File size: 107,955 Bytes
6fa4bc9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 | {
"paper_id": "P15-1015",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T09:08:25.794164Z"
},
"title": "Learning Dynamic Feature Selection for Fast Sequential Prediction",
"authors": [
{
"first": "Emma",
"middle": [],
"last": "Strubell",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Massachusetts Amherst Amherst",
"location": {
"postCode": "01003",
"region": "MA",
"country": "USA"
}
},
"email": "strubell@cs.umass.edu"
},
{
"first": "Luke",
"middle": [],
"last": "Vilnis",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Massachusetts Amherst Amherst",
"location": {
"postCode": "01003",
"region": "MA",
"country": "USA"
}
},
"email": ""
},
{
"first": "Kate",
"middle": [],
"last": "Silverstein",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Massachusetts Amherst Amherst",
"location": {
"postCode": "01003",
"region": "MA",
"country": "USA"
}
},
"email": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Mccallum",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Massachusetts Amherst Amherst",
"location": {
"postCode": "01003",
"region": "MA",
"country": "USA"
}
},
"email": "mccallum@cs.umass.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We present paired learning and inference algorithms for significantly reducing computation and increasing speed of the vector dot products in the classifiers that are at the heart of many NLP components. This is accomplished by partitioning the features into a sequence of templates which are ordered such that high confidence can often be reached using only a small fraction of all features. Parameter estimation is arranged to maximize accuracy and early confidence in this sequence. Our approach is simpler and better suited to NLP than other related cascade methods. We present experiments in left-to-right part-of-speech tagging, named entity recognition, and transition-based dependency parsing. On the typical benchmarking datasets we can preserve POS tagging accuracy above 97% and parsing LAS above 88.5% both with over a five-fold reduction in run-time, and NER F1 above 88 with more than 2x increase in speed.",
"pdf_parse": {
"paper_id": "P15-1015",
"_pdf_hash": "",
"abstract": [
{
"text": "We present paired learning and inference algorithms for significantly reducing computation and increasing speed of the vector dot products in the classifiers that are at the heart of many NLP components. This is accomplished by partitioning the features into a sequence of templates which are ordered such that high confidence can often be reached using only a small fraction of all features. Parameter estimation is arranged to maximize accuracy and early confidence in this sequence. Our approach is simpler and better suited to NLP than other related cascade methods. We present experiments in left-to-right part-of-speech tagging, named entity recognition, and transition-based dependency parsing. On the typical benchmarking datasets we can preserve POS tagging accuracy above 97% and parsing LAS above 88.5% both with over a five-fold reduction in run-time, and NER F1 above 88 with more than 2x increase in speed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Many NLP tasks such as part-of-speech tagging, parsing and named entity recognition have become sufficiently accurate that they are no longer solely an object of research, but are also widely deployed in production systems. These systems can be run on billions of documents, making the efficiency of inference a significant concern-impacting not only wall-clock running time but also computer hardware budgets and the carbon footprint of data centers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This paper describes a paired learning and inference approach for significantly reducing computation and increasing speed while preserving accuracy in the linear classifiers typically used in many NLP tasks. The heart of the prediction computation in these models is a dot-product between a dense parameter vector and a sparse feature vector. The bottleneck in these models is then often a combination of feature extraction and numerical operations, each of which scale linearly in the size of the feature vector. Feature extraction can be even more expensive than the dot products, involving, for example, walking sub-graphs, lexicon lookup, string concatenation and string hashing. We note, however, that in many cases not all of these features are necessary for accurate prediction. For example, in part-of-speech tagging if we see the word \"the,\" there is no need to perform a large dot product or many string operations; we can accurately label the word a DETERMINER using the word identity feature alone. In other cases two features are sufficient: when we see the word \"hits\" preceded by a CARDINAL (e.g. \"two hits\") we can be confident that it is a NOUN.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We present a simple yet novel approach to improve processing speed by dynamically determining on a per-instance basis how many features are necessary for a high-confidence prediction. Our features are divided into a set of feature templates, such as current-token or previous-tag in the case of POS tagging. At training time, we determine an ordering on the templates such that we can approximate model scores at test time by incrementally calculating the dot product in template ordering. We then use a running confidence estimate for the label prediction to determine how many terms of the sum to compute for a given instance, and predict once confidence reaches a certain threshold.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In similar work, cascades of increasingly complex and high-recall models have been used for both structured and unstructured prediction. Viola and Jones (2001) use a cascade of boosted models to perform face detection. Weiss and Taskar (2010) add increasingly higher-order dependencies to a graphical model while filtering the out-put domain to maintain tractable inference. While most traditional cascades pass instances down to layers with increasingly higher recall, we use a single model and accumulate the scores from each additional template until a label is predicted with sufficient confidence, in a stagewise approximation of the full model score. Our technique applies to any linear classifier-based model over feature templates without changing the model structure or decreasing prediction speed.",
"cite_spans": [
{
"start": 137,
"end": 159,
"text": "Viola and Jones (2001)",
"ref_id": "BIBREF26"
},
{
"start": 219,
"end": 242,
"text": "Weiss and Taskar (2010)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Most similarly to our work, Weiss and Taskar (2013) improve performance for several structured vision tasks by dynamically selecting features at runtime. However, they use a reinforcement learning approach whose computational tradeoffs are better suited to vision problems with expensive features. Obtaining a speedup on tasks with comparatively cheap features, such as part-of-speech tagging or transition-based parsing, requires an approach with less overhead. In fact, the most attractive aspect of our approach is that it speeds up methods that are already among the fastest in NLP.",
"cite_spans": [
{
"start": 28,
"end": 51,
