File size: 134,056 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 | {
"paper_id": "Q17-1024",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T15:12:34.966480Z"
},
"title": "Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation",
"authors": [
{
"first": "Melvin",
"middle": [],
"last": "Johnson",
"suffix": "",
"affiliation": {},
"email": "melvinp@google.com"
},
{
"first": "Mike",
"middle": [],
"last": "Schuster",
"suffix": "",
"affiliation": {},
"email": "schuster@google.com"
},
{
"first": "Quoc",
"middle": [
"V"
],
"last": "Le",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Maxim",
"middle": [],
"last": "Krikun",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Yonghui",
"middle": [],
"last": "Wu",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Zhifeng",
"middle": [],
"last": "Chen",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Nikhil",
"middle": [],
"last": "Thorat",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Fernanda",
"middle": [],
"last": "Vi\u00e9gas",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Martin",
"middle": [],
"last": "Wattenberg",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Greg",
"middle": [],
"last": "Corrado",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Macduff",
"middle": [],
"last": "Hughes",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Dean",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We propose a simple solution to use a single Neural Machine Translation (NMT) model to translate between multiple languages. Our solution requires no changes to the model architecture from a standard NMT system but instead introduces an artificial token at the beginning of the input sentence to specify the required target language. Using a shared wordpiece vocabulary, our approach enables Multilingual NMT systems using a single model. On the WMT'14 benchmarks, a single multilingual model achieves comparable performance for English\u2192French and surpasses state-of-theart results for English\u2192German. Similarly, a single multilingual model surpasses stateof-the-art results for French\u2192English and German\u2192English on WMT'14 and WMT'15 benchmarks, respectively. On production corpora, multilingual models of up to twelve language pairs allow for better translation of many individual pairs. Our models can also learn to perform implicit bridging between language pairs never seen explicitly during training, showing that transfer learning and zeroshot translation is possible for neural translation. Finally, we show analyses that hints at a universal interlingua representation in our models and also show some interesting examples when mixing languages.",
"pdf_parse": {
"paper_id": "Q17-1024",
"_pdf_hash": "",
"abstract": [
{
"text": "We propose a simple solution to use a single Neural Machine Translation (NMT) model to translate between multiple languages. Our solution requires no changes to the model architecture from a standard NMT system but instead introduces an artificial token at the beginning of the input sentence to specify the required target language. Using a shared wordpiece vocabulary, our approach enables Multilingual NMT systems using a single model. On the WMT'14 benchmarks, a single multilingual model achieves comparable performance for English\u2192French and surpasses state-of-theart results for English\u2192German. Similarly, a single multilingual model surpasses stateof-the-art results for French\u2192English and German\u2192English on WMT'14 and WMT'15 benchmarks, respectively. On production corpora, multilingual models of up to twelve language pairs allow for better translation of many individual pairs. Our models can also learn to perform implicit bridging between language pairs never seen explicitly during training, showing that transfer learning and zeroshot translation is possible for neural translation. Finally, we show analyses that hints at a universal interlingua representation in our models and also show some interesting examples when mixing languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "End-to-end Neural Machine Translation (NMT) (Sutskever et al., 2014; Bahdanau et al., 2015; Cho et al., 2014) is an approach to machine * Corresponding authors. translation that has rapidly gained adoption in many large-scale settings (Zhou et al., 2016; Wu et al., 2016; Crego and et al., 2016) . Almost all such systems are built for a single language pair -so far there has not been a sufficiently simple and efficient way to handle multiple language pairs using a single model without making significant changes to the basic NMT architecture.",
"cite_spans": [
{
"start": 44,
"end": 68,
"text": "(Sutskever et al., 2014;",
"ref_id": "BIBREF26"
},
{
"start": 69,
"end": 91,
"text": "Bahdanau et al., 2015;",
"ref_id": "BIBREF1"
},
{
"start": 92,
"end": 109,
"text": "Cho et al., 2014)",
"ref_id": "BIBREF4"
},
{
"start": 235,
"end": 254,
"text": "(Zhou et al., 2016;",
"ref_id": "BIBREF30"
},
{
"start": 255,
"end": 271,
"text": "Wu et al., 2016;",
"ref_id": "BIBREF28"
},
{
"start": 272,
"end": 295,
"text": "Crego and et al., 2016)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper we introduce a simple method to translate between multiple languages using a single model, taking advantage of multilingual data to improve NMT for all languages involved. Our method requires no change to the traditional NMT model architecture. Instead, we add an artificial token to the input sequence to indicate the required target language, a simple amendment to the data only. All other parts of the system -encoder, decoder, attention, and shared wordpiece vocabulary as described in Wu et al., (2016) -stay exactly the same. This method has several attractive benefits:",
"cite_spans": [
{
"start": 504,
"end": 521,
