File size: 87,886 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 | {
"paper_id": "P07-1027",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:49:54.085228Z"
},
"title": "Learning Predictive Structures for Semantic Role Labeling of NomBank",
"authors": [
{
"first": "Chang",
"middle": [],
"last": "Liu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore",
"location": {
"addrLine": "3 Science Drive 2",
"postCode": "117543",
"country": "Singapore"
}
},
"email": "liuchan1@comp.nus.edu.sg"
},
{
"first": "Hwee",
"middle": [
"Tou"
],
"last": "Ng",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore",
"location": {
"addrLine": "3 Science Drive 2",
"postCode": "117543",
"country": "Singapore"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper presents a novel application of Alternating Structure Optimization (ASO) to the task of Semantic Role Labeling (SRL) of noun predicates in NomBank. ASO is a recently proposed linear multi-task learning algorithm, which extracts the common structures of multiple tasks to improve accuracy, via the use of auxiliary problems. In this paper, we explore a number of different auxiliary problems, and we are able to significantly improve the accuracy of the Nom-Bank SRL task using this approach. To our knowledge, our proposed approach achieves the highest accuracy published to date on the English NomBank SRL task.",
"pdf_parse": {
"paper_id": "P07-1027",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper presents a novel application of Alternating Structure Optimization (ASO) to the task of Semantic Role Labeling (SRL) of noun predicates in NomBank. ASO is a recently proposed linear multi-task learning algorithm, which extracts the common structures of multiple tasks to improve accuracy, via the use of auxiliary problems. In this paper, we explore a number of different auxiliary problems, and we are able to significantly improve the accuracy of the Nom-Bank SRL task using this approach. To our knowledge, our proposed approach achieves the highest accuracy published to date on the English NomBank SRL task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The task of Semantic Role Labeling (SRL) is to identify predicate-argument relationships in natural language texts in a domain-independent fashion. In recent years, the availability of large human-labeled corpora such as PropBank (Palmer et al., 2005) and FrameNet (Baker et al., 1998) has made possible a statistical approach of identifying and classifying the arguments of verbs in natural language texts. A large number of SRL systems have been evaluated and compared on the standard data set in the CoNLL shared tasks (Carreras and Marquez, 2004; Carreras and Marquez, 2005) , and many systems have performed reasonably well. Compared to the previous CoNLL shared tasks (noun phrase bracketing, chunking, clause identification, and named entity recognition), SRL represents a significant step towards processing the semantic content of natural language texts.",
"cite_spans": [
{
"start": 230,
"end": 251,
"text": "(Palmer et al., 2005)",
"ref_id": "BIBREF16"
},
{
"start": 265,
"end": 285,
"text": "(Baker et al., 1998)",
"ref_id": "BIBREF3"
},
{
"start": 522,
"end": 550,
"text": "(Carreras and Marquez, 2004;",
"ref_id": "BIBREF5"
},
{
"start": 551,
"end": 578,
"text": "Carreras and Marquez, 2005)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Although verbs are probably the most obvious predicates in a sentence, many nouns are also capable of having complex argument structures, often with much more flexibility than its verb counterpart. For example, compare affect and effect: With the recent release of NomBank (Meyers et al., 2004) , it becomes possible to apply machine learning techniques to the task. So far we are aware of only one English NomBank-based SRL system (Jiang and Ng, 2006) , which uses the maximum entropy classifier, although similar efforts are reported on the Chinese NomBank by (Xue, 2006) and on FrameNet by (Pradhan et al., 2004) using a small set of hand-selected nominalizations. Noun predicates also appear in FrameNet semantic role labeling (Gildea and Jurafsky, 2002) , and many FrameNet SRL systems are evaluated in Senseval-3 (Litkowski, 2004) .",
"cite_spans": [
{
"start": 273,
"end": 294,
"text": "(Meyers et al., 2004)",
"ref_id": "BIBREF14"
},
{
"start": 432,
"end": 452,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
},
{
"start": 562,
"end": 573,
"text": "(Xue, 2006)",
"ref_id": "BIBREF21"
},
{
"start": 593,
"end": 615,
"text": "(Pradhan et al., 2004)",
"ref_id": "BIBREF17"
},
{
"start": 731,
"end": 758,
"text": "(Gildea and Jurafsky, 2002)",
"ref_id": "BIBREF10"
},
{
"start": 819,
"end": 836,
"text": "(Litkowski, 2004)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "[ subj Auto",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Semantic role labeling of NomBank is a multiclass classification problem by nature. Using the one-vs-all arrangement, that is, one binary classifier for each possible outcome, the SRL task can be treated as multiple binary classification problems. In the latter view, we are presented with the opportunity to exploit the common structures of these related problems. This is known as multi-task learning in the machine learning literature (Caruana, 1997; Ben-David and Schuller, 2003; Evgeniou and Pontil, 2004; Micchelli and Pontil, 2005; Maurer, 2006) .",
"cite_spans": [
{
"start": 438,
"end": 453,
"text": "(Caruana, 1997;",
"ref_id": "BIBREF7"
},
{
"start": 454,
"end": 483,
"text": "Ben-David and Schuller, 2003;",
"ref_id": "BIBREF4"
},
{
"start": 484,
"end": 510,
"text": "Evgeniou and Pontil, 2004;",
"ref_id": "BIBREF9"
},
{
"start": 511,
"end": 538,
"text": "Micchelli and Pontil, 2005;",
"ref_id": "BIBREF15"
},
{
"start": 539,
"end": 552,