"text": "Weiss and Taskar (2013)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We apply our method to left-to-right part-ofspeech tagging in which we achieve accuracy above 97% on the Penn Treebank WSJ corpus while running more than five times faster than our 97.2% baseline. We also achieve a five-fold increase in transition-based dependency parsing on the WSJ corpus while achieving an LAS just 1.5% lower than our 90.3% baseline. Named entity recognition also shows significant speed increases. We further demonstrate that our method can be tuned for 2.5 3.5x multiplicative speedups with nearly no loss in accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our algorithm speeds up prediction for multiclass classification problems where the label set can be tractably enumerated and scored, and the per-class scores of input features decompose as a sum over multiple feature templates. Frequently, classification problems in NLP are solved through the use of linear classifiers, which compute scores for inputlabel pairs using a dot product. These meet our additive scoring criteria, and our acceleration methods are directly applicable. However, in this work we are interested in speeding up structured prediction problems, specifically part-of-speech (POS) tagging and dependency parsing. We apply our classification algorithms to these problems by reducing them to sequential prediction (Daum\u00e9 III et al., 2009) . For POS tagging, we describe a sentence's part of speech annotation by the left-to-right sequence of tagging decisions for individual tokens (Gim\u00e9nez and M\u00e0rquez, 2004) . Similarly, we implement our parser with a classifier that generates a sequence of shift-reduce parsing transitions (Nivre, 2009) .",
"cite_spans": [
{
"start": 733,
"end": 757,
"text": "(Daum\u00e9 III et al., 2009)",
"ref_id": "BIBREF4"
},
{
"start": 901,
"end": 928,
"text": "(Gim\u00e9nez and M\u00e0rquez, 2004)",
"ref_id": "BIBREF8"
},
{
"start": 1046,
"end": 1059,
"text": "(Nivre, 2009)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Classification and Structured Prediction",
"sec_num": "2"
},
{
"text": "The use of sequential prediction to solve these problems and others has a long history in practice as well as theory. Searn (Daum\u00e9 III et al., 2009) and DAgger (Ross et al., 2011) are two popular principled frameworks for reducing sequential prediction to classification by learning a classifier on additional synthetic training data. However, as we do in our experiments, practitioners often see good results by training on the gold standard labels with an off-the-shelf classification algorithm, as though classifying IID data (Bengtson and Roth, 2008; Choi and Palmer, 2012) .",
"cite_spans": [
{
"start": 124,
"end": 148,
"text": "(Daum\u00e9 III et al., 2009)",
"ref_id": "BIBREF4"
},
{
"start": 160,
"end": 179,
"text": "(Ross et al., 2011)",
"ref_id": "BIBREF20"
},
{
"start": 529,
"end": 554,
"text": "(Bengtson and Roth, 2008;",
"ref_id": "BIBREF0"
},
{
"start": 555,
"end": 577,
"text": "Choi and Palmer, 2012)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Classification and Structured Prediction",
"sec_num": "2"
},
{
"text": "Classifier-based approaches to structured prediction are faster than dynamic programming since they consider only a subset of candidate output structures in a greedy manner. For example, the Stanford CoreNLP classifier-based partof-speech tagger provides a 6.5x speed advantage over their dynamic programming-based model, with little reduction in accuracy. Because our methods are designed for the greedy sequential prediction regime, we can provide further speed increases to the fastest inference methods in NLP.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Classification and Structured Prediction",
"sec_num": "2"
},
{
"text": "Our base classifier for sequential prediction tasks will be a linear model. Given an input x 2 X , a set of labels Y, a feature map (x, y), and a weight vector w, a linear model predicts the highestscoring label",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linear models",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "y \u21e4 = arg max y2Y w \u2022 (x, y).",
"eq_num": "(1)"
}
],
"section": "Linear models",
"sec_num": "3"
},
{
"text": "The parameter w is usually learned by minimizing a regularized (R) sum of loss functions (`) over the training examples indexed by i",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linear models",
"sec_num": "3"
},
{
"text": "w \u21e4 = arg min w X i`( x i , y i , w) + R(w).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linear models",
"sec_num": "3"
},
{
"text": "In this paper, we partition the features into a set of feature templates, so that the weights, feature function, and dot product factor as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linear models",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "w \u2022 (x, y) = X j w j \u2022 j (x, y)",
"eq_num": "(2)"
}
],
"section": "Linear models",
"sec_num": "3"
},
{
"text": "for some set of feature templates { j (x, y)}. Our goal is to approximate the dot products in (1) sufficiently for purposes of prediction, while using as few terms of the sum in (2) as possible.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linear models",
"sec_num": "3"
},
{
"text": "We accomplish this goal by developing paired learning and inference procedures for featuretemplated classifiers that optimize both accuracy and inference speed, using a process of dynamic feature selection. Since many decisions are easy to make in the presence of strongly predictive features, we would like our model to use fewer templates when it is more confident. For a fixed, learned ordering of feature templates, we build up a vector of class scores incrementally over each prefix of the sequence of templates, which we call the prefix scores. Once we reach a stopping criterion based on class confidence (margin), we stop computing prefix scores, and predict the current highest scoring class. Our aim is to train each prefix to be as good a classifier as possible without the following templates, minimizing the number of templates needed for accurate predictions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method",
"sec_num": "4"
},
{
"text": "Given this method for performing fast inference on an ordered set of feature templates, it remains to choose the ordering. In Section 4.5, we develop several methods for picking template orderings, based on ideas from group sparsity (Yuan and Lin, 2006; Swirszcz et al., 2009) , and other techniques for feature subset-selection (Kohavi and John, 1997) .",
"cite_spans": [
{
"start": 233,
"end": 253,
"text": "(Yuan and Lin, 2006;",
"ref_id": "BIBREF31"
},
{
"start": 254,
"end": 276,
"text": "Swirszcz et al., 2009)",
"ref_id": "BIBREF22"
},
{
"start": 329,
"end": 352,
"text": "(Kohavi and John, 1997)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Method",
"sec_num": "4"
},
{
"text": "Given a model that computes scores additively over template-specific scoring functions as in (2), parameters w, and an observation x 2 X, we can define the i'th prefix score for label y 2 Y as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "P i,y (x, w) = i X j=1 w j \u2022 j (x, y),",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "or P i,y when the choice of observations and weights is clear from context. Abusing notation we also refer to the vector containing all i'th prefix scores for observation x associated to each label in Y as P i (x, w), or P i when this is unambiguous.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "Given a parameter m > 0, called the margin, we define a function h on prefix scores:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "h(P i , y) = max{0, max y 0 6 =y P i,y 0 P i,y + m} Algorithm 1 Inference Input: template parameters {w i } k i=1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": ", margin m and optional (for train time) true label y Initialize:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "i = 1 while l > 0^i \uf8ff k do l = max y 0 h(P i , y 0 ) (test) or h(P i , y) (train) i i + 1 end while return {P j } i j=1 (train) or max y 0 P i,y 0 (test) Algorithm 2 Parameter Learning Input: examples {(x i , y i )} N i , margin m Initialize: parameters w 0 = 0, i = 1 while i \uf8ff N do prefixes Infer(x i , y i , w i , m) g i ComputeGradient(prefixes) w i+1 UpdateParameters(w i , g i ) i i + 1 end while return w N",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "This is the familiar structured hinge loss function as in structured support vector machines (Tsochantaridis et al., 2004) , which has a minimum at 0 if and only if class y is ranked ahead of all other classes by at least m.",