"text": "Wu et al., (2016)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 Simplicity: Since no changes are made to the architecture of the model, scaling to more languages is trivial -any new data is simply added, possibly with over-or under-sampling such that all languages are appropriately represented, and used with a new token if the target language changes. Since no changes are made to the training procedure, the mini-batches for training are just sampled from the overall mixed-language training data just like for the single-language case. Since no a-priori decisions about how to allocate parameters for different languages are made, the system adapts automatically to use the total number of param-eters efficiently to minimize the global loss. A multilingual model architecture of this type also simplifies production deployment significantly since it can cut down the total number of models necessary when dealing with multiple languages. Note that at Google, we support a total of over 100 languages as source and target, so theoretically 100 2 models would be necessary for the best possible translations between all pairs, if each model could only support a single language pair. Clearly this would be problematic in a production environment. Even when limiting to translating to/from English only, we still need over 200 models. Finally, batching together many requests from potentially different source and target languages can significantly improve efficiency of the serving system. In comparison, an alternative system that requires language-dependent encoders, decoders or attention modules does not have any of the above advantages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 Low-resource language improvements: In a multilingual NMT model, all parameters are implicitly shared by all the language pairs being modeled. This forces the model to generalize across language boundaries during training. It is observed that when language pairs with little available data and language pairs with abundant data are mixed into a single model, translation quality on the low resource language pair is significantly improved.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 Zero-shot translation: A surprising benefit of modeling several language pairs in a single model is that the model can learn to translate between language pairs it has never seen in this combination during training (zero-shot translation) -a working example of transfer learning within neural translation models. For example, a multilingual NMT model trained with Portuguese\u2192English and English\u2192Spanish examples can generate reasonable translations for Portuguese\u2192Spanish although it has not seen any data for that language pair. We show that the quality of zero-shot language pairs can easily be improved with little additional data of the language pair in question (a fact that has been pre-viously confirmed for a related approach which is discussed in more detail in the next section).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In the remaining sections of this paper we first discuss related work and explain our multilingual system architecture in more detail. Then, we go through the different ways of merging languages on the source and target side in increasing difficulty (many-to-one, one-to-many, many-to-many), and discuss the results of a number of experiments on WMT benchmarks, as well as on some of Google's large-scale production datasets. We present results from transfer learning experiments and show how implicitly-learned bridging (zero-shot translation) performs in comparison to explicit bridging (i.e., first translating to a common language like English and then translating from that common language into the desired target language) as typically used in machine translation systems. We describe visualizations of the new system in action, which provide early evidence of shared semantic representations (interlingua) between languages. Finally we also show some interesting applications of mixing languages with examples: code-switching on the source side and weighted target language mixing, and suggest possible avenues for further exploration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Interlingual translation is a classic method in machine translation (Richens, 1958; Hutchins and Somers, 1992) . Despite its distinguished history, most practical applications of machine translation have focused on individual language pairs because it was simply too difficult to build a single system that translates reliably from and to several languages.",
"cite_spans": [
{
"start": 68,
"end": 83,
"text": "(Richens, 1958;",
"ref_id": "BIBREF20"
},
{
"start": 84,
"end": 110,
"text": "Hutchins and Somers, 1992)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Neural Machine Translation (NMT) (Kalchbrenner and Blunsom, 2013) was shown to be a promising end-to-end learning approach in Sutskever et al. (2014) ; Bahdanau et al. (2015) ; Cho et al. (2014) and was quickly extended to multilingual machine translation in various ways.",
"cite_spans": [
{
"start": 126,
"end": 149,
"text": "Sutskever et al. (2014)",
"ref_id": "BIBREF26"
},
{
"start": 152,
"end": 174,
"text": "Bahdanau et al. (2015)",
"ref_id": "BIBREF1"
},
{
"start": 177,
"end": 194,
"text": "Cho et al. (2014)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "An early attempt is the work in Dong et al., (2015) , where the authors modify an attention-based encoderdecoder approach to perform multilingual NMT by adding a separate decoder and attention mechanism for each target language. In Luong et al., (2015a) multilingual training in a multitask learning setting is described. This model is also an encoder-decoder network, in this case without an attention mechanism. To make proper use of multilingual data, they extend their model with multiple encoders and decoders, one for each supported source and target language. In Caglayan et al., (2016) the authors incorporate multiple modalities other than text into the encoderdecoder framework.",
"cite_spans": [
{
"start": 32,
"end": 51,
"text": "Dong et al., (2015)",
"ref_id": "BIBREF6"
},
{
"start": 232,
"end": 253,
"text": "Luong et al., (2015a)",
"ref_id": "BIBREF16"
},
{
"start": 570,
"end": 593,
"text": "Caglayan et al., (2016)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Several other approaches have been proposed for multilingual training, especially for low-resource language pairs. For instance, in Zoph and Knight (2016) a form of multi-source translation was proposed where the model has multiple different encoders and different attention mechanisms for each source language. However, this work requires the presence of a multi-way parallel corpus between all the languages involved, which is difficult to obtain in practice. Most closely related to our approach is Firat et al., (2016a) in which the authors propose multi-way multilingual NMT using a single shared attention mechanism but multiple encoders/decoders for each source/target language. Recently in Lee et al., (2016) a CNNbased character-level encoder was proposed which is shared across multiple source languages. However, this approach can only perform translations into a single target language.",
"cite_spans": [
{
"start": 132,
"end": 154,
"text": "Zoph and Knight (2016)",
"ref_id": "BIBREF31"
},
{
"start": 502,
"end": 523,
"text": "Firat et al., (2016a)",
"ref_id": "BIBREF7"
},
{
"start": 698,
"end": 716,
