File size: 95,617 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 |
{
"paper_id": "I17-1027",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:37:49.048510Z"
},
"title": "Coordination Boundary Identification with Similarity and Replaceability",
"authors": [
{
"first": "Hiroki",
"middle": [],
"last": "Teranishi",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Nara Institute of Science and Technology",
"location": {}
},
"email": "teranishi.hiroki.sw5@is.naist.jp"
},
{
"first": "Hiroyuki",
"middle": [],
"last": "Shindo",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Nara Institute of Science and Technology",
"location": {}
},
"email": "shindo@is.naist.jp"
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Nara Institute of Science and Technology",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We propose a neural network model for coordination boundary detection. Our method relies on two common properties-similarity and replaceability in conjuncts-in order to detect both similar and dissimilar pairs of conjuncts. The model improves the identification of clause-level coordination using bidirectional recurrent neural networks incorporating two properties as features. We show that our model outperforms existing stateof-the-art methods for the coordination annotated Penn Treebank and Genia corpus without any syntactic information from parsers.",
"pdf_parse": {
"paper_id": "I17-1027",
"_pdf_hash": "",
"abstract": [
{
"text": "We propose a neural network model for coordination boundary detection. Our method relies on two common properties-similarity and replaceability in conjuncts-in order to detect both similar and dissimilar pairs of conjuncts. The model improves the identification of clause-level coordination using bidirectional recurrent neural networks incorporating two properties as features. We show that our model outperforms existing stateof-the-art methods for the coordination annotated Penn Treebank and Genia corpus without any syntactic information from parsers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Coordination is a common structure and one of major ambiguities in human languages. Although coordination gives a large amount of syntactic or semantic information of coordinated phrases, disambiguating coordination still remains one of the difficult problems that state-of-the-art parsers cannot cope with.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Given a coordinator word, how can we find conjuncts? Coordinate structures are characterized by two properties: (1) similar structures often appear in conjuncts, and (2) one conjunct can be replaced with another conjunct without losing sentence consistency in syntax or semantics. However, many previous studies of coordination disambiguation rely only on the similarities between conjuncts, despite the fact that similarities are not always helpful (Shimbo and Hara, 2007; Hara et al., 2009; Hanamoto, 2012) . For example, the sentence \"[at least two commercial versions have been put on the U.S. market], and [an estimated 500 have been sold].\" does not have sim-ilar phrases around the coordinating conjunction \"and.\" Thus, existing methods sometimes fail to capture coordination. In addition to the case where there is a lack of similarities, many similaritybased methods use handcrafted features, heuristic rules, or external resources such as thesauri.",
"cite_spans": [
{
"start": 450,
"end": 473,
"text": "(Shimbo and Hara, 2007;",
"ref_id": "BIBREF15"
},
{
"start": 474,
"end": 492,
"text": "Hara et al., 2009;",
"ref_id": "BIBREF4"
},
{
"start": 493,
"end": 508,
"text": "Hanamoto, 2012)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To overcome these problems, Ficler and Goldberg (2016) proposed a neural network model with the replaceability feature as well as the similarity feature. Their model produces candidate pairs of conjuncts using probabilities assigned by the Berkeley Parser. All candidate pairs are scored on the basis of the similarity, replaceability and parser-derived features, and then the best scored pair is picked. Their approach outperforms existing constituent parsers for the Penn Treebank and similarity-based coordination disambiguation methods such as those by Shimbo and Hara (2007) and Hara et al. (2009) for the Genia treebank. Although Ficler and Goldberg's (2016) method improves performance significantly, it heavily depends on the syntactic parser. They use the outputs from the parser not only for candidates generation and the feature for scoring, but also for computation of the similarities. The problems of propagated errors from the parser and dependencies on external resources still remain in their work.",
"cite_spans": [
{
"start": 28,
"end": 54,
"text": "Ficler and Goldberg (2016)",
"ref_id": "BIBREF2"
},
{
"start": 557,
"end": 579,
"text": "Shimbo and Hara (2007)",
"ref_id": "BIBREF15"
},
{
"start": 584,
"end": 602,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 636,
"end": 664,
"text": "Ficler and Goldberg's (2016)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work, we propose a neural network model for coordination disambiguation that does not require any external syntactic parser. Our model exploits both the similarity and replaceability properties to avoid suffering from an absence of these properties (Section 2). We use bidirectional recurrent neural networks (RNNs) to obtain the contextual information of candidate conjuncts and then compute similarity and replaceability features without syntactic information (Section 3). We show that our model performs well for both types of coordination: NP coordination (whose conjuncts tend to be similar) and S coordination (whose conjuncts make sense individually) and outperforms the methods by Ficler and Goldberg (2016) and Hara et al. (2009) in Section 4.",
"cite_spans": [
{
"start": 697,
"end": 723,
"text": "Ficler and Goldberg (2016)",
"ref_id": "BIBREF2"
},
{
"start": 728,
"end": 746,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The contributions of our work include the following:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(i) Our model can capture dissimilar conjuncts as well as similar ones using the similarity and replaceability features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(ii) Our model performs better than others without any thesauri, feature engineering, or syntactic parsers to extract conjunct features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "2 Coordinate Structure Analysis",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Coordination is a frequently occurring syntactic structure along with several phrases, known as conjuncts. The task of coordination disambiguation is identifying the boundaries of each conjunct with a single coordinator word as one coordinate structure instance. Given a coordinator word (e.g., \"and,\" \"or,\" or \"but\"), a system must return each conjunct span if the word actually plays the role of a coordinator; otherwise, NONE is output for the absence of coordination. The task sounds simple, yet is difficult because two complex phenomena appear in coordination. In this work, we solve this task by focusing on identifying the beginning and end of an entire coordinate structure. Figure 1 shows our task. We attempt to identify two conjuncts to the left and right sides of a conjunction. We call these conjuncts the preconjunct and post-conjunct, respectively 2 . In addition, we assume that the end of the preconjunct and the beginning of the post-conjunct adjoin a coordinator word; thus it appears that we work on the subtask of coordinate structure span identification. After identifying a coordination span, we recover individual conjuncts within the span.",
