File size: 116,463 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 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 | {
"paper_id": "P14-1012",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T09:04:37.365381Z"
},
"title": "Learning New Semi-Supervised Deep Auto-encoder Features for Statistical Machine Translation",
"authors": [
{
"first": "Shixiang",
"middle": [],
"last": "Lu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "China"
}
},
"email": "shixiang.lu@ia.ac.cn"
},
{
"first": "Zhenbiao",
"middle": [],
"last": "Chen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "China"
}
},
"email": "zhenbiao.chen@ia.ac.cn"
},
{
"first": "Bo",
"middle": [],
"last": "Xu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "China"
}
},
"email": "xubo@ia.ac.cn"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In this paper, instead of designing new features based on intuition, linguistic knowledge and domain, we learn some new and effective features using the deep autoencoder (DAE) paradigm for phrase-based translation model. Using the unsupervised pre-trained deep belief net (DBN) to initialize DAE's parameters and using the input original phrase features as a teacher for semi-supervised fine-tuning, we learn new semi-supervised DAE features, which are more effective and stable than the unsupervised DBN features. Moreover, to learn high dimensional feature representation, we introduce a natural horizontal composition of more DAEs for large hidden layers feature learning. On two Chinese-English tasks, our semi-supervised DAE features obtain statistically significant improvements of 1.34/2.45 (IWSLT) and 0.82/1.52 (NIST) BLEU points over the unsupervised DBN features and the baseline features, respectively.",
"pdf_parse": {
"paper_id": "P14-1012",
"_pdf_hash": "",
"abstract": [
{
"text": "In this paper, instead of designing new features based on intuition, linguistic knowledge and domain, we learn some new and effective features using the deep autoencoder (DAE) paradigm for phrase-based translation model. Using the unsupervised pre-trained deep belief net (DBN) to initialize DAE's parameters and using the input original phrase features as a teacher for semi-supervised fine-tuning, we learn new semi-supervised DAE features, which are more effective and stable than the unsupervised DBN features. Moreover, to learn high dimensional feature representation, we introduce a natural horizontal composition of more DAEs for large hidden layers feature learning. On two Chinese-English tasks, our semi-supervised DAE features obtain statistically significant improvements of 1.34/2.45 (IWSLT) and 0.82/1.52 (NIST) BLEU points over the unsupervised DBN features and the baseline features, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Recently, many new features have been explored for SMT and significant performance have been obtained in terms of translation quality, such as syntactic features, sparse features, and reordering features. However, most of these features are manually designed on linguistic phenomena that are related to bilingual language pairs, thus they are very difficult to devise and estimate.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Instead of designing new features based on intuition, linguistic knowledge and domain, for the first time, Maskey and Zhou (2012) explored the possibility of inducing new features in an unsupervised fashion using deep belief net (DBN) for hierarchical phrase-based trans-lation model. Using the 4 original phrase features in the phrase table as the input features, they pre-trained the DBN by contrastive divergence (Hinton, 2002) , and generated new unsupervised DBN features using forward computation. These new features are appended as extra features to the phrase table for the translation decoder.",
"cite_spans": [
{
"start": 107,
"end": 129,
"text": "Maskey and Zhou (2012)",
"ref_id": "BIBREF22"
},
{
"start": 416,
"end": 430,
"text": "(Hinton, 2002)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "However, the above approach has two major shortcomings. First, the input original features for the DBN feature learning are too simple, the limited 4 phrase features of each phrase pair, such as bidirectional phrase translation probability and bidirectional lexical weighting (Koehn et al., 2003) , which are a bottleneck for learning effective feature representation. Second, it only uses the unsupervised layer-wise pre-training of DBN built with stacked sets of Restricted Boltzmann Machines (RBM) (Hinton, 2002) , does not have a training objective, so its performance relies on the empirical parameters. Thus, this approach is unstable and the improvement is limited. In this paper, we strive to effectively address the above two shortcomings, and systematically explore the possibility of learning new features using deep (multilayer) neural networks (DNN, which is usually referred under the name Deep Learning) for SMT.",
"cite_spans": [
{
"start": 276,
"end": 296,
"text": "(Koehn et al., 2003)",
"ref_id": "BIBREF16"
},
{
"start": 501,
"end": 515,
"text": "(Hinton, 2002)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To address the first shortcoming, we adapt and extend some simple but effective phrase features as the input features for new DNN feature learning, and these features have been shown significant improvement for SMT, such as, phrase pair similarity (Zhao et al., 2004) , phrase frequency, phrase length (Hopkins and May, 2011) , and phrase generative probability (Foster et al., 2010) , which also show further improvement for new phrase feature learning in our experiments.",
"cite_spans": [
{
"start": 248,
"end": 267,
"text": "(Zhao et al., 2004)",
"ref_id": "BIBREF32"
},
{
"start": 302,
"end": 325,
"text": "(Hopkins and May, 2011)",
"ref_id": "BIBREF11"
},
{
"start": 362,
"end": 383,
"text": "(Foster et al., 2010)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To address the second shortcoming, inspired by the successful use of DAEs for handwritten digits recognition (Hinton and Salakhutdinov, 2006; , information retrieval (Salakhutdinov and Hinton, 2009; Mirowski et al., 2010) , and speech spectrograms (Deng et al., 2010) , we propose new feature learning using semi-supervised DAE for phrase-based translation model. By using the input data as the teacher, the \"semi-supervised\" fine-tuning process of DAE addresses the problem of \"back-propagation without a teacher\" (Rumelhart et al., 1986) , which makes the DAE learn more powerful and abstract features (Hinton and Salakhutdinov, 2006) . For our semisupervised DAE feature learning task, we use the unsupervised pre-trained DBN to initialize DAE's parameters and use the input original phrase features as the \"teacher\" for semi-supervised backpropagation. Compared with the unsupervised DBN features, our semi-supervised DAE features are more effective and stable.",
"cite_spans": [
{
"start": 109,
"end": 141,
"text": "(Hinton and Salakhutdinov, 2006;",
"ref_id": "BIBREF9"
},
{
"start": 185,
"end": 198,
"text": "Hinton, 2009;",
"ref_id": "BIBREF29"
},
{
"start": 199,
"end": 221,
"text": "Mirowski et al., 2010)",
"ref_id": "BIBREF23"
},
{
"start": 248,
"end": 267,
"text": "(Deng et al., 2010)",
"ref_id": "BIBREF4"
},
{
"start": 515,
"end": 539,
"text": "(Rumelhart et al., 1986)",
"ref_id": "BIBREF28"
},
{
"start": 604,
"end": 636,
"text": "(Hinton and Salakhutdinov, 2006)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Moreover, to learn high dimensional feature representation, we introduce a natural horizontal composition for DAEs (HCDAE) that can be used to create large hidden layer representations simply by horizontally combining two (or more) DAEs (Baldi, 2012) , which shows further improvement compared with single DAE in our experiments.",
"cite_spans": [
{
"start": 237,
"end": 250,
