File size: 117,407 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 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 | {
"paper_id": "D08-1024",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T16:29:52.310795Z"
},
"title": "Online Large-Margin Training of Syntactic and Structural Translation Features",
"authors": [
{
"first": "David",
"middle": [],
"last": "Chiang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Southern California",
"location": {
"addrLine": "4676 Admiralty Way, Suite 1001 Marina del Rey",
"postCode": "90292",
"region": "CA",
"country": "USA"
}
},
"email": "chiang@isi.edu"
},
{
"first": "Yuval",
"middle": [],
"last": "Marton",
"suffix": "",
"affiliation": {
"laboratory": "UMIACS Laboratory for Computational Linguistics and Information Processing",
"institution": "University of Maryland",
"location": {
"postCode": "20742",
"settlement": "College Park",
"region": "MD",
"country": "USA"
}
},
"email": "ymarton@umiacs.umd.edu"
},
{
"first": "Philip",
"middle": [],
"last": "Resnik",
"suffix": "",
"affiliation": {
"laboratory": "UMIACS Laboratory for Computational Linguistics and Information Processing",
"institution": "University of Maryland",
"location": {
"postCode": "20742",
"settlement": "College Park",
"region": "MD",
"country": "USA"
}
},
"email": "resnik@umiacs.umd.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Minimum-error-rate training (MERT) is a bottleneck for current development in statistical machine translation because it is limited in the number of weights it can reliably optimize. Building on the work of Watanabe et al., we explore the use of the MIRA algorithm of Crammer et al. as an alternative to MERT. We first show that by parallel processing and exploiting more of the parse forest, we can obtain results using MIRA that match or surpass MERT in terms of both translation quality and computational cost. We then test the method on two classes of features that address deficiencies in the Hiero hierarchical phrasebased model: first, we simultaneously train a large number of Marton and Resnik's soft syntactic constraints, and, second, we introduce a novel structural distortion model. In both cases we obtain significant improvements in translation performance. Optimizing them in combination, for a total of 56 feature weights, we improve performance by 2.6 B\uf76c\uf765\uf775 on a subset of the NIST 2006 Arabic-English evaluation data.",
"pdf_parse": {
"paper_id": "D08-1024",
"_pdf_hash": "",
"abstract": [
{
"text": "Minimum-error-rate training (MERT) is a bottleneck for current development in statistical machine translation because it is limited in the number of weights it can reliably optimize. Building on the work of Watanabe et al., we explore the use of the MIRA algorithm of Crammer et al. as an alternative to MERT. We first show that by parallel processing and exploiting more of the parse forest, we can obtain results using MIRA that match or surpass MERT in terms of both translation quality and computational cost. We then test the method on two classes of features that address deficiencies in the Hiero hierarchical phrasebased model: first, we simultaneously train a large number of Marton and Resnik's soft syntactic constraints, and, second, we introduce a novel structural distortion model. In both cases we obtain significant improvements in translation performance. Optimizing them in combination, for a total of 56 feature weights, we improve performance by 2.6 B\uf76c\uf765\uf775 on a subset of the NIST 2006 Arabic-English evaluation data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Since its introduction by Och (2003) , minimum error rate training (MERT) has been widely adopted for training statistical machine translation (MT) systems. However, MERT is limited in the number of feature weights that it can optimize reliably, with folk estimates of the limit ranging from 15 to 30 features.",
"cite_spans": [
{
"start": 26,
"end": 36,
"text": "Och (2003)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One recent example of this limitation is a series of experiments by Marton and Resnik (2008) , in which they added syntactic features to Hiero (Chiang, 2005; Chiang, 2007) , which ordinarily uses no linguistically motivated syntactic information. Each of their new features rewards or punishes a derivation depending on how similar or dissimilar it is to a syntactic parse of the input sentence. They found that in order to obtain the greatest improvement, these features had to be specialized for particular syntactic categories and weighted independently. Not being able to optimize them all at once using MERT, they resorted to running MERT many times in order to test different combinations of features. But it would have been preferable to use a training method that can optimize the features all at once.",
"cite_spans": [
{
"start": 68,
"end": 92,
"text": "Marton and Resnik (2008)",
"ref_id": "BIBREF19"
},
{
"start": 143,
"end": 157,
"text": "(Chiang, 2005;",
"ref_id": "BIBREF5"
},
{
"start": 158,
"end": 171,
"text": "Chiang, 2007)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "There has been much work on improving MERT's performance (Duh and Kirchoff, 2008; Smith and Eisner, 2006; Cer et al., 2008) , or on replacing MERT wholesale (Turian et al., 2007; Blunsom et al., 2008) . This paper continues a line of research on online discriminative training (Tillmann and Zhang, 2006; Liang et al., 2006; Arun and Koehn, 2007) , extending that of Watanabe et al. (2007) , who use the Margin Infused Relaxed Algorithm (MIRA) due to Crammer et al. (2003; . Our guiding principle is practicality: like Watanabe et al., we train on a small tuning set comparable in size to that used by MERT, but by parallel processing and exploiting more of the parse forest, we obtain results using MIRA that match or surpass MERT in terms of both translation quality and computational cost on a large-scale translation task.",
"cite_spans": [
{
"start": 57,
"end": 81,
"text": "(Duh and Kirchoff, 2008;",
"ref_id": "BIBREF12"
},
{
"start": 82,
"end": 105,
"text": "Smith and Eisner, 2006;",
"ref_id": "BIBREF25"
},
{
"start": 106,
"end": 123,
"text": "Cer et al., 2008)",
"ref_id": "BIBREF3"
},
{
"start": 157,
"end": 178,
"text": "(Turian et al., 2007;",
"ref_id": "BIBREF30"
},
{
"start": 179,
"end": 200,
"text": "Blunsom et al., 2008)",
"ref_id": "BIBREF1"
},
{
"start": 277,
"end": 303,
"text": "(Tillmann and Zhang, 2006;",
"ref_id": "BIBREF29"
},
{
"start": 304,
"end": 323,
"text": "Liang et al., 2006;",
"ref_id": "BIBREF17"
},
{
"start": 324,
"end": 345,
"text": "Arun and Koehn, 2007)",
"ref_id": "BIBREF0"
},
{
"start": 366,
"end": 388,
"text": "Watanabe et al. (2007)",
"ref_id": "BIBREF31"
},
{
"start": 450,
"end": 471,
"text": "Crammer et al. (2003;",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Taking this further, we test MIRA on two classes of features that make use of syntactic information and hierarchical structure. First, we generalize Marton and Resnik's (2008) soft syntactic constraints by training all of them simultaneously; and, second, we introduce a novel structural distortion model. We obtain significant improvements in both cases, and further large improvements when the two feature sets are combined.",
"cite_spans": [
{
"start": 160,
"end": 175,
"text": "Resnik's (2008)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The paper proceeds as follows. We describe our training algorithm in section 2; our generalization of Marton and Resnik's soft syntactic constraints in section 3; our novel structural distortion features in section 4; and experimental results in section 5.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The translation model is a standard linear model (Och and Ney, 2002) , which we train using MIRA (Crammer and Singer, 2003; Crammer et al., 2006) , following Watanabe et al. (2007) . We describe the basic algorithm first and then progressively refine it.",
"cite_spans": [
{
"start": 49,
"end": 68,
"text": "(Och and Ney, 2002)",
"ref_id": "BIBREF22"
},
{
"start": 97,
"end": 123,
"text": "(Crammer and Singer, 2003;",
"ref_id": "BIBREF8"
},
{
"start": 124,
"end": 145,
"text": "Crammer et al., 2006)",
"ref_id": "BIBREF9"
},
{
"start": 158,