"cite_spans": [
{
"start": 93,
"end": 122,
"text": "(Tsochantaridis et al., 2004)",
"ref_id": "BIBREF25"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "Using this notation, the condition that some label y be ranked first by a margin can be written as h(P i , y) = 0, and the condition that any class be ranked first by a margin can be written as max y 0 h(P i , y 0 ) = 0.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definitions",
"sec_num": "4.1"
},
{
"text": "As described in Algorithm 1, at test time we compute prefixes until some label is ranked ahead of all other labels with a margin m, then predict with that label. At train time, we predict until the correct label is ranked ahead with margin m, and return the whole set of prefixes for use by the learning algorithm. If no prefix scores have a margin, then we predict with the final prefix score involving all the feature templates.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference",
"sec_num": "4.2"
},
{
"text": "We split learning into two subproblems: first, given an ordered sequence of feature templates and our inference procedure, we wish to learn parameters that optimize accuracy while using as few of those templates as possible. Second, given a method for training feature templated classifiers, we want to learn an ordering of templates that optimizes accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning",
"sec_num": "4.3"
},
{
"text": "We wish to optimize several different objectives during learning: template parameters should have strong predictive power on their own, but also work well when combined with the scores from later templates. Additionally, we want to encourage well-calibrated confidence scores that allow us to stop prediction early without significant reduction in generalization ability.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning",
"sec_num": "4.3"
},
{
"text": "To learn parameters that encourage the use of few feature templates, we look at the model as outputting not a single prediction but a sequence of prefix predictions {P i }. For each training example, each feature template receives a number of hinge-loss gradients equal to its distance from the index where the margin requirement is finally reached. This is equivalent to treating each prefix as its own model for which we have a hinge loss function, and learning all models simultaneously. Our high-level approach is described in Algorithm 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "Concretely, for k feature templates we optimize the following structured max-margin objective (with the dependence of P 's on w written explicitly where helpful):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "w \u21e4 = arg min w X (x,y)`( x, y, w) (x, y, w) = i \u21e4 y X i=1 h(P i (x, w), y) i \u21e4 y = min i2{1..k} i s.t. h(P i , y) = 0",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "The per-example gradient of this objective for weights w j corresponding to feature template j then corresponds to",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "@@ w j = i \u21e4 y X i=j j (x, y loss (P i , y)) j (x, y).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "where we define y loss (P i , y) = arg max",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "y 0 P i,y 0 m \u2022 I(y 0 = y),",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "where I is an indicator function of the label y, used to define loss-augmented inference. We add an`2 regularization term to the objective, and tune the margin m and the regularization strength to tradeoff between speed and accuracy. In our experiments, we used a development set to choose a regularizer and margin that reduced testtime speed as much as possible without decreasing accuracy. We then varied the margin for that same model at test time to achieve larger speed gains at the cost of accuracy. In all experiments, the margin with which the model was trained corresponds to the largest margin reported, i.e. that with the highest accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the parameters",
"sec_num": "4.4"
},
{
"text": "We examine three approaches to learning the template ordering.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the template ordering",
"sec_num": "4.5"
},
{
"text": "The Group Lasso regularizer (Yuan and Lin, 2006) penalizes the sum of`2-norms of weights of feature templates (different from what is commonly called \"`2\" regularization, penalizing squared`2 norms),",
"cite_spans": [
{
"start": 28,
"end": 48,
"text": "(Yuan and Lin, 2006)",
"ref_id": "BIBREF31"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Group Lasso and Group Orthogonal Matching Pursuit",
"sec_num": "4.5.1"
},
{
"text": "P i c i kw i k 2 ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Group Lasso and Group Orthogonal Matching Pursuit",
"sec_num": "4.5.1"
},
{
"text": "where c i is a weight for each template. This regularizer encourages entire groups of weights to be set to 0, whose templates can then be discarded from the model. By varying the strength of the regularizer, we can learn an ordering of the importance of each template for a given model. The included groups for a given regularization strength are nearly always subsets of one another (technical conditions for this to be true are given in Hastie et al. (2007) ). The sequence of solutions for varied regularization strength is called the regularization path, and by slight abuse of terminology we use this to refer to the induced template ordering.",
"cite_spans": [
{
"start": 439,
"end": 459,
"text": "Hastie et al. (2007)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Group Lasso and Group Orthogonal Matching Pursuit",
"sec_num": "4.5.1"
},
{
"text": "An alternative and related approach to learning template orderings is based on the Group Orthogonal Matching Pursuit (GOMP) algorithm for generalized linear models (Swirszcz et al., 2009; Lozano et al., 2011) , with a few modifications for the setting of high-dimensional, sparse NLP data (described in Appendix B). Orthogonal matching pursuit algorithms are a set of stagewise feature selection techniques similar to forward stagewise regression (Hastie et al., 2007) and LARS (Efron et al., 2004) . At each stage, GOMP effectively uses each feature template to perform a linear regression to fit the gradient of the loss function. This attempts to find the correlation of each feature subset with the residual of the model. It then adds the feature template that best fits this gradient, and retrains the model. The main weakness of this method is that it fits the gradient of the training error which can rapidly overfit for sparse, highdimensional data. Ultimately, we would prefer to use a development set for feature selection.",
"cite_spans": [
{
"start": 164,
"end": 187,
"text": "(Swirszcz et al., 2009;",
"ref_id": "BIBREF22"
},
{
"start": 188,
"end": 208,
"text": "Lozano et al., 2011)",
"ref_id": "BIBREF15"
},
{
"start": 447,
"end": 468,
"text": "(Hastie et al., 2007)",
"ref_id": "BIBREF10"
},
{
"start": 473,
"end": 498,
"text": "LARS (Efron et al., 2004)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Group Lasso and Group Orthogonal Matching Pursuit",
"sec_num": "4.5.1"
},
{