"text": "(Baldi, 2012)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "It is encouraging that, non-parametric feature expansion using gaussian mixture model (GMM) (Nguyen et al., 2007) , which guarantees invariance to the specific embodiment of the original features, has been proved as a feasible feature generation approach for SMT. Deep models such as DNN have the potential to be much more representationally efficient for feature learning than shallow models like GMM. Thus, instead of GMM, we use DNN (DBN, DAE and HCDAE) to learn new non-parametric features, which has the similar evolution in speech recognition Hinton et al., 2012) . DNN features are learned from the non-linear combination of the input original features, they strong capture highorder correlations between the activities of the original features, and we believe this deep learning paradigm induces the original features to further reach their potential for SMT.",
"cite_spans": [
{
"start": 92,
"end": 113,
"text": "(Nguyen et al., 2007)",
"ref_id": "BIBREF24"
},
{
"start": 549,
"end": 569,
"text": "Hinton et al., 2012)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Finally, we conduct large-scale experiments on IWSLT and NIST Chinese-English translation tasks, respectively, and the results demonstrate that our solutions solve the two aforementioned shortcomings successfully. Our semi-supervised DAE features significantly outperform the unsupervised DBN features and the baseline features, and our introduced input phrase features significantly improve the performance of DAE feature learning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The remainder of this paper is organized as follows. Section 2 briefly summarizes the recent related work about the applications of DNN for SMT tasks. Section 3 presents our introduced input features for DNN feature learning. Section 4 describes how to learn our semi-supervised DAE features for SMT. Section 5 describes and discusses the large-scale experimental results. Finally, we end with conclusions in section 6.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Recently, there has been growing interest in use of DNN for SMT tasks. Le et al. (2012) improved translation quality of n-gram translation model by using a bilingual neural LM, where translation probabilities are estimated using a continuous representation of translation units in lieu of standard discrete representations. Kalchbrenner and Blunsom (2013) introduced recurrent continuous translation models that comprise a class for purely continuous sentence-level translation models. Auli et al. (2013) presented a joint language and translation model based on a recurrent neural network which predicts target words based on an unbounded history of both source and target words. went beyond the log-linear model for SMT and proposed a novel additive neural networks based translation model, which overcome some of the shortcomings suffered by the log-linear model: linearity and the lack of deep interpretation and representation in features. Li et al. (2013) presented an ITG reordering classifier based on recursive autoencoders, and generated vector space representations for variable-sized phrases, which enable predicting orders to exploit syntactic and semantic information. Lu et al. (2014) adapted and extended the max-margin based RNN (Socher et al., 2011) into HPB translation with force decoding and converting tree, and proposed a RNN based word topology model for HPB translation, which successfully capture the topological structure of the words on the source side in a syntactically and semantically meaningful order.",
"cite_spans": [
{
"start": 71,
"end": 87,
"text": "Le et al. (2012)",
"ref_id": "BIBREF17"
},
{
"start": 324,
"end": 355,
"text": "Kalchbrenner and Blunsom (2013)",
"ref_id": "BIBREF12"
},
{
"start": 486,
"end": 504,
"text": "Auli et al. (2013)",
"ref_id": "BIBREF0"
},
{
"start": 945,
"end": 961,
"text": "Li et al. (2013)",
"ref_id": "BIBREF18"
},
{
"start": 1183,
"end": 1199,
"text": "Lu et al. (2014)",
"ref_id": "BIBREF20"
},
{
"start": 1246,
"end": 1267,
"text": "(Socher et al., 2011)",
"ref_id": "BIBREF30"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "However, none of these above works have focused on learning new features automatically with input data, and while learning suitable features (representations) is the superiority of DNN since it has been proposed. In this paper, we systematically explore the possibility of learning new fea-tures using DNN for SMT.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "The phrase-based translation model (Koehn et al., 2003; Och and Ney, 2004) has demonstrated superior performance and been widely used in current SMT systems, and we employ our implementation on this translation model. Next, we adapt and extend some original phrase features as the input features for DAE feature learning.",
"cite_spans": [
{
"start": 35,
"end": 55,
"text": "(Koehn et al., 2003;",
"ref_id": "BIBREF16"
},
{
"start": 56,
"end": 74,
"text": "Och and Ney, 2004)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Input Features for DNN Feature Learning",
"sec_num": "3"
},
{
"text": "We assume that source phrase",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "f = f 1 , \u2022 \u2022 \u2022 , f l f",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "and target phrase e = e 1 , \u2022 \u2022 \u2022 , e le include l f and l e words, respectively. Following (Maskey and Zhou, 2012) , we use the following 4 phrase features of each phrase pair (Koehn et al., 2003) in the phrase table as the first type of input features, bidirectional phrase translation probability (P (e|f ) and P (f |e)), bidirectional lexical weighting (Lex(e|f ) and Lex(f |e)),",
"cite_spans": [
{
"start": 92,
"end": 115,
"text": "(Maskey and Zhou, 2012)",
"ref_id": "BIBREF22"
},
{
"start": 177,
"end": 197,
"text": "(Koehn et al., 2003)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "X 1 \u2192 P (f |e), Lex(f |e), P (e|f ), Lex(e|f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "3.2 Phrase pair similarity Zhao et al. (2004) proposed a way of using term weight based models in a vector space as additional evidences for phrase pair translation quality. This model employ phrase pair similarity to encode the weights of content and non-content words in phrase translation pairs. Following (Zhao et al., 2004) , we calculate bidirectional phrase pair similarity using cosine distance and BM25 distance as,",
"cite_spans": [
{
"start": 27,
"end": 45,
"text": "Zhao et al. (2004)",
"ref_id": "BIBREF32"
},
{
"start": 309,
"end": 328,
"text": "(Zhao et al., 2004)",
"ref_id": "BIBREF32"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "S cos i (e, f ) = le j=1 l f i=1 w e j p(e j |f i )w f i sqrt( le j=1 w 2 e j )sqrt( le j=1 w e j a 2 ) S cos d (f, e) = l f i=1 le j=1 w f i p(f i |e j )w e j sqrt( l f i=1 w 2 f i )sqrt( l f i=1 w f i a 2 )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "where, p(e j |f i ) and p(f i |e j ) represents bidirectional word translation probability. w f i and w e j are term weights for source and target words, w e j a and w f i a are the transformed weights mapped from all source/target words to the target/source dimension at word e j and f i , respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "S bm25 i (e, f ) = l f i=1 idf f i (k 1 + 1)w f i (k 3 + 1)w f i a (K + w f i )(k 3 + w f i a ) S bm25 d (f, e) = le j=1 idf e j (k 1 + 1)w e j (k 3 + 1)w e j a (K + w e j )(k 3 + w e j a )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "where, k 1 , b, k 3 are set to be 1, 1 and 1000, respectively. K = k 1 ((1 \u2212 b) + J/avg(l)), and J is the phrase length (l e or l f ), avg(l) is the average phrase length. Thus, we have the second type of input features",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "X2 \u2192 S cos i (f, e), S bm25 i (f, e), S cos d (e, f ), S bm25 d (e, f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline phrase features",
"sec_num": "3.1"
},
{
"text": "We adapt and extend bidirectional phrase generative probabilities as the input features, which have been used for domain adaptation (Foster et al., 2010) . According to the background LMs, we estimate the bidirectional (source/target side) forward and backward phrase generative probabilities as",
"cite_spans": [
{
"start": 132,
"end": 153,
"text": "(Foster et al., 2010)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase generative probability",