"cite_spans": [],
"ref_spans": [
{
"start": 684,
"end": 692,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Task Description",
"sec_num": "2.1"
},
{
"text": "Coordination has many unique traits other than its structure. We focus on the key properties between conjuncts that can be helpful to disambiguate a coordination boundary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conjunct Properties",
"sec_num": "2.2"
},
{
"text": "(a) Similarity: Conjuncts in a coordination have a similar structure or meaning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conjunct Properties",
"sec_num": "2.2"
},
{
"text": "(b) Replaceability: A conjunct can be replaced with another conjunct in the same coordination.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conjunct Properties",
"sec_num": "2.2"
},
{
"text": "Conjuncts tend to have similar semantic or syntactic constituents. For example, the three conjuncts \"the high level of performance,\" \"the compositional talents of MR. Douglas,\" and \"the obvious sincerity with which Mr. Stolzman chooses his selection\" have part-of-speech (POS) tag sequences starting with \"DT JJ NN(S) IN NN(P) Figure 2 (a) ). Many previous works exploit this characteristic to detect conjuncts (Shimbo and Hara, 2007; Hara et al., 2009) .",
"cite_spans": [
{
"start": 411,
"end": 434,
"text": "(Shimbo and Hara, 2007;",
"ref_id": "BIBREF15"
},
{
"start": 435,
"end": 453,
"text": "Hara et al., 2009)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 327,
"end": 339,
"text": "Figure 2 (a)",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Conjunct Properties",
"sec_num": "2.2"
},
{
"text": "The replaceability of conjuncts is also often observed. A sentence is still consistent even if one conjunct is replaced with another one. For example, the coordination \"Aside from [the Soviet economic plight] and [talks on cutting (strategic) and (chemical) arms]\" can be transformed into \"Aside from [talks on cutting (chemical) and (strategic) arms] and [the Soviet economic plight]\" by exchanging conjuncts. Using this property, we can expand a coordinate structure as one sentence by one conjunct (Figure 2 (b)). Replaceability has recently been used to capture conjuncts (Ficler and Goldberg, 2016 ).",
"cite_spans": [
{
"start": 576,
"end": 602,
"text": "(Ficler and Goldberg, 2016",
"ref_id": "BIBREF2"
}
],
"ref_spans": [
{
"start": 501,
"end": 510,
"text": "(Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Conjunct Properties",
"sec_num": "2.2"
},
{
"text": "The two properties described above are essential clues to identify conjunct spans; however, they are not always available. Coordination sometimes has different types of conjuncts or an ellipsis in conjuncts. For similarity, when conjuncts belong to the S type or are different types of syntactic categories, their semantic and syntactic structures can be apart from each other (e.g., \"[We turned the trading system on]S, and [it did whatever it was programmed to do]S.\" ; \"Bill is [in trouble]PP and [trying to come up with an excuse]VP.\"). For replaceability, when words are omitted in a latter conjunct, we cannot replace one conjunct with another unless we supplement omitted words (e.g., \"[Honeywell's contract totaled $69.7 million], and [IBM's $68.8 million].\"). To cope with the case where there is a lack of similarity or replaceability, our proposed method incorporates both features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conjunct Properties",
"sec_num": "2.2"
},
{
"text": "Our proposed model calculates the scores of all possible preconjunct and post-conjunct pairs. Given a sentence x = {x 1 , x 2 , x 3 , . . . , x N } and coordinator word x k , the preconjunct and postconjunct can be written as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": "s 1 = {x i , . . . , x k\u22121 } (1 \u2264 i \u2264 k \u2212 1) and s 2 = {x k+1 , . . . , x j } (k + 1 \u2264 j \u2264 N )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": ", respectively. As we mentioned in Section 2, we fix the end of the preconjunct at k \u2212 1 and the beginning of the post-conjunct at k + 1. Thus, our model learns and predicts a set of spans (i, j), which indicate the two positions of the beginning and end of a coordination. We identify preconjuncts and post-conjuncts by picking the highest scoring pairs as predicted conjunct spans. Figure 3 shows an overview of our neural network architecture. This model consists of four components.",
"cite_spans": [],
"ref_spans": [
{
"start": 384,
"end": 392,
"text": "Figure 3",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": "Input Layer: Map a sequence of one-hot words and POS tags onto their representations from embeddings.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": "RNN Layer: Produce a sequence of sentencelevel representations based on contexts using a bidirectional RNN.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": "Feature Extractor: Generate the conjunct phrase representations and feature vectors of possible pairs of conjuncts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": "Output Layer: Calculate the scores of pairs of conjuncts using MLP.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": "In the following subsections, we describe these components in detail.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Method",
"sec_num": "3"
},
{
"text": "The first step of our neural network model is to represent a sequence of words and POS tags in distributed vectors, known as embeddings (Bengio et al., 2003) . Our model receives a sequence of one-hot encoded words and POS tags {x word n } N n=1 , {x tag n } N n=1 and then looks them up in the matrices ",
"cite_spans": [
{
"start": 136,
"end": 157,
"text": "(Bengio et al., 2003)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Input Layer",
"sec_num": "3.1"
},
{
"text": "E word \u2208 R d word \u00d7|v word | , E tag \u2208 R dtag\u00d7|vtag| , resulting in a sequence of real-valued vectors h word n \u2208 R d , h tag n \u2208 R d ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input Layer",