"text": "The wrapper method (Kohavi and John, 1997) is a meta-algorithm for feature selection, usually based on a validation set. We employ it in a stagewise approach to learning a sequence of templates. Given an ordering of the initial sub-sequence and a learning procedure, we add each remaining template to our ordering and estimate parameters, selecting as the next template the one that gives the highest increase in development set performance. We begin the procedure with no templates, and repeat the procedure until we have a total ordering over the set of feature templates. When learning the ordering we use the same hyperparameters as will be used during final training.",
"cite_spans": [
{
"start": 19,
"end": 42,
"text": "(Kohavi and John, 1997)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Wrapper Method",
"sec_num": "4.5.2"
},
{
"text": "While simpler than the Lasso and Matching Pursuit approaches, we empirically found this approach to outperform the others, due to the necessity of using a development set to select features for our high-dimensional application areas.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Wrapper Method",
"sec_num": "4.5.2"
},
{
"text": "Our work is primarily inspired by previous research on cascades of classifiers; however, it differs significantly by approximating the score of a single linear model-scoring as few of its features as possible to obtain sufficient confidence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "We pose and address the question of whether a single, interacting set of parameters can be learned such that they efficiently both (1) provide high accuracy and (2) good confidence estimates throughout their use in the lengthening prefixes of the feature template sequence. (These two requirements are both incorporated into our novel parameter estimation algorithm.) In contrast, other work (Weiss and Taskar, 2013; He et al., 2013 ) learns a separate classifier to determine when to add features. Such heavier-weight approaches are unsuitable for our setting, where the core classifier's features and scoring are already so cheap that adding complex decision-making would cause too much computational overhead.",
"cite_spans": [
{
"start": 392,
"end": 416,
"text": "(Weiss and Taskar, 2013;",
"ref_id": "BIBREF28"
},
{
"start": 417,
"end": 432,
"text": "He et al., 2013",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "Other previous work on cascades uses a series of increasingly complex models, such as the Viola-Jones face detection cascade of classifiers (2001), which applies boosted trees trained on subsets of features in increasing order of complexity as needed, aiming to reject many sub-image windows early in processing. We allow scores from each layer to directly affect the final prediction, avoiding duplicate incorporation of evidence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "Our work is also related to the field of learning and inference under test-time budget constraints (Grubb and Bagnell, 2012; Trapeznikov and Saligrama, 2013) . However, common approaches to this problem also employ auxiliary models to rank which feature to add next, and are generally suited for problems where features are expensive to compute (e.g vision) and the extra computation of an auxiliary pruning-decision model is offset by substantial reduction in feature computations (Weiss and Taskar, 2013) . Our method uses confidence scores directly from the model, and so requires no additional computation, making it suitable for speeding up classifier-based NLP methods that are already very fast and have relatively cheap features.",
"cite_spans": [
{
"start": 99,
"end": 124,
"text": "(Grubb and Bagnell, 2012;",
"ref_id": "BIBREF9"
},
{
"start": 125,
"end": 157,
"text": "Trapeznikov and Saligrama, 2013)",
"ref_id": "BIBREF24"
},
{
"start": 482,
"end": 506,
"text": "(Weiss and Taskar, 2013)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "Some cascaded approaches strive at each stage to prune the number of possible output structures under consideration, whereas in our case we focus on pruning the input features. For example, Xu et al. (2013) learn a tree of classifiers that subdivides the set of classes to minimize average testtime cost. Chen et al. (2012) similarly use a linear cascade instead of a tree. Weiss and Taskar (2010) prune output labels in the context of structured prediction through a cascade of increasingly complex models, and Rush and Petrov (2012) successfully apply these structured prediction cascades to the task of graph-based dependency parsing.",
"cite_spans": [
{
"start": 305,
"end": 323,
"text": "Chen et al. (2012)",
"ref_id": "BIBREF1"
},
{
"start": 374,
"end": 397,
"text": "Weiss and Taskar (2010)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "In the context of NLP, He et al. (2013) describe a method for dynamic feature template selection at test time in graph-based dependency parsing. Their technique is particular to the parsing taskmaking a binary decision about whether to lock in edges in the dependency graph at each stage, and enforcing parsing-specific, hard-coded constraints on valid subsequent edges. Furthermore, as described above, they employ an auxiliary model to select features. He and Eisner (2012) share our goal to speed test time prediction by dynamically selecting features, but they also learn an additional model on top of a fixed base model, rather than using the training objective of the model itself.",
"cite_spans": [
{
"start": 455,
"end": 475,
"text": "He and Eisner (2012)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "While our comparisons above focus on other methods of dynamic feature selection, there also exists related work in the field of general (static) feature selection. The most relevant results come from the applications of group sparsity, such as the work of Martins et al. (2011) in Group Lasso for NLP problems. The Group Lasso regularizer (Yuan and Lin, 2006) sparsifies groups of feature weights (e.g. feature templates), and has been used to speed up test-time prediction by removing entire templates from the model. The key difference between this work and ours is that we select our templates based on the test-time difficulty of the inference problem, while the Group Lasso must do so at train time. In Appendix A, we compare against Group Lasso and show improvements in accuracy and speed.",
"cite_spans": [
{
"start": 256,
"end": 277,
"text": "Martins et al. (2011)",
"ref_id": "BIBREF17"
},
{
"start": 339,
"end": 359,
"text": "(Yuan and Lin, 2006)",
"ref_id": "BIBREF31"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "Note that non-grouped approaches to selecting sparse feature subsets, such as boosting and`1 regularization, do not achieve our goal of fast testtime prediction in NLP models, as they would not zero-out entire templates, and still require the computation of a feature for every template for every test instance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "We present experiments on three NLP tasks for which greedy sequence labeling has been a successful solution: part-of-speech tagging, transition-based dependency parsing and named entity recognition. In all cases our method achieves multiplicative speedups at test time with little loss in accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Results",
"sec_num": "6"
},
{
"text": "We conduct our experiments on classifier-based greedy part-of-speech tagging. Our baseline tagger uses the same features described in Choi and Palmer (2012) . We evaluate our models on the Penn Treebank WSJ corpus (Marcus et al., 1993) , employing the typical split of sections used for part-of-speech tagging: 0-18 train, 19-21 development, 22-24 test. The parameters of our models are learned using AdaGrad (Duchi et al., 2011) with`2 regularization via regularized dual averaging (Xiao, 2009) , and we used random search on the development set to select hyperparameters.",
