File size: 112,347 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 | {
"paper_id": "P13-1003",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T09:37:07.046767Z"
},
"title": "Training Nondeficient Variants of IBM-3 and IBM-4 for Word Alignment",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Schoenemann",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Heinrich-Heine-Universit\u00e4t D\u00fcsseldorf",
"location": {
"country": "Germany Universit\u00e4tsstr"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We derive variants of the fertility based models IBM-3 and IBM-4 that, while maintaining their zero and first order parameters, are nondeficient. Subsequently, we proceed to derive a method to compute a likely alignment and its neighbors as well as give a solution of EM training. The arising M-step energies are non-trivial and handled via projected gradient ascent. Our evaluation on gold alignments shows substantial improvements (in weighted Fmeasure) for the IBM-3. For the IBM-4 there are no consistent improvements. Training the nondeficient IBM-5 in the regular way gives surprisingly good results. Using the resulting alignments for phrasebased translation systems offers no clear insights w.r.t. BLEU scores.",
"pdf_parse": {
"paper_id": "P13-1003",
"_pdf_hash": "",
"abstract": [
{
"text": "We derive variants of the fertility based models IBM-3 and IBM-4 that, while maintaining their zero and first order parameters, are nondeficient. Subsequently, we proceed to derive a method to compute a likely alignment and its neighbors as well as give a solution of EM training. The arising M-step energies are non-trivial and handled via projected gradient ascent. Our evaluation on gold alignments shows substantial improvements (in weighted Fmeasure) for the IBM-3. For the IBM-4 there are no consistent improvements. Training the nondeficient IBM-5 in the regular way gives surprisingly good results. Using the resulting alignments for phrasebased translation systems offers no clear insights w.r.t. BLEU scores.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "While most people think of the translation and word alignment models IBM-3 and IBM-4 as inherently deficient models (i.e. models that assign non-zero probability mass to impossible events), in this paper we derive nondeficient variants maintaining their zero order (IBM-3) and first order (IBM-4) parameters. This is possible as IBM-3 and IBM-4 are very special cases of general loglinear models: they are properly derived by the chain rule of probabilities. Deficiency is only introduced by ignoring a part of the history to be conditioned on in the individual factors of the chain rule factorization. While at first glance this seems necessary to obtain zero and first order de- Figure 1 : Plot of the negative log. likelihoods (the quantity to be minimized) arising in training deficient and nondeficient models (for Europarl German | English, training scheme 1 5 H 5 3 5 4 5 ). 1/3/4=IBM-1/3/4, H=HMM, T=Transfer iteration. The curves are identical up to iteration 11.",
"cite_spans": [],
"ref_spans": [
{
"start": 681,
"end": 689,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Iteration 11 shows that merely 5.14% of the (HMM) probability mass are covered by the Viterbi alignment and its neighbors. With deficient models (and deficient empty words) the final negative log likelihood is higher than the initial HMM one, with nondeficient models it is lower than for the HMM, as it should be for a better model. pendencies, we show that with proper renormalization all factors can be made nondeficient. Having introduced the model variants, we proceed to derive a hillclimbing method to compute a likely alignment (ideally the Viterbi alignment) and its neighbors. As for the deficient models, this plays an important role in the E-step of the subsequently derived expectation maximization (EM) training scheme. As usual, expectations in EM are approximated, but we now also get non-trivial Mstep energies. We deal with these via projected gradient ascent.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The downside of our method is its resource consumption, but still we present results on corpora with 100.000 sentence pairs. The source code of this project is available in our word alignment software RegAligner 1 , version 1.2 and later. Figure 1 gives a first demonstration of how much the proposed variants differ from the standard models by visualizing the resulting negative log likelihoods 2 , the quantity to be minimized in EM-training. The nondeficient IBM-4 derives a lower negative log likelihood than the HMM, the regular deficient variant only a lower one than the IBM-1. As an aside, the transfer iteration from HMM to IBM3 (iteration 11) reveals that only 5.14% of the probability mass 3 are preserved when using the Viterbi alignment and its neighbors instead of all alignments.",
"cite_spans": [],
"ref_spans": [
{
"start": 239,
"end": 247,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Indeed, it is widely recognized that -with proper initialization -fertility based models outperform sequence based ones. In particular, sequence based models can simply ignore a part of the sentence to be conditioned on, while fertility based models explicitly factor in a probability of words in this sentence to have no aligned words (or any other number of aligned words, called the fertility). Hence, it is encouraging to see that the nondeficient IBM-4 indeed derives a higher likelihood than the sequence based HMM.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Related Work Today's most widely used models for word alignment are still the models IBM 1-5 of Brown et al. (1993) and the HMM of Vogel et al. (1996) , thoroughly evaluated in (Och and Ney, 2003) . While it is known that fertilitybased models outperform sequence-based ones, the large bulk of word alignment literature following these publications has mostly ignored fertilitybased models. This is different in the present paper which deals exclusively with such models.",
"cite_spans": [
{
"start": 96,
"end": 115,
"text": "Brown et al. (1993)",
"ref_id": "BIBREF1"
},
{
"start": 131,
"end": 150,
"text": "Vogel et al. (1996)",
"ref_id": "BIBREF22"
},
{
"start": 177,
"end": 196,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One reason for the lack of interest is surely that computing expectations and Viterbi alignments for these models is a hard problem (Udupa and Maji, 2006) . Nevertheless, computing Viterbi align-1 https://github.com/Thomas1205/RegAligner, for the reported results we used a slightly earlier version.",
"cite_spans": [
{
"start": 132,
"end": 154,
"text": "(Udupa and Maji, 2006)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "2 Note that the figure slightly favors IBM-1 and HMM as for them the length J of the foreign sequence is assumed to be known whereas IBM-3 and IBM-4 explicitly predict it.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "3 This number regards the corpus probability as in (9) to the power of 1/S, i.e. the objective function in maximum likelihood training. The number is not entirely fair as alignments where more than half the words align to the empty word are assigned a probability of 0. Still, this is an issue only for short sentences. ments for the IBM-3 has been shown to often be practicable (Ravi and Knight, 2010; Schoenemann, 2010) .",
"cite_spans": [
{
"start": 379,
"end": 402,
"text": "(Ravi and Knight, 2010;",
"ref_id": "BIBREF15"
},
{
"start": 403,
"end": 421,
"text": "Schoenemann, 2010)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Much work has been spent on HMM-based formulations, focusing on the computationally tractable side (Toutanova et al., 2002; Sumita et al., 2004; Deng and Byrne, 2005) . In addition, some rather complex models have been proposed that usually aim to replace the fertility based models (Wang and Waibel, 1998; Fraser and Marcu, 2007a) .",