"end": 180,
"text": "Watanabe et al. (2007)",
"ref_id": "BIBREF31"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Learning algorithm",
"sec_num": "2"
},
{
"text": "Let e, by abuse of notation, stand for both output strings and their derivations. We represent the feature vector for derivation e as h(e). Initialize the feature weights w. Then, repeatedly:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "\u2022 Select a batch of input sentences f 1 , . . . , f m .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "\u2022 Decode each f i to obtain a set of hypothesis translations e i1 , . . . , e in .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "\u2022 For each i, select one of the e i j to be the oracle translation e * i , by a criterion described below. Let \u2206h i j = h(e * i ) \u2212 h(e i j ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "\u2022 For each e i j , compute the loss i j , which is some measure of how bad it would be to guess e i j instead of e * i .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "\u2022 Update w to the value of w that minimizes:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "1 2 w \u2212 w 2 + C m i=1 max 1\u2264 j\u2264n ( i j \u2212 \u2206h i j \u2022 w ) (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "where we set C = 0.01. The first term means that we want w to be close to w, and second term (the generalized hinge loss) means that we want w to score e * i higher than each e i j by a margin at least as wide as the loss i j .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "When training is finished, the weight vectors from all iterations are averaged together. (If multiple passes through the training data are made, we only average the weight vectors from the last pass.) The technique of averaging was introduced in the context of perceptrons as an approximation to taking a vote among all the models traversed during training, and has been shown to work well in practice (Freund and Schapire, 1999; Collins, 2002) . We follow McDonald et al. (2005) in applying this technique to MIRA.",
"cite_spans": [
{
"start": 402,
"end": 429,
"text": "(Freund and Schapire, 1999;",
"ref_id": "BIBREF13"
},
{
"start": 430,
"end": 444,
"text": "Collins, 2002)",
"ref_id": "BIBREF7"
},
{
"start": 457,
"end": 479,
"text": "McDonald et al. (2005)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "Note that the objective (1) is not the same as that used by Watanabe et al.; rather, it is the same as that used by Crammer and Singer (2003) and related to that of Taskar et al. (2005) . We solve this optimization problem using a variant of sequential minimal optimization (Platt, 1998) : for each i, initialize \u03b1 i j = C for a single value of j such that e i j = e * i , and initialize \u03b1 i j = 0 for all other values of j. Then, repeatedly choose a sentence i and a pair of hypotheses j, j , and let",
"cite_spans": [
{
"start": 60,
"end": 76,
"text": "Watanabe et al.;",
"ref_id": null
},
{
"start": 116,
"end": 141,
"text": "Crammer and Singer (2003)",
"ref_id": "BIBREF8"
},
{
"start": 165,
"end": 185,
"text": "Taskar et al. (2005)",
"ref_id": "BIBREF28"
},
{
"start": 274,
"end": 287,
"text": "(Platt, 1998)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "w \u2190 w + \u03b4(\u2206h i j \u2212 \u2206h i j ) (2) \u03b1 i j \u2190 \u03b1 i j + \u03b4 (3) \u03b1 i j \u2190 \u03b1 i j \u2212 \u03b4",
"eq_num": "(4)"
}
],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03b4 = clip [\u2212\u03b1 i j ,\u03b1 i j ] ( i j \u2212 i j ) \u2212 (\u2206h i j \u2212 \u2206h i j ) \u2022 w \u2206h i j \u2212 \u2206h i j 2",
"eq_num": "(5)"
}
],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "where the function clip [x,y] (z) gives the closest number to z in the interval [x, y].",
"cite_spans": [
{
"start": 24,
"end": 29,
"text": "[x,y]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Basic algorithm",
"sec_num": "2.1"
},
{
"text": "Assuming B\uf76c\uf765\uf775 as the evaluation criterion, the loss i j of e i j relative to e * i should be related somehow to the difference between their B\uf76c\uf765\uf775 scores. However, B\uf76c\uf765\uf775 was not designed to be used on individual sentences; in general, the highest-B\uf76c\uf765\uf775 translation of a sentence depends on what the other sentences in the test set are. Sentence-level approximations to B\uf76c\uf765\uf775 exist (Lin and Och, 2004; Liang et al., 2006 ), but we found it most effective to perform B\uf76c\uf765\uf775 computations in the context of a set O of previously-translated sentences, following Watanabe et al. (2007) . However, we don't try to accumulate translations for the entire dataset, but simply maintain an exponentially-weighted moving average of previous translations.",
"cite_spans": [
{
"start": 377,
"end": 396,
"text": "(Lin and Och, 2004;",
"ref_id": "BIBREF18"
},
{
"start": 397,
"end": 415,
"text": "Liang et al., 2006",
"ref_id": "BIBREF17"
},
{
"start": 551,
"end": 573,
"text": "Watanabe et al. (2007)",
"ref_id": "BIBREF31"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Loss function",
"sec_num": "2.2"
},
{
"text": "More precisely: For an input sentence f, let e be some hypothesis translation and let {r k } be the set of reference translations for f. Let c(e; {r k }), or simply c(e) for short, be the vector of the following counts: |e|, the effective reference length min k |r k |, and, for 1 \u2264 n \u2264 4, the number of n-grams in e, and the number of n-gram matches between e and {r k }. These counts are sufficient to calculate a B\uf76c\uf765\uf775 score, which we write as B\uf76c\uf765\uf775(c(e)). The pseudo-document O is an exponentially-weighted moving average of these vectors. That is, for each training sentence, let\u00ea be the 1-best translation; after processing the sentence, we update O, and its input length O f :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Loss function",
"sec_num": "2.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "O \u2190 0.9(O + c(\u00ea)) (6) O f \u2190 0.9(O f + |f|)",
"eq_num": "(7)"
}
],
"section": "Loss function",
"sec_num": "2.2"
},
{
"text": "We can then calculate the B\uf76c\uf765\uf775 score of hypotheses e in the context of O. But the larger O is, the smaller the impact the current sentence will have on the B\uf76c\uf765\uf775 score. To correct for this, and to bring the loss function roughly into the same range as typical margins, we scale the B\uf76c\uf765\uf775 score by the size of the input:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Loss function",
"sec_num": "2.2"
},
{
"text": "B(e; f, {r k }) = (O f + |f|) \u00d7 B\uf76c\uf765\uf775(O + c(e; {r k })) (8)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Loss function",
"sec_num": "2.2"
},
{
"text": "which we also simply write as B(e). Finally, the loss function is defined to be:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Loss function",
"sec_num": "2.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "i j = B(e * i ) \u2212 B(e i j )",
"eq_num": "(9)"
}
],
"section": "Loss function",
"sec_num": "2.2"
},
{
"text": "We now describe the selection of e * . We know of three approaches in previous work. The first is to force the decoder to output the reference sentence exactly, and select the derivation with the highest model score, which Liang et al. (2006) call bold updating. The second uses the decoder to search for the highest-B\uf76c\uf765\uf775 translation (Tillmann and Zhang, 2006) , which Arun and Koehn (2007) ",
"cite_spans": [
{
"start": 223,
"end": 242,
"text": "Liang et al. (2006)",
"ref_id": "BIBREF17"
},
{
"start": 334,
"end": 360,
"text": "(Tillmann and Zhang, 2006)",
"ref_id": "BIBREF29"
},
{
"start": 369,
"end": 390,
"text": "Arun and Koehn (2007)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "\u00b5 = 0 \u00b5 = 0.5 \u00b5 = 1 \u00b5 = \u221e",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "Figure 1: Scatter plot of 10-best unique translations of a single sentence obtained by forest rescoring using various values of \u00b5 in equation 11.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "the search for the highest-B\uf76c\uf765\uf775 translation is limited to the n translations with the highest model score, where n must be determined experimentally.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "Here, we introduce a new oracle-translation selection method, formulating the intuition behind local updating as an optimization problem:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "e * = arg max e (B(e) + h(e) \u2022 w)",
