, all parsers miss some connective types altogether, such as finally, specifically, rather, hence (all parsers), earlier (e2e and CONLL), and consequently (CONLL and Discopy). Both of these issues are likely related to token frequency: these connectives all occur very infrequently in the test set and relatively infrequently in the training data. The poor performance might therefore be a reflection of the lack of training data; it remains a question whether the performance on these connectives will generalize if the test set would include more instances of them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Both PDTB e2e and CONLL2015 (but not Discopy) also show poor performance on as, which occurs more frequently in the PDTB test set (n=40). The e2e parser shows higher precision than recall, whereas the CONLL2015 parser shows lower precision than recall. Manual inspection of the instances of as that were not identified did not reveal any pattern indicating why they might have been missed. As we will see in the next paragraphs, as proved to be difficult to identify in the other datasets as well. The divergence between the parsers on this connective is particularly interesting.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Genre-specific findings: BioDRB versus PDTB Performance on the following connectives in Bio-DRB was relatively poor in all parsers: as, once, still, except, and after. Note that performance for as, once, and after was also comparatively lower in PDTB (albeit higher than in BioDRB). Difficulty with these connective types is likely due to their frequent usage as non-discourse connectives. Regarding still, the e2e parser seems to have a specific issue: it only identified those instances of still in BioDRB that occurred argument-initially.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Second, we find an issue with the BioDRB gold label for until: the gold BioDRB dataset contains 11 instances of until, but all three parsers only identified one instance. For the \"false negatives\", the Arg2 only contains a verb or noun phrase, as in (7). Such fragments are generally not considered to be full relational arguments and the connectives are therefore usually not annotated. It appears that BioDRB has more relaxed restrictions on what can constitute a relational argument. This can possibly explain the general trend of the parsers displaying relatively low recall compared to precision scores for BioDRB.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "(7) E14.5 fetal thymic lobes were collected and stored in the TRIzol (GIBCO BRL) at -70C until RNA isolation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "We also find an issue with the gold label for instances of also in BioDRB: all parsers perform well on this connective in the PDTB but in the BioDRB, they show low precision. The gold BioDRB lists 92 instances of also, PDTB e2e identified 183 instances, CONLL 126 instances and Discopy 156 instances. Consider Example (8), a true positive identified by all parsers, and Example (9), a \"false positive\" identified by e2e and Discopy. Both relations are very similar, and so it is unclear why one instance of also was part of the gold dataset and the other was not.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "(8) A much lower, but still significant increase was seen in fetal TN2 cells (25% increase, p < 0.01). Proliferation was also significantly higher at the fetal TN3 stage compared to adult (50% increase, p < 0.01). (9) In the peripheral blood of OX35-treated rats, the percentage of CD3+ cells was significantly lower than in PBS-treated animals. The percentage of CD4+ cells in the OX35-treated group was also significantly lower than that of the PBS control.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Similar to performance in the PDTB, the parsers show low performance on previously in BioDRB, which appears to be a difficult connective to identify accurately. In BioDRB, the errors can be attributed to both an issue with the gold label as well as an issue with the parsers. The gold contained no instances of previously, but the parsers all identified a different number of occurrences (3 to 8 instances). Some of these cases, such as Example (10), appear to be valid.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "(10) In agreement with this, western blot analysis demonstrated an upregulation of Id1 protein, while the amount of Id2 and Id3 protein levels remained unchanged. Previously, Id1 has been considered not to be expressed in later developmental stages than pro-B cells (...).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Genre-specific findings: spoken versus written Spoken data is characterised by certain frequent connectives displaying a higher rate of nonconnective usage compared to written data (e.g., but, so, and). As expected, these connectives show poorer performance in the spoken domain compared to the written domain. We also find lower performance on as, when and then in the spoken domain. Regarding as, we can see this is a consistently difficult connective candidate to identify across all datasets. The poor performance of the parsers on when in TED-MDB can be attributed to an issue with the gold labels. Certain instances of when were not included in the gold TED-MDB, but the parsers were accurate in identifying these \"false positives\". Example (11) presents an instance of when that was not part of the gold dataset but was identified accurately by the parsers. 