"cite_spans": [
{
"start": 134,
"end": 156,
"text": "Choi and Palmer (2012)",
"ref_id": "BIBREF3"
},
{
"start": 214,
"end": 235,
"text": "(Marcus et al., 1993)",
"ref_id": "BIBREF16"
},
{
"start": 409,
"end": 429,
"text": "(Duchi et al., 2011)",
"ref_id": "BIBREF6"
},
{
"start": 483,
"end": 495,
"text": "(Xiao, 2009)",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Part-of-speech tagging",
"sec_num": "6.1"
},
{
"text": "This baseline model (baseline) tags at a rate of approximately 23,000 tokens per second on a 2010 2.1GHz AMD Opteron machine with accuracy comparable to similar taggers (Gim\u00e9nez and M\u00e0rquez, 2004; Choi and Palmer, 2012; Toutanova et al., 2003 Table 1 : Comparison of our models using different margins m, with speeds measured relative to the baseline. We train a model as accurate as the baseline while tagging 3.4x tokens/sec, and in another model maintain > 97% accuracy while tagging 5.2x, and > 96% accuracy with a speedup of 10.3x. the greedy Stanford CoreNLP left3words part-ofspeech tagger also tags at approximately 23,000 tokens per second. Significantly higher absolute speeds for all methods can be attained on more modern machines.",
"cite_spans": [
{
"start": 169,
"end": 196,
"text": "(Gim\u00e9nez and M\u00e0rquez, 2004;",
"ref_id": "BIBREF8"
},
{
"start": 197,
"end": 219,
"text": "Choi and Palmer, 2012;",
"ref_id": "BIBREF3"
},
{
"start": 220,
"end": 242,
"text": "Toutanova et al., 2003",
"ref_id": "BIBREF23"
}
],
"ref_spans": [
{
"start": 243,
"end": 250,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Part-of-speech tagging",
"sec_num": "6.1"
},
{
"text": "We include additional baselines that divide the features into templates, but train the templates' parameters more simply than our algorithm. The stagewise baseline learns the model parameters for each of the templates in order, starting with only one template-once each template has been trained for a fixed number of iterations, that template's parameters are fixed and we add the next one. We also create a separately-trained baseline model for each fixed prefix of the feature templates (fixed). This shows that our speedups are not simply due to superfluous features in the later templates.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Part-of-speech tagging",
"sec_num": "6.1"
},
{
"text": "Our main results are shown in Table 1 . We increase the speed of our baseline POS tagger by a factor of 5.2x without falling below 97% test accuracy. By tuning our training method to more aggressively prune templates, we achieve speedups of over 10x while providing accuracy higher than 96%. It is worth noting that the results for our method (dynamic) are all obtained from a single trained model (with hyperparameters optimized for m = 50, which we observed gave a good speedup with nearly no lossin accuracy on the development set), the only difference being Figure 1 : Left-hand plot depicts test accuracy as a function of the average number of templates used to predict. Right-hand plot shows speedup as a function of accuracy. Our model consistently achieves higher accuracy while using fewer templates resulting in the best ratio of speed to accuracy. that we varied the margin at test time. Superior results for m 6 = 50 could likely be obtained by optimizing hyperparameters for the desired margin.",
"cite_spans": [],
"ref_spans": [
{
"start": 30,
"end": 37,
"text": "Table 1",
"ref_id": null
},
{
"start": 562,
"end": 570,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Part-of-speech tagging",
"sec_num": "6.1"
},
{
"text": "Results show our method (dynamic) learns to dynamically select the number of templates, often using only a small fraction. The majority of test tokens can be tagged using only the first few templates: just over 40% use one template, and 75% require at most four templates, while maintaining 97.17% accuracy. On average 6.71 out of templates are used, though a small set of complicated instances never surpass the margin and use all 46 templates. The right hand plot of Figure 1 shows speedup vs. accuracy for various settings of the confidence margin m.",
"cite_spans": [],
"ref_spans": [
{
"start": 469,
"end": 477,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Part-of-speech tagging",
"sec_num": "6.1"
},
{
"text": "The left plot in Figure 1 depicts accuracy as a function of the number of templates used at test time. We present results for both varying the number of templates directly (dashed) and margin (solid). The baseline model trained on all templates performs very poorly when using marginbased inference, since its training objective does not learn to predict with only prefixes. When predicting using a fixed subset of templates, we use a different baseline model for each one of the 46 total template prefixes, learned with only those features; we then compare the test accuracy of our dynamic model using template prefix i to the baseline model trained on the fixed prefix i. Our model performs just as well as these separately trained models, demonstrating that our objective learns weights that allow each prefix to act as its own high-quality classifier.",
"cite_spans": [],
"ref_spans": [
{
"start": 17,
"end": 25,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Part-of-speech tagging",
"sec_num": "6.1"
},
{
"text": "As described in Section 4.5, we experimented on part-of-speech tagging with three different algorithms for learning an ordering of feature templates: Group Lasso, Group Orthogonal Matching Pursuit (GOMP), and the wrapper method. For the case of Group Lasso, this corresponds to the experimental setup used when evaluating Group Lasso for NLP in Martins et al. (2011) . As detailed in the part-of-speech tagging experiments of Appendix A, we found the wrapper method to work best in our dynamic prediction setting. Therefore, we use it in our remaining experiments in parsing and named entity recognition. Essentially, the Group Lasso picks small templates too early in the ordering by penalizing template norms, and GOMP picks large templates too early by overfitting the train error.",
"cite_spans": [
{
"start": 345,
"end": 366,
"text": "Martins et al. (2011)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Learning the template ordering",
"sec_num": "6.1.1"
},
{
"text": "We base our parsing experiments on the greedy, non-projective transition-based dependency parser described in Choi and Palmer (2011) . Our model uses a total of 60 feature templates based mainly on the word form, POS tag, lemma and assigned head label of current and previous input and stack tokens, and parses about 300 sentences/second on a modest 2.1GHz AMD Opteron machine.",
"cite_spans": [
{
"start": 110,
"end": 132,
"text": "Choi and Palmer (2011)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transition-based dependency parsing",
"sec_num": "6.2"
},
{
"text": "We train our parser on the English Penn Tree-Bank, learning the parameters using AdaGrad and the parsing split, training on sections 2-21, testing on section 23 and using section 22 for development and the Stanford dependency framework (de Figure 2 : Parsing speedup as a function of accuracy. Our model achieves the highest accuracy while using the fewest feature templates. Marneffe and Manning, 2008) . POS tags were automatically generated via 10-way jackknifing using the baseline POS model described in the previous section, trained with AdaGrad using`2 regularization, with parameters tuned on the development set to achieve 97.22 accuracy on WSJ sections 22-24. Lemmas were automatically generated using the ClearNLP morphological analyzer. We measure accuracy using labeled and unlabeled attachment scores excluding punctuation, achieving a labeled score of 90.31 and unlabeled score of 91.83, which are comparable to similar greedy parsers (Choi and Palmer, 2011; Honnibal and Goldberg, 2013) .",