"sec_num": "3.3"
},
{
"text": "P f (f ) = P (f1)P (f2|f1) \u2022 \u2022 \u2022 P (f l f |f l f \u2212n+1 , \u2022 \u2022 \u2022 , f l f \u22121 ) P f (e) = P (e1)P (e2|e1) \u2022 \u2022 \u2022 P (e le |e le\u2212n+1 , \u2022 \u2022 \u2022 , e le\u22121 ) P b (f ) = P (f l f )P (f l f \u22121 |f l f ) \u2022 \u2022 \u2022 P (f1|fn, \u2022 \u2022 \u2022 , f2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase generative probability",
"sec_num": "3.3"
},
{
"text": "P b (e) = P (e le )P (e le\u22121 |e le )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase generative probability",
"sec_num": "3.3"
},
{
"text": "\u2022 \u2022 \u2022 P (e1|en, \u2022 \u2022 \u2022 , e2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase generative probability",
"sec_num": "3.3"
},
{
"text": "where, the bidirectional forward and backward 1 background 4-gram LMs are trained by the corresponding side of bilingual corpus 2 . Then, we have the third type of input features X 3 \u2192 P f (e), P b (e), P f (f ), P b (f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase generative probability",
"sec_num": "3.3"
},
{
"text": "We consider bidirectional phrase frequency as the input features, and estimate them as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase frequency",
"sec_num": "3.4"
},
{
"text": "P (f ) = count(f ) |f i |=|f | count(f i )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase frequency",
"sec_num": "3.4"
},
{
"text": "P (e) = count(e) |e j |=|e| count(e j ) where, the count(f )/count(e) are the total number of phrase f/e appearing in the source/target side of the bilingual corpus, and the denominator are the total number of the phrases whose length are equal to |f |/|e|, respectively. Then, we have the forth type of input features X 4 \u2192 P (f ), P (e) 3.5 Phrase length Phrase length plays an important role in the translation process (Koehn, 2010; Hopkins and May, 2011) . We normalize bidirectional phrase length by the maximum phrase length, and introduce them as the last type of input features",
"cite_spans": [
{
"start": 422,
"end": 435,
"text": "(Koehn, 2010;",
"ref_id": "BIBREF14"
},
{
"start": 436,
"end": 458,
"text": "Hopkins and May, 2011)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase frequency",
"sec_num": "3.4"
},
{
"text": "X 5 \u2192 l n e , l n f",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase frequency",
"sec_num": "3.4"
},
{
"text": "In summary, except for the first type of phrase feature X 1 which is used by (Maskey and Zhou, 2012) , we introduce another four types of effective phrase features X 2 , X 3 , X 4 and X 5 . Now, the input original phrase features X includes 16 features in our experiments, as follows,",
"cite_spans": [
{
"start": 77,
"end": 100,
"text": "(Maskey and Zhou, 2012)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase frequency",
"sec_num": "3.4"
},
{
"text": "X \u2192 X 1 , X 2 , X 3 , X 4 , X 5",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase frequency",
"sec_num": "3.4"
},
{
"text": "We build the DAE network where the first layer with visible nodes equaling to 16, and each visible node v i corresponds to the above original features X in each phrase pair.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase frequency",
"sec_num": "3.4"
},
{
"text": "Each translation rule in the phrase-based translation model has a set number of features that are combined in the log-linear model (Och and Ney, 2002) , and our semi-supervised DAE features can also be combined in this model. In this section, we design our DAE network with various network structures for new feature learning.",
"cite_spans": [
{
"start": 131,
"end": 150,
"text": "(Och and Ney, 2002)",
"ref_id": "BIBREF26"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semi-Supervised Deep Auto-encoder Features Learning for SMT",
"sec_num": "4"
},
{
"text": "Inspired by (Maskey and Zhou, 2012) , we first learn a deep generative model for feature learning using DBN. DBN is composed of multiple layers of latent variables with the first layer representing the visible feature vectors, which is built with stacked sets of RBMs (Hinton, 2002) . For a RBM, there is full connectivity between layers, but no connections within either layer. The connection weight W , hidden layer biases c and visible layer biases b can be learned efficiently using the contrastive divergence (Hinton, 2002; Carreira-Perpinan and Hinton, 2005) . When given a hidden layer h, factorial conditional distribution of visible layer v can be estimated by",
"cite_spans": [
{
"start": 12,
"end": 35,
"text": "(Maskey and Zhou, 2012)",
"ref_id": "BIBREF22"
},
{
"start": 268,
"end": 282,
"text": "(Hinton, 2002)",
"ref_id": "BIBREF6"
},
{
"start": 514,
"end": 528,
"text": "(Hinton, 2002;",
"ref_id": "BIBREF6"
},
{
"start": 529,
"end": 564,
"text": "Carreira-Perpinan and Hinton, 2005)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "P (v = 1|h) = \u03c3(b + h T W T )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "where \u03c3 denotes the logistic sigmoid. Given v, the element-wise conditional distribution of h is The two conditional distributions can be shown to correspond to the generative model,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "P (h = 1|v) = \u03c3(c + v T W )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "P (v, h) = 1 Z exp(\u2212E(v, h))",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": ", Z = v,h e \u2212E(v,h) E(v, h) = \u2212b T v \u2212 c T h \u2212 v T W h",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "After learning the first RBM, we treat the activation probabilities of its hidden units, when they are being driven by data, as the data for training a second RBM. Similarly, a n th RBM is built on the output of the n \u2212 1 th one and so on until a sufficiently deep architecture is created. These n RBMs can then be composed to form a DBN in which it is easy to infer the states of the n th layer of hidden units from the input in a single forward pass , as shown in Figure 1 . This greedy, layer-by-layer pre-training can be repeated several times to learn a deep, hierarchical model (DBN) in which each layer of features captures strong high-order correlations between the activities of features in the layer below.",
"cite_spans": [],
"ref_spans": [
{
"start": 466,
"end": 474,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "To deal with real-valued input features X in our task, we use an RBM with Gaussian visible units (GRBM) with a variance of 1 on each dimension. Hence, P (v|h) and E(v, h) in the first RBM of DBN need to be modified as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "P (v|h) = N (v; b + h T W T , I) E(v, h) = 1 2 (v \u2212 b) T (v \u2212 b) \u2212 c T h \u2212 v T W h",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "where I is the appropriate identity matrix. Figure 2 : After the unsupervised pre-training, the DBNs are \"unrolled\" to create a semisupervised DAE, which is then fine-tuned using back-propagation of error derivatives.",
"cite_spans": [],
"ref_spans": [
{
"start": 44,
"end": 52,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "To speed-up the pre-training, we subdivide the entire phrase pairs (with features X) in the phrase table into small mini-batches, each containing 100 cases, and update the weights after each minibatch. Each layer is greedily pre-trained for 50 epochs through the entire phrase pairs. The weights are updated using a learning rate of 0.1, momentum of 0.9, and a weight decay of 0.0002 \u00d7 weight \u00d7 learning rate. The weight matrix W are initialized with small random values sampled from a zero-mean normal distribution with variance 0.01.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "After the pre-training, for each phrase pair in the phrase table, we generate the DBN features (Maskey and Zhou, 2012) by passing the original phrase features X through the DBN using forward computation.",
"cite_spans": [
{
"start": 95,
"end": 118,
"text": "(Maskey and Zhou, 2012)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Learning a Deep Belief Net",
"sec_num": "4.1"