11We thrive when we stay at our own leading edge.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Poor performance on then in DiscoSPICE can be attributed in part to missed instances in the gold dataset, but also to the common usage of then as a non-discourse connective in DiscoSPICE, as in Example (12). Such instances, where no clear arguments or relation sense can be identified for the connective, were not annotated in DiscoSPICE but the parsers did identify them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "(12) Yeah so, so hopefully just the three people are alright and they're not. Cos then like with eleven people you'd be assured to have a few good people there but three people you're just.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "We also note a peculiarity for if in TED-MDB. The parsers identified false positives for if in TED-MDB, an issue which can be attributed to annotation standards. The instances were in fact rhetorical \"what if\" relations (see Example 13), whereby the first argument could be taken to be what (or rather, what substitutes Arg1). Annotation frameworks would likely not consider such instances as true connectives, since the what cannot constitute a full relational argument according to most segmentation conventions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "(13) What if they used that firepower to allocate more of their capital to companies working the hardest at solving these challenges or at least not exacerbating them?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Finally, we note two issues specific to transcribed spoken data. First, some connectives are sometimes spelled differently in transcribed text than in normal written text, such as because being transcribed as cos. Such phonetic spelling variants lead to a higher rate of false negatives. Second, some false negatives in DiscoSPICE can be attributed to disfluencies, interjections or fillers such as uh and ehm. For example, both PDTB e2e and CONLL (but not Discopy) missed an instance of the connective after in DiscoSPICE (Example 14) because the connective is immediately followed by a filler, which affected the syntactic parse of the argument. Both of these issues might be solved by additional preprocessing of the spoken text (correcting spelling variants and removing disfluencies), but a better solution would be to develop a parser for spoken data that can handle such characteristics. 14Fintan rang me actually right after uhm I put down the phone to you.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Connective-specific findings for DisSent Of the 18 connectives that DisSent aims to identify, it shows poor performance in most datasets on four connectives in particular: and, so, as, and though. Furthermore, DisSent fails to identify any instances of but in both the PDTB and BioDRB. These results can be explained by DisSent's method:it does not identify connectives based on a heuristic search, but rather based on the syntactic pattern that the connective occurs in. The possible patterns that are provided for every connective are, in some cases, too restricted or coarse-grained, which is why the parser misses many instances of and, and all instances of but. This parser is hence extremely sensitive to the dependency parse of the dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analyses of specific connectives",
"sec_num": "4.2"
},
{
"text": "Explicit connective identification can be done relatively reliably by existing parsers, but gains can still be made in this area. We therefore aimed to evaluate existing parsers and uncover more finegrained errors. The results showed that Discopy (Knaebel and Stede, 2020) outperformed the other parsers in three out of four datasets: PDTB, Bio-DRB, and TED-MDB. This indicates that the contextualized embeddings used by Discopy are more flexible predictors of discourse connective usage than the syntactic and lexical features used by other parsers, even on out-of-domain data. The exception to this is DiscoSPICE, for which the PDTB e2e parser performed best. However, even e2e's performance on this dataset was not sufficient. DiscoSPICE contains features and syntactic patterns that are specific to spoken data, such as disfluencies, incomplete sentence structures, and increased ambiguity of connectives (e.g., whether so is used as a connective or marker). These features can explain the low performance of all parsers on this dataset. There is still room for improvement in this area.",
"cite_spans": [
{
"start": 247,
"end": 272,
"text": "(Knaebel and Stede, 2020)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "The performance of all parsers was lower on outof-domain text compared to PDTB. This reaffirms earlier findings, which showed that a connective identification classifier trained on PDTB does not perform well on BioDRB even with domain adaptation techniques, compared to a classifier trained on the BioDRB alone (Ramesh and Yu, 2010; Prasad et al., 2011) . Of course, these results cannot be considered surprising, given that prior work on discourse parsing has heavily focused on the written domain, with a strong bias towards newspaper text. Similar evaluations have not been done on other domains, nor have the parsers been applied frequently to other domains (but see Laali and Kosseim, 2014; Marchal et al., 2021 , for an application of the e2e parser to spoken translated data). This underlines the importance of evaluation of existing parsers on other domains and the need for domain adaption of connective identification models and classifiers.",