"text": "Lee et al., (2016)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Our approach is related to the multitask learning framework (Caruana, 1998) . Despite its promise, this framework has seen limited practical success in real world applications. In speech recognition, there have been many successful reports of modeling multiple languages using a single model ( (Schultz and Kirchhoff, 2006) for an extensive reference and references therein). Multilingual language processing has also shown to be successful in domains other than translation (Gillick et al., 2016; Tsvetkov et al., 2016) .",
"cite_spans": [
{
"start": 60,
"end": 75,
"text": "(Caruana, 1998)",
"ref_id": "BIBREF3"
},
{
"start": 294,
"end": 323,
"text": "(Schultz and Kirchhoff, 2006)",
"ref_id": "BIBREF21"
},
{
"start": 475,
"end": 497,
"text": "(Gillick et al., 2016;",
"ref_id": "BIBREF12"
},
{
"start": 498,
"end": 520,
"text": "Tsvetkov et al., 2016)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "There have been other approaches similar to ours in spirit, but used for very different purposes. In Sennrich et al.,(2016a), the NMT framework has been extended to control the politeness level of the target translation by adding a special token to the source sentence. The same idea was used in Yamagishi et al., (2016) to add the distinction between 'active' and 'passive' tense to the generated target sentence.",
"cite_spans": [
{
"start": 296,
"end": 320,
"text": "Yamagishi et al., (2016)",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Our method has an additional benefit not seen in other systems: it gives the system the ability to perform zero-shot translation, meaning the system can translate from a source language to a target language without having seen explicit examples from this spe-cific language pair during training. Zero-shot translation was the direct goal of Firat et al., (2016c) . Although they were not able to achieve this direct goal, they were able to do what they call \"zeroresource\" translation by using their pre-trained multiway multilingual model and later fine-tuning it with pseudo-parallel data generated by the model. It should be noted that the difference between \"zeroshot\" and \"zero-resource\" translation is the additional fine-tuning step which is required in the latter approach.",
"cite_spans": [
{
"start": 341,
"end": 362,
"text": "Firat et al., (2016c)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "To the best of our knowledge, our work is the first to validate the use of true multilingual translation using a single encoder-decoder model, and is incidentally also already used in a production setting. It is also the first work to demonstrate the possibility of zero-shot translation, a successful example of transfer learning in machine translation, without any additional steps.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "The multilingual model architecture is identical to Google's Neural Machine Translation (GNMT) system (Wu et al., 2016 ) (with the optional addition of direct connections between encoder and decoder layers which we have used for some of our experiments).",
"cite_spans": [
{
"start": 102,
"end": 118,
"text": "(Wu et al., 2016",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "To be able to make use of multilingual data within a single system, we propose one simple modification to the input data, which is to introduce an artificial token at the beginning of the input sentence to indicate the target language the model should translate to. For instance, consider the following En\u2192Es pair of sentences:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "How are you? -> \u00bfC\u00f3mo est\u00e1s?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "It will be modified to:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "<2es> How are you? -> \u00bfC\u00f3mo est\u00e1s?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "to indicate that Spanish is the target language. Note that we don't specify the source language -the model will learn this automatically.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "After adding the token to the input data, we train the model with all multilingual data consisting of multiple language pairs at once, possibly after overor undersampling some of the data to adjust for the relative ratio of the language data available. To address the issue of translation of unknown words and to limit the vocabulary for computational efficiency, we use a shared wordpiece model (Schuster and Nakajima, 2012) across all the source and target data used for training, usually with 32,000 word pieces. The segmentation algorithm used here is very similar (with small differences) to Byte-Pair-Encoding (BPE) which was described in Gage (1994) and was also used in Sennrich et al., (2016b) for machine translation. All training is carried out similar to (Wu et al., 2016) and implemented in TensorFlow (Abadi and et al., 2016) .",
"cite_spans": [
{
"start": 396,
"end": 425,
"text": "(Schuster and Nakajima, 2012)",
"ref_id": "BIBREF22"
},
{
"start": 645,
"end": 656,
"text": "Gage (1994)",
"ref_id": "BIBREF11"
},
{
"start": 678,
"end": 702,
"text": "Sennrich et al., (2016b)",
"ref_id": "BIBREF25"
},
{
"start": 767,
"end": 784,
"text": "(Wu et al., 2016)",
"ref_id": "BIBREF28"
},
{
"start": 815,
"end": 839,
"text": "(Abadi and et al., 2016)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "In summary, this approach is the simplest among the alternatives that we are aware of. During training and inference, we only need to add one additional token to each sentence of the source data to specify the desired target language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "3"
},
{
"text": "In this section, we apply our proposed method to train multilingual models in several different configurations. Since we can have models with either single or multiple source/target languages we test three interesting cases for mapping languages: 1) many to one, 2) one to many, and 3) many to many. As already discussed in Section 2, other models have been used to explore some of these cases already, but for completeness we apply our technique to these interesting use cases again to give a full picture of the effectiveness of our approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "4"
},
{
"text": "We will also show results and discuss benefits of bringing together many (un)related languages in a single large-scale model trained on production data. Finally, we will present our findings on zero-shot translation where the model learns to translate between pairs of languages for which no explicit parallel examples existed in the training data, and show results of experiments where adding additional data improves zero-shot translation quality further.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "4"