"eq_num": "(10)"
}
],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "Instead of choosing the highest-B\uf76c\uf765\uf775 translation from an n-best list, we choose the translation that maximizes a combination of (approximate) B\uf76c\uf765\uf775 and the model. We can also interpret (10) in the following way: we want e * to be the max-B\uf76c\uf765\uf775 translation, but we also want to minimize (1). So we balance these two criteria against each other:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "e * = arg max e (B(e) \u2212 \u00b5(B(e) \u2212 h(e) \u2022 w))",
"eq_num": "(11)"
}
],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "where (B(e) \u2212 h(e) \u2022 w) is that part of (1) that depends on e * , and \u00b5 is a parameter that controls how much we are willing to allow some translations to have higher B\uf76c\uf765\uf775 than e * if we can better minimize (1). Setting \u00b5 = 0 would reduce to max-B\uf76c\uf765\uf775 updating; setting \u00b5 = \u221e would never update w at all. Setting \u00b5 = 0.5 reduces to equation (10). Figure 1 shows the 10-best unique translations for a single input sentence according to equation 11under various settings of \u00b5. The points at far right are the translations that are scored highest according to the model. The \u00b5 = 0 points in the upper-left corner are typical of oracle translations that would be selected under the max-B\uf76c\uf765\uf775 policy: they indeed have a very high B\uf76c\uf765\uf775 score, but are far removed from the translations preferred by the model; thus they would cause violent updates to w. Local updating would select the topmost point labeled \u00b5 = 1. Our scheme would select one of the \u00b5 = 0.5 points, which have B\uf76c\uf765\uf775 scores almost as high as the max-B\uf76c\uf765\uf775 translations, yet are not very far from the translations preferred by the model.",
"cite_spans": [],
"ref_spans": [
{
"start": 346,
"end": 354,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Oracle translations",
"sec_num": "2.3"
},
{
"text": "What is the set {e i j } of translation hypotheses? Ideally we would let it be the set of all possible translations, and let the objective function (1) take all of them into account. This is the approach taken by Taskar et al. (2004) , but their approach assumes that the loss function can be decomposed into local loss functions. Since our loss function cannot be so decomposed, we select:",
"cite_spans": [
{
"start": 213,
"end": 233,
"text": "Taskar et al. (2004)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "\u2022 the 10-best translations according to the model; we then rescore the forest to obtain",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "\u2022 the 10-best translations according to equation (11) with \u00b5 = 0.5, the first of which is the oracle translation, and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "\u2022 the 10-best translations with \u00b5 = \u221e, to serve as negative examples.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "The last case is what Crammer et al. (2006) call max-loss updating (where \"loss\" refers to the generalized hinge loss) and Taskar et al. (2005) call lossaugmented inference. The rationale here is that since the objective (1) tries to minimize max",
"cite_spans": [
{
"start": 22,
"end": 43,
"text": "Crammer et al. (2006)",
"ref_id": "BIBREF9"
},
{
"start": 123,
"end": 143,
"text": "Taskar et al. (2005)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "j ( i j \u2212 \u2206h i j \u2022 w )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": ", we should include the translations that have the highest ( i j \u2212 \u2206h i j \u2022 w) in order to approximate the effect of using the whole forest. See Figure 1 again for an illustration of the hypotheses selected for a single sentence. The max-B\uf76c\uf765\uf775 points in the upper left are not included (and would have no effect even if they were included). The \u00b5 = \u221e points in the lower-right are the negative examples: they are poor translations that are scored too high by the model, and the learning algorithm attempts to shift them to the left.",
"cite_spans": [],
"ref_spans": [
{
"start": 145,
"end": 153,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "To perform the forest rescoring, we need to use several approximations, since an exact search for B\uf76c\uf765\uf775-optimal translations is NP-hard (Leusch et al., 2008) . For every derivation e in the forest, we calculate a vector c(e) of counts as in Section 2.2 except using unclipped counts of n-gram matches (Dreyer et al., 2007) , that is, the number of matches for an ngram can be greater than the number of occurrences of the n-gram in any reference translation. This can be done efficiently by calculating c for every hyperedge (rule application) in the forest:",
"cite_spans": [
{
"start": 135,
"end": 156,
"text": "(Leusch et al., 2008)",
"ref_id": "BIBREF16"
},
{
"start": 300,
"end": 321,
"text": "(Dreyer et al., 2007)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "\u2022 the number of output words generated by the rule",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "\u2022 the effective reference length scaled by the fraction of the input sentence consumed by the rule",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "\u2022 the number of n-grams formed by the application of the rule (1 \u2264 n \u2264 4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "\u2022 the (unclipped) number of n-gram matches formed by the application of the rule (1 \u2264 n \u2264 4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "We keep track of n-grams using the same scheme used to incorporate an n-gram language model into the decoder (Wu, 1996; Chiang, 2007) . To find the best derivation in the forest, we traverse it bottom-up as usual, and for every set of alternative subtranslations, we select the one with the highest score. But here a rough approximation lurks, because we need to calculate B on the nodes of the forest, but B does not have the optimal substructure property, i.e., the optimal score of a parent node cannot necessarily be calculated from the optimal scores of its children. Nevertheless, we find that this rescoring method is good enough for generating high-B\uf76c\uf765\uf775 oracle translations and low-B\uf76c\uf765\uf775 negative examples.",
"cite_spans": [
{
"start": 109,
"end": 119,
"text": "(Wu, 1996;",
"ref_id": "BIBREF32"
},
{
"start": 120,
"end": 133,
"text": "Chiang, 2007)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Selecting hypothesis translations",
"sec_num": "2.4"
},
{
"text": "One convenient property of MERT is that it is embarrassingly parallel: we decode the entire tuning set sending different sentences to different processors, and during optimization of feature weights, different random restarts can be sent to different processors. In order to make MIRA comparable in efficiency to MERT, we must parallelize it. But with an online learning algorithm, parallelization requires a little more coordination. We run MIRA on each processor simultaneously, with each maintaining its own weight vector. A master process distributes different sentences from the tuning set to each of the processors; when each processor finishes decoding a sentence, it transmits the resulting hypotheses, with their losses, to all the other processors and receives any hypotheses waiting from other processors. Those hypotheses were generated from different weight vectors, but can still provide useful information. The sets of hypotheses thus collected are then processed as one batch. When the whole training process is finished, we simply average all the weight vectors from all the processors.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parallelization",
"sec_num": "2.5"
},
{