"cite_spans": [
{
"start": 99,
"end": 123,
"text": "(Toutanova et al., 2002;",
"ref_id": "BIBREF19"
},
{
"start": 124,
"end": 144,
"text": "Sumita et al., 2004;",
"ref_id": "BIBREF18"
},
{
"start": 145,
"end": 166,
"text": "Deng and Byrne, 2005)",
"ref_id": "BIBREF4"
},
{
"start": 283,
"end": 306,
"text": "(Wang and Waibel, 1998;",
"ref_id": "BIBREF23"
},
{
"start": 307,
"end": 331,
"text": "Fraser and Marcu, 2007a)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Another line of models (Melamed, 2000; Marcu and Wong, 2002; Cromi\u00e8res and Kurohashi, 2009) focuses on joint probabilities to get around the garbage collection effect (i.e. that for conditional models, rare words in the given language align to too many words in the predicted language). The downside is that these models are computationally harder to handle.",
"cite_spans": [
{
"start": 23,
"end": 38,
"text": "(Melamed, 2000;",
"ref_id": "BIBREF11"
},
{
"start": 39,
"end": 60,
"text": "Marcu and Wong, 2002;",
"ref_id": "BIBREF10"
},
{
"start": 61,
"end": 91,
"text": "Cromi\u00e8res and Kurohashi, 2009)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A more recent line of work introduces various forms of regularity terms, often in the form of symmetrization (Liang et al., 2006; Gra\u00e7a et al., 2010; Bansal et al., 2011) and recently by using L 0 norms (Vaswani et al., 2012) .",
"cite_spans": [
{
"start": 109,
"end": 129,
"text": "(Liang et al., 2006;",
"ref_id": "BIBREF9"
},
{
"start": 130,
"end": 149,
"text": "Gra\u00e7a et al., 2010;",
"ref_id": "BIBREF7"
},
{
"start": 150,
"end": 170,
"text": "Bansal et al., 2011)",
"ref_id": "BIBREF0"
},
{
"start": 203,
"end": 225,
"text": "(Vaswani et al., 2012)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We begin with a short review of fertility-based models in general and IBM-3, IBM-4 and IBM-5 specifically. All are due to (Brown et al., 1993) who proposed to use the deficient models IBM-3 and IBM-4 to initialize the nondeficient IBM-5.",
"cite_spans": [
{
"start": 122,
"end": 142,
"text": "(Brown et al., 1993)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "For a foreign sentence f = f J 1 = (f 1 , . . . , f J ) with J words and an English one e = e I 1 = (e 1 , . . . , e I ) with I words, the (conditional) probability p(f J 1 |e I 1 ) of getting the foreign sentence as a translation of the English one is modeled by introducing the word alignment a as a hidden variable:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "p(f J 1 |e I 1 ) = a p(f J 1 , a|e I 1 )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "All IBM models restrict the space of alignments to those where a foreign word can align to at most one target word. The resulting alignment is then written as a vector a J 1 , where each a j takes integral values between 0 and I, with 0 indicating that f j has no English correspondence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "The fertility-based models IBM-3, IBM-4 and IBM-5 factor the (conditional) probability p(f J 1 , a J 1 |e I 1 ) of obtaining an alignment and a translation given an English sentence according to the following generative story:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "1. For i = 1, 2, . . . , I, decide on the number \u03a6 i of foreign words aligned to e i . This number is called the fertility of e i . Choose with probability p(",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "\u03a6 i |e I 1 , \u03a6 i\u22121 1 ) = p(\u03a6 i |e i ). 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "Choose the number \u03a6 0 of unaligned words in the (still unknown) foreign sequence. Choose with probability p(\u03a6 0 |e I 1 ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "\u03a6 I 1 ) = p(\u03a6 0 | I i=1 \u03a6 i ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "Since each foreign word belongs to exactly one English position (including 0), the foreign sequence is now known to be of length J = I i=0 \u03a6 i . 3. For each i = 1, 2, . . . , I, and k = 1, . . . , \u03a6 i decide on (a) the identity f i,k of the next foreign word aligned to e i . Choose with probability",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "p(f i,k |e I 1 , \u03a6 I 0 , d i\u22121 1 , d i,1 , . . . , d i,k\u22121 , f i,k ) = p(f i,k |e i ), where d i comprises all d i,k",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "for word i (see point b) below) and f i,k comprises all foreign words known at that point. (b) the position d i,k of the just generated foreign word f i,k , with probability , where e 0 is an artificial \"empty word\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "p(d i,k |e I 1 , \u03a6 I 0 , d i\u22121 1 , d i,1 , . . . , d i,k\u22121 , f i,k , f i,k ) = p(d i,k |e i , d i\u22121 1 , d i,1 , . . . , d i,k\u22121 , f i,k , J).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "To model the probability for the number of unaligned words in step 2, each of the I i=1 \u03a6 i properly aligned foreign words generates an unaligned foreign word with probability p 0 , resulting in",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "p \u03a6 0 I i=1 \u03a6 i = \uf8eb \uf8ec \uf8ed I i=1 \u03a6 i \u03a6 0 \uf8f6 \uf8f7 \uf8f8p \u03a6 i 0 (1\u2212p 0 ) ( i \u03a6 i )\u2212\u03a6 0 ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "with a base probability p 0 and the combinato-",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "rial coefficients n k = n! k!(n\u2212k)! , where n! = n k=1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "k denotes the factorial of n. The main difference between IBM-3, IBM-4 and IBM-5 is the choice of probability model in step 3 b), called a distortion model. The choices are now detailed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The models IBM-3, IBM-4 and IBM-5",
"sec_num": "2"
},
{
"text": "The IBM-3 implements a zero order distortion model, resulting in",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "p(d i,k |i, J) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "Since most of the context to be conditioned on is ignored, this allows invalid configurations to occur with non-zero probability: some foreign positions can be chosen several times, while others remain empty. One says that the model is deficient. On the other hand, the model for p(\u03a6",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "0 | I i=1 \u03a6 i )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "is nondeficient, and in training this often results in very high probabilities p 0 . To prevent this it is common to make this model deficient as well (Och and Ney, 2003) , which improves performance immensely and gives much better results than simply fixing p 0 in the original model. As for each i the d i,k can appear in any order (i.e. need not be in ascending order), there are I i=1 \u03a6 i ! ways to generate the same alignment a J 1 (where the \u03a6 i are the fertilities induced by a J 1 ). In total, the IBM-3 has the following probability model:",
"cite_spans": [
{
"start": 151,
"end": 170,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "p(f J 1 , a J 1 |e I 1 ) = J j=1 p(f j |e a j ) \u2022 p(j|a j , J) (1) \u2022 p \u03a6 0 | I i=1 \u03a6 i \u2022 I i=1 \u03a6 i ! p(\u03a6 i |e i ) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "Reducing the Number of Parameters While using non-parametric models p(j|i, J) is convenient for closed-form M-steps in EM training, these parameters are not very intuitive. Instead, in this paper we use the parametric model",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "p(j|i, J) = p(j|i) J j=1 p(j|i)",
"eq_num": "(2)"
}