},
{
"text": "For WMT, we train our models on the WMT'14 En\u2192Fr and the WMT'14 En\u2192De datasets. In both cases, we use newstest2014 as the test sets to be able to compare against previous work (Luong et al., 2015c; S\u00e9bastien et al., 2015; Zhou et al., 2016; Wu et al., 2016) . For WMT Fr\u2192En and De\u2192En we use newstest2014 and newstest2015 as test sets. De-spite training on WMT'14 data, which is somewhat smaller than WMT'15, we test our De\u2192En model on newstest2015, similar to Luong et al., (2015b) . The combination of newstest2012 and newstest2013 is used as the development set. In addition to WMT, we also evaluate the multilingual approach on some Google-internal large-scale production datasets representing a wide spectrum of languages with very distinct linguistic properties: En\u2194Japanese(Ja), En\u2194Korean(Ko), En\u2194Es, and En\u2194Pt. These datasets are two to three orders of magnitude larger than the WMT datasets.",
"cite_spans": [
{
"start": 176,
"end": 197,
"text": "(Luong et al., 2015c;",
"ref_id": "BIBREF18"
},
{
"start": 198,
"end": 221,
"text": "S\u00e9bastien et al., 2015;",
"ref_id": "BIBREF23"
},
{
"start": 222,
"end": 240,
"text": "Zhou et al., 2016;",
"ref_id": "BIBREF30"
},
{
"start": 241,
"end": 257,
"text": "Wu et al., 2016)",
"ref_id": "BIBREF28"
},
{
"start": 460,
"end": 481,
"text": "Luong et al., (2015b)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets, Training Protocols and Evaluation Metrics",
"sec_num": "4.1"
},
{
"text": "Our training protocols are mostly identical to those described in Wu et al., (2016) . We find that some multilingual models take a little more time to train than single language pair models, likely because each language pair is seen only for a fraction of the training process. We use larger batch sizes with a slightly higher initial learning rate to speed up the convergence of these models.",
"cite_spans": [
{
"start": 66,
"end": 83,
"text": "Wu et al., (2016)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets, Training Protocols and Evaluation Metrics",
"sec_num": "4.1"
},
{
"text": "We evaluate our models using the standard BLEU score metric and to make our results comparable to previous work (Sutskever et al., 2014; Luong et al., 2015c; Zhou et al., 2016; Wu et al., 2016) , we report tokenized BLEU score as computed by the multi-bleu.pl script, which can be downloaded from the public implementation of Moses. 1 To test the influence of varying amounts of training data per language pair we explore two strategies when building multilingual models: a) where we oversample the data from all language pairs to be of the same size as the largest language pair, and b) where we mix the data as is without any change. The wordpiece model training is done after the optional oversampling taking into account all the changed data ratios. For the WMT models we report results using both of these strategies. For the production models, we always balance the data such that the ratios are equal.",
"cite_spans": [
{
"start": 112,
"end": 136,
"text": "(Sutskever et al., 2014;",
"ref_id": "BIBREF26"
},
{
"start": 137,
"end": 157,
"text": "Luong et al., 2015c;",
"ref_id": "BIBREF18"
},
{
"start": 158,
"end": 176,
"text": "Zhou et al., 2016;",
"ref_id": "BIBREF30"
},
{
"start": 177,
"end": 193,
"text": "Wu et al., 2016)",
"ref_id": "BIBREF28"
},
{
"start": 333,
"end": 334,
"text": "1",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets, Training Protocols and Evaluation Metrics",
"sec_num": "4.1"
},
{
"text": "One benefit of the way we share all the components of the model is that the mini-batches can contain data from different language pairs during training and inference, which are typically just random samples from the final training and test data distributions. This is a simple way of preventing \"catastrophic forgetting\" -tendency for knowledge of previously learned task(s) (e.g. language pair A) to be abruptly forgotten as information relevant to the current task (e.g. language pair B) is incorporated (French, 1999) . Other approaches to multilingual translation require complex update scheduling mechanisms to prevent this effect (Firat et al., 2016b) .",
"cite_spans": [
{
"start": 506,
"end": 520,
"text": "(French, 1999)",
"ref_id": "BIBREF10"
},
{
"start": 636,
"end": 657,
"text": "(Firat et al., 2016b)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets, Training Protocols and Evaluation Metrics",
"sec_num": "4.1"
},
{
"text": "In this section we explore having multiple source languages and a single target language -the simplest way of combining language pairs. Since there is only a single target language no additional source token is required. We perform three sets of experiments:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Many to One",
"sec_num": "4.2"
},
{
"text": "\u2022 The first set of experiments is on the WMT datasets, where De\u2192En and Fr\u2192En are combined to train a multilingual model. Our baselines are two single language pair models: De\u2192En and Fr\u2192En trained independently. We perform these experiments once with oversampling and once without.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Many to One",
"sec_num": "4.2"
},
{
"text": "\u2022 The second set of experiments is on production data where we combine Ja\u2192En and Ko\u2192En, with oversampling. The baselines are two single language pair models trained independently.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Many to One",
"sec_num": "4.2"
},
{
"text": "\u2022 Finally, the third set of experiments is on production data where we combine Es\u2192En and Pt\u2192En, with oversampling. The baselines are again two single language pair models trained independently.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Many to One",
"sec_num": "4.2"
},
{