},
{
"text": "To learn a semi-supervised DAE, we first \"unroll\" the above n layer DBN by using its weight matrices to create a deep, 2n-1 layer network whose lower layers use the matrices to \"encode\" the input and whose upper layers use the matrices in reverse order to \"decode\" the input (Hinton and Salakhutdinov, 2006; Salakhutdinov and Hinton, 2009; Deng et al., 2010) , as shown in Figure 2 . The layer-wise learning of DBN as above must be treated as a pre-training stage that finds a good region of the parameter space, which is used to initialize our DAE's parameters. Starting in this region, the DAE is then fine-tuned using average squared error (between the output and input) backpropagation to minimize reconstruction error, as to make its output as equal as possible to its input.",
"cite_spans": [
{
"start": 275,
"end": 307,
"text": "(Hinton and Salakhutdinov, 2006;",
"ref_id": "BIBREF9"
},
{
"start": 308,
"end": 339,
"text": "Salakhutdinov and Hinton, 2009;",
"ref_id": "BIBREF29"
},
{
"start": 340,
"end": 358,
"text": "Deng et al., 2010)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 373,
"end": 381,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "From DBN to Deep Auto-encoder",
"sec_num": "4.2"
},
{
"text": "For the fine-tuning of DAE, we use the method of conjugate gradients on larger mini-batches of 1000 cases, with three line searches performed for each mini-batch in each epoch. To determine an adequate number of epochs and to avoid overfitting, we fine-tune on a fraction phrase table and test performance on the remaining validation phrase table, and then repeat fine-tuning on the entire phrase table for 100 epochs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "From DBN to Deep Auto-encoder",
"sec_num": "4.2"
},
{
"text": "We experiment with various values for the noise variance and the threshold, as well as the learning rate, momentum, and weight-decay parameters used in the pre-training, the batch size and epochs in the fine-tuning. Our results are fairly robust to variations in these parameters. The precise weights found by the pre-training do not matter as long as it finds a good region of the parameter space from which to start the fine-tuning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "From DBN to Deep Auto-encoder",
"sec_num": "4.2"
},
{
"text": "The fine-tuning makes the feature representation in the central layer of the DAE work much better (Salakhutdinov and Hinton, 2009) . After the fine-tuning, for each phrase pair in the phrase table, we estimate our DAE features by passing the original phrase features X through the \"encoder\" part of the DAE using forward computation.",
"cite_spans": [
{
"start": 117,
"end": 130,
"text": "Hinton, 2009)",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "From DBN to Deep Auto-encoder",
"sec_num": "4.2"
},
{
"text": "To combine these learned features (DBN and DAE feature) into the log-linear model, we need to eliminate the impact of the non-linear learning mechanism. Following (Maskey and Zhou, 2012) , these learned features are normalized by the average of each dimensional respective feature set. Then, we append these features for each phrase pair to the phrase table as extra features.",
"cite_spans": [
{
"start": 163,
"end": 186,
"text": "(Maskey and Zhou, 2012)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "From DBN to Deep Auto-encoder",
"sec_num": "4.2"
},
{
"text": "Auto-encoders (HCDAE)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Horizontal Composition of Deep",
"sec_num": "4.3"
},
{
"text": "Although DAE can learn more powerful and abstract feature representation, the learned features usually have smaller dimensionality compared with the dimensionality of the input features, such as the successful use for handwritten digits recognition (Hinton and Salakhutdinov, 2006; , information retrieval (Salakhutdinov and Hinton, 2009; Mirowski et al., 2010) , and speech spectrograms (Deng et al., 2010) . Moreover, although we have introduced another four types of phrase features (X 2 , X 3 , X 4 and X 5 ), the only 16 features in X are a bottleneck for learning large hidden layers feature representation, because it has limited information, the performance of the high-dimensional DAE features which are directly learned from single DAE is not very satisfactory.",
"cite_spans": [
{
"start": 249,
"end": 281,
"text": "(Hinton and Salakhutdinov, 2006;",
"ref_id": "BIBREF9"
},
{
"start": 325,
"end": 338,
"text": "Hinton, 2009;",
"ref_id": "BIBREF29"
},
{
"start": 339,
"end": 361,
"text": "Mirowski et al., 2010)",
"ref_id": "BIBREF23"
},
{
"start": 388,
"end": 407,
"text": "(Deng et al., 2010)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Horizontal Composition of Deep",
"sec_num": "4.3"
},
{
"text": "To learn high-dimensional feature representation and to further improve the performance, we introduce a natural horizontal composition for DAEs that can be used to create large hidden layer representations simply by horizontally combining two (or more) DAEs (Baldi, 2012) , as shown in Figure 3 . Two single DAEs with architectures 16/m 1 /16 and 16/m 2 /16 can be trained and the hidden layers can be combined to yield an expanded hidden feature representation of size m 1 + m 2 , which can then be fed to the subsequent layers of the overall architecture. Thus, these new m 1 + m 2 -dimensional DAE features are added as extra features to the phrase table.",
"cite_spans": [
{
"start": 258,
"end": 271,
"text": "(Baldi, 2012)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 286,
"end": 294,
"text": "Figure 3",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Horizontal Composition of Deep",
"sec_num": "4.3"
},
{
"text": "Differences in m 1 -and m 2 -dimensional hidden representations could be introduced by many different mechanisms (e.g., learning algorithms, initializations, training samples, learning rates, or distortion measures) (Baldi, 2012) . In our task, we introduce differences by using different initializations and different fractions of the phrase table. -32-16-2-16-32-16 . Correspondingly, 4-16-8-2 and 6(8)-16-8-2 represent the input features are X 1 and X 1 +X i .",
"cite_spans": [
{
"start": 216,
"end": 229,
"text": "(Baldi, 2012)",
"ref_id": "BIBREF1"
},
{
"start": 350,
"end": 367,
"text": "-32-16-2-16-32-16",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Horizontal Composition of Deep",
"sec_num": "4.3"
},
{
"text": "We now test our DAE features on the following two Chinese-English translation tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "5.1"
},
{
"text": "IWSLT. The bilingual corpus is the Chinese-English part of Basic Traveling Expression corpus (BTEC) and China-Japan-Korea (CJK) corpus (0.38M sentence pairs with 3.5/3.8M Chinese/English words). The LM corpus is the English side of the parallel data (BTEC, CJK and CWMT08 3 ) (1.34M sentences). Our development set is IWSLT 2005 test set (506 sentences), and our test set is IWSLT 2007 test set (489 sentences).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "5.1"
},
{
"text": "NIST. The bilingual corpus is LDC 4 (3.4M sentence pairs with 64/70M Chinese/English words). The LM corpus is the English side of the parallel data as well as the English Gigaword corpus (LDC2007T07) (11.3M sentences). Our development set is NIST 2005 MT evaluation set (1084 sentences), and our test set is NIST 2006 MT evaluation set (1664 sentences).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "5.1"
},
{
"text": "We choose the Moses (Koehn et al., 2007) (Maskey and Zhou, 2012) , our proposed DAE and HCDAE feature) as extra features to the phrase table on two tasks. \"DBN X 1 xf\", \"DBN X xf\", \"DAE X 1 xf\" and \"DAE X xf\" represent that we use DBN and DAE, input features X 1 and X, to learn x-dimensional features, respectively. \"HCDAE X x+xf\" represents horizontally combining two DAEs and each DAE has the same x-dimensional learned features. All improvements on two test sets are statistically significant by the bootstrap resampling (Koehn, 2004) . *: significantly better than the baseline (p < 0.05), **: significantly better than \"DBN X 1 xf\" or \"DBN X xf\" (p < 0.01), ***: significantly better than \"DAE X 1 xf\" or \"DAE X xf\" (p < 0.01), ****: significantly better than \"HCDAE X x+xf\" (p < 0.01), +: significantly better than \"X 2 +X 3 +X 4 +X 5 \" (p < 0.01).",