"cite_spans": [
{
"start": 311,
"end": 332,
"text": "(Ramesh and Yu, 2010;",
"ref_id": "BIBREF20"
},
{
"start": 333,
"end": 353,
"text": "Prasad et al., 2011)",
"ref_id": "BIBREF19"
},
{
"start": 671,
"end": 695,
"text": "Laali and Kosseim, 2014;",
"ref_id": "BIBREF12"
},
{
"start": 696,
"end": 716,
"text": "Marchal et al., 2021",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "The results further indicated that performance is affected by the connectives' usage as non-discourse connectives and the connectives' frequency. One solution, as suggested by Lin et al. (2014) , is to separately train a model for each highly ambiguous connective and another generic model to identify the remaining connectives. Another solution could be to provide parsers with more training data for infrequent connectives, which can be obtained via connective generation; this approach has recently been applied to address the lack of training data for implicit relations (Shi and Demberg, 2019; Kurfal\u0131 and \u00d6stling, 2021) .",
"cite_spans": [
{
"start": 176,
"end": 193,
"text": "Lin et al. (2014)",
"ref_id": "BIBREF14"
},
{
"start": 575,
"end": 598,
"text": "(Shi and Demberg, 2019;",
"ref_id": "BIBREF23"
},
{
"start": 599,
"end": 625,
"text": "Kurfal\u0131 and \u00d6stling, 2021)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "When analyzing connective-specific results, we found that poor performance could often be attributed to issues with the gold label. Some of the false positives seemed to be valid connectives, and might therefore actually be missed instances in the gold dataset. This highlights that in manual annotation, errors are still prevalent and maybe inevitable to some extent, which can affect performance of parsers. One way to remedy this is to support manual annotation with the output of SOTA parsers or even a simple heuristic, so that inconsistencies or false negatives would be less likely to occur.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "We also found divergences between what frameworks and parsers consider relational arguments: BioDRB identified instances of until with only a noun phrase as the second argument, which the parsers consistently did not consider a relational connective. Conversely, Discopy consistently identified instances of what if -relations, which the gold standard did not consider to be relations. Discourse segmentation therefore has an impact on connective identification as well. Unfortunately, the issue of segmentation is still not entirely resolved in the field of discourse relation annotation (see, e.g., Hoek et al., 2018) .",
"cite_spans": [
{
"start": 601,
"end": 619,
"text": "Hoek et al., 2018)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Many connectives annotated in the gold datasets could not be identified at all by the parsers because they rely on PDTB2's connective list. This list is not exhaustive and would need to be expanded for the parsers to provide more coverage. One place to start would be PDTB 3.0's connective list (Webber et al., 2019) . However, even this list will not cover all connective types that might also occur in other genres (e.g. hereafter occurs in BioDRB but not PDTB). Extending connective lists with a general lexicon of English connectives (Das et al., 2018) can provide new connective candidates as well.",
"cite_spans": [
{
"start": 295,
"end": 316,
"text": "(Webber et al., 2019)",
"ref_id": "BIBREF25"
},
{
"start": 538,
"end": 556,
"text": "(Das et al., 2018)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "With regards to running the models, we observed a trade-off between accuracy and computational cost: the training and testing of Discopy (a bertbased model) required much more computational energy than any traditional parser (Bender et al., 2021) . When comparing the success of neural methods and the other three parsers, it is important to be clear about the context in which they are used (Bender and Koller, 2020) .",
"cite_spans": [
{
"start": 225,
"end": 246,
"text": "(Bender et al., 2021)",
"ref_id": "BIBREF0"
},
{
"start": 392,
"end": 417,
"text": "(Bender and Koller, 2020)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Finally, we note that our results did not perfectly replicate those of the original authors. This is likely due to a difference in the experimental set-up used (e.g., the dependency parse method used by Wang and Lan (2015) was not publicly available) and a difference in evaluation methods. To ensure that the lack of replicability was not due to incorrect implementation, the second author and another, unrelated researcher independently implemented all parse methods, and neither were able to perfectly replicate the results. Note that Han et al. (2020) also obtained different results than the original when replicating the e2e parser (although they observed improved scores), and attributed these divergences to minor variations between experimental set-ups in terms of implementations, hyperparameter settings and/or choice of the type of F1 score reported on. These results emphasize the general need in the field for more transparent reporting and a more consistent approach to evaluation (see also Kim et al., 2020 , for implicit relation classification).",
"cite_spans": [
{
"start": 203,
"end": 222,
"text": "Wang and Lan (2015)",
"ref_id": "BIBREF24"
},
{
"start": 538,