"text": "Maurer, 2006)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we apply Alternating Structure Optimization (ASO) (Ando and Zhang, 2005a) to the semantic role labeling task on NomBank. ASO is a recently proposed linear multi-task learning algorithm based on empirical risk minimization. The method requires the use of multiple auxiliary problems, and its effectiveness may vary depending on the specific auxiliary problems used. ASO has been shown to be effective on the following natural language processing tasks: text categorization, named entity recognition, part-of-speech tagging, and word sense disambiguation (Ando and Zhang, 2005a; Ando and Zhang, 2005b; Ando, 2006) . This paper makes two significant contributions. First, we present a novel application of ASO to the SRL task on NomBank. We explore the effect of different auxiliary problems, and show that learning predictive structures with ASO results in significantly improved SRL accuracy. Second, we achieve accuracy higher than that reported in (Jiang and Ng, 2006) and advance the state of the art in SRL research.",
"cite_spans": [
{
"start": 65,
"end": 88,
"text": "(Ando and Zhang, 2005a)",
"ref_id": "BIBREF0"
},
{
"start": 568,
"end": 591,
"text": "(Ando and Zhang, 2005a;",
"ref_id": "BIBREF0"
},
{
"start": 592,
"end": 614,
"text": "Ando and Zhang, 2005b;",
"ref_id": "BIBREF1"
},
{
"start": 615,
"end": 626,
"text": "Ando, 2006)",
"ref_id": "BIBREF2"
},
{
"start": 964,
"end": 984,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The rest of this paper is organized as follows. We give an overview of NomBank and ASO in Sections 2 and 3 respectively. The baseline linear classifier is described in detail in Section 4, followed by the description of the ASO classifier in Section 5, where we focus on exploring different auxiliary problems. We provide discussions in Section 6, present related work in Section 7, and conclude in Section 8.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "NomBank annotates the set of arguments of noun predicates, just as PropBank annotates the arguments of verb predicates. As many noun predicates are nominalizations (e.g., replacement vs. replace), the same frames are shared with PropBank as much as possible, thus achieving some consistency with the latter regarding the accepted arguments and the meanings of each label.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NomBank",
"sec_num": "2"
},
{
"text": "Unlike in PropBank, arguments in NomBank can overlap with each other and with the predicate. For example:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NomBank",
"sec_num": "2"
},
{
"text": "[ location U.S.] [ pred,subj,obj steelmakers]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NomBank",
"sec_num": "2"
},
{
"text": "have supplied the steel.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NomBank",
"sec_num": "2"
},
{
"text": "Here the predicate make has subject steelmakers and object steel, analogous to Steelmakers make steel. The difference is that here make and steel are both part of the word steelmaker.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NomBank",
"sec_num": "2"
},
{
"text": "Each argument in NomBank is given one or more labels, out of the following 20: ARG0, ARG1, ARG2, ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NomBank",
"sec_num": "2"
},
{
"text": "This section gives a brief overview of ASO as implemented in this work. For a more complete description, see (Ando and Zhang, 2005a) .",
"cite_spans": [
{
"start": 109,
"end": 132,
"text": "(Ando and Zhang, 2005a)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Alternating structure optimization",
"sec_num": "3"
},
{
"text": "Given a set of training samples consisting of n feature vectors and their corresponding binary labels, {X i , Y i } for i \u2208 {1, . . . , n} where each X i is a p-dimensional vector, a binary linear classifier attempts to approximate the unknown relation by Y i = u T X i . The outcome is considered +1 if u T X is positive, or -1 otherwise. A well-established way to find the weight vector u is empirical risk minimization with least square regularization:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "u = arg min u 1 n n i=1 L u T X i , Y i + \u03bb u 2 (1) Function L(p, y)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "is known as the loss function. It encodes the penalty for a given discrepancy between the predicted label and the true label. In this work, we use a modification of Huber's robust loss function, similar to that used in (Ando and Zhang, 2005a) :",
"cite_spans": [
{
"start": 219,
"end": 242,
"text": "(Ando and Zhang, 2005a)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(p, y) = \uf8f1 \uf8f2 \uf8f3 \u22124py if py < \u22121 (1 \u2212 py) 2 if \u22121 \u2264 py < 1 0 if py \u2265 1",
"eq_num": "(2)"
}
],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "We fix the regularization parameter \u03bb to 10 \u22124 , similar to that used in (Ando and Zhang, 2005a) . The expression u 2 is defined as p i=1 u 2 p . When m binary classification problems are to be solved together, a h\u00d7p matrix \u0398 may be used to capture the common structures of the m weight vectors u l for l \u2208 {1, . . . , m} (h \u2264 m). We mandate that the rows of \u0398 be orthonormal, i.e., \u0398\u0398 T = I h\u00d7h . The h rows of \u0398 represent the h most significant components shared by all the u's. This relationship is modeled by",
"cite_spans": [
{
"start": 73,
"end": 96,
"text": "(Ando and Zhang, 2005a)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "u l = w l + \u0398 T v l (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "The parameters [{w l , v l }, \u0398] may then be found by joint empirical risk minimization over all the m problems, i.e., their values should minimize the combined empirical risk:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "m l=1 1 n n i=1 L (w l + \u0398 T v l ) T X l i , Y l i + \u03bb w l 2 (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-task linear classifier",
"sec_num": "3.1"
},
{
"text": "An important observation in (Ando and Zhang, 2005a) is that the binary classification problems used to derive \u0398 are not necessarily those problems we are aiming to solve. In fact, new problems can be invented for the sole purpose of obtaining a better \u0398. Thus, we distinguish between two types of problems in ASO: auxiliary problems, which are used to obtain \u0398, and target problems, which are the problems we are aiming to solve 1 . For instance, in the argument identification task, the only target problem is to identify arguments vs. non-arguments, whereas in the argument classification task, there are 20 binary target problems, one to identify each of the 20 labels (ARG0, ARG1, . . . ).",