"text": "All of the multilingual and single language pair models have the same total number of parameters as the baseline NMT models trained on a single language pair (using 1024 nodes, 8 LSTM layers and a shared wordpiece model vocabulary of 32k, a total of 255M parameters per model). A side effect of this equal choice of parameters is that it is presumably unfair to the multilingual models as the number of parameters available per language pair is reduced by a factor of N compared to the single language pair models, if N is the number of language pairs combined in the multilingual model. The multilingual model also has to handle the combined vocabulary of all the single models. We chose to keep the number of parameters constant for all models to simplify experimentation. We relax this constraint for some of the large-scale experiments shown further below. The results are presented in Table 1 . For all experiments the multilingual models outperform the baseline single systems despite the above mentioned disadvantage with respect to the number of parameters available per language pair. One possible hypothesis explaining the gains is that the model has been shown more English data on the target side, and that the source languages belong to the same language families, so the model has learned useful generalizations.",
"cite_spans": [],
"ref_spans": [
{
"start": 890,
"end": 897,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Many to One",
"sec_num": "4.2"
},
{
"text": "For the WMT experiments, we obtain a maximum gain of +1.27 BLEU for Fr\u2192En. Note that the results on both the WMT test sets are better than other published state-of-the-art results for a single model, to the best of our knowledge.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Many to One",
"sec_num": "4.2"
},
{
"text": "In this section, we explore the application of our method when there is a single source language and multiple target languages. Here we need to prepend the input with an additional token to specify the target language. We perform three sets of experiments very similar to the previous section. Table 2 summarizes the results when performing translations into multiple target languages. We see that the multilingual models are comparable to, and in some cases outperform, the baselines, but not always. We obtain a large gain of +0.9 BLEU for En\u2192Es. Unlike the previous set of results, there are less significant gains in this setting. This is perhaps due to the fact that the decoder has a more difficult time translating into multiple target languages which may even have different scripts, which are combined into a single shared wordpiece vocabulary. Note that even for languages with entirely different scripts (e.g., Korean and Japanese) there is significant overlap in wordpieces when real data is used, as often numbers, dates, names, websites, punctuation etc. are actually using a shared script (ASCII). We observe that oversampling helps the smaller language pair (En\u2192De) at the cost of lower quality for the larger language pair (En\u2192Fr). The model without oversampling achieves better results on the larger language compared to the smaller one as expected. We also find that this effect is more prominent on smaller datasets (WMT) and much less so on our much larger production datasets.",
"cite_spans": [],
"ref_spans": [
{
"start": 294,
"end": 301,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "One to Many",
"sec_num": "4.3"
},
{
"text": "In this section, we report on experiments when there are multiple source languages and multiple target languages within a single model -the most difficult setup. Since multiple target languages are given, the input needs to be prepended with the target language token as above.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Many to Many",
"sec_num": "4.4"
},
{
"text": "The results are presented in Table 3 . We see that the multilingual production models with the same model size and vocabulary size as the single language models are quite close to the baselines -the average relative loss in BLEU score across all experiments is only approximately 2.5%.",
"cite_spans": [],
"ref_spans": [
{
"start": 29,
"end": 36,
"text": "Table 3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Many to Many",
"sec_num": "4.4"
},
{
"text": "Although there are some significant losses in quality from training many languages jointly using a model with the same total number of parameters as the single language pair models, these models reduce the total complexity involved in training and productionization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Many to Many",
"sec_num": "4.4"
},
{
"text": "This section shows the result of combining 12 production language pairs having a total of 3B parameters (255M per single model) into a single multilingual model. A range of multilingual models were trained, starting from the same size as a single language pair model with 255M parameters (1024 nodes) up to 650M parameters (1792 nodes). As above, the input needs to be prepended with the target language token. We oversample the examples from the smaller language pairs to balance the data as explained above.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Large-scale Experiments",
"sec_num": "4.5"
},
{
"text": "The results for single language pair models versus multilingual models with increasing numbers of parameters are summarized in Table 4 . We find that the multilingual models are on average worse than the single models (about 5.6% to 2.5% relative depending on size, however, some actually get better) and as expected the average difference gets smaller when going to larger multilingual models. It should be noted that the largest multilingual model we have trained has still about five times less parameters than the combined single models.",
"cite_spans": [],
"ref_spans": [
{
"start": 127,
"end": 134,
"text": "Table 4",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Large-scale Experiments",
"sec_num": "4.5"
},
{
"text": "The multilingual model also requires only roughly 1/12-th of the training time (or computing resources) to converge compared to the combined single models (total training time for all our models is still in the order of weeks). Another important point is that since we only train for a little longer than a standard single model, the individual language pairs can see as little as 1/12-th of the data in comparison to their single language pair models but still produce satisfactory results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Large-scale Experiments",
"sec_num": "4.5"
},
{
"text": "In summary, multilingual NMT enables us to group languages with little loss in quality while having the benefits of better training efficiency, smaller number of models, and easier productionization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Large-scale Experiments",
"sec_num": "4.5"
},
{