"cite_spans": [
{
"start": 20,
"end": 40,
"text": "(Koehn et al., 2007)",
"ref_id": "BIBREF15"
},
{
"start": 41,
"end": 64,
"text": "(Maskey and Zhou, 2012)",
"ref_id": "BIBREF22"
},
{
"start": 525,
"end": 538,
"text": "(Koehn, 2004)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "5.1"
},
{
"text": "discounting. We perform pairwise ranking optimization (Hopkins and May, 2011) to tune feature weights. The translation quality is evaluated by case-insensitive IBM BLEU-4 metric. The baseline translation models are generated by Moses with default parameter settings. In the contrast experiments, our DAE and HCDAE features are appended as extra features to the phrase table. The details of the used network structure in our experiments are shown in Table 1 . Table 2 presents the main translation results. We use DBN, DAE and HCDAE (with 6 layers' network depth), input features X 1 and X, to learn 2-, 4-and 8-dimensional features, respectively. From the results, we can get some clear trends:",
"cite_spans": [
{
"start": 54,
"end": 77,
"text": "(Hopkins and May, 2011)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 449,
"end": 456,
"text": "Table 1",
"ref_id": "TABREF0"
},
{
"start": 459,
"end": 466,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "5.1"
},
{
"text": "1. Adding new DNN features as extra features significantly improves translation accuracy (row 2-17 vs. 1), with the highest increase of 2.45 (IWSLT) and 1.52 (NIST) (row 14 vs. 1) BLEU points over the baseline features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "2. Compared with the unsupervised DBN features, our semi-supervised DAE features are more effective for translation decoder (row 3 vs. 2; row 5 vs. 4; row 8 vs. 7; row 11 vs. 10; row 13 vs. 12; row 16 vs. 15). Specially, Table 3 shows the variance distributions of the learned each dimensional DBN and DAE feature, our DAE features have bigger variance distributions which means Table 4 : The effectiveness of our introduced input features. \"DAE X 1 +X i 4f\" represents that we use DAE, input features X 1 + X i , to learn 4dimensional features. *: significantly better than \"DAE X 1 4f\" (p < 0.05).",
"cite_spans": [],
"ref_spans": [
{
"start": 221,
"end": 228,
"text": "Table 3",
"ref_id": "TABREF4"
},
{
"start": 379,
"end": 386,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "IWSLT NIST \u03c3 1 \u03c3 2 \u03c3 3 \u03c3 4 \u03c3 1 \u03c3 2 \u03c3 3 \u03c3 4 DBN X 1 4f",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "our DAE features have more discriminative power, and also their variance distributions are more stable.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "3. HCDAE outperforms single DAE for high dimensional feature learning (row 6 vs. 5; row 9 vs. 8; row 14 vs. 13; row 17 vs. 16), and further improve the performance of DAE feature learning, which can also somewhat address the bring shortcoming of the limited input features. 4. Except for the phrase feature X 1 (Maskey and Zhou, 2012), our introduced input features X significantly improve the DAE feature learning (row 11 vs. 3; row 13 vs. 5; row 16 vs. 8). Specially, Table 4 shows the detailed effectiveness of our introduced input features for DAE feature learning, and the results show that each type of features are very effective for DAE feature learning.",
"cite_spans": [],
"ref_spans": [
{
"start": 470,
"end": 477,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "5. Adding the original features (X 2 , X 3 , X 4 and X 5 ) and DAE/HCDAE features together can further improve translation performance (row 19-23 vs. 18), with the highest increase of 3.16 (IWSLT) and 2.06 (NIST) (row 21 vs. 1) BLEU points over the baseline features. DAE and HCDAE features are learned from the non-linear combination of the original features, they strong capture high-order correlations between the activities of the original features, as to be further interpreted to reach their potential for SMT. We suspect these learned fea-tures are complementary to the original features. Figure 5 : The compared results of using single DAE and the HCDAE for feature learning on two development sets. Figure 4 shows our DAE features are not only more effective but also more stable than DBN features with various network structures. Also, adding more input features (X vs. X 1 ) not only significantly improves the performance of DAE feature learning, but also slightly improves the performance of DBN feature learning. Figure 5 shows there is little change in the performance of using single DAE to learn different dimensional DAE features, but the 4-dimensional features work more better and more stable. HC-DAE outperforms the single DAE and learns highdimensional representation more effectively, especially for the peak point in each condition.",
"cite_spans": [],
"ref_spans": [
{
"start": 596,
"end": 604,
"text": "Figure 5",
"ref_id": null
},
{
"start": 708,
"end": 716,
"text": "Figure 4",
"ref_id": null
},
{
"start": 1027,
"end": 1035,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5.2"
},
{
"text": "Figures 5 also shows the best network depth for DAE feature learning is 6 layers. When the network depth of DBN is 7 layers, the network depth of corresponding DAE during the fine-tuning is 13 layers. Although we have pre-trained the corresponding DBN, this DAE network is so deep, the fine-tuning does not work very well and typically finds poor local minima. We suspect this leads to the decreased performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analysis",
"sec_num": "5.3"
},
{
"text": "In this paper, instead of designing new features based on intuition, linguistic knowledge and domain, we have learned new features using the DAE for the phrase-based translation model. Using the unsupervised pre-trained DBN to initialize DAE's parameters and using the input original phrase features as the \"teacher\" for semi-supervised backpropagation, our semi-supervised DAE features are more effective and stable than the unsupervised DBN features (Maskey and Zhou, 2012) . Moreover, to further improve the performance, we introduce some simple but effective features as the input features for feature learning. Lastly, to learn high dimensional feature representation, we introduce a natural horizontal composition of two DAEs for large hidden layers feature learning.",
"cite_spans": [
{
"start": 452,
"end": 475,
"text": "(Maskey and Zhou, 2012)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "On two Chinese-English translation tasks, the results demonstrate that our solutions solve the two aforementioned shortcomings successfully. Firstly, our DAE features obtain statistically significant improvements of 1.34/2.45 (IWSLT) and 0.82/1.52 (NIST) BLEU points over the DBN features and the baseline features, respectively. Secondly, compared with the baseline phrase features X 1 , our introduced input original phrase features X significantly improve the performance of not only our DAE features but also the DBN features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "The results also demonstrate that DNN (DAE and HCDAE) features are complementary to the original features for SMT, and adding them together obtain statistically significant improvements of 3.16 (IWSLT) and 2.06 (NIST) BLEU points over the baseline features. Compared with the original features, DNN (DAE and HCDAE) features are learned from the non-linear combination of the original features, they strong capture high-order correlations between the activities of the original features, and we believe this deep learning paradigm induces the original features to further reach their potential for SMT.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "Backward LM has been introduced byXiong et al. (2011), which successfully capture both the preceding and succeeding contexts of the current word, and we estimate the backward LM by inverting the order in each sentence in the training data from the original order to the reverse order.2 This corpus is used to train the translation model in our experiments, and we will describe it in detail in section 5.1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "the 4th China Workshop on Machine Translation 4 LDC2002E18, LDC2002T01, LDC2003E07, LDC2003E14, LDC2003T17, LDC2004T07, LDC2004T08, LDC2005T06, LDC2005T10, LDC2005T34, LDC2006T04, LDC2007T09",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This work was supported by 863 program in China (No. 2011AA01A207). We would like to thank Xingyuan Peng, Lichun Fan and Hongyan Li for their helpful discussions. We also thank the anonymous reviewers for their insightful comments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Joint language and translation modeling with recurrent neural networks",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Auli",