"end": 555,
"text": "Han et al. (2020)",
"ref_id": "BIBREF5"
},
{
"start": 1006,
"end": 1022,
"text": "Kim et al., 2020",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "A comparison of different parse methods revealed that Discopy, a neural parser using sentence embeddings, generally outperforms parsers using syntactic and lexical features. The results also showed a severe performance drop when applying the parsers to other domains, especially spontaneous spoken discourse. This can be attributed to genrespecific syntactic structures, issues with the gold standards, and differences between connective lexicons. These results emphasise the need for out-ofdomain training and evaluation, and provide insight as to where gains can be made. exception of TED-MDB, as displayed in Table 2 . The top three parsers considered here because these were most competitive. Chi 2 tests confirm that the difference in performance (TP and FP distribution) between the datasets is significant in each parser, as shown in Table 3 . Rhetorical \"what if\" relations present false positives; New connective types not on PDTB2's list.",
"cite_spans": [
{
"start": 752,
"end": 776,
"text": "(TP and FP distribution)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 612,
"end": 619,
"text": "Table 2",
"ref_id": "TABREF2"
},
{
"start": 841,
"end": 848,
"text": "Table 3",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "and, as, but, so, then after, also, finally, for, later, otherwise, still then",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Disco-SPICE",
"sec_num": null
},
{
"text": "Phonetic spelling of connectives cannot be identified; Syntactic structures affect performance; PDTB e2e and CONLL are affected by interjections. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Disco-SPICE",
"sec_num": null
},
{
"text": "Recall on PDTB is not 1 because there was one instance of if...if...then, where the gold standard attributed a single instance of then to both if 's.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://osf.io/xzsa7/?view_only= 0cc5d937fa264fb59b5fe5e035c1ac75",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This research was funded by the German Research Foundation (DFG) as part of SFB 1102 \"Information Density and Linguistic Encoding\". We are grateful to Ida Novindasari and Anam Sadiq for their contributions to the implementations of the parsers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "On the dangers of stochastic parrots: Can language models be too big",
"authors": [
{
"first": "Emily",
"middle": [
"M"
],
"last": "Bender",
"suffix": ""
},
{
"first": "Timnit",
"middle": [],
"last": "Gebru",
"suffix": ""
},
{
"first": "Angelina",
"middle": [],
"last": "Mcmillan-Major",
"suffix": ""
},
{
"first": "Shmargaret",
"middle": [],
"last": "Shmitchell",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Emily M. Bender, Timnit Gebru, Angelina McMillan- Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? . Proceedings of the 2021 ACM Confer- ence on Fairness, Accountability, and Transparency.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Climbing towards NLU: On meaning, form, and understanding in the age of data",
"authors": [
{
"first": "Emily",
"middle": [
"M"
],
"last": "Bender",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Koller",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "5185--5198",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.463"
]
},
"num": null,
"urls": [],
"raw_text": "Emily M. Bender and Alexander Koller. 2020. Climb- ing towards NLU: On meaning, form, and under- standing in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, pages 5185-5198, Online. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Discourse markers in speech: characteristics and challenges for corpus annotation",
"authors": [
{
"first": "Ludivine",
"middle": [],
"last": "Crible",
"suffix": ""
},
{
"first": "Maria-Josep",
"middle": [],
"last": "Cuenca",
"suffix": ""
}
],
"year": 2017,
"venue": "Dialogue and Discourse",
"volume": "8",
"issue": "2",
"pages": "149--166",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ludivine Crible and Maria-Josep Cuenca. 2017. Dis- course markers in speech: characteristics and chal- lenges for corpus annotation. Dialogue and Dis- course, 8(2):149-166.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Constructing a lexicon of English discourse connectives",
"authors": [
{
"first": "Debopam",
"middle": [],
"last": "Das",
"suffix": ""
},
{
"first": "Tatjana",
"middle": [],
"last": "Scheffler",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "Bourgonje",
"suffix": ""
},
{
"first": "Manfred",
"middle": [],
"last": "Stede",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue",
"volume": "",
"issue": "",
"pages": "360--365",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Debopam Das, Tatjana Scheffler, Peter Bourgonje, and Manfred Stede. 2018. Constructing a lexicon of En- glish discourse connectives. In Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dia- logue, pages 360-365.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "BioDCA Identifier: A System for Automatic Identification of Discourse Connective and Arguments from Biomedical Text",
"authors": [
{
"first": "Sindhuja",
"middle": [],
"last": "Gopalan",
"suffix": ""
},
{
"first": "Devi",
"middle": [],
"last": "Sobha Lalitha",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the Fifth Workshop on Building and Evaluating Resources for Biomedical Text Mining (BioTxtM2016)",