"text": "The most straight-forward approach of translating between languages where no or little parallel data is available is to use explicit bridging, meaning to translate to an intermediate language first and then to translate to the desired target language. The intermediate language is often English as xx\u2192En and En\u2192yy data is more readily available. The two potential disadvantages of this approach are: a) total translation time doubles, b) the potential loss of quality by translating to/from the intermediate language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Zero-Shot Translation",
"sec_num": "4.6"
},
{
"text": "An interesting benefit of our approach is that it allows to perform directly implicit bridging (zero-shot translation) between a language pair for which no explicit parallel training data has been seen without any modification to the model. Obviously, the model will only be able to do zero-shot translation between languages it has seen individually as source and target languages during training at some point, not for entirely new ones.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Zero-Shot Translation",
"sec_num": "4.6"
},
{
"text": "To demonstrate this we will use two multilingual models -a model trained with examples from two different language pairs, Pt\u2192En and En\u2192Es (Model 1), and a model trained with examples from four different language pairs, En\u2194Pt and En\u2194Es (Model 2). As with the previous multilingual models, both of these models perform comparable to or even slightly better than the baseline single models for the language pairs explicitly seen. Additionally, we show that both of these models can generate reasonable quality Pt\u2192Es translations (BLEU scores above 20) without ever having seen Pt\u2192Es data during training. To our knowledge this is the first successful demonstration of true multilingual zero-shot translation. Table 5 summarizes our results for the Pt\u2192Es translation experiments. Rows (a) and (b) show the performance of the phrase-based machine translation (PBMT) system and the NMT system through explicit bridging (Pt\u2192En, then En\u2192Es). It can be seen that the NMT system outperforms the PBMT system by close to 2 BLEU points. For comparison, we also built a single NMT model on all available Pt\u2192Es parallel sentences (see (c) in Table 5 ). The most interesting observation is that both Model 1 and Model 2 can perform zero-shot translation with reasonable quality (see (d) and (e)) compared to the initial expectation that this would not work at all. Note that Model 2 outperforms Model 1 by close to 3 BLEU points although Model 2 was trained with four language pairs as opposed to with only two for Model 1 (with both models having the same number of total parameters). In this case the addition of Spanish on the source side and Portuguese on the target side helps Pt\u2192Es zero-shot translation (which is the opposite direction of where we would expect it to help). We believe that this unexpected effect is only possible because our shared architecture enables the model to learn a form of interlingua between all these languages. We explore this hypothesis in more detail in Section 5.",
"cite_spans": [
{
"start": 1267,
"end": 1270,
"text": "(d)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 706,
"end": 713,
"text": "Table 5",
"ref_id": "TABREF4"
},
{
"start": 1127,
"end": 1134,
"text": "Table 5",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Zero-Shot Translation",
"sec_num": "4.6"
},
{
"text": "Finally we incrementally train zero-shot Model 2 with a small amount of true Pt\u2192Es parallel data (an order of magnitude less than Table 5 (c)) and obtain the best quality and half the decoding time compared to explicit bridging (Table 5 (b) ). The resulting model cannot be called zero-shot anymore since some true parallel data has been used to improve it. Overall this shows that the proposed approach of implicit bridging using zero-shot translation via multilingual models can serve as a good baseline for further incremental training with relatively small amounts of true parallel data of the zero-shot direction. This result is especially significant for non-English lowresource language pairs where it might be easier to obtain parallel data with English but much harder to obtain parallel data for language pairs where neither the source nor the target language is English. We explore the effect of using parallel data in more detail in Section 4.7.",
"cite_spans": [
{
"start": 228,
"end": 240,
"text": "(Table 5 (b)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Zero-Shot Translation",
"sec_num": "4.6"
},
{
"text": "Since Portuguese and Spanish are of the same language family, an interesting question is how well zero-shot translation works for less related languages. Table 6 shows the results for explicit and implicit bridging from Spanish to Japanese using the largescale model from Table 4 -Spanish and Japanese can be regarded as quite unrelated. As expected zeroshot translation works worse than explicit bridging and the quality drops relatively more (roughly 50% drop in BLEU score) than for the case of more related languages as shown above. Despite the quality drop, this proves that our approach enables zero-shot translation even between unrelated languages. ",
"cite_spans": [],
"ref_spans": [
{
"start": 154,
"end": 161,
"text": "Table 6",
"ref_id": "TABREF5"
},
{
"start": 272,
"end": 279,
"text": "Table 4",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Zero-Shot Translation",
"sec_num": "4.6"
},
{
"text": "In this section, we explore two ways of leveraging available parallel data to improve zero-shot translation quality, similar in spirit to what was reported in Firat et al., (2016c) . For our multilingual architecture we consider:",
"cite_spans": [
{
"start": 159,
"end": 180,
"text": "Firat et al., (2016c)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of Direct Parallel Data",
"sec_num": "4.7"
},
{
"text": "\u2022 Incrementally training the multilingual model on the additional parallel data for the zero-shot directions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of Direct Parallel Data",
"sec_num": "4.7"
},
{
"text": "\u2022 Training a new multilingual model with all available parallel data mixed equally.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of Direct Parallel Data",
"sec_num": "4.7"
},
{