"sec_num": "3.1"
},
{
"text": "h word t = W word x word t h tag t = W tag x tag t h (0) t = [h word t ; h tag t ] h (0) = h (0) 1 , . . . , h (0) N (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input Layer",
"sec_num": "3.1"
},
{
"text": "A sequence of distributed vectors is transformed into hidden state vectors using stacked bidirectional RNNs. Bidirectional RNNs process a time series of inputs from the past to a future direction and from the future to a past direction. We can make use of left-to-right (forward) and rightto-left (backward) contexts using these networks. The output of the -th layer of stacked bidirectional RNNs at a time step t in the forward direction, which is denoted as h f ,t , is computed as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "RNN Layer",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "h f ,t = f (h f ,t\u22121 , h \u22121,t )",
"eq_num": "(2)"
}
],
"section": "RNN Layer",
"sec_num": "3.2"
},
{
"text": "where h f ,t\u22121 is the hidden state vector of the same layer at the previous time step t \u2212 1 in the same direction and h \u22121,t is the hidden state vector of the previous bidirectional layer at the same time step t. The hidden vector of the -th layer of stacked bidirectional RNNs at a time step t in the backward direction is also computed in the same way. The stacked bidirectional RNNs that we use in this work output hidden state vectors by concatenating the vectors {h f ,t } T t=1 from the forward direction and {h b ,t } T t=1 from the backward direction at each time step t in every layer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "RNN Layer",
"sec_num": "3.2"
},
{
"text": "In general, an RNN has a function f (\u2022) expressed as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "RNN Layer",
"sec_num": "3.2"
},
{
"text": "f (x t , h t\u22121 ) = g(W x t + U h t\u22121 )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "RNN Layer",
"sec_num": "3.2"
},
{
"text": "where g(\u2022) is an arbitrary nonlinear function such as the hyperbolic tangent tanh(\u2022) or rectified linear unit (ReLU). We use the long short term memory (LSTM) (Hochreiter and Schmidhuber, 1997) as the function f (\u2022) to prevent backpropagated errors from vanishing or exploding, which arise in RNNs (Pascanu et al., 2013) .",
"cite_spans": [
{
"start": 159,
"end": 193,
"text": "(Hochreiter and Schmidhuber, 1997)",
"ref_id": "BIBREF6"
},
{
"start": 298,
"end": 320,
"text": "(Pascanu et al., 2013)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "RNN Layer",
"sec_num": "3.2"
},
{
"text": "This component produces a feature vector based on a representation of a preconjunct and postconjunct and a series of vectors {h t } T t=1 from bidirectional RNNs. We compute the preconjunct representation v pre i and post-conjunct v post j using the function g(\u2022). In this work, we define elementwise averaging as the function g(",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extractor",
"sec_num": "3.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u2022). g(h l:m ) = average h l , h l+1 , . . . , h m\u22121 , h m (3) Thus, v pre i and v post j are expressed as v pre i = g(h i:k\u22121 ) (1 \u2264 i \u2264 k \u2212 1) v post j = g(h k+1:j ) (k + 1 \u2264 j \u2264 N )",
"eq_num": "(4)"
}
],
"section": "Feature Extractor",
"sec_num": "3.3"
},
{
"text": "Then v pre i and v post j are fed into the following two feature extraction functions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extractor",
"sec_num": "3.3"
},
{
"text": "In order to capture the similarity between the preconjunct and the post-conjunct, the feature vector is computed as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity feature vector",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "f sim (v pre i , v post j ) = |v pre i \u2212 v post j |; v pre i v post j",
"eq_num": "(5)"
}
],
"section": "Similarity feature vector",
"sec_num": null
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity feature vector",
"sec_num": null
},
{
"text": "|v pre i \u2212 v post j",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity feature vector",
"sec_num": null
},
{
"text": "| is the absolute value of element-wise subtraction, and v pre i v post j is element-wise multiplication. These subtraction and multiplication operations are intended to model the semantic distance and relatedness (Ji and Eisenstein, 2013; Tai et al., 2015; Hashimoto et al., 2016) .",
"cite_spans": [
{
"start": 214,
"end": 239,
"text": "(Ji and Eisenstein, 2013;",
"ref_id": "BIBREF7"
},
{
"start": 240,
"end": 257,
"text": "Tai et al., 2015;",
"ref_id": "BIBREF17"
},
{
"start": 258,
"end": 281,
"text": "Hashimoto et al., 2016)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity feature vector",
"sec_num": null
},
{
"text": "We define a feature vector based on the conjunct replaceability as follows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replaceability feature vector",
"sec_num": null
},
{
"text": "f repl (h 1:N , i, j, k) = |h i\u22121 h i \u2212 h i\u22121 h k+1 |; |h j h j+1 \u2212 h k\u22121 h j+1 | (6)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replaceability feature vector",
"sec_num": null
},
{
"text": "where h i\u22121 is the context vector that is linked to the heads of conjuncts and h j+1 is the context vector that is linked to the tails of conjuncts. The first subtraction |h i\u22121 h i \u2212h i\u22121 h k+1 | is the difference between two context-conjunct connections at the beginning of coordination. The second subtraction |h j h j+1 \u2212 h k\u22121 h j+1 | is the difference between two context-conjunct connections at the end of coordination. These distance measures can be interpreted as difficulty in replacing conjuncts. Note that the function f repl (h 1:N , i, j, k) returns a zero vector when i = 0 or j = N .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replaceability feature vector",
"sec_num": null
},
{
"text": "This layer computes the scores of pairs of conjuncts based on the similarity feature vectors and the replaceability feature vectors. The network is a multilayered perceptron (MLP) that consists of multiple layers of computational units interconnected in a feed-forward way. The score of a preconjunct (i, k \u2212 1) and post-conjunct (k + 1, j) candidate pair is calculated as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Output Layer",
"sec_num": "3.4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Score(i, j) = MLP f sim (v pre i , v post j ); f repl (h 1:N , i, j, k)",
"eq_num": "(7)"
}
],
"section": "Output Layer",
"sec_num": "3.4"
},
{
"text": "To cope with the absence of coordination against a coordinator, we also calculate the score for a candidate of NONE. The score NONE is simply computed as the product of a weight vector and the sentence-level representation of the coordinator from the RNN layer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Output Layer",