"cite_spans": [
{
"start": 28,
"end": 51,
"text": "(Ando and Zhang, 2005a)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The ASO algorithm",
"sec_num": "3.2"
},
{
"text": "The target problems can also be used as an auxiliary problem. In addition, we can invent new auxiliary problems, e.g., in the argument identification stage, we can predict whether there are three words between the constituent and the predicate using the features of argument identification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The ASO algorithm",
"sec_num": "3.2"
},
{
"text": "Assuming there are k target problems and m auxiliary problems, it is shown in (Ando and Zhang, 2005a ) that by performing one round of minimization, an approximate solution of \u0398 can be obtained from (4) by the following algorithm:",
"cite_spans": [
{
"start": 78,
"end": 100,
"text": "(Ando and Zhang, 2005a",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The ASO algorithm",
"sec_num": "3.2"
},
{
"text": "1. For each of the m auxiliary problems, learn u l as described by (1).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The ASO algorithm",
"sec_num": "3.2"
},
{
"text": "U = [u 1 , u 2 , . . . , u m ], a p \u00d7 m matrix.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "This is a simplified version of the definition in (Ando and Zhang, 2005a) , made possible because the same \u03bb is used for all auxiliary problems.",
"cite_spans": [
{
"start": 50,
"end": 73,
"text": "(Ando and Zhang, 2005a)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "3. Perform Singular Value Decomposition (SVD) on U :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "U = V 1 DV T 2 , where V 1 is a p \u00d7 m ma- trix.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "The first h columns of V 1 are stored as rows of \u0398.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "4. Given \u0398, we learn w and v for each of the k target problems by minimizing the empirical risk of the associated training samples:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "1 n n i=1 L (w + \u0398 T v) T X i , Y i + \u03bb w 2 (5)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "5. The weight vector of each target problem can be found by:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "u = w + \u0398 T v",
"eq_num": "(6)"
}
],
"section": "Find",
"sec_num": "2."
},
{
"text": "By choosing a convex loss function, e.g., (2), steps 1 and 4 above can be formulated as convex optimization problems and are efficiently solvable.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "The procedure above can be considered as a Principal Component Analysis in the predictor space.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "Step (3) above extracts the most significant components shared by the predictors of the auxiliary problems and hopefully, by the predictors of the target problems as well. The hint of potential significant components helps (5) to outperform the simple linear predictor (1).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Find",
"sec_num": "2."
},
{
"text": "The SRL task is typically separated into two stages: argument identification and argument classification. During the identification stage, each constituent in a sentence's parse tree is labeled as either argument or non-argument. During the classification stage, each argument is given one of the 20 possible labels (ARG0, ARG1, . . . ). The linear classifier described by (1) is used as the baseline in both stages. For comparison, the F1 scores of a maximum entropy classifier are also reported here.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline classifier",
"sec_num": "4"
},
{
"text": "Eighteen baseline features and six additional features are proposed in (Jiang and Ng, 2006) for Nom-Bank argument identification. As the improvement of the F1 score due to the additional features is not statistically significant, we use the set of eighteen baseline features for simplicity. These features are reproduced in Table 1 for easy reference.",
"cite_spans": [
{
"start": 71,
"end": 91,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 324,
"end": 331,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Argument identification",
"sec_num": "4.1"
},
{
"text": "Unlike in (Jiang and Ng, 2006) , we do not prune arguments dominated by other arguments or those that overlap with the predicate in the training data. Accordingly, we do not maximize the probability of the entire labeled parse tree as in (Toutanova et al., 2005) . After the features of every constituent are extracted, each constituent is simply classified independently as either argument or non-argument.",
"cite_spans": [
{
"start": 10,
"end": 30,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
},
{
"start": 238,
"end": 262,
"text": "(Toutanova et al., 2005)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Argument identification",
"sec_num": "4.1"
},
{
"text": "The linear classifier described above is trained on sections 2 to 21 and tested on section 23. A maximum entropy classifier is trained and tested in the same manner. The F1 scores are presented in the first row of Table 3 , in columns linear and maxent respectively. The J&N column presents the result reported in (Jiang and Ng, 2006) using both baseline and additional features. The last column aso presents the best result from this work, to be explained in Section 5.",
"cite_spans": [
{
"start": 314,
"end": 334,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 214,
"end": 221,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Argument identification",
"sec_num": "4.1"
},
{