"text": "For our experiments, we use a baseline model which we call \"Zero-Shot\" trained on a combined parallel corpus of English\u2194{Belarusian(Be), Russian(Ru), Ukrainian(Uk)}. We trained a second model on the above corpus together with additional Ru\u2194{Be, Uk} data. We call this model \"From-Scratch\". Both models support four target languages, and are evaluated on our standard test sets. As done previously we oversample the data such that all language pairs are represented equally. Finally, we take the best checkpoint of the \"Zero-Shot\" model, and run incremental training on a small portion of the data used to train the \"From-Scratch\" model for a short period of time until convergence (in this case 3% of \"Zero-Shot\" model total training time). We call this model \"Incremental\". As can be seen from Table 7 , for the English\u2194X directions, all three models show comparable scores. On the Ru\u2194{Be, Uk} directions, the \"Zero-Shot\" model already achieves relatively high BLEU scores for all directions except one, without any explicit parallel data. This could be because these languages are linguistically related. In the \"From-Scratch\" column, we see that training a new model from scratch improves the zero-shot translation directions further. However, this strategy has a slightly negative effect on the En\u2194X directions because our oversampling strategy will reduce the frequency of the data from these directions. In the final column, we see that incremental training with direct parallel data recovers most of the BLEU score difference between the first two columns on the zero-shot language pairs. In summary, our shared architecture models the zero-shot language pairs quite well and hence enables us to easily improve their quality with a small amount of additional parallel data.",
"cite_spans": [],
"ref_spans": [
{
"start": 795,
"end": 802,
"text": "Table 7",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Effect of Direct Parallel Data",
"sec_num": "4.7"
},
{
"text": "The results of this paper -that training a model across multiple languages can enhance performance at the individual language level, and that zero-shot translation can be effective -raise a number of questions about how these tasks are handled inside the model. Is the network learning some sort of shared representation, in which sentences with the same meaning are represented in similar ways regardless of language? Does the model operate on zero-shot translations in the same way as it treats language pairs it has been trained on? One way to study the representations used by the network is to look at the activations of the network during translation. A starting point for investigation is the set of context vectors, i.e., the sum of internal encoder states weighted by their attention probabilities per step (Eq. 5in (Bahdanau et al., 2015)) .",
"cite_spans": [
{
"start": 825,
"end": 849,
"text": "(Bahdanau et al., 2015))",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Visual Analysis",
"sec_num": "5"
},
{
"text": "A translation of a single sentence generates a sequence of context vectors. In this context, our original questions about shared representation can be studied by looking at how the vector sequences of different sentences relate. We could then ask for example: Do sentences cluster together depending on the source or target language? Or instead do sentences with similar meanings cluster, regardless of language? We try to find answers to these questions by looking at lower-dimensional representations of internal embeddings of the network that humans can more easily interpret.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Visual Analysis",
"sec_num": "5"
},
{
"text": "Several trained networks indeed show strong visual evidence of a shared representation. For example, Figure 1 below was produced from a manyto-many model trained on four language pairs, English\u2194Japanese and English\u2194Korean. To visualize the model in action we began with a small corpus of 74 triples of semantically identical cross-language phrases. That is, each triple contained phrases in English, Japanese and Korean with the same underlying meaning. To compile these triples, we searched a ground-truth database for English sentences which were paired with both Japanese and Korean translations.",
"cite_spans": [],
"ref_spans": [
{
"start": 101,
"end": 109,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Evidence for an Interlingua",
"sec_num": "5.1"
},
{
"text": "We then applied the trained model to translate each sentence of each triple into the two other possible languages. Performing this process yielded six new sentences based on each triple, for a total of 74 * 6 = 444 total translations with 9,978 steps corresponding to the same number of context vectors. Since context vectors are high-dimensional, we use the TensorFlow Embedding Projector 2 to map them into more accessible 3D space via t-SNE (Maaten and Hinton, 2008) . In the following diagrams, each point represents a single decoding step during the translation process. Points that represent steps for a given sentence are connected by line segments. Figure 1 shows a global view of all 9,978 context vectors. Points produced from the same original sentence triple are all given the same (random) color. Inspection of these clusters shows that each strand represents a single sentence, and clusters of strands generally represent a set of translations of the same underlying sentence, but with different source and target languages.",
"cite_spans": [
{
"start": 444,
"end": 469,
"text": "(Maaten and Hinton, 2008)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [
{
"start": 657,
"end": 665,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Evidence for an Interlingua",
"sec_num": "5.1"
},
{
"text": "At right are two close-ups: one of an individual cluster, still coloring based on membership in the same triple, and one where we have colored by source language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evidence for an Interlingua",
"sec_num": "5.1"
},
{
"text": "Not all models show such clean semantic clustering. Sometimes we observed joint embeddings in some regions of space coexisting with separate large clusters which contained many context vectors from just one language pair.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Partially Separated Representations",
"sec_num": "5.2"
},
{
"text": "For example, Figure 2a shows a t-SNE projection of context vectors from a model that was trained on Portuguese\u2192English (blue) and English\u2192Spanish (yellow) and performing zero-shot translation from Portuguese\u2192Spanish (red). This projection shows 153 semantically identical triples translated as described above, yielding 459 total translations. The large red region on the left primarily contains zeroshot Portuguese\u2192Spanish translations. In other words, for a significant number of sentences, the zero-shot translation has a different embedding than the two trained translation directions. On the other hand, some zero-shot translation vectors do seem to fall near the embeddings found in other languages, as on the large region on the right.",