"sec_num": "3.4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Score(NONE) = w \u2022 h k + b",
"eq_num": "(8)"
}
],
"section": "Output Layer",
"sec_num": "3.4"
},
{
"text": "Using these scoring functions, we assign scores to all possible pairs of conjuncts. Thus, when the length of a sentence is N and a coordinator appears as the k-th word, we obtain (k \u2212 1) \u00d7 (N \u2212 k) + 1 candidates and choose the pair with the best score as the predicted conjuncts with the softmax function. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Output Layer",
"sec_num": "3.4"
},
{
"text": "The loss function is the negative log-likelihood of the true pair of conjuncts y (k) :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning",
"sec_num": "3.5"
},
{
"text": "J(\u03b8) = \u2212 D d=1 logp \u03b8 (y (d) |x (d) ) + \u03bb 2 \u03b8 2 (10)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning",
"sec_num": "3.5"
},
{
"text": "where D is the number of occurrences of coordinator words in a training dataset, \u03b8 is a set of model parameters, and the hyperparameter \u03bb adjusts the regularization strength. The model parameters are optimized by minimizing the loss using the stochastic gradient descent (SGD).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning",
"sec_num": "3.5"
},
{
"text": "We evaluate our proposed model using the coordination annotated Penn Treebank (Ficler, 2016) and the Genia treebank beta (Kim et al., 2003) . We present the number of occurrences of coordinator words and the number of sentences with coordination in We use the coordination annotated Penn Treebank and divide it into wsj 2-21 as the training set, wsj 22 as the development set, and wsj 23 as the testing set. We use pretrained 200-dimensional word embeddings from the New York Times section in English Gigaword (fifth edition) (Parker et al., 2011) using Word2Vec 4 with its default parameter. For the POS tags, we use 10-way jackknifing using the Stanford POS Tagger (Toutanova et al., 2003) and initialize the 50-dimensional embeddings with the uniform distribution within [\u22121, 1]. We use three-layer bidirectional LSTMs as an RNN layer. The dimensionality of the LSTM hidden vectors in each direction is selected from {400, 600}. Our MLP consists of one hidden layer with ReLU activation, and an output layer. The number of the hidden layer units is selected from {1200, 2400}. The model parameters are optimized by the minibatched SGD with a batch size of 20. The learning rate is automatically tuned by Adam (Kingma and Ba, 2014) . When training, we apply dropout (Srivastava et al., 2014) to the embeddings, input vectors of each LSTM in bidirectional LSTMs (except the first layer), and the hidden layer of the MLP. Dropout ratio is selected from {0.33, 0.50}. We choose the regularization strength \u03bb from {0.0001, 0.0005, 0.001}. We train our model for 50 iterations and choose the model that achieves the best F1 score 5 on the development set and evaluate it with the testing set. We present the final hyperparameters choice in Table 2 .",
"cite_spans": [
{
"start": 78,
"end": 92,
"text": "(Ficler, 2016)",
"ref_id": "BIBREF1"
},
{
"start": 121,
"end": 139,
"text": "(Kim et al., 2003)",
"ref_id": "BIBREF9"
},
{
"start": 526,
"end": 547,
"text": "(Parker et al., 2011)",
"ref_id": null
},
{
"start": 667,
"end": 691,
"text": "(Toutanova et al., 2003)",
"ref_id": "BIBREF18"
},
{
"start": 1212,
"end": 1233,
"text": "(Kingma and Ba, 2014)",
"ref_id": "BIBREF10"
},
{
"start": 1268,
"end": 1293,
"text": "(Srivastava et al., 2014)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [
{
"start": 1737,
"end": 1744,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "We evaluate our model on the basis of the ability to predict the beginning and end of each co-Parameter Value Dimension of the LSTM hidden vector 600 MLP units in the hidden layer 2400 Dropout ratio (all) 0.50 Regularization term \u03bb 0.0001 Table 2 : The final hyperparameters in the experiment for the Penn Treebank.",
"cite_spans": [],
"ref_spans": [
{
"start": 239,
"end": 246,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation Metrics",
"sec_num": "4.1.2"
},
{
"text": "ordination (whole) with the precision, recall, and F1 measures. In another setup, we focus on NP coordination 6 . To compare the performance with Ficler and Goldberg (2016) , we also evaluate our model with two conjunct spans that are adjacent to the coordinator (inner), the first and last conjuncts (outer), and all complete conjuncts (exact). Furthermore, in order to investigate the effectiveness of our proposed features, we perform the experiment with a simple baseline model that uses two averaged vectors as features (Eq. 3) and feeds them into the MLP instead of the similarity and replaceability features (Eq. 7).",
"cite_spans": [
{
"start": 146,
"end": 172,
"text": "Ficler and Goldberg (2016)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Metrics",
"sec_num": "4.1.2"
},
{
"text": "Note that our proposed model learns and predicts the coordinate structure boundaries and not each conjunct; thus, when evaluating the inner, outer, and exact metrics, we simply divide the preconjuncts into subconjuncts using the character \",\" as the divider.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Metrics",
"sec_num": "4.1.2"
},
{
"text": "We present the results in Table 3 . For all metrics, the recall values are low compared with the precision values. Our model is likely to produce NONE for some coordinators by mistake. The proposed model suffers from a worse outer metric than the inner metric. Intuitively, this is because the preconjunct for the inner prediction is placed next to a coordinator and it is easier to identify its span, while outer conjuncts occur apart from the coordinators. Table 4 summarizes the performance of different uses of features. The similarity and replaceability features work better than the baseline independently. However, the joint model performs the best by exploiting both features. Table 4 : Performance of different sets of features for the PTB development set for the outer metric. \"f sim ,\" \"f repl ,\" and \"Both\" indicate the use of similarity feature vectors, replaceability feature vectors, and both feature vectors, respectively.",
"cite_spans": [],
"ref_spans": [
{
"start": 26,
"end": 33,
"text": "Table 3",
"ref_id": "TABREF4"
},
{
"start": 459,
"end": 466,
"text": "Table 4",
"ref_id": null
},
{
"start": 685,
"end": 692,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "4.1.3"
},
{
"text": "the previously reported result). For NP coordination, our model achieves competitive results, despite the rough extraction of conjuncts from preconjuncts, even for inner-conjunct prediction.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "4.1.3"
},
{