"text": "In NomBank, some constituents have more than one label. For simplicity, we always assign exactly one label to each identified argument in this step. For the 0.16% arguments with multiple labels in the training A diverse set of 28 features is used in (Jiang and Ng, 2006) for argument classification. In this work, the number of features is pruned to 11, so that we can work with reasonably many auxiliary problems in later experiments with ASO.",
"cite_spans": [
{
"start": 250,
"end": 270,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "4.2"
},
{
"text": "To find a smaller set of effective features, we start with all the features considered in (Jiang and Ng, 2006) , in (Xue and Palmer, 2004) , and various combinations of them, for a total of 52 features. These features are then pruned by the following algorithm:",
"cite_spans": [
{
"start": 90,
"end": 110,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
},
{
"start": 116,
"end": 138,
"text": "(Xue and Palmer, 2004)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "4.2"
},
{
"text": "1. For each feature in the current feature set, do step (2).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "4.2"
},
{
"text": "2. Remove the selected feature from the feature set. Obtain the F1 score of the remaining features when applied to the argument classification task, on development data section 24 with gold identification. Table 2 : Features used in argument classification step (2). The corresponding feature is removed from the current feature set if its F1 score is the same as or higher than the F1 score of retaining all features. 4. Repeat steps (1)-(3) until the F1 score starts to drop.",
"cite_spans": [],
"ref_spans": [
{
"start": 206,
"end": 213,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "4.2"
},
{
"text": "The 11 features so obtained are presented in Table 2. Using these features, a linear classifier and a maximum entropy classifier are trained on sections 2 to 21, and tested on section 23. The F1 scores are presented in the second row of Table 3 , in columns linear and maxent respectively. The J&N column presents the result reported in (Jiang and Ng, 2006) .",
"cite_spans": [
{
"start": 337,
"end": 357,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 237,
"end": 244,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "4.2"
},
{
"text": "In the combined task, we run the identification task with gold parse trees, and then the classification task with the output of the identification task. This way the combined effect of errors from both stages on the final classification output can be assessed. The scores of this complete SRL system are presented in the third row of Table 3 .",
"cite_spans": [],
"ref_spans": [
{
"start": 334,
"end": 341,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Further experiments and discussion",
"sec_num": "4.3"
},
{
"text": "To test the performance of the combined task on automatic parse trees, we employ two different configurations. First, we train the various classifiers on sections 2 to 21 using gold argument labels and automatic parse trees produced by Charniak's reranking parser (Charniak and Johnson, 2005) , and test them on section 23 with automatic parse trees. This is the same configuration as reported in (Pradhan et al., 2005; Jiang and Ng, 2006) . The scores are presented in the fourth row auto parse (t&t) in Table 3 .",
"cite_spans": [
{
"start": 264,
"end": 292,
"text": "(Charniak and Johnson, 2005)",
"ref_id": "BIBREF8"
},
{
"start": 397,
"end": 419,
"text": "(Pradhan et al., 2005;",
"ref_id": "BIBREF18"
},
{
"start": 420,
"end": 439,
"text": "Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 505,
"end": 512,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Further experiments and discussion",
"sec_num": "4.3"
},
{
"text": "Next, we train the various classifiers on sections 2 to 21 using gold argument labels and gold parse trees. To minimize the discrepancy between gold and automatic parse trees, we remove all the nodes in the gold trees whose POS are -NONE-, as they do not span any word and are thus never generated by the automatic parser. The resulting classifiers are then tested on section 23 using automatic parse trees. The scores are presented in the last row auto parse (test) of Table 3 . We note that auto parse (test) consistently outperforms auto parse (t&t) .",
"cite_spans": [
{
"start": 547,
"end": 552,
"text": "(t&t)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 470,
"end": 477,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Further experiments and discussion",
"sec_num": "4.3"
},
{
"text": "We believe that auto parse (test) is a more realistic setting in which to test the performance of SRL on automatic parse trees. When presented with some previously unseen test data, we are forced to rely on its automatic parse trees. However, for the best results we should take advantage of gold parse trees whenever possible, including those of the labeled training data. Our maximum entropy classifier consistently outperforms (Jiang and Ng, 2006) , which also uses a maximum entropy classifier. The primary difference is that we use a later version of NomBank (September 2006 release vs. September 2005 release). In addition, we use somewhat different features and treat overlapping arguments differently.",
"cite_spans": [
{
"start": 430,
"end": 450,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Further experiments and discussion",
"sec_num": "4.3"
},
{
"text": "Our ASO classifier uses the same features as the baseline linear classifier. The defining characteristic, and also the major challenge in successfully applying the ASO algorithm is to find related auxiliary problems that can reveal common structures shared with the target problem. To organize our search for good auxiliary problems for SRL, we separate them into two categories, unobservable auxiliary problems and observable auxiliary problems.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Applying ASO to SRL",
"sec_num": "5"
},
{