],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "with the more intuitive parameters p(j|i). The arising M-step energy is addressed by projected gradient ascent (see below). These parameters are also used for the nondeficient variants. Using the original non-parametric ones can be handled in a very similar manner to the methods set forth below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-3",
"sec_num": "2.1"
},
{
"text": "The distortion model of the IBM-4 is a first order one that generates the d i,k of each English position i in ascending order (i.e. for 1 < k \u2264 \u03a6 i we have",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "d i,k > d i,k\u22121 ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "There is then a one-to-one correspondence between alignments a J 1 and (valid) distortion parameters (d i,k ) i=1,...,I, k=1,...,\u03a6 i and therefore no longer a factor of I i=1 \u03a6 i ! . The IBM-4 has two sub-distortion models, one for the first aligned word (k = 1) of an English position and one for all following words (k > 1, only 4 the closest preceding English word that has aligned foreign words. The aligned foreign positions of [i] are combined into a center position [i] , the rounded average of the positions. Now, the distortion probability for the first word (k = 1) is",
"cite_spans": [
{
"start": 331,
"end": 332,
"text": "4",
"ref_id": null
},
{
"start": 473,
"end": 476,
"text": "[i]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "if \u03a6 i > 1). For position i, let [i] = arg max{i |1 \u2264 i < i, \u03a6 i > 0} denote",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "p =1 (d i,1 | [i] , A(f i,1 ), B(e [i] ), J) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "where A gives the word class of a foreign word and B the word class of an English word (there are typically 50 classes per language, derived by machine learning techniques). The probability is further reduced to a dependency on the difference of the positions, i.e.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "p =1 (d i,1 \u2212 [i] | A(f i,1 ), B(e [i] )). For k > 1 the model is p >1 (d i,k |d i,k\u22121 , A(f i,k ), J) , which is likewise reduced to p >1 (d i,k \u2212 d i,k\u22121 | A(f i,k )).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "Note that in both differencebased formulations the dependence on J has to be dropped to get closed-form solutions of the M-step in EM training, and Brown et al. note themselves that the IBM-4 can place words before the start and after the end of the sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "Reducing Deficiency In this paper, we also investigate the effect of reducing the amount of wasted probability mass by enforcing the dependence on J by proper renormalization, i.e. using",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "p =1 (j|j , A(f i,1 ), B(e [i] ), J) = (3) p =1 (j \u2212 j |A(f i,1 ), B(e [i] )) J j =1 p =1 (j \u2212 j |A(f i,1 ), B(e [i] ))",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": ", for the first aligned word and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "p >1 (j|j , A(f i,k ), J) = (4) p >1 (j \u2212 j | A(f i,k )) J j =1 p >1 (j \u2212 j | A(f i,k )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": ") for all following words, again handling the M-step in EM training via projected gradient ascent. With this strategy words can no longer be placed outside the sentence, but a lot of probability mass is still wasted on configurations where at least one foreign (or predicted) position j aligns to two or more positions i, i in the English (or given) language (and consequently there are more unaligned source words than the generated \u03a6 0 ). Therefore, here, too, the probability for \u03a6 0 has to be made deficient to get good performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "In summary, the base model for the IBM-4 is:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "p(f J 1 , a J 1 |e I 1 ) = p \u03a6 0 | I i=1 \u03a6 i (5) \u2022 J j=1 p(f j |e a j ) \u2022 I i=1 p(\u03a6 i |e i ) \u2022 i:\u03a6 i >0 p =1 (d i,1 \u2212 [i] |A(f i,1 ), B(e [i] )) \u2022 \u03a6 i k=2 p >1 (d i,k \u2212 d i,k\u22121 |A(f i,k )) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "where empty products are understood to be 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-4",
"sec_num": "2.2"
},
{
"text": "We note in passing that the distortion model of the IBM-5 is nondeficient and has parameters for filling the nth open gap in the foreign sequence given that there are N positions to choose from -see the next section for exactly what positions one can choose from. There is also a dependence on word classes for the foreign language. This is neither a zero order nor a first order dependence, and in (Och and Ney, 2003) the first order model of the IBM-4, though deficient, outperformed the IBM-5. The IBM-5 is therefore rarely used in practice. This motivated us to instead reformulate IBM-3 and IBM-4 as nondeficient models. In our results, however, the IBM-5 gave surprisingly good results and was often superior to all variants of the IBM-4.",
"cite_spans": [
{
"start": 399,
"end": 418,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "IBM-5",
"sec_num": "2.3"
},
{
"text": "From now on we always enforce that for each position i the indices d i,k are generated in ascending",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient Variants of IBM-and IBM-4",
"sec_num": "3"
},
{
"text": "order (d i,k > d i,k\u22121 for k > 1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient Variants of IBM-and IBM-4",
"sec_num": "3"
},
{
"text": ". A central concept for the generation of d i,k in step 3(b) is the set of positions in the foreign sequence that are still without alignment. We denote the set of these positions by",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient Variants of IBM-and IBM-4",
"sec_num": "3"
},
{
"text": "J i,k,J = {1, . . . , J} \u2212 {d i,k | 1 \u2264 k < k} \u2212 {d i ,k | 1 \u2264 i < i, 1 \u2264 k \u2264 \u03a6 i }",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient Variants of IBM-and IBM-4",
"sec_num": "3"
},
{
"text": "where the dependence on the various d i ,k is not made explicit in the following. It is tempting to think that in a nondeficient model all members of J i,k,J can be chosen for d i,k , but this holds only \u03a6 i = 1. Otherwise, the requirement of generating the d i,k in ascending order prevents us from choosing the (\u03a6 i \u2212k) largest entries in J i,k,J . For k > 1 we also have to remove all positions smaller than d i,k\u22121 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient Variants of IBM-and IBM-4",
"sec_num": "3"
},
{
"text": "Let J \u03a6 i i,k,J denote the set where these positions have been removed. With that, we can state the nondeficient variants of IBM-3 and IBM-4.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient Variants of IBM-and IBM-4",
"sec_num": "3"
},
{
"text": "For the IBM-3, we define the auxiliary quantity",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-3",
"sec_num": "3.1"
},
{
"text": "q(d i,k = j | i, J \u03a6 i i,k,J ) = p(j|i) if j \u2208 J \u03a6 i i,k,J 0 else ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-3",
"sec_num": "3.1"
},
{
"text": "where we use the zero order parameters p(j|i) we also use for the standard (deficient) IBM-3, compare (2). To get a nondeficient variant, it remains to renormalize, resulting in",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-3",
"sec_num": "3.1"
},
{
"text": "p(d i,k = j|i, J \u03a6 i i,k,J ) = q(j|i, J \u03a6 i i,k,J ) J j=1 q(j|i, J \u03a6 i i,k,J )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-3",
"sec_num": "3.1"
},
{
"text": ". 6Further, note that the factors \u03a6 i ! now have to be removed from (1) as the d i,k are generated in ascending order. Lastly, here we use the original nondeficient empty word model p",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-3",