"suffix": ""
},
{
"first": "Michel",
"middle": [],
"last": "Galley",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of EMNLP",
"volume": "",
"issue": "",
"pages": "1044--1054",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Auli, Michel Galley, Chris Quirk and Geoffrey Zweig. 2013. Joint language and translation model- ing with recurrent neural networks. In Proceedings of EMNLP, pages 1044-1054.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Autoencoders, unsupervised learning, and deep architectures",
"authors": [
{
"first": "Pierre",
"middle": [],
"last": "Baldi",
"suffix": ""
}
],
"year": 2012,
"venue": "",
"volume": "27",
"issue": "",
"pages": "37--50",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pierre Baldi. 2012. Autoencoders, unsupervised learn- ing, and deep architectures. JMLR: workshop on un- supervised and transfer learning, 27:37-50.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "On contrastive divergence learning",
"authors": [
{
"first": "Miguel",
"middle": [
"A"
],
"last": "Carreira-Perpinan",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of AI and Statistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Miguel A. Carreira-Perpinan and Geoffrey E. Hinton. 2005. On contrastive divergence learning. In Pro- ceedings of AI and Statistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Context-dependent pre-trained deep neural networks for large vocabulary speech recognition",
"authors": [
{
"first": "George",
"middle": [],
"last": "Dahl",
"suffix": ""
},
{
"first": "Dong",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "Li",
"middle": [],
"last": "Deng",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Acero",
"suffix": ""
}
],
"year": 2012,
"venue": "IEEE Transactions on Audio, Speech, and Language Processing",
"volume": "20",
"issue": "1",
"pages": "30--42",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "George Dahl, Dong Yu, Li Deng, and Alex Acero. 2012. Context-dependent pre-trained deep neural networks for large vocabulary speech recognition. IEEE Transactions on Audio, Speech, and Language Processing, 20(1):30-42.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Binary coding of speech spectrograms using a deep auto-encoder",
"authors": [
{
"first": "Li",
"middle": [],
"last": "Deng",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Seltzer",
"suffix": ""
},
{
"first": "Dong",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Acero",
"suffix": ""
},
{
"first": "Abdelrahman",
"middle": [],
"last": "Mohamed",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of INTERSPEECH",
"volume": "",
"issue": "",
"pages": "1692--1695",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Li Deng, Mike Seltzer, Dong Yu, Alex Acero, Abdel- rahman Mohamed, and Geoffrey E. Hinton. 2010. Binary coding of speech spectrograms using a deep auto-encoder. In Proceedings of INTERSPEECH, pages 1692-1695.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Discriminative instance weighting for domain adaptation in statistical machine translation",
"authors": [
{
"first": "George",
"middle": [],
"last": "Foster",
"suffix": ""
},
{
"first": "Cyril",
"middle": [],
"last": "Goutte",
"suffix": ""
},
{
"first": "Roland",
"middle": [],
"last": "Kuhn",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of EMNLP",
"volume": "",
"issue": "",
"pages": "451--459",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "George Foster, Cyril Goutte, and Roland Kuhn. 2010. Discriminative instance weighting for domain adap- tation in statistical machine translation. In Proceed- ings of EMNLP, pages 451-459.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Training products of experts by minimizing contrastive divergence",
"authors": [
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
}
],
"year": 2002,
"venue": "Neural Computation",
"volume": "14",
"issue": "8",
"pages": "1771--1800",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geoffrey E. Hinton. 2002. Training products of ex- perts by minimizing contrastive divergence. Neural Computation, 14(8):1771-1800.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Deep neural networks for acoustic modeling in speech tecognition",
"authors": [
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "Li",
"middle": [],
"last": "Deng",
"suffix": ""
},
{
"first": "Dong",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "George",
"middle": [],
"last": "Dahl",
"suffix": ""
},
{
"first": "Abdel-Rahman",
"middle": [],
"last": "Mohamed",
"suffix": ""
},
{
"first": "Navdeep",
"middle": [],
"last": "Jaitly",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Senior",
"suffix": ""
},
{
"first": "Vincent",
"middle": [],
"last": "Vanhoucke",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Nguyen",
"suffix": ""
},
{
"first": "Tara",
"middle": [],
"last": "Sainath",
"suffix": ""
},
{
"first": "Brian",
"middle": [],
"last": "Kingsbury",
"suffix": ""
}
],
"year": 2012,
"venue": "IEEE Signal Processing Magazine",
"volume": "29",
"issue": "6",
"pages": "82--97",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geoffrey E. Hinton, Li Deng, Dong Yu, George Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury. 2012. Deep neural networks for acoustic modeling in speech tecogni- tion. IEEE Signal Processing Magazine, 29(6):82- 97.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Transforming auto-encoders",
"authors": [
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Krizhevsky",
"suffix": ""
},
{
"first": "Sida",
"middle": [
"D"
],
"last": "Wang",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of ANN",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geoffrey E. Hinton, Alex Krizhevsky, and Sida D. Wang. 2001. Transforming auto-encoders. In Pro- ceedings of ANN.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Reducing the dimensionality of data with neural networks",
"authors": [
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [
"R"
],
"last": "Salakhutdinov",
"suffix": ""
}
],
"year": 2006,
"venue": "Science",
"volume": "313",
"issue": "",
"pages": "504--507",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geoffrey E. Hinton and Ruslan R. Salakhutdinov. 2006. Reducing the dimensionality of data with neural networks. Science, 313:504-507.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A fast learning algorithm for deep belief nets",
"authors": [
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "Simon",
"middle": [],
"last": "Osindero",
"suffix": ""
},
{
"first": "Yee-Whye",
"middle": [],
"last": "Teh",
"suffix": ""
}
],
"year": 2006,
"venue": "Neural Computation",
"volume": "18",
"issue": "",
"pages": "1527--1544",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geoffrey E. Hinton, Simon Osindero, and Yee-Whye Teh. 2006. A fast learning algorithm for deep belief nets. Neural Computation, 18:1527-1544.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Tuning as ranking",
"authors": [
{
"first": "Mark",
"middle": [],
"last": "Hopkins",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of EMNLP",
"volume": "",
"issue": "",
"pages": "1352--1362",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mark Hopkins and Jonathan May 2011. Tuning as ranking. In Proceedings of EMNLP, pages 1352- 1362.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Recurrent continuous translation models",