"text": "Unobservable auxiliary problems are problems whose true outcome cannot be observed from a raw text corpus but must come from another source, e.g., human labeling. For instance, predicting the argument class (i.e., ARG0, ARG1, . . . ) of a constituent is an unobservable auxiliary problem (which is also the only usable unobservable auxiliary problem here), because the true outcomes (i.e., the argument classes) are only available from human labels annotated in NomBank.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "For argument identification, we invent the following 20 binary unobservable auxiliary problems to take advantage of information previously unused at this stage:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "To predict the outcome of argument classification (i.e., ARG0, ARG1, . . . ) using the features of argument identification (pred, subcat, . . . ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "Thus for argument identification, we have 20 auxiliary problems (one auxiliary problem for predicting each of the argument classes ARG0, ARG1, . . . ) and one target problem (predicting whether a constituent is an argument) for the ASO algorithm described in Section 3.2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "In the argument classification task, the 20 binary target problems are also the unobservable auxiliary problems (one auxiliary problem for predicting each of the argument classes ARG0, ARG1, . . . ). Thus, we use the same 20 problems as both auxiliary problems and target problems.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "We train an ASO classifier on sections 2 to 21 and test it on section 23. With the 20 unobservable auxiliary problems, we obtain the F1 scores reported in the last column of Table 3 . In all the experiments, we keep h = 20, i.e., all the 20 columns of V 1 are kept.",
"cite_spans": [],
"ref_spans": [
{
"start": 174,
"end": 181,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "Comparing the F1 score of ASO against that of the linear classifier in every task (i.e., identification, classification, combined, both auto parse configurations), the improvement achieved by ASO is statistically significant (p < 0.05) based on the \u03c7 2 test.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "Comparing the F1 score of ASO against that of the maximum entropy classifier, the improvement in all but one task (argument classification) is statistically significant (p < 0.05). For argument classification, the improvement is not statistically significant (p = 0.08).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unobservable auxiliary problems",
"sec_num": "5.1"
},
{
"text": "Observable auxiliary problems are problems whose true outcome can be observed from a raw text corpus without additional externally provided labels. An example is to predict whether hw=trader from a constituent's other features, since the head word of a constituent can be obtained from the raw text alone. By definition, an observable auxiliary problem can always be formulated as predicting a feature of the training data. Depending on whether the baseline linear classifier already uses the feature to be predicted, we face two possibilities:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Observable auxiliary problems",
"sec_num": "5.2"
},
{
"text": "Predicting a used feature In auxiliary problems of this type, we must take care to remove the feature itself from the training data. For example, we must not use the feature path or pred&path to predict path itself.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Observable auxiliary problems",
"sec_num": "5.2"
},
{
"text": "Predicting an unused feature These auxiliary problems provide information that the classifier was previously unable to incorporate. The desirable characteristics of such a feature are:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Observable auxiliary problems",
"sec_num": "5.2"
},
{
"text": "1. The feature, although unused, should have been considered for the target problem so it is probably related to the target problem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Observable auxiliary problems",
"sec_num": "5.2"
},
{
"text": "2. The feature should not be highly correlated with a used feature, e.g., since the lastword feature is used in argument identification, we will not consider predicting lastword.pos as an auxiliary problem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Observable auxiliary problems",
"sec_num": "5.2"
},
{
"text": "Each chosen feature can create thousands of binary auxiliary problems. E.g., by choosing to predict hw, we can create auxiliary problems predicting whether hw=to, whether hw=trader, etc. To have more positive training samples, we only predict the most frequent features. Thus we will probably predict whether hw=to, but not whether hw=trader, since to occurs more frequently than trader as a head word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Observable auxiliary problems",
"sec_num": "5.2"
},
{
"text": "In argument identification using gold parse trees, we experiment with predicting three unused features as auxiliary problems: distance (distance between the predicate and the constituent), parent.lsis.hw (head word of the parent constituent's left sister) and parent.rsis.hw (head word of the parent constituent's right sister). We then experiment with predicting four used features: hw, lastword, ptype and path.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Argument identification",
"sec_num": "5.2.1"
},
{
"text": "The ASO classifier is trained on sections 2 to 21, and tested on section 23. Due to the large data size, we are unable to use more than 20 binary auxiliary problems or to experiment with combinations of them. The F1 scores are presented in Table 4 .",
"cite_spans": [],
"ref_spans": [
{
"start": 240,
"end": 247,
"text": "Table 4",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Argument identification",
"sec_num": "5.2.1"
},
{
"text": "In argument classification using gold parse trees and gold identification, we experiment with predicting three unused features path, partialpath, and chunkseq (concatenation of the phrase types of text chunks between the predicate and the constituent). We then experiment with predicting three used features hw, lastword, and ptype.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "5.2.2"