"text": "We also evaluate our model with the Genia treebank beta to compare with the previous work of Hara et al. (2009) and Ficler and Goldberg (2016) . The settings of this experiment are based on those presented in Section 4.1.1, except for the following hyperparameters: Word embeddings are initialized by the pretrained 200-dimensional representation that BioASQ (Tsatsaronis et al., 2012) provides. These embeddings are trained from biomedical abstracts by using Word2Vec. We use gold POS as in Hara et al. (2009) , and the dimension of the POS embeddings is 50. For regularization, we set \u03bb = 0.0005 and train our model for 20 iterations.",
"cite_spans": [
{
"start": 93,
"end": 111,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 116,
"end": 142,
"text": "Ficler and Goldberg (2016)",
"ref_id": "BIBREF2"
},
{
"start": 359,
"end": 385,
"text": "(Tsatsaronis et al., 2012)",
"ref_id": "BIBREF19"
},
{
"start": 492,
"end": 510,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "4.2.1"
},
{
"text": "As in Hara et al. (2009) , we measure the recall values of coordinate structure boundary prediction, disregarding individual conjunct spans 7 . Thus, we do not decode conjuncts because our model can be compared directly. Coordination phrases in the Table 6 : Recall with Genia treebank beta. The numbers in the columns \"Ficler16\" and \"Hara09\" are taken from their papers; Ficler16 : (Ficler and Goldberg, 2016) ; Hara09 : (Hara et al., 2009) .",
"cite_spans": [
{
"start": 6,
"end": 24,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 383,
"end": 410,
"text": "(Ficler and Goldberg, 2016)",
"ref_id": "BIBREF2"
},
{
"start": 422,
"end": 441,
"text": "(Hara et al., 2009)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 249,
"end": 256,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation Metrics",
"sec_num": "4.2.2"
},
{
"text": "Genia treebank are explicitly annotated with a special label (COOD). Making use of this label, we also measure the performance for each type of coordination, as reported in previous work. We evaluate our model by five-fold cross-validation, as in Hara et al. (2009) .",
"cite_spans": [
{
"start": 247,
"end": 265,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Metrics",
"sec_num": "4.2.2"
},
{
"text": "We present the results in Table 6 . For all coordination, our model outperforms the scores reported by Hara et al. (2009) and Ficler and Goldberg (2016) . In the evaluation of each type, our method greatly improves the performance for VP, SBAR, and especially the S type of coordination compared with the similarity-based method of Hara et al. (2009) . Regarding the S type, our results are considerably better than those of Ficler and Goldberg (2016) .",
"cite_spans": [
{
"start": 103,
"end": 121,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 126,
"end": 152,
"text": "Ficler and Goldberg (2016)",
"ref_id": "BIBREF2"
},
{
"start": 332,
"end": 350,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 425,
"end": 451,
"text": "Ficler and Goldberg (2016)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [
{
"start": 26,
"end": 33,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "4.2.3"
},
{
"text": "As presented in Table 4 , our proposed replaceability feature significantly contributes to the detection of this type of coordination, where only the similarity feature does not work because of a collapse of similarity between conjuncts. The results for NP coordination, which accounts for nearly 65% of all coordination, are fairly good for the Genia corpus; however, the model proposed by Ficler and Goldberg (2016) exhibits better performance than ours for the PTB for the inner metric.",
"cite_spans": [],
"ref_spans": [
{
"start": 16,
"end": 23,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "4.2.3"
},
{
"text": "Approaches using the similarity property between conjuncts have been developed in previous works.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "5"
},
{
"text": "Regarding the task of coordination identification in Japanese, Kurohashi and Nagao (1994) used a chart to compute the similarity between conjuncts and identify conjunct spans with a dynamic programming technique. Shimbo and Hara (2007) proposed a sequence alignment model with dynamic programming to capture locally similar structures in two conjuncts on the basis of the set of features including word surfaces, POS tags, and morphological characteristics. The similarity score in their work is computed by a weighted linear combination (perceptron) of manually designed features assigned to edges and nodes in graphs, while the score in the work of Kurohashi and Nagao (1994) is calculated from a score function that uses some rules based on the observation of coordination. Although the method of Shimbo and Hara (2007) could not handle nested coordinate structures, Hara et al. (2009) extended their work to cope with nested coordination as well as three or more than consecutive conjuncts. Their proposed method defined several production rules to build consistent coordination trees with discriminative functions based on the similarity score. Hanamoto (2012) used dual decomposition to combine an HPSG parser with the model of Hara et al. (2009) . The method of use of the replaceability property has recently been adopted by Ficler and Goldberg (2016) . They incorporated the replaceability property between conjuncts into the feature representations, as well as the similarity property. They made use of these properties to assign scores to candidate pairs of conjuncts. Their method consists of three components: a binary classifier to detect the presence of coordination, the parser extended from the Berkeley Parser (Petrov et al., 2006) to generate candidate pairs, and a discriminative neural network to identify conjuncts. As similarity features, they compute the Euclidean distance between the two representations of con-juncts, which are computed from syntactic trees generated by the parser, and this is more efficient with respect to the time complexity compared with the methods with graphs. The replaceability feature vectors are produced from bidirectional LSTMs by processing two sentences that are produced by leaving out one of two conjuncts. Their model then scores all candidate pairs of conjuncts from feature vectors including similarities, replaceabilities, and additional three values derived from the probabilities assigned by the parser. The best scored pair is selected as the most probable conjuncts. For the Genia corpus, their model outperformed the method of Hara et al. (2009) which only relied on the similarity property. Using neural networks, they overcame the problems of manually elaborated features and of access to external sources such as thesauri. However, their method heavily depends on their extension of the Berkeley Parser. Therefore, the problem of error propagation between components and the parser still remains. Kawahara and Kurohashi (2008) tried to resolve coordination disambiguation without any similarities on the basis of the dependency relations and generative probabilities of phrases including conjuncts. Yoshimoto et al. (2015) extended the graph-based dependency parsing algorithm to handle coordinations.",