"authors": [
{
"first": "Nal",
"middle": [],
"last": "Kalchbrenner",
"suffix": ""
},
{
"first": "Phil",
"middle": [],
"last": "Blunsom",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of EMNLP",
"volume": "",
"issue": "",
"pages": "1700--1709",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nal Kalchbrenner and Phil Blunsom. 2013. Recur- rent continuous translation models. In Proceedings of EMNLP, pages 1700-1709.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Statistical significance tests from achine translation evaluation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of ACL",
"volume": "",
"issue": "",
"pages": "388--395",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn. 2004. Statistical significance tests from achine translation evaluation. In Proceedings of ACL, pages 388-395.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Statistical machine translation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn. 2010. Statistical machine translation. Cambridge University Press.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Moses: Open source toolkit for statistical machine translation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "Hieu",
"middle": [],
"last": "Hoang",
"suffix": ""
},
{
"first": "Alexandra",
"middle": [],
"last": "Birch",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
},
{
"first": "Marcello",
"middle": [],
"last": "Federico",
"suffix": ""
},
{
"first": "Nicola",
"middle": [],
"last": "Bertoldi",
"suffix": ""
},
{
"first": "Brooke",
"middle": [],
"last": "Cowan",
"suffix": ""
},
{
"first": "Wade",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "Christine",
"middle": [],
"last": "Moran",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Zens",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Dyer",
"suffix": ""
},
{
"first": "Ondrej",
"middle": [],
"last": "Bojar",
"suffix": ""
},
{
"first": "Alexandra",
"middle": [],
"last": "Constantin",
"suffix": ""
},
{
"first": "Evan",
"middle": [],
"last": "Herbst",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of ACL, Demonstration Session",
"volume": "",
"issue": "",
"pages": "177--180",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of ACL, Demonstration Session, pages 177-180.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Statistical phrase-based translation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "Franz",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of NAACL",
"volume": "",
"issue": "",
"pages": "48--54",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn, Franz J. Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of NAACL, pages 48-54.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Continuous space translation models with neural networks",
"authors": [
{
"first": "Hai-Son",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Alexandre",
"middle": [],
"last": "Allauzen",
"suffix": ""
},
{
"first": "Fran\u00e7ois",
"middle": [],
"last": "Yvon",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of NAACL",
"volume": "",
"issue": "",
"pages": "39--48",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hai-Son Le, Alexandre Allauzen, and Fran\u00e7ois Yvon. 2012. Continuous space translation models with neural networks. In Proceedings of NAACL, pages 39-48.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Recursive autoencoders for ITG-based translation",
"authors": [
{
"first": "Peng",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Yang",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Maosong",
"middle": [],
"last": "Sun",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of EMNLP",
"volume": "",
"issue": "",
"pages": "567--577",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Peng Li, Yang Liu, Maosong Sun. 2013. Recursive autoencoders for ITG-based translation. In Proceed- ings of EMNLP, pages 567-577.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Additive neural networks for statistical machine translation",
"authors": [
{
"first": "Lemao",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Taro",
"middle": [],
"last": "Watanabe",
"suffix": ""
},
{
"first": "Eiichiro",
"middle": [],
"last": "Sumita",
"suffix": ""
},
{
"first": "Tiejun",
"middle": [],
"last": "Zhao",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of ACL",
"volume": "",
"issue": "",
"pages": "791--801",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lemao Liu, Taro Watanabe, Eiichiro Sumita, and Tiejun Zhao. 2013. Additive neural networks for statistical machine translation. In Proceedings of ACL, pages 791-801.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Recursive neural network based word topology model for hierarchical phrase-based speech translation",
"authors": [
{
"first": "Shixiang",
"middle": [],
"last": "Lu",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Wei",
"suffix": ""
},
{
"first": "Xiaoyin",
"middle": [],
"last": "Fu",
"suffix": ""
},
{
"first": "Bo",
"middle": [],
"last": "Xu",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of ICASSP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shixiang Lu, Wei Wei, Xiaoyin Fu and Bo Xu. 2014. Recursive neural network based word topology model for hierarchical phrase-based speech transla- tion. In Proceedings of ICASSP.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Soft syntactic constraints for hierarchical phrase-based translation",
"authors": [
{
"first": "Yuval",
"middle": [],
"last": "Marton",
"suffix": ""
},
{
"first": "Philip",
"middle": [],
"last": "Resnik",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL",
"volume": "",
"issue": "",
"pages": "1003--1011",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuval Marton and Philip Resnik. 2008. Soft syntactic constraints for hierarchical phrase-based translation. In Proceedings of ACL, pages 1003-1011.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Unsupervised deep belief features for speech translation",
"authors": [
{
"first": "Sameer",
"middle": [],
"last": "Maskey",
"suffix": ""
},
{
"first": "Bowen",
"middle": [],
"last": "Zhou",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of INTERSPEECH",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sameer Maskey and Bowen Zhou. 2012. Unsuper- vised deep belief features for speech translation. In Proceedings of INTERSPEECH.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Dynamic auto-encoders for semantic indexing",
"authors": [
{
"first": "Piotr",
"middle": [],
"last": "Mirowski",
"suffix": ""
},
{
"first": "Marcaurelio",
"middle": [],
"last": "Ranzato",
"suffix": ""
},
{
"first": "Yann",
"middle": [],
"last": "Le-Cun",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of NIPS-2010 Workshop on Deep Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Piotr Mirowski, MarcAurelio Ranzato, and Yann Le- Cun. 2010. Dynamic auto-encoders for semantic indexing. In Proceedings of NIPS-2010 Workshop on Deep Learning.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Training non-parametric features for statistical machine translation",
"authors": [
{
"first": "Patrick",
"middle": [],
"last": "Nguyen",
"suffix": ""
},
{
"first": "Milind",
"middle": [],
"last": "Mahajan",
"suffix": ""
},
{
"first": "Xiaodong",
"middle": [],
"last": "He",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of WMT",
"volume": "",
"issue": "",
"pages": "72--79",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Patrick Nguyen, Milind Mahajan, and Xiaodong He. 2007. Training non-parametric features for statis- tical machine translation. In Proceedings of WMT, pages 72-79.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Improved statistical alignment models",
"authors": [
{
"first": "J",
"middle": [],
"last": "Franz",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of ACL",
"volume": "",
"issue": "",
"pages": "440--447",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz J. Och and Hermann Ney. 2000. Improved sta- tistical alignment models. In Proceedings of ACL, pages 440-447.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Discriminative training and maximum entropy models for statistical machine translation",