"cite_spans": [
{
"start": 376,
"end": 403,
"text": "Marneffe and Manning, 2008)",
"ref_id": "BIBREF5"
},
{
"start": 950,
"end": 973,
"text": "(Choi and Palmer, 2011;",
"ref_id": "BIBREF2"
},
{
"start": 974,
"end": 1002,
"text": "Honnibal and Goldberg, 2013)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [
{
"start": 240,
"end": 248,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Transition-based dependency parsing",
"sec_num": "6.2"
},
{
"text": "Our experimental setup is the same as for partof-speech tagging. We compare our model (dynamic) to both a single baseline model trained on all features, and a set of 60 models each trained on a prefix of feature templates. Our experiments vary the margin used during prediction (solid) as well as the number of templates used (dashed).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Transition-based dependency parsing",
"sec_num": "6.2"
},
{
"text": "As in part-of-speech tagging, we observe significant test-time speedups when applying our method of dynamic feature selection to dependency parsing. With a loss of only 0.04 labeled attachment score (LAS), our model produces parses 2.7 times faster than the baseline. As listed in Table 2, with a more aggressive margin our model can parse more than 3 times faster while remaining above 90% LAS, and more than 5 times faster while maintaining accuracy above 88.5%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Transition-based dependency parsing",
"sec_num": "6.2"
},
{
"text": "In Figure 2 we see not only that our dynamic model consistently achieves higher accuracy while Table 2 : Comparison of our baseline and templated models using varying margins m and numbers of templates. using fewer templates, but also that our model (dynamic, dashed) performs exactly as well as separate models trained on each prefix of templates (baseline, dashed), demonstrating again that our training objective is successful in learning a single model that can predict as well as possible using any prefix of feature templates while successfully selecting which of these prefixes to use on a perexample basis.",
"cite_spans": [],
"ref_spans": [
{
"start": 3,
"end": 11,
"text": "Figure 2",
"ref_id": null
},
{
"start": 95,
"end": 102,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Transition-based dependency parsing",
"sec_num": "6.2"
},
{
"text": "We implement a greedy left-to-right named entity recognizer based on Ratinov and Roth (2009) using a total of 46 feature templates, including surface features such as lemma and capitalization, gazetteer look-ups, and each token's extended prediction history, as described in (Ratinov and Roth, 2009) . Training, tuning, and evaluation are performed on the CoNLL 2003 English data set with the BILOU encoding to denote label spans.",
"cite_spans": [
{
"start": 69,
"end": 92,
"text": "Ratinov and Roth (2009)",
"ref_id": "BIBREF19"
},
{
"start": 275,
"end": 299,
"text": "(Ratinov and Roth, 2009)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity recognition",
"sec_num": "6.3"
},
{
"text": "Our baseline model achieves F1 scores of 88.35 and 93.37 on the test and development sets, respectively, and tags at a rate of approximately 5300 tokens per second on the hardware described in the experiments above. We achieve a 2.3x speedup while maintaining F1 score above 88 on the test set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity recognition",
"sec_num": "6.3"
},
{
"text": "By learning to dynamically select the most predictive features at test time, our algorithm provides significant speed improvements to classifier-based structured prediction algorithms, which themselves already comprise the fastest methods in NLP. Further, these speed gains come at very lit- Table 3 : Comparison of our baseline and templated NER models using varying margin m and number of templates.",
"cite_spans": [],
"ref_spans": [
{
"start": 292,
"end": 299,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "tle extra implementation cost and can easily be combined with existing state-of-the-art systems. Future work will remove the fixed ordering for feature templates, and dynamically add additional features based on the current scores of different labels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
}
],
"back_matter": [
{
"text": "This work was supported in part by the Center for Intelligent Information Retrieval, in part by DARPA under agreement number FA8750-13-2-0020, and in part by NSF grant #CNS-0958392. The U.S. Government is authorized to reproduce and distribute reprint for Governmental purposes notwithstanding any copyright annotation thereon. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": "8"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Understanding the value of features for coreference resolution",
"authors": [
{
"first": "Eric",
"middle": [],
"last": "Bengtson",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Roth",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "294--303",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eric Bengtson and Dan Roth. 2008. Understanding the value of features for coreference resolution. In Pro- ceedings of the Conference on Empirical Methods in Natural Language Processing, pages 294-303. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Classifier cascade for minimizing feature evaluation cost",
"authors": [
{
"first": "Minmin",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "\"",
"middle": [],
"last": "Zhixiang",
"suffix": ""
},
{
"first": "\"",
"middle": [],
"last": "Eddie",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Q",
"middle": [],
"last": "Kilian",
"suffix": ""
},
{
"first": "Olivier",
"middle": [],
"last": "Weinberger",
"suffix": ""
},
{
"first": "Dor",
"middle": [],
"last": "Chappele",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kedem",
"suffix": ""
}
],
"year": 2012,
"venue": "AISTATS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Minmin Chen, Zhixiang \"Eddie\" Xu, Kilian Q Wein- berger, Olivier Chappele, and Dor Kedem. 2012. Classifier cascade for minimizing feature evaluation cost. In AISTATS.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Getting the Most out of Transition-based Dependency Parsing. Association for Computational Linguistics",
"authors": [
{
"first": "Jinho",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "",
"issue": "",
"pages": "687--692",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jinho Choi and Martha Palmer. 2011. Getting the Most out of Transition-based Dependency Parsing. Asso- ciation for Computational Linguistics, pages 687- 692.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Fast and robust part-of-speech tagging using dynamic model selection",
"authors": [
{
"first": "Jinho",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
}
],
"year": 2012,
"venue": "Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jinho Choi and Martha Palmer. 2012. Fast and robust part-of-speech tagging using dynamic model selec- tion. In Association for Computational Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Search-based structured prediction",
"authors": [
{
"first": "Hal",
"middle": [],
"last": "Daum\u00e9",
"suffix": ""
},
{
"first": "Iii",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "John",
"middle": [],
"last": "Langford",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2009,
"venue": "Machine Learning",
"volume": "75",
"issue": "",