"cite_spans": [
{
"start": 63,
"end": 89,
"text": "Kurohashi and Nagao (1994)",
"ref_id": "BIBREF11"
},
{
"start": 213,
"end": 235,
"text": "Shimbo and Hara (2007)",
"ref_id": "BIBREF15"
},
{
"start": 651,
"end": 677,
"text": "Kurohashi and Nagao (1994)",
"ref_id": "BIBREF11"
},
{
"start": 800,
"end": 822,
"text": "Shimbo and Hara (2007)",
"ref_id": "BIBREF15"
},
{
"start": 870,
"end": 888,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 1234,
"end": 1252,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 1333,
"end": 1359,
"text": "Ficler and Goldberg (2016)",
"ref_id": "BIBREF2"
},
{
"start": 1728,
"end": 1749,
"text": "(Petrov et al., 2006)",
"ref_id": "BIBREF14"
},
{
"start": 2597,
"end": 2615,
"text": "Hara et al. (2009)",
"ref_id": "BIBREF4"
},
{
"start": 2970,
"end": 2999,
"text": "Kawahara and Kurohashi (2008)",
"ref_id": "BIBREF8"
},
{
"start": 3172,
"end": 3195,
"text": "Yoshimoto et al. (2015)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "5"
},
{
"text": "We propose a neural network model to disambiguate coordinate structure boundaries. Our method relies on two properties: (i) conjuncts tend to have a similar structure in syntax or semantics and (ii) conjuncts can be replaced with each other, maintaining sentence consistency. On the basis of these observations, we compute two feature vectors from a sequence of vectors produced by bidirectional RNNs. Our model can capture the connections between conjuncts and other parts of sentences and sentence-level coordination. As a result, our model outperforms existing methods and achieves state-of-the-art performance. The biggest contribution of our work is resolving dependency on information from syntactic parsers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "We plan to improve our model to handle three or more conjuncts in future work. In addition, since our method treats nested coordinate structures individually, we expect to create constraints to build non-overlapping coordination spans.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "We write coordinator words with their position in a sentence in the form of wordposition to distinguish them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "If two or more conjuncts appear before a conjunction, we regard them as one conjunct.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We consider \"and,\" \"or,\" \"but,\" \"nor,\" and \"and/or\" in the PTB and \"and,\" \"or,\" and \"but\" in the Genia as coordinator words followingFicler and Goldberg (2016) andHara et al. (2009).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://code.google.com/archive/p/word2vec/ 5 This F1 score is measured for the whole criterion, which is mentioned later.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We consider that NP and NX are NP coordination as inFicler and Goldberg (2016).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "In the Genia corpus, all coordinator words are associated with conjuncts; thus, there is no absence of coordination, as described inTable 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This work was partly supported by JST CREST Grant Number JPMJCR1513, Japan. We are grateful to our colleagues in the NAIST Computational Linguistics Laboratory and the anonymous reviewers for their helpful insights and comments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A neural probabilistic language model",
"authors": [
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
},
{
"first": "R\u00e9jean",
"middle": [],
"last": "Ducharme",
"suffix": ""
},
{
"first": "Pascal",
"middle": [],
"last": "Vincent",
"suffix": ""
},
{
"first": "Christian",
"middle": [],
"last": "Jauvin",
"suffix": ""
}
],
"year": 2003,
"venue": "Journal of machine learning research",
"volume": "3",
"issue": "",
"pages": "1137--1155",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoshua Bengio, R\u00e9jean Ducharme, Pascal Vincent, and Christian Jauvin. 2003. A neural probabilistic lan- guage model. Journal of machine learning research, 3(Feb):1137-1155.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Coordination Annotation Extension in the Penn Tree Bank",
"authors": [
{
"first": "Jessica",
"middle": [],
"last": "Ficler",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL 2016)",
"volume": "",
"issue": "",
"pages": "834--842",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jessica Ficler. 2016. Coordination Annotation Exten- sion in the Penn Tree Bank. Proceedings of the 54th Annual Meeting of the Association for Compu- tational Linguistics (ACL 2016), pages 834-842.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "A neural network for coordination boundary prediction",
"authors": [
{
"first": "Jessica",
"middle": [],
"last": "Ficler",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1610.03946"
]
},
"num": null,
"urls": [],
"raw_text": "Jessica Ficler and Yoav Goldberg. 2016. A neural net- work for coordination boundary prediction. arXiv preprint arXiv:1610.03946.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Coordination Structure Analysis using Dual Decomposition",
"authors": [
{
"first": "Atsushi",
"middle": [],
"last": "Hanamoto",
"suffix": ""
}
],
"year": 2012,
"venue": "",
"volume": "",
"issue": "",
"pages": "430--438",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Atsushi Hanamoto. 2012. Coordination Structure Analysis using Dual Decomposition. pages 430- 438.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Coordinate Structure Analysis with Global Structural Constraints and Alignment-Based Local Features",
"authors": [
{
"first": "Kazuo",
"middle": [],
"last": "Hara",
"suffix": ""
},
{
"first": "Masashi",
"middle": [],
"last": "Shimbo",
"suffix": ""
},
{
"first": "Hideharu",
"middle": [],
"last": "Okuma",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "1",
"issue": "",
"pages": "967--975",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kazuo Hara, Masashi Shimbo, Hideharu Okuma, and Yuji Matsumoto. 2009. Coordinate Structure Analysis with Global Structural Constraints and Alignment-Based Local Features. 1(August):967- 975.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A joint many-task model: Growing a neural network for multiple nlp tasks",
"authors": [
{
"first": "Kazuma",
"middle": [],
"last": "Hashimoto",
"suffix": ""
},
{
"first": "Caiming",
"middle": [],
"last": "Xiong",
"suffix": ""
},
{
"first": "Yoshimasa",
"middle": [],
"last": "Tsuruoka",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1611.01587"
]
},
"num": null,
"urls": [],