"sec_num": "3.1"
},
{
"text": "(\u03a6 0 | I i=1 \u03a6 i )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-3",
"sec_num": "3.1"
},
{
"text": ", resulting in a totally nondeficient model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-3",
"sec_num": "3.1"
},
{
"text": "With the notation set up, it is rather straightforward to derive a nondeficient variant of the IBM-4. Here, there are the two cases k = 1 and k > 1. We begin with the case k = 1. Abbreviating \u03b1 = A(f i,1 ) and \u03b2 = B(e [i] ), we define the auxiliary quantity",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "q =1 (d i,1 = j| [i] , \u03b1, \u03b2, J \u03a6 i i,k,J ) = (7) p =1 (j \u2212 [i] |\u03b1, \u03b2) if j \u2208 J \u03a6 i i,k,J 0 else ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "again using the -now first order -parameters of the base model. The nondeficient distribution",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "p =1 (d i,1 = j| [i] , \u03b1, \u03b2, J \u03a6 i i,k,J )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "is again obtained by renormalization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "For the case k > 1, we abbreviate \u03b1 = A(f i,k ) and introduce the auxiliary quantity",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "q >1 (d i,k = j|d i,k\u22121 , \u03b1, J \u03a6 i i,k,J ) = (8) p >1 (j \u2212 d i,k\u22121 |\u03b1) if j \u2208 J \u03a6 i i,k,J 0 else ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "from which the nondeficient distribution",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "p >1 (d i,k =j|d i,k\u22121 , \u03b1, J \u03a6 i i,k,J )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "is again obtained by renormalization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nondeficient IBM-4",
"sec_num": "3.2"
},
{
"text": "For the task of word alignment, we infer the parameters of the models using the maximum likelihood criterion",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training the New Variants",
"sec_num": "4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "max \u03b8 S s=1 p \u03b8 (f s |e s )",
"eq_num": "(9)"
}
],
"section": "Training the New Variants",
"sec_num": "4"
},
{
"text": "on a set of training data (i.e. sentence pairs s = 1, . . . , S). Here, \u03b8 comprises all base parameters of the respective model (e.g. for the IBM-3 all p(f |e), all p(\u03a6, e) and all p(j|i) ) and p \u03b8 signifies the dependence of the model on the parameters. Note that (9) is truly a constrained optimization problem as the parameters \u03b8 have to satisfy a number of probability normalization constraints. When p \u03b8 (\u2022) denotes a fertility based model the resulting problem is a non-concave maximization problem with many local minima and no (known) closed-form solutions. Hence, it is handled by computational methods, which typically apply the logarithm to the above function.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training the New Variants",
"sec_num": "4"
},
{
"text": "Our method of choice to attack the maximum likelihood problem is expectation maximization (EM), the standard in the field, which we explain below. Due to non-concaveness the starting point for EM is of extreme importance. As is common, we first train an IBM-1 and then an HMM before proceeding to the IBM-3 and finally the IBM-4.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training the New Variants",
"sec_num": "4"
},
{
"text": "As in the training of the deficient IBM-3 and IBM-4 models, we approximate the expectations in the E-step by a set of likely alignments, ideally centered around the Viterbi alignment, but already for the regular deficient variants computing it is NP-hard (Udupa and Maji, 2006) . A first task is therefore to compute such a set. This task is also needed for the actual task of word alignment (annotating a given sentence pair with an alignment).",
"cite_spans": [
{
"start": 255,
"end": 277,
"text": "(Udupa and Maji, 2006)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training the New Variants",
"sec_num": "4"
},
{
"text": "For computing alignments, we use the common procedure of hillclimbing where we start with an alignment, then iteratively compute the probabilities of all alignments differing by a move or a swap (Brown et al., 1993) and move to the best of these if it beats the current alignment.",
"cite_spans": [
{
"start": 195,
"end": 215,
"text": "(Brown et al., 1993)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Alignment Computation",
"sec_num": "4.1"
},
{
"text": "Since we cannot ignore parts of the history and still get a nondeficient model, computing the probabilities of the neighbors cannot be handled incrementally (or rather only partially, for the dictionary and fertility models). While this does increase running times, in practice the M-steps take longer than the E-steps.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Alignment Computation",
"sec_num": "4.1"
},
{
"text": "For self-containment, we recall here that for an alignment a J 1 applying the move a J 1 [j \u2192 i] results in the alignment\u00e2 J 1 defined by\u00e2 j = i and\u00e2 j = a j for j = j. Applying the swap a J 1 [j 1 \u2194 j 2 ] results in the alignment\u00e2 J 1 defined by\u00e2 j 1 = a j 2 ,\u00e2 j 2 = a j 1 and\u00e2 j = a j elsewhere. If a J 1 is the alignment produced by hillclimbing, the move matrix m \u2208 IR J\u00d7I+1 is defined by m j,i being the probability of a J 1 [j \u2192 i] as long as a j = i, otherwise 0. Likewise the swap matrix s \u2208 IR J\u00d7J is defined as s j 1 ,j 2 being the probability of a J 1 [j 1 \u2194 j 2 ] for a j 1 = a j 2 , 0 otherwise. The move and swap matrices are used to approximate expectations in EM training (see below).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Alignment Computation",
"sec_num": "4.1"
},
{
"text": "Naive Scheme It is tempting to account for the changes in the model in hillclimbing, but to otherwise use the regular M-step procedures (closed form solution when not conditioning on J for the IBM-4 and for the non-parametric IBM-3, otherwise projected gradient ascent) for the deficient models. However, we verified that this is not a good idea: not only can the likelihood go down in the process (even if we could compute expectations exactly), but these schemes also heavily increase p 0 in each iteration, i.e. the same problem Och and Ney (2003) found for the deficient models. There is therefore the need to execute the Mstep properly, and when done the problem is indeed resolved.",
"cite_spans": [
{
"start": 532,
"end": 550,
"text": "Och and Ney (2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "Proper EM The expectation maximization (EM) framework (Dempster et al., 1977; Neal and Hinton, 1998 ) is a class of template procedures (rather than a proper algorithm) that iteratively requires solving the task",
"cite_spans": [
{
"start": 54,
"end": 77,
"text": "(Dempster et al., 1977;",
"ref_id": "BIBREF3"
},
{
"start": 78,
"end": 99,
"text": "Neal and Hinton, 1998",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "max \u03b8 k S s=1 as p \u03b8 k\u22121 (a s |f s , e s ) log p \u03b8 k (f s , a s |e s )",
"eq_num": "(10"
}