},
{
"text": "Combinations of these auxiliary problems are also tested. In all combined, we use the first 100 problems from each of the six groups of observable auxiliary problems. In selected combined, we use the first 100 problems from each of path, chunkseq, lastword and ptype problems.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "5.2.2"
},
{
"text": "The ASO classifier is trained on sections 2 to 21, and tested on section 23. The F1 scores are shown in Table 5 ",
"cite_spans": [],
"ref_spans": [
{
"start": 104,
"end": 111,
"text": "Table 5",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Argument classification",
"sec_num": "5.2.2"
},
{
"text": "Some of our experiments are limited by the extensive computing resources required for a fuller exploration. For instance, \"predicting unused features\" type of auxiliary problems might hold some hope for further improvement in argument identification, if a larger number of auxiliary problems can be used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussions",
"sec_num": "6"
},
{
"text": "ASO has been demonstrated to be an effective semi-supervised learning algorithm (Ando and Zhang, 2005a; Ando and Zhang, 2005b; Ando, 2006) . However, we have been unable to use unlabeled data to improve the accuracy. One possible reason is the cumulative noise from the many cascading steps involved in automatic SRL of unlabeled data: syntactic parse, predicate identification (where we identify nouns with at least one argument), argument identification, and finally argument classification, which reduces the effectiveness of adding unlabeled data using ASO. (Caruana, 1997) discusses configurations where both used inputs and unused inputs (due to excessive noise) are utilized as additional outputs. In contrast, our work concerns linear predictors using empirical risk minimization.",
"cite_spans": [
{
"start": 80,
"end": 103,
"text": "(Ando and Zhang, 2005a;",
"ref_id": "BIBREF0"
},
{
"start": 104,
"end": 126,
"text": "Ando and Zhang, 2005b;",
"ref_id": "BIBREF1"
},
{
"start": 127,
"end": 138,
"text": "Ando, 2006)",
"ref_id": "BIBREF2"
},
{
"start": 562,
"end": 577,
"text": "(Caruana, 1997)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussions",
"sec_num": "6"
},
{
"text": "A variety of auxiliary problems are tested in (Ando and Zhang, 2005a; Ando and Zhang, 2005b) in the semi-supervised settings, i.e., their auxiliary problems are generated from unlabeled data. This differs significantly from the supervised setting in our work, where only labeled data is used. While (Ando and Zhang, 2005b) uses \"predicting used features\" (previous/current/next word) as auxiliary problems with good results in named entity recognition, the use of similar observable auxiliary problems in our work gives no statistically significant improvements.",
"cite_spans": [
{
"start": 46,
"end": 69,
"text": "(Ando and Zhang, 2005a;",
"ref_id": "BIBREF0"
},
{
"start": 70,
"end": 92,
"text": "Ando and Zhang, 2005b)",
"ref_id": "BIBREF1"
},
{
"start": 299,
"end": 322,
"text": "(Ando and Zhang, 2005b)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussions",
"sec_num": "6"
},
{
"text": "More recently, for the word sense disambiguation (WSD) task, (Ando, 2006) experimented with both supervised and semi-supervised auxiliary problems, although the auxiliary problems she used are different from ours.",
"cite_spans": [
{
"start": 61,
"end": 73,
"text": "(Ando, 2006)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussions",
"sec_num": "6"
},
{
"text": "In this paper, we have presented a novel application of Alternating Structure Optimization (ASO) to the Semantic Role Labeling (SRL) task on NomBank. The possible auxiliary problems are categorized and tested extensively. Our results outperform those reported in (Jiang and Ng, 2006) . To the best of our knowledge, we achieve the highest SRL accuracy published to date on the English NomBank.",
"cite_spans": [
{
"start": 263,
"end": 283,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "Note that this definition deviates slightly from the one in(Ando and Zhang, 2005a). We find the definition here more convenient for our subsequent discussion.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": ". Select the highest of all the scores obtained in",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Multi-output neural networks learn several tasks simultaneously. In addition to the target outputs,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A framework for learning predictive structures from multiple tasks and unlabeled data",
"authors": [
{
"first": "R",
"middle": [
"K"
],
"last": "Ando",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2005,
"venue": "Journal of Machine Learning Research",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. K. Ando and T. Zhang. 2005a. A framework for learning predictive structures from multiple tasks and unlabeled data. Journal of Machine Learning Research.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A high-performance semisupervised learning method for text chunking",
"authors": [
{
"first": "R",
"middle": [
"K"
],
"last": "Ando",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. K. Ando and T. Zhang. 2005b. A high-performance semi- supervised learning method for text chunking. In Proc. of ACL.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Applying alternating structure optimization to word sense disambiguation",
"authors": [
{
"first": "R",
"middle": [
"K"
],
"last": "Ando",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. of CoNLL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. K. Ando. 2006. Applying alternating structure optimization to word sense disambiguation. In Proc. of CoNLL.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "The Berkeley FrameNet project",
"authors": [
{
"first": "C",
"middle": [
"F"
],
"last": "Baker",
"suffix": ""
},
{
"first": "C",
"middle": [
"J"
],
"last": "Fillmore",
"suffix": ""
},
{
"first": "J",
"middle": [
"B"
],
"last": "Lowe",
"suffix": ""
}
],
"year": 1998,