"raw_text": "Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsu- ruoka, and Richard Socher. 2016. A joint many-task model: Growing a neural network for multiple nlp tasks. arXiv preprint arXiv:1611.01587.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Long short-term memory",
"authors": [
{
"first": "Sepp",
"middle": [],
"last": "Hochreiter",
"suffix": ""
},
{
"first": "J\u00fcrgen",
"middle": [],
"last": "Schmidhuber",
"suffix": ""
}
],
"year": 1997,
"venue": "Neural computation",
"volume": "9",
"issue": "8",
"pages": "1735--1780",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735-1780.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Discriminative improvements to distributional sentence similarity",
"authors": [
{
"first": "Yangfeng",
"middle": [],
"last": "Ji",
"suffix": ""
},
{
"first": "Jacob",
"middle": [],
"last": "Eisenstein",
"suffix": ""
}
],
"year": 2013,
"venue": "EMNLP",
"volume": "",
"issue": "",
"pages": "891--896",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yangfeng Ji and Jacob Eisenstein. 2013. Discrimina- tive improvements to distributional sentence similar- ity. In EMNLP, pages 891-896.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Coordination disambiguation without any similarities",
"authors": [
{
"first": "Daisuke",
"middle": [],
"last": "Kawahara",
"suffix": ""
},
{
"first": "Sadao",
"middle": [],
"last": "Kurohashi",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 22nd International Conference on Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "425--432",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daisuke Kawahara and Sadao Kurohashi. 2008. Co- ordination disambiguation without any similarities. In Proceedings of the 22nd International Conference on Computational Linguistics-Volume 1, pages 425- 432. Association for Computational Linguistics.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Genia corpusa semantically annotated corpus for bio-textmining",
"authors": [
{
"first": "J-D",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Tomoko",
"middle": [],
"last": "Ohta",
"suffix": ""
},
{
"first": "Yuka",
"middle": [],
"last": "Tateisi",
"suffix": ""
},
{
"first": "Junichi",
"middle": [],
"last": "Tsujii",
"suffix": ""
}
],
"year": 2003,
"venue": "Bioinformatics",
"volume": "19",
"issue": "1",
"pages": "180--182",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J-D Kim, Tomoko Ohta, Yuka Tateisi, and Junichi Tsu- jii. 2003. Genia corpusa semantically annotated cor- pus for bio-textmining. Bioinformatics, 19(suppl 1):i180-i182.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Adam: A method for stochastic optimization",
"authors": [
{
"first": "Diederik",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Ba",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1412.6980"
]
},
"num": null,
"urls": [],
"raw_text": "Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A syntactic analysis method of long japanese sentences based on the detection of conjunctive structures",
"authors": [
{
"first": "Sadao",
"middle": [],
"last": "Kurohashi",
"suffix": ""
},
{
"first": "Makoto",
"middle": [],
"last": "Nagao",
"suffix": ""
}
],
"year": 1994,
"venue": "Computational Linguistics",
"volume": "20",
"issue": "4",
"pages": "507--534",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sadao Kurohashi and Makoto Nagao. 1994. A syntac- tic analysis method of long japanese sentences based on the detection of conjunctive structures. Compu- tational Linguistics, 20(4):507-534.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "On the difficulty of training recurrent neural networks",
"authors": [
{
"first": "Razvan",
"middle": [],
"last": "Pascanu",
"suffix": ""
},
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2013,
"venue": "International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "1310--1318",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. 2013. On the difficulty of training recurrent neural networks. In International Conference on Machine Learning, pages 1310-1318.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Learning accurate, compact, and interpretable tree annotation",
"authors": [
{
"first": "Slav",
"middle": [],
"last": "Petrov",
"suffix": ""
},
{
"first": "Leon",
"middle": [],
"last": "Barrett",
"suffix": ""
},
{
"first": "Romain",
"middle": [],
"last": "Thibaux",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "433--440",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. 2006. Learning accurate, compact, and inter- pretable tree annotation. In Proceedings of the 21st International Conference on Computational Lin- guistics and the 44th annual meeting of the Associa- tion for Computational Linguistics, pages 433-440. Association for Computational Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "A Discriminative Learning Model for Coordinate Conjunctions",
"authors": [
{
"first": "Masashi",
"middle": [],
"last": "Shimbo",
"suffix": ""
},
{
"first": "Kazuo",
"middle": [],
"last": "Hara",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "610--619",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Masashi Shimbo and Kazuo Hara. 2007. A Discrimi- native Learning Model for Coordinate Conjunctions. (June):610-619.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Dropout: a simple way to prevent neural networks from overfitting",
"authors": [
{
"first": "Nitish",
"middle": [],
"last": "Srivastava",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Krizhevsky",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
}
],
"year": 2014,
"venue": "Journal of Machine Learning Research",
"volume": "15",
"issue": "1",
"pages": "1929--1958",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Re- search, 15(1):1929-1958.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Improved semantic representations from tree-structured long short-term memory networks",
"authors": [
{
"first": "Kai Sheng",
"middle": [],
"last": "Tai",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
},
{
"first": "Christopher D",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1503.00075"
]
},
"num": null,
"urls": [],
"raw_text": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved semantic representations from tree-structured long short-term memory net- works. arXiv preprint arXiv:1503.00075.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Feature-rich part-ofspeech tagging with a cyclic dependency network",
"authors": [
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Christopher",
"suffix": ""
},
{
"first": "Yoram",
"middle": [],
"last": "Manning",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Singer",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology",
"volume": "1",
"issue": "",
"pages": "173--180",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kristina Toutanova, Dan Klein, Christopher D Man- ning, and Yoram Singer. 2003. Feature-rich part-of- speech tagging with a cyclic dependency network. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computa- tional Linguistics on Human Language Technology- Volume 1, pages 173-180. Association for Compu- tational Linguistics.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Bioasq: A challenge on large-scale biomedical semantic indexing and question answering",