],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": ") by appropriate means. Here, \u03b8 k\u22121 are the parameters from the previous iteration, while \u03b8 k are those derived in the current iteration. Of course, here and in the following the normalization constraints on \u03b8 apply, as already in (9). On explicit request of a reviewer we give a detailed account for our setting here. Readers not interested in the details can safely move on to the next section.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "Details on EM For the corpora occurring in practice, the function (10) has many more terms than there are atoms in the universe. The trick is that p \u03b8 k (f s , a s |e s ) is a product of factors, where each factor depends on very few components of \u03b8 k only. Taking the logarithm gives a sum of logarithms, and in the end we are left with the problem of computing the weights of each factor, which turn out to be expectations. To apply this to the (deficient) IBM-3 model with parametric distortion we simplify p \u03b8 k\u22121 (a s |f s , e s ) = p(a s ) and define the counts n f,e (a s ) = Js j=1 \u03b4(f s j , f ) \u2022 \u03b4(e s a s j , e), n \u03a6,e (a s ) = Is i=1 \u03b4(e s i , e)\u2022\u03b4(\u03a6 i (a s ), \u03a6) and n j,i (a s ) = \u03b4(a s j , i). We also use short hand notations for sets, e.g. {p(f |e)} is meant as the set of all translation probabilities induced by the given corpus. With this notation, after reordering the terms problem (10) can be written as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "max {p(f |e)},{p(\u03a6|e)},{p(j|i)} (11) e,f S s=1 as p(a s ) n f,e (a s ) log p(f |e) + e,\u03a6 S s=1 as p(a s ) n \u03a6,e (a s ) log p(\u03a6, e) + i,j S s=1 as p(a s ) n j,i (a s ) log p(j|i, J) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "Indeed, the weights in each line turn out to be nothing else than expectations of the respective factor under the distribution p \u03b8 k\u22121 (a s |f s , e s ) and will henceforth be written as w f,e , w \u03a6,e and w j,i,J . Therefore, executing an iteration of EM requires first calculating all expectations (E-step) and then solving the maximization problems (M-step). For models such as IBM-1 and HMM the expectations can be calculated efficiently, so the enormous sum of terms in (10) is equivalently written as a manageable one. In this case it can be shown 5 that the new \u03b8 k must have a higher likelihood (9) than \u03b8 k\u22121 (unless a stationary point is reached). In fact, any \u03b8 that has a higher value in the auxiliary function (11) than \u03b8 k\u22121 must also have a higher likelihood. This is an important background for parametric models such as (2) where the M-step cannot be solved exactly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "For IBM-3/4/5 computing exact expectations is intractable (Udupa and Maji, 2006) and approximations have to be used (in fact, even computing the likelihood for a given \u03b8 is intractable). We use the common procedure based on hillclimbing and the move/swap matrices. The likelihood is not guaranteed to increase but it (or rather its approximation) always did in each of the five run iterations. Nevertheless, the main advantage of EM is preserved: problem (11) decomposes into several smaller problems, one for each probability distribution since the parameters are tied by the normalization constraints. The result is one problem for each e involving all p(f |e), one for each e involving all p(\u03a6|e) and one for each i involving all p(j|i).",
"cite_spans": [
{
"start": 58,
"end": 80,
"text": "(Udupa and Maji, 2006)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "The problems for the translation probabilities and the fertility probabilities yield the known standard update rules. The most interesting case is the problem for the (parametric) distortion models. In the deficient setting, the problem for each i is",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "max {p(j|i)} J w i,j,J log p(j|i) J j =1 p(j |i)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "In the nondeficient setting, we now drop the subscripts i, k, J and the superscript \u03a6 from the sets defined in the previous sections, i.e. we write J instead of J \u03a6 i,k,J . The M-step problem is then",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "max {p(j|i)} E i = j J :j\u2208J w j,i,J log p(j|i, J ) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "where w j,i,J (with j \u2208 J ) is the expectation for aligning j to i when one can choose among the positions in J , and with p(j|i, J ) as in (6). In principle there is an exponential number of expectations w j,i,J . However, since we approximate expectations from the move and swap matrices, and hence by O((I + J) \u2022 J) alignments per sentence pair, in the end we get a polynomial number of terms. Currently we only consider alignments with (approximated) p \u03b8 k\u22121 (a s |f s , e s ) > 10 \u22126 . Importantly, the fact that we get separate M-step problems for different i allows us to reduce memory consumption by using refined data structures when storing the expectations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "For both the deficient and the nondeficient variants, the M-step problems for the distortion parameters p(j|i) are non-trivial, non-concave and have no (known) closed form solutions. We approach them via the method of projected gradient ascent (PGA), where the gradient for the nondeficient problem is",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "\u2202E i \u2202p(j|i) = J :j\u2208J w j,J p(j|i) \u2212 j \u2208J w j ,J j \u2208J p(j |i)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": ".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "When running PGA we guarantee that the resulting {p(j|i)} has a higher function value E i than the input ones (unless a stationary point is input). We stop when a cutoff criterion indicates a local maximum or 250 iterations are used up.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "Projected Gradient Ascent This method is used in a couple of recent papers, notably (Schoenemann, 2011; Vaswani et al., 2012) and is briefly sketched here for self-containment (see those papers for more details). To solve a maximization problem",
"cite_spans": [
{
"start": 84,
"end": 103,
"text": "(Schoenemann, 2011;",
"ref_id": "BIBREF17"
},
{
"start": 104,
"end": 125,
"text": "Vaswani et al., 2012)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "max p(j|i)\u22650, j p(j|i)=1 E i ({p(j|i)})",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "for some (differentiable) function E i (\u2022), one iteratively starts at the current point {p k (j|i)}, computes the gradient \u2207E i ({p k (j|i)}) and goes to the point",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "q(j|i) = p k (j|i) + \u03b1\u2207E i (p k (j|i)) , j = 1, . . . , J",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "for some step-length \u03b1. This point is generally not a probability distribution, so one computes the nearest probability distribution",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "min q (j|i)\u22650, j q (j|i)=1 J j=1 q (j|i) \u2212 q(j|i) 2 ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "a step known as projection which we solve with the method of (Michelot, 1986) . The new distribution {q (j|i)} is not guaranteed to have a higher E i (\u2022), but (since the constraint set is a convex one) a suitable interpolation of {p k (j|i)} and {q (j|i)} is guaranteed to have a higher value (unless {p k (j|i)} is a local maximum or minimum of E i (\u2022)). Such a point is computed by backtracking line search and defines the next iterate {p k+1 (j|i)}.",
"cite_spans": [
{
"start": 61,
"end": 77,
"text": "(Michelot, 1986)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "IBM-4 When moving from the IBM-3 to the IBM-4, only the last line in (11) changes. In the end one gets two new kinds of problems, for",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "p =1 (\u2022) and p >1 (\u2022). For p =1 (\u2022)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "we have one problem for each foreign class \u03b1 and each English class \u03b2, of the form",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "max {p =1 (j|j ,\u03b1,\u03b2)} j,j ,J w j,j ,J,\u03b1,\u03b2 log p =1 (j|j , \u03b1, \u03b2, J)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "for reduced deficiency (with p =1 (j|j , \u03b1, \u03b2, J) as in (3) ) and of the form Table 1 : Alignment accuracy (weighted F-measure times 100, \u03b1 = 0.1) on Europarl with 100.000 sentence pairs. Reduced deficiency means renormalization as in 3and 4, so that words cannot be placed before or after the sentence. For the IBM-3, the nondeficient variant is clearly best. For the IBM-4 it is better in roughly half the cases, both with and without word classes.",