"text": "Having described our training algorithm, which includes several practical improvements to Watanabe et al.'s usage of MIRA, we proceed in the remainder of the paper to demonstrate the utility of the our training algorithm on models with large numbers of structurally sensitive features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parallelization",
"sec_num": "2.5"
},
{
"text": "The first features we explore are based on a line of research introduced by Chiang (2005) and improved on by Marton and Resnik (2008) . A hierarchical phrase-based translation model is based on synchronous context-free grammar, but does not normally use any syntactic information derived from linguistic knowledge or treebank data: it uses translation rules that span any string of words in the input sentence, without regard for parser-defined syntactic constituency boundaries. Chiang (2005) experimented with a constituency feature that rewarded rules whose source language side exactly spans a syntactic constituent according to the output of an external source-language parser. This feature can be viewed as a soft syntactic constraint: it biases the model toward translations that respect syntactic structure, but does not force it to use them. However, this more syntactically aware model, when tested in Chinese-English translation, did not improve translation performance.",
"cite_spans": [
{
"start": 76,
"end": 89,
"text": "Chiang (2005)",
"ref_id": "BIBREF5"
},
{
"start": 109,
"end": 133,
"text": "Marton and Resnik (2008)",
"ref_id": "BIBREF19"
},
{
"start": 480,
"end": 493,
"text": "Chiang (2005)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Soft syntactic constraints",
"sec_num": "3"
},
{
"text": "Recently, Marton and Resnik (2008) revisited the idea of constituency features, and succeeded in showing that finer-grained soft syntactic constraints yield substantial improvements in B\uf76c\uf765\uf775 score for both Chinese-English and Arabic-English translation. In addition to adding separate features for dif-ferent syntactic nonterminals, they introduced a new type of constraint that penalizes rules when the source language side crosses the boundaries of a source syntactic constituent, as opposed to simply rewarding rules when they are consistent with the source-language parse tree.",
"cite_spans": [
{
"start": 10,
"end": 34,
"text": "Marton and Resnik (2008)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Soft syntactic constraints",
"sec_num": "3"
},
{
"text": "Marton and Resnik optimized their features' weights using MERT. But since MERT does not scale well to large numbers of feature weights, they were forced to test individual features and manually selected feature combinations each in a separate model. Although they showed gains in translation performance for several such models, many larger, potentially better feature combinations remained unexplored. Moreover, the best-performing feature subset was different for the two language pairs, suggesting that this labor-intensive feature selection process would have to be repeated for each new language pair.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Soft syntactic constraints",
"sec_num": "3"
},
{
"text": "Here, we use MIRA to optimize Marton and Resnik's finer-grained single-category features all at once. We define below two sets of features, a coarsegrained class that combines several constituency categories, and a fine-grained class that puts different categories into different features. Both kinds of features were used by Marton and Resnik, but only a few at a time. Crucially, our training algorithm provides the ability to train all the fine-grained features, a total of 34 feature weights, simultaneously.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Soft syntactic constraints",
"sec_num": "3"
},
{
"text": "As the basis for coarsegrained syntactic features, we selected the following nonterminal labels based on their frequency in the tuning data, whether they frequently cover a span of more than one word, and whether they represent linguistically relevant constituents: NP, PP, S, VP, SBAR, ADJP, ADVP, and QP. We define two new features, one which fires when a rule's source side span in the input sentence matches any of the above-mentioned labels in the input parse, and another which fires when a rule's source side span crosses a boundary of one of these labels (e.g., its source side span only partially covers the words in a VP subtree, and it also covers some or all or the words outside the VP subtree). These two features are equivalent to Marton and Resnik's XP = and XP + feature combinations, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Coarse-grained features",
"sec_num": null
},
{
"text": "Fine-grained features We selected the following nonterminal labels that appear more than 100 times in the tuning data: NP, PP, S, VP, SBAR, ADJP, WHNP, PRT, ADVP, PRN, and QP. The labels that were excluded were parts of speech, nonconstituent labels like FRAG, or labels that occurred only two or three times. For each of these labels X, we added a separate feature that fires when a rule's source side span in the input sentence matches X, and a second feature that fires when a span crosses a boundary of X. These features are similar to Marton and Resnik's X = and X + , except that our set includes features for WHNP, PRT, and PRN.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Coarse-grained features",
"sec_num": null
},
{
"text": "In addition to parser-based syntactic constraints, which were introduced in prior work, we introduce a completely new set of features aimed at improving the modeling of reordering within Hiero. Again, the feature definition gives rise to a larger number of features than one would expect to train successfully using MERT.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Structural distortion features",
"sec_num": "4"
},
{
"text": "In a phrase-based model, reordering is performed both within phrase pairs and by the phrasereordering model. Both mechanisms are able to learn that longer-distance reorderings are more costly than shorter-distance reorderings: phrase pairs, because phrases that involve more extreme reorderings will (presumably) have a lower count in the data, and phrase reordering, because models are usually explicitly dependent on distance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Structural distortion features",
"sec_num": "4"
},
{
"text": "By contrast, in a hierarchical model, all reordering is performed by a single mechanism, the rules of the grammar. In some cases, the model will be able to learn a preference for shorter-distance reorderings, as in a phrase-based system, but in the case of a word being reordered across a nonterminal, or two nonterminals being reordered, there is no dependence in the model on the size of the nonterminal or nonterminals involved in reordering.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Structural distortion features",
"sec_num": "4"
},
{
"text": "So, for example, if we have rules",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Structural distortion features",
"sec_num": "4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "X \u2192 (il dit X 1 , he said X 1 ) (12) X \u2192 (il dit X 1 , X 1 he said)",
"eq_num": "(13)"
}
],
"section": "Structural distortion features",
"sec_num": "4"
},
{
"text": "we might expect that rule (12) is more common in general, but that rule (13) becomes more and more To address this defect, we can classify every nonterminal pair occurring on the right-hand side of each grammar rule as \"reordered\" or \"not reordered\", that is, whether it intersects any other word alignment link or nonterminal pair (see Figure 2) . We then define coarse-and fine-grained versions of the structural distortion model. Coarse-grained features Let R be a binaryvalued random variable that indicates whether a nonterminal occurrence is reordered, and let S be an integer-valued random variable that indicates how many source words are spanned by the nonterminal occurrence. We can estimate P(R | S ) via relativefrequency estimation from the rules as they are extracted from the parallel text, and incorporate this probability as a new feature of the model. Fine-grained features A difficulty with the coarse-grained reordering features is that the grammar extraction process finds overlapping rules in the training data and might not give a sensible probability estimate; moreover, reordering statistics from the training data might not carry over perfectly into the translation task (in particular, the training data may have some very freely-reordering translations that one might want to avoid replicating in translation). As an alternative, we introduce a fine-grained version of our distortion model that can be trained directly in the translation task as follows: define a separate binary feature for each value of (R, S ), where R is as above and S \u2208 { , 1, . . . , 9, \u226510} and means any size. For example, if a nonterminal with span 11 has its contents reordered, then the features (true, \u226510) and (true, ) would both fire. Grouping all sizes of 10 or more into a single feature is designed to avoid overfitting.",