"authors": [
{
"first": "George",
"middle": [],
"last": "Tsatsaronis",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Schroeder",
"suffix": ""
},
{
"first": "Georgios",
"middle": [],
"last": "Paliouras",
"suffix": ""
},
{
"first": "Yannis",
"middle": [],
"last": "Almirantis",
"suffix": ""
},
{
"first": "Ion",
"middle": [],
"last": "Androutsopoulos",
"suffix": ""
},
{
"first": "Eric",
"middle": [],
"last": "Gaussier",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Gallinari",
"suffix": ""
},
{
"first": "Thierry",
"middle": [],
"last": "Artieres",
"suffix": ""
},
{
"first": "Matthias",
"middle": [],
"last": "Michael R Alvers",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Zschunke",
"suffix": ""
}
],
"year": 2012,
"venue": "AAAI fall symposium: Information retrieval and knowledge discovery in biomedical text",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "George Tsatsaronis, Michael Schroeder, Georgios Paliouras, Yannis Almirantis, Ion Androutsopoulos, Eric Gaussier, Patrick Gallinari, Thierry Artieres, Michael R Alvers, Matthias Zschunke, et al. 2012. Bioasq: A challenge on large-scale biomedical se- mantic indexing and question answering. In AAAI fall symposium: Information retrieval and knowl- edge discovery in biomedical text.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Coordination-aware dependency parsing",
"authors": [
{
"first": "Akifumi",
"middle": [],
"last": "Yoshimoto",
"suffix": ""
},
{
"first": "Kazuo",
"middle": [],
"last": "Hara",
"suffix": ""
},
{
"first": "Masashi",
"middle": [],
"last": "Shimbo",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Akifumi Yoshimoto, Kazuo Hara, Masashi Shimbo, and Yuji Matsumoto. 2015. Coordination-aware de- pendency parsing (preliminary report). IWPT 2015, page 66.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"num": null,
"text": "The coordination identification task and our subtask.",
"type_str": "figure"
},
"FIGREF1": {
"uris": null,
"num": null,
"text": "from [the Soviet economic plight], one other . . . 2. Aside from [talks on cutting (strategic) arms], one other . . . 3. Aside from [talks on cutting (chemical) arms], one other . . . (b) Replaceability Characteristic of conjuncts . . . \" in common. At a phrase level, they all are categorized as NP and have identical tree structures (",
"type_str": "figure"
},
"FIGREF2": {
"uris": null,
"num": null,
"text": "respectively. These realvalued vectors are concatenated as the input of the",
"type_str": "figure"
},
"FIGREF3": {
"uris": null,
"num": null,
"text": "Overview of the architecture for coordination analysis.",
"type_str": "figure"
},
"FIGREF4": {
"uris": null,
"num": null,
"text": "next layer.",
"type_str": "figure"
},
"FIGREF5": {
"uris": null,
"num": null,
"text": "s = [Score(NONE); Score(1, k + 1); . . . ; Score(1, N); . . . ; Score(k \u2212 1, N )] p \u03b8 (y|x) = softmax(s)",
"type_str": "figure"
},
"TABREF1": {
"content": "<table><tr><td>3 .</td></tr></table>",
"html": null,
"text": "",
"num": null,
"type_str": "table"
},
"TABREF2": {
"content": "<table><tr><td>: The number of coordinators in the</td></tr><tr><td>datasets. (#count) indicates the number of actual</td></tr><tr><td>presences of coordination.</td></tr><tr><td>4.1 Evaluation Using the Penn Treebank</td></tr><tr><td>4.1.1 Experimental Setup</td></tr></table>",
"html": null,
"text": "",
"num": null,
"type_str": "table"
},
"TABREF3": {
"content": "<table><tr><td>presents a comparison with existing</td></tr><tr><td>methods. For all coordination, our proposed</td></tr><tr><td>method outperforms the state-of-the-art models</td></tr><tr><td>with a test set F1 score of 72.81 (0.11 better than</td></tr></table>",
"html": null,
"text": "72.87 74.36 77.90 75.05 76.45 outer 72.48 69.57 70.99 76.24 73.45 74.82 inner 74.07 71.10 72.56 77.43 74.59 75.99 exact 72.11 69.22 70.63 75.77 72.99 74.35",
"num": null,
"type_str": "table"
},
"TABREF4": {
"content": "<table><tr><td/><td/><td>All</td><td/><td/><td>NP</td></tr><tr><td/><td>P</td><td>R</td><td>F</td><td>P</td><td>R</td><td>F</td></tr><tr><td colspan=\"7\">Baseline 70.83 68.75 69.77 74.27 72.87 73.57</td></tr><tr><td>fsim</td><td colspan=\"6\">71.79 69.92 70.84 74.76 73.22 73.98</td></tr><tr><td>f repl</td><td colspan=\"6\">74.29 71.58 72.91 76.12 73.68 74.88</td></tr><tr><td>Both</td><td colspan=\"6\">75.92 72.87 74.36 77.90 75.05 76.45</td></tr></table>",
"html": null,
"text": "Performance difference by the metrics for the PTB development set.",
"num": null,
"type_str": "table"
},
"TABREF5": {
"content": "<table><tr><td/><td/><td>Dev</td><td/><td/><td>Test</td></tr><tr><td/><td>P</td><td>R</td><td>F</td><td>P</td><td>R</td><td>F</td></tr><tr><td/><td/><td/><td colspan=\"2\">All Coordination</td><td/></tr><tr><td>Berkeley</td><td colspan=\"4\">70.14 NP Coordination</td><td/></tr><tr><td>Berkeley</td><td colspan=\"6\">67.53 70.93 69.18 69.51 72.61 71.02</td></tr><tr><td>Zpar</td><td colspan=\"6\">69.14 72.31 70.68 69.81 72.92 71.33</td></tr><tr><td>Ficler16</td><td colspan=\"6\">75.17 74.82 74.99 76.91 75.31 76.1</td></tr><tr><td>Ours</td><td colspan=\"6\">77.43 74.59 75.99 75.87 74.76 75.31</td></tr></table>",
"html": null,
"text": "70.72 70.42 68.52 69.33 68.92 Zpar 72.21 72.72 72.46 68.24 69.42 68.82 Ficler16 72.34 72.25 72.29 72.81 72.61 72.7 Ours 74.07 71.10 72.56 73.46 72.16 72.81",
"num": null,
"type_str": "table"
},
"TABREF6": {
"content": "<table><tr><td colspan=\"5\">: Performance of inner-conjunct prediction</td></tr><tr><td colspan=\"5\">on all coordination and on NP coordination for the</td></tr><tr><td colspan=\"5\">PTB. The results for the three methods other than</td></tr><tr><td colspan=\"5\">our method are reported in Ficler16 : (Ficler and</td></tr><tr><td colspan=\"2\">Goldberg, 2016).</td><td/><td/><td/></tr><tr><td>COOD</td><td>#</td><td>Ours</td><td>Ficler16</td><td>Hara09</td></tr><tr><td>Overall</td><td>3598</td><td>65.98</td><td>64.14</td><td>61.5</td></tr><tr><td>NP</td><td>2317</td><td>66.59</td><td>65.08</td><td>64.2</td></tr><tr><td>VP</td><td>465</td><td>63.87</td><td>71.82</td><td>54.2</td></tr><tr><td>ADJP</td><td>321</td><td>78.50</td><td>74.76</td><td>80.4</td></tr><tr><td>S</td><td>188</td><td>52.65</td><td>17.02</td><td>22.9</td></tr><tr><td>PP</td><td>167</td><td>53.89</td><td>56.28</td><td>59.9</td></tr><tr><td>UCP</td><td>60</td><td>50.00</td><td>51.66</td><td>36.7</td></tr><tr><td>SBAR</td><td>56</td><td>78.57</td><td>91.07</td><td>51.8</td></tr><tr><td>ADVP</td><td>21</td><td>85.71</td><td>80.95</td><td>85.7</td></tr><tr><td>Others</td><td>3</td><td>33.33</td><td>33.33</td><td>66.7</td></tr></table>",
"html": null,
"text": "",
"num": null,
"type_str": "table"
}
}
}
} |