"cite_spans": [],
"ref_spans": [
{
"start": 78,
"end": 85,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "max {p =1 (j|j ,\u03b1,\u03b2)} j,j ,J w j,j ,J ,\u03b1,\u03b2 log p =1 (j|j , \u03b1, \u03b2, J )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "for the nondeficient variant, with",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "p =1 (j|j , \u03b1, \u03b2, J ) based on (7).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "For p >1 (\u2022) we have one problem per foreign class \u03b1, of the form",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "max {p >1 (j|j ,\u03b1)} j,j ,J w j,j ,J,\u03b1 log p >1 (j|j , \u03b1, J)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "for reduced deficiency, with p >1 (j|j , \u03b1, J) based on (4), and for the nondeficient variant it has the form max",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "{p >1 (j|j ,\u03b1)} j,j ,J w j,j ,J ,\u03b1 log p >1 (j|j , \u03b1, J ) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "with p >1 (j|j , \u03b1, J ) based on (8). Calculating the gradients is analogous to the IBM-3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parameter Update",
"sec_num": "4.2"
},
{
"text": "We test the proposed methods on subsets of the Europarl corpus for German and English as well as Spanish and English, using lower-cased corpora. We evaluate alignment accuracies on gold alignments 6 in the form of weighted F-measures with \u03b1 = 0.1, which performed well in (Fraser and Marcu, 2007b) . In addition we evaluate the effect on phrase-based translation on one of the tasks. We implement the proposed methods in our own framework RegAligner rather than GIZA++, which is only rudimentally maintained. Therefore, we compare to the deficient models in our own software as well as to those in GIZA++.",
"cite_spans": [
{
"start": 272,
"end": 297,
"text": "(Fraser and Marcu, 2007b)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "We run 5 iterations of IBM-1, followed by 5 iterations of HMM, 5 of IBM-3 and finally 5 of IBM-4. The first iteration of the IBM-3 collects counts from the HMM, and likewise the first iteration of the IBM-4 collects counts from the IBM-3 (in both cases the move and swap matrices are filled with probabilities of the former model, then theses matrices are used as in a regular model iteration). A nondeficient IBM-4 is always initialized by a nondeficient IBM-3. We did not set a fertility limit (except for GIZA++).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "Experiments were run on a Core i5 with 2.5 GHz and 8 GB of memory. The latter was the main reason why we did not use still larger corpora 7 . The running times for the entire training were half a day without word classes and a day with word classes. With 50 instead of 250 PGA iterations in all M-steps we get only half these running times, but the resulting F-measures deteriorate, especially for the IBM-4 with classes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "The running times of our implementation of the IBM-5 are much more favorable: the entire training then runs in little more than an hour.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "The alignment accuracies -weighted F-measures with \u03b1 = 0.1 -for the tested corpora and model variants are given in Table 1 . Clearly, nondeficiency greatly improves the accuracy of the IBM-3, both compared to our deficient implementation and that of GIZA++. For the IBM-4 we get improvements for the nondeficient variant in roughly half the cases, both with and without word classes. We think this is an issue of local minima, inexactly solved M-steps and sensitiveness to initialization.",
"cite_spans": [],
"ref_spans": [
{
"start": 115,
"end": 122,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Alignment Accuracy",
"sec_num": "5.1"
},
{
"text": "Interestingly, also the reduced deficient IBM-4 is not always better than the fully deficient variant. Again, we think this is due to problems with the non-concave nature of the models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Alignment Accuracy",
"sec_num": "5.1"
},
{
"text": "There is also quite some surprise regarding the IBM-5: contrary to the findings of (Och and Ney, 2003) the IBM-5 in GIZA++ performs best in three out of four cases -when competing with both deficient and nondeficient variants of IBM-3 and IBM-4. Our own implementation gives slightly different results (as we do not use smoothing), but it, too, performs very well.",
"cite_spans": [
{
"start": 83,
"end": 102,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Alignment Accuracy",
"sec_num": "5.1"
},
{
"text": "We also check the effect of the various alignments (all produced by RegAligner) on translation performance for phrase-based translation, randomly choosing translation from German to English. We use MOSES with a 5-gram language model (trained on 500.000 sentence pairs) and the standard setup in the MOSES Experiment Management System: training is run in both directions, the alignments are combined using diag-grow-final-and (Och and Ney, 2003) and the parameters of MOSES are optimized on 750 development sentences.",
"cite_spans": [
{
"start": 425,
"end": 444,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Effect on Translation Performance",
"sec_num": "5.2"
},
{
"text": "The resulting BLEU-scores are shown in Table 2 . However, the table shows no clear trends and even the IBM-3 is not clearly inferior to the IBM-4. We think that one would need to handle larger corpora (or run multiple instances of Minimum Error Rate Training with different random seeds) to get more meaningful insights. Hence, at present our paper is primarily of theoretical value.",
"cite_spans": [],
"ref_spans": [
{
"start": 39,
"end": 47,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Effect on Translation Performance",
"sec_num": "5.2"
},
{
"text": "We have shown that the word alignment models IBM-3 and IBM-4 can be turned into nondeficient Table 2 : Evaluation of phrase-based translation from German to English with the obtained alignments (for 100.000 sentence pairs). Training is run in both directions and the resulting alignments are combined via diag-grow-final-and. The table shows no clear superiority of any method.",
"cite_spans": [],
"ref_spans": [
{
"start": 93,