"cite_spans": [],
"ref_spans": [
{
"start": 337,
"end": 346,
"text": "Figure 2)",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Structural distortion features",
"sec_num": "4"
},
{
"text": "Again, using MIRA makes it practical to train with the full fine-grained feature set-coincidentally also a total of 34 features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Structural distortion features",
"sec_num": "4"
},
{
"text": "We now describe our experiments to test MIRA and our features, the soft-syntactic constraints and the structural distortion features, on an Arabic-English translation task. It is worth noting that this experimentation is on a larger scale than Watanabe et al.'s (2007) , and considerably larger than Marton and Resnik's (2008) .",
"cite_spans": [
{
"start": 244,
"end": 268,
"text": "Watanabe et al.'s (2007)",
"ref_id": null
},
{
"start": 300,
"end": 326,
"text": "Marton and Resnik's (2008)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment and results",
"sec_num": "5"
},
{
"text": "The baseline model was Hiero with the following baseline features (Chiang, 2005; Chiang, 2007) :",
"cite_spans": [
{
"start": 66,
"end": 80,
"text": "(Chiang, 2005;",
"ref_id": "BIBREF5"
},
{
"start": 81,
"end": 94,
"text": "Chiang, 2007)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "\u2022 two language models",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "\u2022 phrase translation probabilities p( f | e) and p(e | f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "\u2022 lexical weighting in both directions (Koehn et al., 2003) \u2022 word penalty",
"cite_spans": [
{
"start": 39,
"end": 59,
"text": "(Koehn et al., 2003)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "\u2022 penalties for:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "automatically extracted rules identity rules (translating a word into itself) two classes of number/name translation rules glue rules The probability features are base-100 logprobabilities.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "The rules were extracted from all the allowable parallel text from the NIST 2008 evaluation (152+175 million words of Arabic+English), aligned by IBM Model 4 using GIZA++ (union of both directions). Hierarchical rules were extracted from the most in-domain corpora (4.2+5.4 million words) and phrases were extracted from the remainder. We trained the coarse-grained distortion model on 10,000 sentences of the training data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "Two language models were trained, one on data similar to the English side of the parallel text and one on 2 billion words of English. Both were 5gram models with modified Kneser-Ney smoothing, lossily compressed using a perfect-hashing scheme similar to that of Talbot and Brants (2008) but using minimal perfect hashing (Botelho et al., 2005) .",
"cite_spans": [
{
"start": 262,
"end": 286,
"text": "Talbot and Brants (2008)",
"ref_id": "BIBREF26"
},
{
"start": 321,
"end": 343,
"text": "(Botelho et al., 2005)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "We partitioned the documents of the NIST 2004 (newswire) and 2005 Arabic-English evaluation data into a tuning set (1178 sentences) and a development set (1298 sentences). The test data was the NIST 2006 Arabic-English evaluation data (NIST part, newswire and newsgroups, 1529 sentences).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "To obtain syntactic parses for this data, we tokenized it according to the Arabic Treebank standard using AMIRA (Diab et al., 2004) , parsed it with the Stanford parser (Klein and Manning, 2003) , and then forced the trees back into the MT system's tokenization. 1 We ran both MERT and MIRA on the tuning set using 20 parallel processors. We stopped MERT when the score on the tuning set stopped increasing, as is common practice, and for MIRA, we used the development set to decide when to stop training. 2 In our runs, MERT took an average of 9 passes through the tuning set and MIRA took an average of 8 passes. (For comparison, Watanabe et al. report decoding their tuning data of 663 sentences 80 times.) Table 1 shows the results of our experiments with the training methods and features described above. All significance testing was performed against the first line (MERT baseline) using paired bootstrap resampling (Koehn, 2004) .",
"cite_spans": [
{
"start": 112,
"end": 131,
"text": "(Diab et al., 2004)",
"ref_id": "BIBREF10"
},
{
"start": 169,
"end": 194,
"text": "(Klein and Manning, 2003)",
"ref_id": "BIBREF14"
},
{
"start": 263,
"end": 264,
"text": "1",
"ref_id": null
},
{
"start": 506,
"end": 507,
"text": "2",
"ref_id": null
},
{
"start": 923,
"end": 936,
"text": "(Koehn, 2004)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 710,
"end": 717,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "5.1"
},
{
"text": "First of all, we find that MIRA is competitive with MERT when both use the baseline feature set. In- Table 1 : Comparison of MERT and MIRA on various feature sets. Key: # = number of features; nw = newswire, ng = newsgroups; + or ++ = significantly better than MERT baseline (p < 0.05 or p < 0.01, respectively), \u2212 = significantly worse than MERT baseline (p < 0.05).",
"cite_spans": [],
"ref_spans": [
{
"start": 101,
"end": 108,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "deed, the MIRA system scores significantly higher on the test set; but if we break the test set down by genre, we see that the MIRA system does slightly worse on newswire and better on newsgroups. (This is largely attributable to the fact that the MIRA translations tend to be longer than the MERT translations, and the newsgroup references are also relatively longer than the newswire references.) When we add more features to the model, the two training methods diverge more sharply. When training with MERT, the coarse-grained pair of syntax features yields a small improvement, but the finegrained syntax features do not yield any further improvement. By contrast, when the fine-grained features are trained using MIRA, they yield substantial improvements. We observe similar behavior for the structural distortion features: MERT is not able to take advantage of the finer-grained features, but MIRA is. Finally, using MIRA to combine both classes of features, 56 in all, produces the largest improvement, 2.6 B\uf76c\uf765\uf775 points over the MERT baseline on the full test set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "We also tested some of the differences between our training method and Watanabe et al.'s (2007) ; the results are shown in Table 2 . Compared with local updating (line 2), our method of selecting the oracle translation and negative examples does better by 0.5 B\uf76c\uf765\uf775 points on the development data. Using lossaugmented inference to add negative examples to local updating (line 3) does not appear to help. Nevertheless, the negative examples are important: for if we use our method for selecting the oracle translation without the additional negative examples (line 4), the algorithm fails, generating very long translations and unable to find a weight setting to shorten them. It appears, then, that the additional negative examples enable the algorithm to reliably learn from the enhanced oracle translations. Finally, we compared our parallelization method against a simpler method in which all processors learn independently and their weight vectors are all averaged together (line 5). We see that sharing information among the processors makes a significant difference.",