"volume": "",
"issue": "",
"pages": "89--98",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sindhuja Gopalan and Sobha Lalitha Devi. 2016. BioDCA Identifier: A System for Automatic Iden- tification of Discourse Connective and Arguments from Biomedical Text. In Proceedings of the Fifth Workshop on Building and Evaluating Resources for Biomedical Text Mining (BioTxtM2016), pages 89- 98.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Rencontre des \u00c9tudiants Chercheurs en Informatique pour le Traitement Automatique des Langues (R\u00c9CITAL, 22e \u00e9dition)",
"authors": [
{
"first": "Kelvin",
"middle": [],
"last": "Han",
"suffix": ""
},
{
"first": "Phyllicia",
"middle": [],
"last": "Leavitt",
"suffix": ""
},
{
"first": "Srilakshmi",
"middle": [],
"last": "Balard",
"suffix": ""
}
],
"year": 2020,
"venue": "Actes de la 6e conf\u00e9rence conjointe Journ\u00e9es d'\u00c9tudes sur la Parole (JEP, 33e \u00e9dition), Traitement Automatique des Langues Naturelles (TALN, 27e \u00e9dition)",
"volume": "3",
"issue": "",
"pages": "123--136",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kelvin Han, Phyllicia Leavitt, and Srilakshmi Balard. 2020. Comparing PTB and UD information for PDTB discourseconnective identification. In Actes de la 6e conf\u00e9rence conjointe Journ\u00e9es d'\u00c9tudes sur la Parole (JEP, 33e \u00e9dition), Traitement Automa- tique des Langues Naturelles (TALN, 27e \u00e9dition), Rencontre des \u00c9tudiants Chercheurs en Informa- tique pour le Traitement Automatique des Langues (R\u00c9CITAL, 22e \u00e9dition). Volume 3: Rencontre des \u00c9tudiants Chercheurs en Informatique pour le TAL, pages 123-136.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Segmenting discourse: Incorporating interpretation into segmentation? Corpus Linguistics and Linguistic Theory",
"authors": [
{
"first": "Jet",
"middle": [],
"last": "Hoek",
"suffix": ""
},
{
"first": "Jacqueline",
"middle": [],
"last": "Evers-Vermeul",
"suffix": ""
},
{
"first": "Ted Jm",
"middle": [],
"last": "Sanders",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "14",
"issue": "",
"pages": "357--386",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jet Hoek, Jacqueline Evers-Vermeul, and Ted JM Sanders. 2018. Segmenting discourse: Incorporat- ing interpretation into segmentation? Corpus Lin- guistics and Linguistic Theory, 14(2):357-386.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "A latent variable recurrent neural network for discourse-driven language models",
"authors": [
{
"first": "Yangfeng",
"middle": [],
"last": "Ji",
"suffix": ""
},
{
"first": "Gholamreza",
"middle": [],
"last": "Haffari",
"suffix": ""
},
{
"first": "Jacob",
"middle": [],
"last": "Eisenstein",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "332--342",
"other_ids": {
"DOI": [
"10.18653/v1/N16-1037"
]
},
"num": null,
"urls": [],
"raw_text": "Yangfeng Ji, Gholamreza Haffari, and Jacob Eisenstein. 2016. A latent variable recurrent neural network for discourse-driven language models. In Proceedings of the 2016 Conference of the North American Chap- ter of the Association for Computational Linguis- tics: Human Language Technologies, pages 332- 342, San Diego, California. Association for Compu- tational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "ICE-Ireland: A User's Guide: Documentation to Accompany the Ireland Component of the International Corpus of English (ICE-Ireland). Cl\u00f3 Ollscoil na Banr\u00edona",
"authors": [
{
"first": "L",
"middle": [],
"last": "Jeffrey",
"suffix": ""
},
{
"first": "John Monfries",
"middle": [],
"last": "Kallen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kirk",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jeffrey L Kallen and John Monfries Kirk. 2008. ICE- Ireland: A User's Guide: Documentation to Accom- pany the Ireland Component of the International Corpus of English (ICE-Ireland). Cl\u00f3 Ollscoil na Banr\u00edona.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Implicit discourse relation classification: We need to talk about evaluation",
"authors": [
{
"first": "Najoung",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Song",
"middle": [],
"last": "Feng",
"suffix": ""
},
{
"first": "Chulaka",
"middle": [],
"last": "Gunasekara",
"suffix": ""
},
{
"first": "Luis",
"middle": [],
"last": "Lastras",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "5404--5414",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Najoung Kim, Song Feng, Chulaka Gunasekara, and Luis Lastras. 2020. Implicit discourse relation clas- sification: We need to talk about evaluation. In Pro- ceedings of the 58th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 5404- 5414. Association for Computational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Contextualized embeddings for connective disambiguation in shallow discourse parsing",
"authors": [
{
"first": "Ren\u00e9",
"middle": [],
"last": "Knaebel",
"suffix": ""
},
{
"first": "Manfred",
"middle": [],
"last": "Stede",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the First Workshop on Computational Approaches to Discourse",
"volume": "",
"issue": "",
"pages": "65--75",
"other_ids": {
"DOI": [
"10.18653/v1/2020.codi-1.7"
]
},
"num": null,
"urls": [],
"raw_text": "Ren\u00e9 Knaebel and Manfred Stede. 2020. Contextu- alized embeddings for connective disambiguation in shallow discourse parsing. In Proceedings of the First Workshop on Computational Approaches to Discourse, pages 65-75, Online. Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Let's be explicit about that: Distant supervision for implicit discourse relation classification via connective prediction",