"end": 100,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "In fact, the IBM-4 is not superior to the IBM-3 and the HMM is about equal to the IBM-3. We think that one needs to handle larger corpora to get clearer insights.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "variants, an important aim of probabilistic modeling for word alignment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "Here we have exploited that the models are proper applications of the chain rule of probabilities, where deficiency is only introduced by ignoring parts of the history for the distortion factors in the factorization. By proper renormalization the desired nondeficient variants are obtained.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "The arising models are trained via expectation maximization. In the E-step we use hillclimbing to get a likely alignment (ideally the Viterbi alignment). While this cannot be handled fully incrementally, it is still fast enough in practice. The M-step energies are non-concave and have no (known) closed-form solutions. They are handled via projected gradient ascent.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "For the IBM-3 nondeficiency clearly improves alignment accuracy. For the IBM-4 we get improved accuracies in roughly half the cases, both with and without word classes. The IBM-5 performs surprisingly well, it is often best and hence much better than its reputation. An evaluation of phrase based translation showed no clear insights.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "Nevertheless, we think that nondeficiency in fertility based models is an important issue, and that at the very least our paper is of theoretical value. The implementations are publicly available in RegAligner 1.2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "If the set is empty, instead a sentence start probability is used. Note that we differ slightly in notation compared to(Brown et al., 1993).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "See e.g. the author's course notes (in German), currently http://user.phil-fak.uni-duesseldorf.de/ tosch/downloads/statmt/wordalign.pdf.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "from(Lambert et al., 2005) and from http://user.phil-fak.uni-duesseldorf.de/ tosch/downloads.html.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The main memory bottleneck is the IBM-4 (6 GB without classes, 8 GB with). Using refined data structures should reduce this bottleneck.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Gappy phrasal alignment by agreement",
"authors": [
{
"first": "M",
"middle": [],
"last": "Bansal",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Quirk",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Moore",
"suffix": ""
}
],
"year": 2011,
"venue": "Annual Meeting of the Association for Computational Linguistics (ACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Bansal, C. Quirk, and R. Moore. 2011. Gappy phrasal alignment by agreement. In Annual Meet- ing of the Association for Computational Linguistics (ACL), Portland, Oregon, June.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "The mathematics of statistical machine translation: Parameter estimation",
"authors": [
{
"first": "P",
"middle": [
"F"
],
"last": "Brown",
"suffix": ""
},
{
"first": "S",
"middle": [
"A"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "V",
"middle": [
"J"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "R",
"middle": [
"L"
],
"last": "Mercer",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "263--311",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P.F. Brown, S.A. Della Pietra, V.J. Della Pietra, and R.L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Compu- tational Linguistics, 19(2):263-311, June.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "An alignment algorithm using Belief Propagation and a structurebased distortion model",
"authors": [
{
"first": "F",
"middle": [],
"last": "Cromi\u00e8res",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Kurohashi",
"suffix": ""
}
],
"year": 2009,
"venue": "Conference of the European Chapter of the Association for Computational Linguistics (EACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. Cromi\u00e8res and S. Kurohashi. 2009. An alignment algorithm using Belief Propagation and a structure- based distortion model. In Conference of the Euro- pean Chapter of the Association for Computational Linguistics (EACL), Athens, Greece, April.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Maximum likelihood from incomplete data via the EM algorithm",
"authors": [
{
"first": "A",
"middle": [
"P"
],
"last": "Dempster",
"suffix": ""
},
{
"first": "N",
"middle": [
"M"
],
"last": "Laird",
"suffix": ""
},
{
"first": "D",
"middle": [
"B"
],
"last": "Rubin",
"suffix": ""
}
],
"year": 1977,
"venue": "Journal of the Royal Statistical Society, Series B",
"volume": "39",
"issue": "1",
"pages": "1--38",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A.P. Dempster, N.M. Laird, and D.B. Rubin. 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical So- ciety, Series B, 39(1):1-38.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "HMM word and phrase alignment for statistical machine translation",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Deng",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Byrne",
"suffix": ""
}
],
"year": 2005,
"venue": "HLT-EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Deng and W. Byrne. 2005. HMM word and phrase alignment for statistical machine translation. In HLT-EMNLP, Vancouver, Canada, October.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Getting the structure right for word alignment: LEAF",
"authors": [
{
"first": "A",
"middle": [],
"last": "Fraser",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2007,
"venue": "Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Fraser and D. Marcu. 2007a. Getting the structure right for word alignment: LEAF. In Conference on Empirical Methods in Natural Language Processing (EMNLP), Prague, Czech Republic, June.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Measuring word alignment quality for statistical machine translation",
"authors": [
{
"first": "A",
"middle": [],
"last": "Fraser",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2007,
"venue": "Computational Linguistics",
"volume": "33",
"issue": "3",
"pages": "293--303",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Fraser and D. Marcu. 2007b. Measuring word alignment quality for statistical machine translation. Computational Linguistics, 33(3):293-303, Septem- ber.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Learning tractable word alignment models with complex constraints",
"authors": [
{
"first": "J",
"middle": [],
"last": "Gra\u00e7a",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Ganchev",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Taskar",
"suffix": ""
}
],
"year": 2010,
"venue": "Computational Linguistics",
"volume": "36",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Gra\u00e7a, K. Ganchev, and B. Taskar. 2010. Learning tractable word alignment models with complex con- straints. Computational Linguistics, 36, September.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Guidelines for word alignment evaluation and manual alignment",
"authors": [
{
"first": "P",
"middle": [],
"last": "Lambert",
"suffix": ""
},
{
"first": "A",
"middle": [
"D"
],
"last": "Gispert",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Banchs",
"suffix": ""
},
{
"first": "J",
"middle": [
"B"
],
"last": "Marino",
"suffix": ""
}
],
"year": 2005,
"venue": "Language Resources and Evaluation",
"volume": "39",
"issue": "4",
"pages": "267--285",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Lambert, A.D. Gispert, R. Banchs, and J.B. Marino. 2005. Guidelines for word alignment evaluation and manual alignment. Language Resources and Evalu- ation, 39(4):267-285.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Alignment by agreement",
"authors": [
{
"first": "P",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Taskar",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Klein",
"suffix": ""
}
],
"year": 2006,
"venue": "Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Liang, B. Taskar, and D. Klein. 2006. Alignment by agreement. In Human Language Technology Con- ference of the North American Chapter of the As- sociation of Computational Linguistics, New York, New York, June.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A phrase-based, joint probability model for statistical machine translation",
"authors": [
{
"first": "D",
"middle": [],
"last": "Marcu",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Wong",
"suffix": ""
}
],
"year": 2002,
"venue": "Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Marcu and W. Wong. 2002. A phrase-based, joint probability model for statistical machine trans- lation. In Conference on Empirical Methods in Nat- ural Language Processing (EMNLP), Philadelphia, Pennsylvania, July.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Models of translational equivalence among words",