"cite_spans": [
{
"start": 71,
"end": 95,
"text": "Watanabe et al.'s (2007)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 123,
"end": 130,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "In this paper, we have brought together two existing lines of work: the training method of Watanabe et al. (2007) , and the models of Chiang (2005) and Marton and Resnik (2008) . Watanabe et al.'s work showed that large-margin training with MIRA can be made feasible for state-of-the-art MT systems by using a manageable tuning set; we have demonstrated that parallel processing and exploiting more of the parse forest improves MIRA's performance and that, even using the same set of features, MIRA's performance compares favorably to MERT in terms of both translation quality and computational cost. Marton and Resnik (2008) showed that it is possible to improve translation in a data-driven framework by incorporating source-side syntactic analysis in the form of soft syntactic constraints. This work joins a growing body of work demonstrating the utility of syntactic information in statistical MT. In the area of source-side syntax, recent research has continued to improve tree-to-string translation models, soften the constraints of the input tree in various ways (Mi et al., 2008; Zhang et al., 2008) , and extend phrase-based translation with sourceside soft syntactic constraints (Cherry, 2008) . All this work shows strong promise, but Marton and Resnik's soft syntactic constraint approach is particularly appealing because it can be used unobtrusively with any hierarchically-structured translation model. Here, we have shown that using MIRA to weight all the constraints at once removes the crucial drawback of the approach, the problem of feature selection.",
"cite_spans": [
{
"start": 91,
"end": 113,
"text": "Watanabe et al. (2007)",
"ref_id": "BIBREF31"
},
{
"start": 134,
"end": 147,
"text": "Chiang (2005)",
"ref_id": "BIBREF5"
},
{
"start": 152,
"end": 176,
"text": "Marton and Resnik (2008)",
"ref_id": "BIBREF19"
},
{
"start": 601,
"end": 625,
"text": "Marton and Resnik (2008)",
"ref_id": "BIBREF19"
},
{
"start": 1071,
"end": 1088,
"text": "(Mi et al., 2008;",
"ref_id": "BIBREF21"
},
{
"start": 1089,
"end": 1108,
"text": "Zhang et al., 2008)",
"ref_id": "BIBREF33"
},
{
"start": 1190,
"end": 1204,
"text": "(Cherry, 2008)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "Finally, we have introduced novel structural distortion features to fill a notable gap in the hierarchical phrase-based approach. By capturing how reordering depends on constituent length, these features improve translation quality significantly. In sum, we have shown that removing the bottleneck of MERT opens the door to many possibilities for better translation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "The only notable consequence this had for our experimentation is that proclitic Arabic prepositions were fused onto the first word of their NP object, so that the PP and NP brackets were coextensive.2 We chose this policy for MIRA to avoid overfitting. However, we could have used the tuning set for this purpose, just as with MERT: in none of our runs would this change have made more than a 0.2 B\uf76c\uf765\uf775 difference on the development set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Thanks to Michael Bloodgood for performing initial simulations of parallelized perceptron training. Thanks also to John DeNero, Kevin Knight, Daniel Marcu, and Fei Sha for valuable discussions and suggestions. This research was supported in part by DARPA contract HR0011-06-C-0022 under subcontract to BBN Technologies and HR0011-06-02-001 under subcontract to IBM.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Online learning methods for discriminative training of phrase based statistical machine translation",
"authors": [
{
"first": "Abhishek",
"middle": [],
"last": "Arun",
"suffix": ""
},
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2007,
"venue": "Proc. MT Summit XI",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Abhishek Arun and Philipp Koehn. 2007. Online learning methods for discriminative training of phrase based statistical machine translation. In Proc. MT Summit XI.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A discriminative latent variable model for statistical machine translation",
"authors": [
{
"first": "Phil",
"middle": [],
"last": "Blunsom",
"suffix": ""
},
{
"first": "Trevor",
"middle": [],
"last": "Cohn",
"suffix": ""
},
{
"first": "Miles",
"middle": [],
"last": "Osborne",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Phil Blunsom, Trevor Cohn, and Miles Osborne. 2008. A discriminative latent variable model for statistical ma- chine translation. In Proc. ACL-08: HLT.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "A practical minimal perfect hashing method",
"authors": [
{
"first": "C",
"middle": [],
"last": "Fabiano",
"suffix": ""
},
{
"first": "Yoshiharu",
"middle": [],
"last": "Botelho",
"suffix": ""
},
{
"first": "Nivio",
"middle": [],
"last": "Kohayakawa",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ziviani",
"suffix": ""
}
],
"year": 2005,
"venue": "4th International Workshop on Efficient and Experimental Algorithms (WEA05)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fabiano C. Botelho, Yoshiharu Kohayakawa, and Nivio Ziviani. 2005. A practical minimal perfect hashing method. In 4th International Workshop on Efficient and Experimental Algorithms (WEA05).",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Regularization and search for minimum error rate training",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Cer",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Jurafsky",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. Third Workshop on Statistical Machine Translation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Cer, Daniel Jurafsky, and Christopher D. Man- ning. 2008. Regularization and search for minimum error rate training. In Proc. Third Workshop on Statis- tical Machine Translation.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Cohesive phrase-based decoding for statistical machine translation",
"authors": [
{
"first": "Colin",
"middle": [],
"last": "Cherry",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Colin Cherry. 2008. Cohesive phrase-based decoding for statistical machine translation. In Proc. ACL-08: HLT.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A hierarchical phrase-based model for statistical machine translation",
"authors": [
{
"first": "David",
"middle": [],
"last": "Chiang",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Chiang. 2005. A hierarchical phrase-based model for statistical machine translation. In Proc. ACL 2005.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Hierarchical phrase-based translation",
"authors": [
{
"first": "David",
"middle": [],
"last": "Chiang",
"suffix": ""
}
],
"year": 2007,
"venue": "Computational Linguistics",
"volume": "",
"issue": "2",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Chiang. 2007. Hierarchical phrase-based transla- tion. Computational Linguistics, 33(2).",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Discriminative training methods for Hidden Markov Models: Theory and experiments with perceptron algorithms",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Collins",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Collins. 2002. Discriminative training methods for Hidden Markov Models: Theory and experiments with perceptron algorithms. In Proc. EMNLP 2002.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Ultraconservative online algorithms for multiclass problems",
"authors": [
{
"first": "Koby",
"middle": [],
"last": "Crammer",
"suffix": ""
},
{
"first": "Yoram",
"middle": [],
"last": "Singer",
"suffix": ""
}
],
"year": 2003,
"venue": "Journal of Machine Learning Research",
"volume": "3",
"issue": "",