"pages": "297--325",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hal Daum\u00e9 III, John Langford, and Daniel Marcu. 2009. Search-based structured prediction. Machine Learning, 75(3):297-325.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "The stanford typed dependencies representation",
"authors": [
{
"first": "Marie-Catherine",
"middle": [],
"last": "De Marneffe",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2008,
"venue": "COLING 2008 Workshop on Crossframework and Cross-domain Parser Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marie-Catherine de Marneffe and Christopher D. Man- ning. 2008. The stanford typed dependencies rep- resentation. In COLING 2008 Workshop on Cross- framework and Cross-domain Parser Evaluation.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization",
"authors": [
{
"first": "John",
"middle": [],
"last": "Duchi",
"suffix": ""
},
{
"first": "Elad",
"middle": [],
"last": "Hazan",
"suffix": ""
},
{
"first": "Yoram",
"middle": [],
"last": "Singer",
"suffix": ""
}
],
"year": 2011,
"venue": "JMLR",
"volume": "12",
"issue": "",
"pages": "2121--2159",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John Duchi, Elad Hazan, and Yoram Singer. 2011. Adaptive Subgradient Methods for Online Learning and Stochastic Optimization. JMLR, 12:2121-2159.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Least angle regression",
"authors": [
{
"first": "Trevor",
"middle": [],
"last": "Bradley Efron",
"suffix": ""
},
{
"first": "Iain",
"middle": [],
"last": "Hastie",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Johnstone",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Tibshirani",
"suffix": ""
}
],
"year": 2004,
"venue": "The Annals of Statistics",
"volume": "32",
"issue": "2",
"pages": "407--499",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bradley Efron, Trevor Hastie, Iain Johnstone, Robert Tibshirani, et al. 2004. Least angle regression. The Annals of Statistics, 32(2):407-499.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Svmtool: A general pos tagger generator based on support vector machines",
"authors": [
{
"first": "Jes\u00fas",
"middle": [],
"last": "Gim\u00e9nez",
"suffix": ""
},
{
"first": "Llu\u00eds",
"middle": [],
"last": "M\u00e0rquez",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 4th LREC",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jes\u00fas Gim\u00e9nez and Llu\u00eds M\u00e0rquez. 2004. Svmtool: A general pos tagger generator based on support vector machines. In Proceedings of the 4th LREC, Lisbon, Portugal.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "SpeedBoost: Anytime Prediction with Uniform Near-Optimality",
"authors": [
{
"first": "Alexander",
"middle": [],
"last": "Grubb",
"suffix": ""
},
{
"first": "J",
"middle": [
"Andrew"
],
"last": "Bagnell",
"suffix": ""
}
],
"year": 2012,
"venue": "AISTATS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexander Grubb and J. Andrew Bagnell. 2012. SpeedBoost: Anytime Prediction with Uniform Near-Optimality. In AISTATS.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Forward stagewise regression and the monotone lasso",
"authors": [
{
"first": "Trevor",
"middle": [],
"last": "Hastie",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "Taylor",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Tibshirani",
"suffix": ""
},
{
"first": "Guenther",
"middle": [],
"last": "Walther",
"suffix": ""
}
],
"year": 2007,
"venue": "Electronic Journal of Statistics",
"volume": "1",
"issue": "",
"pages": "1--29",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Trevor Hastie, Jonathan Taylor, Robert Tibshirani, and Guenther Walther. 2007. Forward stagewise regres- sion and the monotone lasso. Electronic Journal of Statistics, 1:1-29.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Cost-sensitive dynamic feature selection",
"authors": [
{
"first": "He",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Eisner",
"suffix": ""
}
],
"year": 2012,
"venue": "ICML Workshop on Inferning: Interactions between Inference and Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "He He and Jason Eisner. 2012. Cost-sensitive dynamic feature selection. In ICML Workshop on Inferning: Interactions between Inference and Learning.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Dynamic feature selection for dependency parsing",
"authors": [
{
"first": "He",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Hal",
"middle": [],
"last": "Daum\u00e9",
"suffix": ""
},
{
"first": "Iii",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Eisner",
"suffix": ""
}
],
"year": 2013,
"venue": "EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "He He, Hal Daum\u00e9 III, and Jason Eisner. 2013. Dy- namic feature selection for dependency parsing. In EMNLP.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "A Non-Monotonic Arc-Eager Transition System for Dependency Parsing",
"authors": [
{
"first": "M",
"middle": [],
"last": "Honnibal",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M Honnibal and Y Goldberg. 2013. A Non-Monotonic Arc-Eager Transition System for Dependency Pars- ing. CoNLL.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Wrappers for feature subset selection",
"authors": [
{
"first": "Ron",
"middle": [],
"last": "Kohavi",
"suffix": ""
},
{
"first": "George H John",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 1997,
"venue": "Artificial Intelligence",
"volume": "97",
"issue": "1",
"pages": "273--324",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ron Kohavi and George H John. 1997. Wrappers for feature subset selection. Artificial Intelligence, 97(1):273-324.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Group orthogonal matching pursuit for logistic regression",
"authors": [
{
"first": "C",
"middle": [],
"last": "Aur\u00e9lie",
"suffix": ""
},
{
"first": "Grzegorz",
"middle": [],
"last": "Lozano",
"suffix": ""
},
{
"first": "Naoki",
"middle": [],
"last": "Swirszcz",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Abe",
"suffix": ""
}
],
"year": 2011,
"venue": "International Conference on Artificial Intelligence and Statistics",
"volume": "",
"issue": "",
"pages": "452--460",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aur\u00e9lie C Lozano, Grzegorz Swirszcz, and Naoki Abe. 2011. Group orthogonal matching pursuit for logis- tic regression. In International Conference on Arti- ficial Intelligence and Statistics, pages 452-460.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Building a Large Annotated Corpus of English: The Penn Treebank",
"authors": [
{
"first": "Mitchell",
"middle": [
"P"
],
"last": "Marcus",
"suffix": ""
},
{
"first": "Beatrice",
"middle": [],
"last": "Santorini",
"suffix": ""
},
{
"first": "Mary",
"middle": [
"Ann"
],
"last": "Marcinkiewicz",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "313--330",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a Large Annotated Corpus of English: The Penn Treebank. Computa- tional Linguistics, 19(2):313-330.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Structured sparsity in structured prediction",
"authors": [
{
"first": "Andr\u00e9",
"middle": [],
"last": "Martins",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "Smith",
"suffix": ""
},
{
"first": "Pedro",
"middle": [],
"last": "Aguiar",
"suffix": ""
},
{
"first": "M\u00e1rio",
"middle": [],
"last": "Figueiredo",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1500--1511",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andr\u00e9 Martins, Noah Smith, Pedro Aguiar, and M\u00e1rio Figueiredo. 2011. Structured sparsity in structured prediction. In Proceedings of the Conference on Empirical Methods in Natural Language Process- ing, pages 1500-1511. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Non-projective dependency parsing in expected linear time",
"authors": [
{