"venue": "Proc. of COLING-ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. F. Baker, C. J. Fillmore, and J. B. Lowe. 1998. The Berkeley FrameNet project. In Proc. of COLING-ACL.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Exploiting task relatedness for multiple task learning",
"authors": [
{
"first": "S",
"middle": [],
"last": "Ben-David",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Schuller",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. of COLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Ben-David and R. Schuller. 2003. Exploiting task related- ness for multiple task learning. In Proc. of COLT.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Introduction to the CoNLL-2004 shared task: Semantic role labeling",
"authors": [
{
"first": "X",
"middle": [],
"last": "Carreras",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Marquez",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of CoNLL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "X. Carreras and L. Marquez. 2004. Introduction to the CoNLL- 2004 shared task: Semantic role labeling. In Proc. of CoNLL.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Introduction to the CoNLL-2005 shared task: Semantic role labeling",
"authors": [
{
"first": "X",
"middle": [],
"last": "Carreras",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Marquez",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. of CoNLL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "X. Carreras and L. Marquez. 2005. Introduction to the CoNLL- 2005 shared task: Semantic role labeling. In Proc. of CoNLL.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Multitask Learning",
"authors": [
{
"first": "R",
"middle": [],
"last": "Caruana",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Caruana. 1997. Multitask Learning. Ph.D. thesis, School of Computer Science, CMU.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Coarse-to-fine n-best parsing and MaxEnt discriminative reranking",
"authors": [
{
"first": "E",
"middle": [],
"last": "Charniak",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Johnson",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "E. Charniak and M. Johnson. 2005. Coarse-to-fine n-best pars- ing and MaxEnt discriminative reranking. In Proc. of ACL.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Regularized multitask learning",
"authors": [
{
"first": "T",
"middle": [],
"last": "Evgeniou",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Pontil",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of KDD",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Evgeniou and M. Pontil. 2004. Regularized multitask learn- ing. In Proc. of KDD.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Automatic labeling of semantic roles",
"authors": [
{
"first": "D",
"middle": [],
"last": "Gildea",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2002,
"venue": "Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Gildea and D. Jurafsky. 2002. Automatic labeling of seman- tic roles. Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Semantic role labeling of Nom-Bank: A maximum entropy approach",
"authors": [
{
"first": "Z",
"middle": [
"P"
],
"last": "Jiang",
"suffix": ""
},
{
"first": "H",
"middle": [
"T"
],
"last": "Ng",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. of EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Z. P. Jiang and H. T. Ng. 2006. Semantic role labeling of Nom- Bank: A maximum entropy approach. In Proc. of EMNLP.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Senseval-3 task: automatic labeling of semantic roles",
"authors": [
{
"first": "K",
"middle": [
"C"
],
"last": "Litkowski",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of SENSEVAL-3",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. C. Litkowski. 2004. Senseval-3 task: automatic labeling of semantic roles. In Proc. of SENSEVAL-3.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Bounds for linear multitask learning",
"authors": [
{
"first": "A",
"middle": [],
"last": "Maurer",
"suffix": ""
}
],
"year": 2006,
"venue": "Journal of Machine Learning Research",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Maurer. 2006. Bounds for linear multitask learning. Journal of Machine Learning Research.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "The NomBank project: An interim report",
"authors": [
{
"first": "A",
"middle": [],
"last": "Meyers",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Reeves",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Macleod",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Szekeley",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Zielinska",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Young",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Grishman",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of HLT/NAACL Workshop on Frontiers in Corpus Annotation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Meyers, R. Reeves, C. Macleod, R. Szekeley, V. Zielinska, B. Young, and R. Grishman. 2004. The NomBank project: An interim report. In Proc. of HLT/NAACL Workshop on Frontiers in Corpus Annotation.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Kernels for multitask learning",
"authors": [
{
"first": "C",
"middle": [
"A"
],
"last": "Micchelli",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Pontil",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. of NIPS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. A. Micchelli and M. Pontil. 2005. Kernels for multitask learning. In Proc. of NIPS.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "The Proposition Bank: an annotated corpus of semantic roles",
"authors": [
{
"first": "M",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Gildea",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Kingsbury",
"suffix": ""
}
],
"year": 2005,
"venue": "Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Palmer, D. Gildea, and P. Kingsbury. 2005. The Proposition Bank: an annotated corpus of semantic roles. Computational Linguistics.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Parsing arguments of nominalizations in English and Chinese",