"authors": [
{
"first": "Murathan",
"middle": [],
"last": "Kurfal\u0131",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "\u00d6stling",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2106.03192"
]
},
"num": null,
"urls": [],
"raw_text": "Murathan Kurfal\u0131 and Robert \u00d6stling. 2021. Let's be explicit about that: Distant supervision for implicit discourse relation classification via connective pre- diction. arXiv preprint arXiv:2106.03192.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Inducing discourse connectives from parallel texts",
"authors": [
{
"first": "Majid",
"middle": [],
"last": "Laali",
"suffix": ""
},
{
"first": "Leila",
"middle": [],
"last": "Kosseim",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers",
"volume": "",
"issue": "",
"pages": "610--619",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Majid Laali and Leila Kosseim. 2014. Inducing dis- course connectives from parallel texts. In Proceed- ings of COLING 2014, the 25th International Con- ference on Computational Linguistics: Technical Pa- pers, pages 610-619.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Multi-task attentionbased neural networks for implicit discourse relationship representation and identification",
"authors": [
{
"first": "Man",
"middle": [],
"last": "Lan",
"suffix": ""
},
{
"first": "Jianxiang",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Yuanbin",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Zheng-Yu",
"middle": [],
"last": "Niu",
"suffix": ""
},
{
"first": "Haifeng",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1299--1308",
"other_ids": {
"DOI": [
"10.18653/v1/D17-1134"
]
},
"num": null,
"urls": [],
"raw_text": "Man Lan, Jianxiang Wang, Yuanbin Wu, Zheng-Yu Niu, and Haifeng Wang. 2017. Multi-task attention- based neural networks for implicit discourse rela- tionship representation and identification. In Pro- ceedings of the 2017 Conference on Empirical Meth- ods in Natural Language Processing, pages 1299- 1308, Copenhagen, Denmark. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A PDTB-styled end-to-end discourse parser",
"authors": [
{
"first": "Ziheng",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Min-Yen",
"middle": [],
"last": "Hwee Tou Ng",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kan",
"suffix": ""
}
],
"year": 2014,
"venue": "Natural Language Engineering",
"volume": "20",
"issue": "2",
"pages": "151--184",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ziheng Lin, Hwee Tou Ng, and Min-Yen Kan. 2014. A PDTB-styled end-to-end discourse parser. Natural Language Engineering, 20(2):151-184.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Semi-automatic discourse annotation in a low-resource language: Developing a connective lexicon for Nigerian Pidgin",
"authors": [
{
"first": "Marian",
"middle": [],
"last": "Marchal",
"suffix": ""
},
{
"first": "C",
"middle": [
"J"
],
"last": "Merel",
"suffix": ""
},
{
"first": "Vera",
"middle": [],
"last": "Scholman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Demberg",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the Second Workshop on Computational Approaches to Discourse",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marian Marchal, Merel CJ Scholman, and Vera Dem- berg. 2021. Semi-automatic discourse annotation in a low-resource language: Developing a connective lexicon for Nigerian Pidgin. In Proceedings of the Second Workshop on Computational Approaches to Discourse.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "DisSent: Learning sentence representations from explicit discourse relations",
"authors": [
{
"first": "Allen",
"middle": [],
"last": "Nie",
"suffix": ""
},
{
"first": "Erin",
"middle": [],
"last": "Bennett",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "Goodman",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "4497--4510",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1442"
]
},
"num": null,
"urls": [],
"raw_text": "Allen Nie, Erin Bennett, and Noah Goodman. 2019. DisSent: Learning sentence representations from ex- plicit discourse relations. In Proceedings of the 57th Annual Meeting of the Association for Com- putational Linguistics, pages 4497-4510, Florence, Italy. Association for Computational Linguistics.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Using syntax to disambiguate explicit discourse connectives in text",
"authors": [
{
"first": "Emily",
"middle": [],
"last": "Pitler",
"suffix": ""
},
{
"first": "Ani",
"middle": [],
"last": "Nenkova",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the ACL-IJCNLP 2009 Conference Short Papers",
"volume": "",
"issue": "",
"pages": "13--16",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Emily Pitler and Ani Nenkova. 2009. Using syntax to disambiguate explicit discourse connectives in text. In Proceedings of the ACL-IJCNLP 2009 Confer- ence Short Papers, pages 13-16, Suntec, Singapore. Association for Computational Linguistics.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "The Penn Discourse TreeBank 2.0",
"authors": [
{
"first": "Rashmi",
"middle": [],
"last": "Prasad",
"suffix": ""
},
{
"first": "Nikhil",
"middle": [],
"last": "Dinesh",
"suffix": ""
},
{
"first": "Alan",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Eleni",
"middle": [],
"last": "Miltsakaki",
"suffix": ""
},
{
"first": "Livio",