"authors": [
{
"first": "J",
"middle": [],
"last": "Franz",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of ACL",
"volume": "",
"issue": "",
"pages": "295--302",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz J. Och and Hermann Ney. 2002. Discriminative training and maximum entropy models for statistical machine translation. In Proceedings of ACL, pages 295-302.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "The alignment template approach to statistical machine translation",
"authors": [
{
"first": "J",
"middle": [],
"last": "Franz",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2004,
"venue": "Computational Linguistics",
"volume": "30",
"issue": "4",
"pages": "417--449",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz J. Och and Hermann Ney. 2004. The alignment template approach to statistical machine translation. Computational Linguistics, 30(4):417-449.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Learning internal representations by back-propagation errors",
"authors": [
{
"first": "David",
"middle": [],
"last": "Rumelhart",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "Ronale",
"middle": [],
"last": "Williams",
"suffix": ""
}
],
"year": 1986,
"venue": "Parallel Distributed Processing",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Rumelhart, Geoffrey E. Hinton, and Ronale Williams. 1986. Learning internal representations by back-propagation errors. Parallel Distributed Processing, Vol 1: Foundations, MIT Press.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Semantic hashing",
"authors": [
{
"first": "R",
"middle": [],
"last": "Ruslan",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hinton",
"suffix": ""
}
],
"year": 2009,
"venue": "International Journal of Approximate Reasoning",
"volume": "50",
"issue": "7",
"pages": "969--978",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ruslan R. Salakhutdinov and Geoffrey E. Hinton. 2009. Semantic hashing. International Journal of Approximate Reasoning, 50(7):969-978.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Parsing natural scenes and natural language with recursive neural networks",
"authors": [
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
},
{
"first": "Cliff",
"middle": [
"C"
],
"last": "Lin",
"suffix": ""
},
{
"first": "Andrew",
"middle": [
"Y"
],
"last": "Ng",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of ICML",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Richard Socher, Cliff C. Lin, Andrew Y. Ng, and Christopher D. Manning. 2011. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of ICML.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Enhancing language models in statistical machine translation with backward n-grams and mutual information triggers",
"authors": [
{
"first": "Deyi",
"middle": [],
"last": "Xiong",
"suffix": ""
},
{
"first": "Min",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Haizhou",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of ACL",
"volume": "",
"issue": "",
"pages": "1288--1297",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Deyi Xiong, Min Zhang, and Haizhou Li. 2011. Enhancing language models in statistical machine translation with backward n-grams and mutual in- formation triggers. In Proceedings of ACL, pages 1288-1297.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Phrase pair rescoring with term weightings for statistical machine translation",
"authors": [
{
"first": "Bing",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Stephan",
"middle": [],
"last": "Vogel",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Waibel",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bing Zhao, Stephan Vogel, and Alex Waibel. 2004. Phrase pair rescoring with term weightings for sta- tistical machine translation. In Proceedings of EMNLP.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"text": "Pre-training consists of learning a stack of RBMs, and these RBMs create an unsupervised DBN.",
"num": null,
"uris": null
},
"FIGREF1": {
"type_str": "figure",
"text": "Horizontal composition of DAEs to expand high-dimensional features learning.",
"num": null,
"uris": null
},
"TABREF0": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table><tr><td>For example, the architecture 16-32-16-2 (4 lay-</td></tr><tr><td>ers' network depth) corresponds to the DAE with</td></tr><tr><td>16-dimensional input features (X) (input layer),</td></tr><tr><td>32/16 hidden units (first/second hidden layer), and</td></tr><tr><td>2-dimensional output features (new DAE features)</td></tr><tr><td>(output layer). During the fine-tuning, the DAE's</td></tr><tr><td>network structure becomes 16</td></tr></table>",
"text": "Details of the used network structure."
},
"TABREF1": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table><tr><td>#</td><td/><td>Features</td><td>IWSLT Dev Test</td><td>NIST Dev Test</td></tr><tr><td>1</td><td/><td>Baseline</td><td>50.81 41.13</td><td>36.12 32.59</td></tr><tr><td>2 3 4 5 6 7 8 9</td><td>X 1</td><td>+DBN X 1 2f +DAE X 1 2f +DBN X 1 4f +DAE X 1 4f +HCDAE X 1 2+2f +DBN X 1 8f +DAE X 1 8f +HCDAE X 1 4+4f</td><td>51.92 42.07 * 52.49 43.22 * * 51.45 41.78 * 52.45 43.06 * * 53.69 43.23 * * * 51.74 41.85 * 52.33 42.98 * * 52.52 43.26 * * *</td><td>36.33 33.11 * 36.92 33.44 * * 36.45 33.12 * 36.88 33.47 * * 37.06 33.68 * * * 36.61 33.24 * 36.81 33.36 * * 37.01 33.63 * * *</td></tr><tr><td>10</td><td/><td>+DBN X 2f</td><td>52.21 42.24 *</td><td>36.72 33.21 *</td></tr><tr><td>11</td><td/><td>+DAE X 2f</td><td>52.86 43.45 * *</td><td>37.39 33.83 * *</td></tr><tr><td>12</td><td/><td>+DBN X 4f</td><td>51.83 42.08 *</td><td>34.45 33.07 *</td></tr><tr><td>13 14</td><td>X</td><td>+DAE X 4f +HCDAE X 2+2f</td><td>52.81 43.47 * * 53.05 43.58 * * *</td><td>37.48 33.92 * * 37.59 34.11 * * *</td></tr><tr><td>15</td><td/><td>+DBN X 8f</td><td>51.93 42.01 *</td><td>36.74 33.29 *</td></tr><tr><td>16</td><td/><td>+DAE X 8f</td><td>52.69 43.26 * *</td><td>37.36 33.75 * *</td></tr><tr><td>17</td><td/><td>+HCDAE X 4+4f</td><td>52.93 43.49 * * *</td><td>37.53 34.02 * * *</td></tr><tr><td>18</td><td/><td>+(X 2 +X 3 +X 4 +X 5 )</td><td>52.23 42.91 *</td><td>36.96 33.65 *</td></tr><tr><td>19 20 21 22</td><td/><td colspan=\"3\">53.55 44.17 + * * * 53.61 44.22 + * * * +(X 2 + * * * * 38.23 34.50 + * * * +(X 2 +X 3 +X 4 +X 5 )+DAE X 2f 38.28 34.47 + * * * +(X 2 +X 3 +X 4 +X 5 )+DAE X 4f +(X 2 +X 3 +X 4 +X 5 )+DAE X 8f 53.47 44.19 + * * * 38.26 34.46 + * * *</td></tr><tr><td>23</td><td/><td>+(X 2</td><td/><td/></tr><tr><td/><td/><td/><td colspan=\"2\">framework to implement our phrase-based ma-</td></tr><tr><td/><td/><td/><td colspan=\"2\">chine system. The 4-gram LMs are estimated</td></tr><tr><td/><td/><td/><td colspan=\"2\">by the SRILM toolkit with modified Kneser-Ney</td></tr></table>",
"text": "+X 3 +X 4 +X 5 )+HCDAE X 2+2f 53.75 44.28 + * * * * 38.35 34.65 +X 3 +X 4 +X 5 )+HCDAE X 4+4f 53.62 44.29 + * * * * 38.39 34.57 + * * * *"
},
"TABREF2": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table/>",
"text": ""
},
"TABREF3": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table/>",
"text": "0.1678 0.2873 0.2037 0.1622 0.0691 0.1813 0.0828 0.1637 DBN X 4f 0.2010 0.1590 0.2793 0.1692 0.1267 0.1146 0.2147 0.1051 DAE X 1 4f 0.5072 0.4486 0.1309 0.6012 0.2136 0.2168 0.2047 0.2526 DAE X 4f 0.5215 0.4594 0.2371 0.6903 0.2421 0.2694 0.3034 0.2642"
},
"TABREF4": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table><tr><td colspan=\"3\">Figure 4: The compared results of feature learning with different network structures on two development</td></tr><tr><td>sets.</td><td/><td/></tr><tr><td>Features</td><td>IWSLT Dev Test</td><td>NIST Dev Test</td></tr><tr><td>+DAE X1 4f</td><td>52.45 43.06</td><td>36.88 33.47</td></tr><tr><td>+DAE</td><td/><td/></tr></table>",
"text": "The variance distributions of each dimensional learned DBN feature and DAE feature on the two tasks. X1+X2 4f 52.76 43.38 * 37.28 33.80 * +DAE X1+X3 4f 52.61 43.27 * 37.13 33.66 * +DAE X1+X4 4f 52.52 43.24 * 36.96 33.58 * +DAE X1+X5 4f 52.49 43.13 * 36.96 33.56 * +DAE X 4f 52.81 43.47 * 37.48 33.92 *"
}
}
}
} |