"pages": "951--991",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Koby Crammer and Yoram Singer. 2003. Ultraconserva- tive online algorithms for multiclass problems. Jour- nal of Machine Learning Research, 3:951-991.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Shai Shalev-Shwartz, and Yoram Singer",
"authors": [
{
"first": "Koby",
"middle": [],
"last": "Crammer",
"suffix": ""
},
{
"first": "Ofer",
"middle": [],
"last": "Dekel",
"suffix": ""
},
{
"first": "Joseph",
"middle": [],
"last": "Keshet",
"suffix": ""
}
],
"year": 2006,
"venue": "Journal of Machine Learning Research",
"volume": "7",
"issue": "",
"pages": "551--585",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Koby Crammer, Ofer Dekel, Joseph Keshet, Shai Shalev- Shwartz, and Yoram Singer. 2006. Online passive- aggressive algorithms. Journal of Machine Learning Research, 7:551-585.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Automatic tagging of Arabic text: From raw text to base phrase chunks",
"authors": [
{
"first": "Mona",
"middle": [],
"last": "Diab",
"suffix": ""
},
{
"first": "Kadri",
"middle": [],
"last": "Hacioglu",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. HLT/NAACL 2004. Companion volume",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mona Diab, Kadri Hacioglu, and Daniel Jurafsky. 2004. Automatic tagging of Arabic text: From raw text to base phrase chunks. In Proc. HLT/NAACL 2004. Companion volume.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Comparing reordering constraints for SMT using efficient B\uf76c\uf765\uf775 oracle computation",
"authors": [
{
"first": "Markus",
"middle": [],
"last": "Dreyer",
"suffix": ""
},
{
"first": "Keith",
"middle": [],
"last": "Hall",
"suffix": ""
},
{
"first": "Sanjeev",
"middle": [],
"last": "Khudanpur",
"suffix": ""
}
],
"year": 2007,
"venue": "Proc. 2007 Workshop on Syntax and Structure in Statistical Translation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Markus Dreyer, Keith Hall, and Sanjeev Khudanpur. 2007. Comparing reordering constraints for SMT us- ing efficient B\uf76c\uf765\uf775 oracle computation. In Proc. 2007 Workshop on Syntax and Structure in Statistical Trans- lation.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Beyond log-linear models: Boosted minimum error rate training for nbest re-ranking",
"authors": [
{
"first": "Kevin",
"middle": [],
"last": "Duh",
"suffix": ""
},
{
"first": "Katrin",
"middle": [],
"last": "Kirchoff",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. ACL-08: HLT, Short Papers",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kevin Duh and Katrin Kirchoff. 2008. Beyond log-linear models: Boosted minimum error rate training for n- best re-ranking. In Proc. ACL-08: HLT, Short Papers.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Large margin classification using the perceptron algorithm. Machine Learning",
"authors": [
{
"first": "Yoav",
"middle": [],
"last": "Freund",
"suffix": ""
},
{
"first": "Robert",
"middle": [
"E"
],
"last": "Schapire",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "37",
"issue": "",
"pages": "277--296",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoav Freund and Robert E. Schapire. 1999. Large mar- gin classification using the perceptron algorithm. Ma- chine Learning, 37:277-296.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Fast exact inference with a factored model for natural language parsing",
"authors": [
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Chris",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2002,
"venue": "Advances in Neural Information Processing Systems",
"volume": "15",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dan Klein and Chris D. Manning. 2003. Fast exact infer- ence with a factored model for natural language pars- ing. In Advances in Neural Information Processing Systems 15 (NIPS 2002).",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Statistical significance tests for machine translation evaluation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "Franz",
"middle": [
"Josef"
],
"last": "Och",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. HLT-NAACL 2003. Philipp Koehn",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proc. HLT-NAACL 2003. Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proc. EMNLP 2004.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Complexity of finding the BLEU-optimal hypothesis in a confusion network",
"authors": [
{
"first": "Gregor",
"middle": [],
"last": "Leusch",
"suffix": ""
},
{
"first": "Evgeny",
"middle": [],
"last": "Matusov",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. EMNLP 2008. This volume",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gregor Leusch, Evgeny Matusov, and Hermann Ney. 2008. Complexity of finding the BLEU-optimal hy- pothesis in a confusion network. In Proc. EMNLP 2008. This volume.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "An end-to-end discriminative approach to machine translation",
"authors": [
{
"first": "Percy",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "Alexandre",
"middle": [],
"last": "Bouchard-C\u00f4t\u00e9",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Ben",
"middle": [],
"last": "Taskar",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. COLING-ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Percy Liang, Alexandre Bouchard-C\u00f4t\u00e9, Dan Klein, and Ben Taskar. 2006. An end-to-end discriminative ap- proach to machine translation. In Proc. COLING-ACL 2006.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "ORANGE: a method for evaluating automatic evaluation metrics for machine translation",
"authors": [
{
"first": "Chin-Yew",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Franz Josef",
"middle": [],
"last": "Och",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. COLING",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chin-Yew Lin and Franz Josef Och. 2004. ORANGE: a method for evaluating automatic evaluation metrics for machine translation. In Proc. COLING 2004.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Soft syntactic constraints for hierarchical phrased-based translation",
"authors": [
{
"first": "Yuval",
"middle": [],
"last": "Marton",
"suffix": ""
},
{
"first": "Philip",
"middle": [],
"last": "Resnik",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuval Marton and Philip Resnik. 2008. Soft syntactic constraints for hierarchical phrased-based translation. In Proc. ACL-08: HLT.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Online large-margin training of dependency parsers",
"authors": [
{
"first": "Ryan",
"middle": [],
"last": "Mcdonald",
"suffix": ""
},
{
"first": "Koby",
"middle": [],
"last": "Crammer",
"suffix": ""
},
{
"first": "Fernando",
"middle": [],
"last": "Pereira",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ryan McDonald, Koby Crammer, and Fernando Pereira. 2005. Online large-margin training of dependency parsers. In Proc. ACL 2005.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Forestbased translation",
"authors": [
{
"first": "Haitao",
"middle": [],
"last": "Mi",
"suffix": ""
},
{
"first": "Liang",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Qun",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Haitao Mi, Liang Huang, and Qun Liu. 2008. Forest- based translation. In Proc. ACL-08: HLT.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Discriminative training and maximum entropy models for statistical machine translation",
"authors": [
{
"first": "Josef",
"middle": [],
"last": "Franz",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz Josef Och and Hermann Ney. 2002. Discrimina- tive training and maximum entropy models for statis- tical machine translation. In Proc. ACL 2002.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Minimum error rate training in statistical machine translation",
"authors": [
{
"first": "Franz Josef",
"middle": [],
"last": "Och",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In Proc. ACL 2003.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Fast training of support vector machines using sequential minimal optimization",
"authors": [
{
"first": "C",
"middle": [],
"last": "John",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Platt",
"suffix": ""
}
],
"year": 1998,