"middle": [],
"last": "Robaldo",
"suffix": ""
},
{
"first": "Aravind",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Bonnie",
"middle": [],
"last": "Webber",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC'08)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt- sakaki, Livio Robaldo, Aravind Joshi, and Bon- nie Webber. 2008. The Penn Discourse TreeBank 2.0. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC'08), Marrakech, Morocco. European Lan- guage Resources Association (ELRA).",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "The biomedical discourse relation bank",
"authors": [
{
"first": "Rashmi",
"middle": [],
"last": "Prasad",
"suffix": ""
},
{
"first": "Susan",
"middle": [],
"last": "Mcroy",
"suffix": ""
},
{
"first": "Nadya",
"middle": [],
"last": "Frid",
"suffix": ""
},
{
"first": "Aravind",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Hong",
"middle": [],
"last": "Yu",
"suffix": ""
}
],
"year": 2011,
"venue": "BMC bioinformatics",
"volume": "12",
"issue": "1",
"pages": "1--18",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rashmi Prasad, Susan McRoy, Nadya Frid, Aravind Joshi, and Hong Yu. 2011. The biomedical dis- course relation bank. BMC bioinformatics, 12(1):1- 18.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Identifying discourse connectives in biomedical text",
"authors": [
{
"first": "Ramesh",
"middle": [],
"last": "Balaji Polepalli",
"suffix": ""
},
{
"first": "Hong",
"middle": [],
"last": "Yu",
"suffix": ""
}
],
"year": 2010,
"venue": "AMIA Annual Symposium Proceedings",
"volume": "2010",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Balaji Polepalli Ramesh and Hong Yu. 2010. Iden- tifying discourse connectives in biomedical text. In AMIA Annual Symposium Proceedings, volume 2010, page 657. American Medical Informatics As- sociation.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Annotating discourse relations in spoken language: A comparison of the PDTB and CCR frameworks",
"authors": [
{
"first": "Ines",
"middle": [],
"last": "Rehbein",
"suffix": ""
},
{
"first": "Merel",
"middle": [],
"last": "Scholman",
"suffix": ""
},
{
"first": "Vera",
"middle": [],
"last": "Demberg",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16)",
"volume": "",
"issue": "",
"pages": "1039--1046",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ines Rehbein, Merel Scholman, and Vera Demberg. 2016. Annotating discourse relations in spoken language: A comparison of the PDTB and CCR frameworks. In Proceedings of the Tenth Inter- national Conference on Language Resources and Evaluation (LREC'16), pages 1039-1046, Portoro\u017e, Slovenia. European Language Resources Associa- tion (ELRA).",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Linguistic cohesion in middle-school texts: A comparison of logical connectives usage in science and social studies textbooks",
"authors": [
{
"first": "Diego",
"middle": [],
"last": "Xavier Roman",
"suffix": ""
},
{
"first": "Allison",
"middle": [],
"last": "Brice\u00f1o",
"suffix": ""
},
{
"first": "Hannah",
"middle": [],
"last": "Rohde",
"suffix": ""
},
{
"first": "Stephanie",
"middle": [],
"last": "Hironaka",
"suffix": ""
}
],
"year": 2016,
"venue": "The Electronic Journal for Research in Science & Mathematics Education",
"volume": "",
"issue": "6",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diego Xavier Roman, Allison Brice\u00f1o, Hannah Ro- hde, and Stephanie Hironaka. 2016. Linguistic co- hesion in middle-school texts: A comparison of log- ical connectives usage in science and social studies textbooks. The Electronic Journal for Research in Science & Mathematics Education, 20(6).",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Learning to explicitate connectives with Seq2Seq network for implicit discourse relation classification",
"authors": [
{
"first": "Wei",
"middle": [],
"last": "Shi",
"suffix": ""
},
{
"first": "Vera",
"middle": [],
"last": "Demberg",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 13th International Conference on Computational Semantics -Long Papers",
"volume": "",
"issue": "",
"pages": "188--199",
"other_ids": {
"DOI": [
"10.18653/v1/W19-0416"
]
},
"num": null,
"urls": [],
"raw_text": "Wei Shi and Vera Demberg. 2019. Learning to explic- itate connectives with Seq2Seq network for implicit discourse relation classification. In Proceedings of the 13th International Conference on Computational Semantics -Long Papers, pages 188-199, Gothen- burg, Sweden. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "A refined endto-end discourse parser",
"authors": [
{
"first": "Jianxiang",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Man",
"middle": [],
"last": "Lan",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the Nineteenth Conference on Computational Natural Language Learning -Shared Task",
"volume": "",
"issue": "",
"pages": "17--24",
"other_ids": {
"DOI": [
"10.18653/v1/K15-2002"
]
},
"num": null,
"urls": [],
"raw_text": "Jianxiang Wang and Man Lan. 2015. A refined end- to-end discourse parser. In Proceedings of the Nine- teenth Conference on Computational Natural Lan- guage Learning -Shared Task, pages 17-24, Beijing, China. Association for Computational Linguistics.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "The Penn Discourse Treebank 3.0 annotation manual",
"authors": [
{
"first": "Bonnie",
"middle": [],
"last": "Webber",
"suffix": ""
},
{
"first": "Rashmi",
"middle": [],
"last": "Prasad",
"suffix": ""
},
{