"first": "Joakim",
"middle": [],
"last": "Nivre",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP",
"volume": "1",
"issue": "",
"pages": "351--359",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Joakim Nivre. 2009. Non-projective dependency pars- ing in expected linear time. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, vol- ume 1, pages 351-359. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Design challenges and misconceptions in named entity recognition",
"authors": [
{
"first": "Lev",
"middle": [],
"last": "Ratinov",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Roth",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Thirteenth Conference on Computational Natural Language Learning",
"volume": "",
"issue": "",
"pages": "147--155",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lev Ratinov and Dan Roth. 2009. Design challenges and misconceptions in named entity recognition. In Proceedings of the Thirteenth Conference on Com- putational Natural Language Learning, pages 147- 155. Association for Computational Linguistics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "A reduction of imitation learning and structured prediction to no-regret online learning",
"authors": [
{
"first": "St\u00e9phane",
"middle": [],
"last": "Ross",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"J"
],
"last": "Gordon",
"suffix": ""
},
{
"first": "Drew",
"middle": [],
"last": "Bagnell",
"suffix": ""
}
],
"year": 2011,
"venue": "JMLR Proceedings",
"volume": "15",
"issue": "",
"pages": "627--635",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "St\u00e9phane Ross, Geoffrey J. Gordon, and Drew Bagnell. 2011. A reduction of imitation learning and struc- tured prediction to no-regret online learning. In Ge- offrey J. Gordon, David B. Dunson, and Miroslav Dud\u00edk, editors, AISTATS, volume 15 of JMLR Pro- ceedings, pages 627-635.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Vine pruning for efficient multi-pass dependency parsing",
"authors": [
{
"first": "M",
"middle": [],
"last": "Alexander",
"suffix": ""
},
{
"first": "Slav",
"middle": [],
"last": "Rush",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Petrov",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "498--507",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexander M Rush and Slav Petrov. 2012. Vine prun- ing for efficient multi-pass dependency parsing. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, pages 498-507. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Grouped orthogonal matching pursuit for variable selection and prediction",
"authors": [
{
"first": "Grzegorz",
"middle": [],
"last": "Swirszcz",
"suffix": ""
},
{
"first": "Naoki",
"middle": [],
"last": "Abe",
"suffix": ""
},
{
"first": "Aurelie",
"middle": [
"C"
],
"last": "Lozano",
"suffix": ""
}
],
"year": 2009,
"venue": "Advances in Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "1150--1158",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Grzegorz Swirszcz, Naoki Abe, and Aurelie C Lozano. 2009. Grouped orthogonal matching pursuit for variable selection and prediction. In Advances in Neural Information Processing Systems, pages 1150-1158.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"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": "HLT-NAACL",
"volume": "",
"issue": "",
"pages": "",
"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 HLT-NAACL.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Supervised sequential classification under budget constraints",
"authors": [
{
"first": "Kirill",
"middle": [],
"last": "Trapeznikov",
"suffix": ""
},
{
"first": "Venkatesh",
"middle": [],
"last": "Saligrama",
"suffix": ""
}
],
"year": 2013,
"venue": "AISTATS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kirill Trapeznikov and Venkatesh Saligrama. 2013. Supervised sequential classification under budget constraints. In AISTATS.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Support vector machine learning for interdependent and structured output spaces",
"authors": [
{
"first": "Ioannis",
"middle": [],
"last": "Tsochantaridis",
"suffix": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Hofmann",
"suffix": ""
},
{
"first": "Thorsten",
"middle": [],
"last": "Joachims",
"suffix": ""
},
{
"first": "Yasemin",
"middle": [],
"last": "Altun",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the Twenty-first International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ioannis Tsochantaridis, Thomas Hofmann, Thorsten Joachims, and Yasemin Altun. 2004. Support vector machine learning for interdependent and structured output spaces. In Proceedings of the Twenty-first In- ternational Conference on Machine Learning, page 104.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Rapid object detection using a boosted cascade of simple features",
"authors": [
{
"first": "Paul",
"middle": [],
"last": "Viola",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Jones",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Paul Viola and Michael Jones. 2001. Rapid object de- tection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recog- nition, volume 1, pages I-511. IEEE.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Structured prediction cascades",
"authors": [
{
"first": "David",
"middle": [],
"last": "Weiss",
"suffix": ""
},
{
"first": "Ben",
"middle": [],
"last": "Taskar",
"suffix": ""
}
],
"year": 2010,
"venue": "AISTATS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Weiss and Ben Taskar. 2010. Structured predic- tion cascades. In AISTATS.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Learning adaptive value of information for structured prediction",
"authors": [
{
"first": "David",
"middle": [],
"last": "Weiss",
"suffix": ""
},
{
"first": "Ben",
"middle": [],
"last": "Taskar",
"suffix": ""
}
],
"year": 2013,
"venue": "NIPS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Weiss and Ben Taskar. 2013. Learning adaptive value of information for structured prediction. In NIPS.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Dual Averaging Method for Regularized Stochastic Learning and Online Optimization",
"authors": [
{
"first": "Lin",
"middle": [],
"last": "Xiao",
"suffix": ""
}
],
"year": 2009,
"venue": "NIPS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lin Xiao. 2009. Dual Averaging Method for Regular- ized Stochastic Learning and Online Optimization. In NIPS.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Cost-sensitive tree of classifiers",
"authors": [
{
"first": "\"eddie\"",
"middle": [],
"last": "Zhixiang",
"suffix": ""
},
{
"first": "Matt",
"middle": [
"J"
],
"last": "Xu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kusner",
"suffix": ""
},
{
"first": "Q",
"middle": [],
"last": "Kilian",
"suffix": ""
},
{
"first": "Minmin",
"middle": [],
"last": "Weinberger",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2013,
"venue": "ICML",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhixiang \"Eddie\" Xu, Matt J Kusner, Kilian Q Wein- berger, and Minmin Chen. 2013. Cost-sensitive tree of classifiers. In ICML.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Model selection and estimation in regression with grouped variables",
"authors": [
{
"first": "Ming",
"middle": [],
"last": "Yuan",
"suffix": ""
},
{
"first": "Yi",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2006,
"venue": "Journal of the Royal Statistical Society: Series B (Statistical Methodology)",
"volume": "68",
"issue": "1",
"pages": "49--67",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ming Yuan and Yi Lin. 2006. Model selection and es- timation in regression with grouped variables. Jour- nal of the Royal Statistical Society: Series B (Statis- tical Methodology), 68(1):49-67.",
"links": null
}
},
"ref_entries": {}
}
} |