"venue": "Advances in Kernel Methods: Support Vector Learning",
"volume": "",
"issue": "",
"pages": "195--208",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John C. Platt. 1998. Fast training of support vector machines using sequential minimal optimization. In Bernhard Sch\u00f6lkopf, Christopher J. C. Burges, and Alexander J. Smola, editors, Advances in Kernel Meth- ods: Support Vector Learning, pages 195-208. MIT Press.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Minimum risk annealing for training log-linear models",
"authors": [
{
"first": "David",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Eisner",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. COLING/ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David A. Smith and Jason Eisner. 2006. Minimum risk annealing for training log-linear models. In Proc. COLING/ACL 2006, Poster Sessions.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Randomized language models via perfect hash functions",
"authors": [
{
"first": "David",
"middle": [],
"last": "Talbot",
"suffix": ""
},
{
"first": "Thorsten",
"middle": [],
"last": "Brants",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Talbot and Thorsten Brants. 2008. Random- ized language models via perfect hash functions. In Proc. ACL-08: HLT.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Max-margin parsing",
"authors": [
{
"first": "Ben",
"middle": [],
"last": "Taskar",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Collins",
"suffix": ""
},
{
"first": "Daphne",
"middle": [],
"last": "Koller",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. EMNLP 2004",
"volume": "",
"issue": "",
"pages": "1--8",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ben Taskar, Dan Klein, Michael Collins, Daphne Koller, and Christopher Manning. 2004. Max-margin pars- ing. In Proc. EMNLP 2004, pages 1-8.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Learning structured prediction models: A large margin approach",
"authors": [
{
"first": "Ben",
"middle": [],
"last": "Taskar",
"suffix": ""
},
{
"first": "Vassil",
"middle": [],
"last": "Chatalbashev",
"suffix": ""
},
{
"first": "Daphne",
"middle": [],
"last": "Koller",
"suffix": ""
},
{
"first": "Carlos",
"middle": [],
"last": "Guestrin",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. ICML",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ben Taskar, Vassil Chatalbashev, Daphne Koller, and Carlos Guestrin. 2005. Learning structured predic- tion models: A large margin approach. In Proc. ICML 2005.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "A discriminative global training algorithm for statistical MT",
"authors": [
{
"first": "Christoph",
"middle": [],
"last": "Tillmann",
"suffix": ""
},
{
"first": "Tong",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. COLING-ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christoph Tillmann and Tong Zhang. 2006. A discrimi- native global training algorithm for statistical MT. In Proc. COLING-ACL 2006.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Scalable discriminative learning for natural language parsing and translation",
"authors": [
{
"first": "Joseph",
"middle": [],
"last": "Turian",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Wellington",
"suffix": ""
},
{
"first": "I",
"middle": [
"Dan"
],
"last": "Melamed",
"suffix": ""
}
],
"year": 2006,
"venue": "Advances in Neural Information Processing Systems",
"volume": "19",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Joseph Turian, Benjamin Wellington, and I. Dan Melamed. 2007. Scalable discriminative learning for natural language parsing and translation. In Advances in Neural Information Processing Systems 19 (NIPS 2006).",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Online large-margin training for statistical machine translation",
"authors": [
{
"first": "Taro",
"middle": [],
"last": "Watanabe",
"suffix": ""
},
{
"first": "Jun",
"middle": [],
"last": "Suzuki",
"suffix": ""
},
{
"first": "Hajime",
"middle": [],
"last": "Tsukuda",
"suffix": ""
},
{
"first": "Hideki",
"middle": [],
"last": "Isozaki",
"suffix": ""
}
],
"year": 2007,
"venue": "Proc. EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Taro Watanabe, Jun Suzuki, Hajime Tsukuda, and Hideki Isozaki. 2007. Online large-margin training for statis- tical machine translation. In Proc. EMNLP 2007.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "A polynomial-time algorithm for statistical machine translation",
"authors": [
{
"first": "Dekai",
"middle": [],
"last": "Wu",
"suffix": ""
}
],
"year": 1996,
"venue": "Proc. 34th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "152--158",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dekai Wu. 1996. A polynomial-time algorithm for statistical machine translation. In Proc. 34th Annual Meeting of the Association for Computational Linguis- tics, pages 152-158.",
"links": null
},
"BIBREF33": {
"ref_id": "b33",
"title": "A tree sequence alignment-based tree-to-tree translation model",
"authors": [
{
"first": "Min",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Hongfei",
"middle": [],
"last": "Jiang",
"suffix": ""
},
{
"first": "Aiti",
"middle": [],
"last": "Aw",
"suffix": ""
},
{
"first": "Haizhou",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Sheng",
"middle": [],
"last": "Chew Lim Tan",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Min Zhang, Hongfei Jiang, Aiti Aw, Haizhou Li, Chew Lim Tan, and Sheng Li. 2008. A tree se- quence alignment-based tree-to-tree translation model. In Proc. ACL-08: HLT.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Classifying nonterminal occurrences for the structural distortion model. rare as X 1 gets larger. The default Hiero features have no way to learn this.",
"type_str": "figure",
"num": null,
"uris": null
},
"TABREF0": {
"text": "call max-B\uf76c\uf765\uf775 updating. Liang et al. and Arun and Koehn experiment with these methods and both opt for a third method, which Liang et al. call local updating: generate an n-best list of translations and select the highest-B\uf76c\uf765\uf775 translation from it. The intuition is that due to noise in the training data or reference transla-",
"num": null,
"type_str": "table",
"html": null,
"content": "<table><tr><td/><td>1</td><td/><td/><td/><td/><td/></tr><tr><td/><td>0.9</td><td/><td/><td/><td/><td/></tr><tr><td>B\uf76c\uf765\uf775 score</td><td>0.7 0.8</td><td/><td/><td/><td/><td/></tr><tr><td/><td>0.6</td><td/><td/><td/><td/><td/></tr><tr><td/><td>0.5</td><td/><td/><td/><td/><td/></tr><tr><td/><td>0.4</td><td/><td/><td/><td/><td/></tr><tr><td/><td>-90</td><td>-85</td><td>-80</td><td>-75</td><td>-70</td><td>-65</td><td>-60</td></tr><tr><td/><td/><td/><td/><td>Model score</td><td/><td/></tr></table>"
},
"TABREF1": {
"text": "++ 51.3 + 34.5 ++ 46.4 ++ distortion (fine) 34 53.3 ++ 51.5 ++ 34.7 ++ 46.7 ++ distortion+syntax (fine) 56 53.6 ++ 52.0 ++ 35.0 ++ 47.2 ++",
"num": null,
"type_str": "table",
"html": null,
"content": "<table><tr><td/><td/><td/><td>Dev</td><td colspan=\"3\">NIST 06 (NIST part)</td></tr><tr><td>Train</td><td>Features</td><td>#</td><td>nw</td><td>nw</td><td>ng</td><td>nw+ng</td></tr><tr><td colspan=\"2\">MERT baseline</td><td colspan=\"2\">12 52.0</td><td colspan=\"3\">50.5 32.4 44.6</td></tr><tr><td/><td>syntax (coarse)</td><td colspan=\"2\">14 52.2</td><td colspan=\"3\">50.9 33.0 + 45.0 +</td></tr><tr><td/><td>syntax (fine)</td><td colspan=\"2\">34 52.1</td><td colspan=\"3\">50.4 33.5 ++ 44.8</td></tr><tr><td/><td>distortion (coarse)</td><td colspan=\"2\">13 52.3</td><td colspan=\"3\">51.3 + 34.3 ++ 45.8 ++</td></tr><tr><td/><td>distortion (fine)</td><td colspan=\"2\">34 52.0</td><td colspan=\"3\">50.9 34.5 ++ 45.5 ++</td></tr><tr><td colspan=\"2\">MIRA baseline</td><td colspan=\"2\">12 52.0</td><td colspan=\"3\">49.8 \u2212 34.2 ++ 45.3 ++</td></tr><tr><td/><td>syntax (fine)</td><td colspan=\"2\">34 53.1</td><td/><td/><td/></tr></table>"
},
"TABREF3": {
"text": "Effect of removing various improvements in learning method. Key: \u2212 or \u2212\u2212 = significantly worse than full system (p < 0.05 or p < 0.01, respectively); LAI = loss-augmented inference for additional negative examples.",
"num": null,
"type_str": "table",
"html": null,
"content": "<table/>"
}
}
}
} |