"authors": [
{
"first": "D",
"middle": [],
"last": "Melamed",
"suffix": ""
}
],
"year": 2000,
"venue": "Computational Linguistics",
"volume": "26",
"issue": "2",
"pages": "221--249",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Melamed. 2000. Models of translational equiv- alence among words. Computational Linguistics, 26(2):221-249.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "A finite algorithm for finding the projection of a point onto the canonical simplex of IR n",
"authors": [
{
"first": "C",
"middle": [],
"last": "Michelot",
"suffix": ""
}
],
"year": 1986,
"venue": "Journal on Optimization Theory and Applications",
"volume": "50",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Michelot. 1986. A finite algorithm for finding the projection of a point onto the canonical simplex of IR n . Journal on Optimization Theory and Applica- tions, 50(1), July.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "A view of the EM algorithm that justifies incremental, sparse, and other variants",
"authors": [
{
"first": "R",
"middle": [
"M"
],
"last": "Neal",
"suffix": ""
},
{
"first": "G",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R.M. Neal and G.E. Hinton. 1998. A view of the EM algorithm that justifies incremental, sparse, and other variants. In M.I. Jordan, editor, Learning in Graphical Models. MIT press.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A systematic comparison of various statistical alignment models",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2003,
"venue": "Computational Linguistics",
"volume": "29",
"issue": "1",
"pages": "19--51",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F.J. Och and H. Ney. 2003. A systematic comparison of various statistical alignment models. Computa- tional Linguistics, 29(1):19-51.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Does GIZA++ make search errors?",
"authors": [
{
"first": "S",
"middle": [],
"last": "Ravi",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Knight",
"suffix": ""
}
],
"year": 2010,
"venue": "Computational Linguistics",
"volume": "",
"issue": "3",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Ravi and K. Knight. 2010. Does GIZA++ make search errors? Computational Linguistics, 36(3).",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Computing optimal alignments for the IBM-3 translation model",
"authors": [
{
"first": "T",
"middle": [],
"last": "Schoenemann",
"suffix": ""
}
],
"year": 2010,
"venue": "Conference on Computational Natural Language Learning (CoNLL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Schoenemann. 2010. Computing optimal align- ments for the IBM-3 translation model. In Confer- ence on Computational Natural Language Learning (CoNLL), Uppsala, Sweden, July.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Regularizing mono-and biword models for word alignment",
"authors": [
{
"first": "T",
"middle": [],
"last": "Schoenemann",
"suffix": ""
}
],
"year": 2011,
"venue": "International Joint Conference on Natural Language Processing (IJCNLP)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Schoenemann. 2011. Regularizing mono-and bi- word models for word alignment. In International Joint Conference on Natural Language Processing (IJCNLP), Chiang Mai, Thailand, November.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "EBMT, SMT, Hybrid and more: ATR spoken language translation system",
"authors": [
{
"first": "E",
"middle": [],
"last": "Sumita",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Akiba",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Doi",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Finch",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Imamura",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Okuma",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Paul",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Shimohata",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Watanabe",
"suffix": ""
}
],
"year": 2004,
"venue": "International Workshop on Spoken Language Translation (IWSLT)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "E. Sumita, Y. Akiba, T. Doi, A. Finch, K. Imamura, H. Okuma, M. Paul, M. Shimohata, and T. Watan- abe. 2004. EBMT, SMT, Hybrid and more: ATR spoken language translation system. In Interna- tional Workshop on Spoken Language Translation (IWSLT), Kyoto, Japan, September.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Extensions to HMM-based statistical word alignment models",
"authors": [
{
"first": "K",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "H",
"middle": [
"T"
],
"last": "Ilhan",
"suffix": ""
},
{
"first": "C",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2002,
"venue": "Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Toutanova, H.T. Ilhan, and C.D. Manning. 2002. Extensions to HMM-based statistical word align- ment models. In Conference on Empirical Meth- ods in Natural Language Processing (EMNLP), Philadelphia, Pennsylvania, July.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Computational complexity of statistical machine translation",
"authors": [
{
"first": "R",
"middle": [],
"last": "Udupa",
"suffix": ""
},
{
"first": "H",
"middle": [
"K"
],
"last": "Maji",
"suffix": ""
}
],
"year": 2006,
"venue": "Conference of the European Chapter of the Association for Computational Linguistics (EACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Udupa and H.K. Maji. 2006. Computational com- plexity of statistical machine translation. In Con- ference of the European Chapter of the Association for Computational Linguistics (EACL), Trento, Italy, April.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Smaller alignment models for better translations: Unsupervised word alignment with the l 0 -norm",
"authors": [
{
"first": "A",
"middle": [],
"last": "Vaswani",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Chiang",
"suffix": ""
}
],
"year": 2012,
"venue": "Annual Meeting of the Association for Computational Linguistics (ACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Vaswani, L. Huang, and D. Chiang. 2012. Smaller alignment models for better translations: Unsuper- vised word alignment with the l 0 -norm. In Annual Meeting of the Association for Computational Lin- guistics (ACL), Jeju, Korea, July.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "HMM-based word alignment in statistical translation",
"authors": [
{
"first": "S",
"middle": [],
"last": "Vogel",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Tillmann",
"suffix": ""
}
],
"year": 1996,
"venue": "International Conference on Computational Linguistics (COLING)",
"volume": "",
"issue": "",
"pages": "836--841",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Vogel, H. Ney, and C. Tillmann. 1996. HMM-based word alignment in statistical translation. In Inter- national Conference on Computational Linguistics (COLING), pages 836-841, Copenhagen, Denmark, August.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Modeling with structures in statistical machine translation",
"authors": [
{
"first": "Y.-Y",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Waibel",
"suffix": ""
}
],
"year": 1998,
"venue": "International Conference on Computational Linguistics (COLING)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y.-Y. Wang and A. Waibel. 1998. Modeling with structures in statistical machine translation. In In- ternational Conference on Computational Linguis- tics (COLING), Montreal, Canada, August.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "4. The remaining \u03a6 0 open positions in the foreign sequence align to position 0. Decide on the corresponding foreign words with p(f d 0,k |e 0 )",
"uris": null,
"type_str": "figure"
}
}
}
} |