"first": "Alan",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Aravind",
"middle": [],
"last": "Joshi",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bonnie Webber, Rashmi Prasad, Alan Lee, and Ar- avind Joshi. 2019. The Penn Discourse Treebank 3.0 annotation manual. Philadelphia, University of Pennsylvania.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Ted multilingual discourse bank (tedmdb): a parallel corpus annotated in the pdtb style",
"authors": [
{
"first": "Deniz",
"middle": [],
"last": "Zeyrek",
"suffix": ""
},
{
"first": "Am\u00e1lia",
"middle": [],
"last": "Mendes",
"suffix": ""
},
{
"first": "Yulia",
"middle": [],
"last": "Grishina",
"suffix": ""
},
{
"first": "Murathan",
"middle": [],
"last": "Kurfal\u0131",
"suffix": ""
},
{
"first": "Samuel",
"middle": [],
"last": "Gibbon",
"suffix": ""
},
{
"first": "Maciej",
"middle": [],
"last": "Ogrodniczuk",
"suffix": ""
}
],
"year": 2019,
"venue": "Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "1--27",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Deniz Zeyrek, Am\u00e1lia Mendes, Yulia Grishina, Mu- rathan Kurfal\u0131, Samuel Gibbon, and Maciej Ogrod- niczuk. 2019. Ted multilingual discourse bank (ted- mdb): a parallel corpus annotated in the pdtb style. Language Resources and Evaluation, pages 1-27.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Aligning books and movies: Towards story-like visual explanations by watching movies and reading books",
"authors": [
{
"first": "Yukun",
"middle": [],
"last": "Zhu",
"suffix": ""
},
{
"first": "Ryan",
"middle": [],
"last": "Kiros",
"suffix": ""
},
{
"first": "Rich",
"middle": [],
"last": "Zemel",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "Raquel",
"middle": [],
"last": "Urtasun",
"suffix": ""
},
{
"first": "Antonio",
"middle": [],
"last": "Torralba",
"suffix": ""
},
{
"first": "Sanja",
"middle": [],
"last": "Fidler",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the IEEE international conference on computer vision",
"volume": "",
"issue": "",
"pages": "19--27",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhut- dinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE inter- national conference on computer vision, pages 19- 27.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "A Model comparisons",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A Model comparisons",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Chi 2 tests confirm that the difference in performance (TP and FN distribution) between the top three parsers is significant in each dataset",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chi 2 tests confirm that the difference in perfor- mance (TP and FN distribution) between the top three parsers is significant in each dataset, with the",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"num": null,
"html": null,
"text": ". Table 1: Precision, Recall and F1 per corpus and parse method.",
"type_str": "table",
"content": "<table><tr><td/><td colspan=\"2\">Heuristic</td><td colspan=\"2\">PDTB e2e</td><td colspan=\"2\">CONLL2015</td><td/><td>DisSent</td><td/><td>Discopy</td></tr><tr><td/><td>P</td><td>R F1</td><td>P</td><td>R F1</td><td>P</td><td>R F1</td><td>P</td><td>R F1</td><td>P</td><td>R F1</td></tr><tr><td>PDTB</td><td>.29</td><td/><td/><td/><td/><td/><td/><td/><td/></tr></table>"
},
"TABREF2": {
"num": null,
"html": null,
"text": "Chi 2 statistics to test whether there is a significant difference between the parsers per dataset.",
"type_str": "table",
"content": "<table><tr><td>Parser</td><td colspan=\"2\">\u03c7 2 df</td></tr><tr><td>e2e</td><td>879.32</td><td>3 ***</td></tr><tr><td>CONLL</td><td>1005</td><td>3 ***</td></tr><tr><td colspan=\"2\">Discopy 1204.8</td><td>3 ***</td></tr></table>"
},
"TABREF3": {
"num": null,
"html": null,
"text": "Chi 2 statistics to test whether there is a significant difference between the datasets per parser.",
"type_str": "table",
"content": "<table/>"
},
"TABREF4": {
"num": null,
"html": null,
"text": "presents the highlights of the connectivespecific results per dataset.",
"type_str": "table",
"content": "<table><tr><td/><td>Poor</td><td>perfor-</td><td colspan=\"2\">Unidentified con-</td><td>Issues with</td><td>Other observations</td></tr><tr><td/><td>mance</td><td/><td>nectives</td><td/><td>gold label</td></tr><tr><td>PDTB2</td><td colspan=\"2\">as, or, nor, once,</td><td>finally,</td><td>specif-</td><td>-</td><td>PDTB e2e and CONLL2015 show</td></tr><tr><td/><td colspan=\"2\">previously</td><td>ically,</td><td>rather,</td><td/><td>low performance on as, but Discopy</td></tr><tr><td/><td/><td/><td>hence</td><td/><td/><td>performs better.</td></tr><tr><td>BioDRB</td><td colspan=\"2\">after, as, except,</td><td>besides</td><td/><td>until, also,</td><td>BioDRB maintains different seg-</td></tr><tr><td/><td colspan=\"2\">once, still</td><td/><td/><td>previously</td><td>mentation rules; New connective</td></tr><tr><td/><td/><td/><td/><td/><td/><td>types not on PDTB2's list; PDTB</td></tr><tr><td/><td/><td/><td/><td/><td/><td>e2e parser only identifies argument-</td></tr><tr><td/><td/><td/><td/><td/><td/><td>initial instances of still.</td></tr><tr><td>TED-</td><td colspan=\"2\">and, as, so,</td><td colspan=\"2\">for, on the one</td><td>when</td></tr><tr><td>MDB</td><td colspan=\"2\">then, when</td><td colspan=\"2\">hand, rather</td><td/></tr></table>"
},
"TABREF5": {
"num": null,
"html": null,
"text": "Summary of connective-specific analysis per dataset.",
"type_str": "table",
"content": "<table/>"
}
}
}
} |