"authors": [
{
"first": "S",
"middle": [
"S"
],
"last": "Pradhan",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Sun",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "J",
"middle": [
"H"
],
"last": "Martin",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of HLT/NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. S. Pradhan, H. Sun, W. Ward, J. H. Martin, and D. Jurafsky. 2004. Parsing arguments of nominalizations in English and Chinese. In Proc. of HLT/NAACL.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Support vector learning for semantic argument classification",
"authors": [
{
"first": "S",
"middle": [],
"last": "Pradhan",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Hacioglu",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Krugler",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "J",
"middle": [
"H"
],
"last": "Martin",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2005,
"venue": "Machine Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Pradhan, K. Hacioglu, V. Krugler, W. Ward, J. H. Martin, and D. Jurafsky. 2005. Support vector learning for semantic argument classification. Machine Learning.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Joint learning improves semantic role labeling",
"authors": [
{
"first": "K",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Haghighi",
"suffix": ""
},
{
"first": "C",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Toutanova, A. Haghighi, and C. D. Manning. 2005. Joint learning improves semantic role labeling. In Proc. of ACL.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Calibrating features for semantic role labeling",
"authors": [
{
"first": "N",
"middle": [],
"last": "Xue",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Palmer",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. Xue and M. Palmer. 2004. Calibrating features for semantic role labeling. In Proc. of EMNLP.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Semantic role labeling of nominalized predicates in Chinese",
"authors": [
{
"first": "N",
"middle": [],
"last": "Xue",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. of HLT/NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. Xue. 2006. Semantic role labeling of nominalized predi- cates in Chinese. In Proc. of HLT/NAACL.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "ARG3, ARG4, ARG5, ARG8, ARG9, ARGM-ADV, ARGM-CAU, ARGM-DIR, ARGM-DIS, ARGM-EXT, ARGM-LOC, ARGM-MNR, ARGM-MOD, ARGM-NEG, ARGM-PNC, ARGM-PRD, and ARGM-TMP. Thus, the above sentence is annotated in NomBank as: [ ARGM-LOC U.S.] [ PRED,ARG0,ARG1 steelmakers] have supplied the steel.",
"uris": null,
"type_str": "figure"
},
"TABREF0": {
"html": null,
"type_str": "table",
"text": "prices] [ arg\u2212ext greatly] [ pred affect] [ obj the PPI]. [ subj Auto prices] have a [ arg\u2212ext big] [ pred effect] [ obj on the PPI]. The [ pred effect] [ subj of auto prices] [ obj on the PPI] is [ arg\u2212ext big]. [ subj The auto prices'] [ pred effect] [ obj on the PPI] is [ arg\u2212ext big]. The arguments of noun predicates can often be more easily omitted compared to the verb predicates: The [ pred effect] [ subj of auto prices] is [ arg\u2212ext big]. The [ pred effect] [ obj on the PPI] is [ arg\u2212ext big]. The [ pred effect] is [ arg\u2212ext big].",
"content": "<table/>",
"num": null
},
"TABREF1": {
"html": null,
"type_str": "table",
"text": "Features used in argument identification data, we pick the first and discard the rest. (Note that the same is not done on the test data.)",
"content": "<table><tr><td>1 pred</td><td>the stemmed predicate</td></tr><tr><td>2 subcat</td><td>grammar rule that expands the</td></tr><tr><td/><td>predicate P's parent</td></tr><tr><td>3 ptype</td><td>syntactic category (phrase</td></tr><tr><td/><td>type) of the constituent C</td></tr><tr><td>4 hw</td><td>syntactic head word of C</td></tr><tr><td>5 path</td><td>syntactic path from C to P</td></tr><tr><td>6 position</td><td>whether C is to the left/right of</td></tr><tr><td/><td>or overlaps with P</td></tr><tr><td>7 firstword</td><td>first word spanned by C</td></tr><tr><td>8 lastword</td><td>last word spanned by C</td></tr><tr><td>9 lsis.ptype</td><td>phrase type of left sister</td></tr><tr><td>10 rsis.hw</td><td>right sister's head word</td></tr><tr><td colspan=\"2\">11 rsis.hw.pos POS of right sister's head word</td></tr><tr><td colspan=\"2\">12 parent.ptype phrase type of parent</td></tr><tr><td>13 parent.hw</td><td>parent's head word</td></tr><tr><td>14 partialpath</td><td>path from C to the lowest com-</td></tr><tr><td/><td>mon ancestor with P</td></tr><tr><td colspan=\"2\">15 ptype & length of path</td></tr><tr><td>16 pred & hw</td><td/></tr><tr><td>17 pred & path</td><td/></tr><tr><td colspan=\"2\">18 pred & position</td></tr></table>",
"num": null
},
"TABREF4": {
"html": null,
"type_str": "table",
"text": "F1 scores of various classifiers on Nom-Bank SRL",
"content": "<table/>",
"num": null
},
"TABREF6": {
"html": null,
"type_str": "table",
"text": "F1 scores of ASO with observable auxiliary problems on argument identification. All h = 20.",
"content": "<table><tr><td>From Table 4 and 5, we observe that although</td></tr><tr><td>the use of observable auxiliary problems consis-</td></tr></table>",
"num": null
},
"TABREF7": {
"html": null,
"type_str": "table",
"text": "F1 scores of ASO with observable auxiliary problems on argument classification. All h = 100.",
"content": "<table><tr><td>tently improves the performance of the classifier,</td></tr><tr><td>the differences are small and not statistically signif-</td></tr><tr><td>icant. Further experiments combining unobservable</td></tr><tr><td>and observable auxiliary problems fail to outperform</td></tr><tr><td>ASO with unobservable auxiliary problems alone.</td></tr><tr><td>In summary, our work shows that unobservable</td></tr><tr><td>auxiliary problems significantly improve the perfor-</td></tr><tr><td>mance of NomBank SRL. In contrast, observable</td></tr><tr><td>auxiliary problems are not effective.</td></tr></table>",
"num": null
}
}
}
} |