"cite_spans": [],
"ref_spans": [
{
"start": 13,
"end": 22,
"text": "Figure 2a",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Partially Separated Representations",
"sec_num": "5.2"
},
{
"text": "It is natural to ask whether the large cluster of \"separated\" zero-shot translations has any significance. A definitive answer requires further investigation, but in this case zero-shot translations in the separated area do tend to have lower BLEU scores. Figure 2b shows a plot of BLEU scores of a zeroshot translation versus the average pointwise distance between it and the same translation from a trained language pair. An interesting area for future research is to find a more reliable correspondence between embedding geometry and model performance to predict the quality of a zero-shot translation during decoding by comparing it to the embedding of the translation through a trained language pair.",
"cite_spans": [],
"ref_spans": [
{
"start": 256,
"end": 265,
"text": "Figure 2b",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Partially Separated Representations",
"sec_num": "5.2"
},
{
"text": "Having a mechanism to translate from a random source language to a single chosen target language using an additional source token made us think about what happens when languages are mixed on the source or target side. In particular, we were interested in the following two experiments: 1) Can a multilingual model successfully handle multi-language input (code-switching) in the middle of a sentence?; 2) What happens when a multilingual model is triggered with a linear mix of two target language tokens?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mixing Languages",
"sec_num": "6"
},
{
"text": "Here we show how multilingual models deal with source language code-switching -an example from a multilingual {Ja,Ko}\u2192En model is below. Mixing Japanese and Korean in the source produces in many cases correct English translations, showing that codeswitching can be handled by this model, although no such code-switching samples were present in the training data. Note that the model can effectively handle the different typographic scripts since the individual characters/wordpieces are present in the shared vocabulary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Source Language Code-Switching",
"sec_num": "6.1"
},
{
"text": "\u2022 Japanese: \u79c1\u306f\u6771\u4eac\u5927\u5b66\u306e\u5b66\u751f\u3067\u3059\u3002 \u2192 I am a student at Tokyo University. \u2022 Korean: \u1102 \u1161\u1102 \u1173 \u11ab \u1103 \u1169\u110f \u116d \u1103 \u1162\u1112 \u1161 \u11a8\u110b \u1174 \u1112 \u1161 \u11a8\u1109 \u1162 \u11bc\u110b \u1175 \u11b8\u1102 \u1175\u1103 \u1161. \u2192 I am a student at Tokyo University. \u2022 Japanese/Korean: \u79c1\u306f\u6771\u4eac\u5927\u5b66\u1112 \u1161 \u11a8\u1109 \u1162 \u11bc\u110b \u1175 \u11b8\u1102 \u1175 \u1103 \u1161. \u2192 I am a student of Tokyo University.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Source Language Code-Switching",
"sec_num": "6.1"
},
{
"text": "Interestingly, the mixed-language translation is slightly different from both single source language translations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Source Language Code-Switching",
"sec_num": "6.1"
},
{
"text": "Here we test what happens when we mix target languages. Using a multilingual En\u2192{Ja, Ko} model, we feed a linear combination (1\u2212w)<2ja>+w<2ko> of the embedding vectors for \"<2ja>\" and \"<2ko>\". Clearly, for w = 0 the model should produce Japanese, for w = 1 it should produce Korean, but what happens in between?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Weighted Target Language Selection",
"sec_num": "6.2"
},
{
"text": "The model may produce some sort of intermediate language (\"Japarean\"), but the results turn out to be less surprising. Most of the time the output just switches from one language to another around w = 0.5. In some cases, for intermediate values of w, the model switches languages mid-sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Weighted Target Language Selection",
"sec_num": "6.2"
},
{
"text": "A possible explanation for this behavior is that the target language model, implicitly learned by the decoder LSTM, may make it very hard to mix words from different languages, especially when they use different scripts. Table 8 shows an example of mixed target languages (Ja/Ko), where we can observe an interesting transition in the script and grammar. At w ko = 0.58, the model translates the source sentence into a mix of Japanese and Korean. At w ko = 0.60, the sentence is translated into full Korean, where all of the source words are captured, however, the ordering of the words is not natural. When w ko is increased to 0.7, the model starts to translate the source sentence into a Korean sentence that sounds more natural. 3",
"cite_spans": [],
"ref_spans": [
{
"start": 221,
"end": 228,
"text": "Table 8",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Weighted Target Language Selection",
"sec_num": "6.2"
},
{
"text": "We present a simple solution to multilingual NMT. We show that we can train multilingual NMT models that can be used to translate between a number of different languages using a single model where all parameters are shared, which as a positive sideeffect also improves the translation quality of lowresource languages in the mix. We also show that zero-shot translation without explicit bridging is possible, which is the first time to our knowledge that a form of true transfer learning has been shown to work for machine translation. To explicitly improve the zero-shot translation quality, we explore two ways of adding available parallel data and find that small additional amounts are sufficient to reach satisfactory results. In our largest experiment we merge 12 language pairs into a single model and achieve only slightly lower translation quality as for the single language pair baselines despite the drastically reduced amount of modeling capacity per language in the multilingual model. Visual interpretation of the results shows that these models learn a form of interlingua representation between all involved language pairs. The simple architecture makes it possible to mix languages on the source or target side to yield some interesting translation examples. Our approach has been shown to work reliably in a Google-scale production setting and enables us to scale to a large number of languages quickly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "http://www.statmt.org/moses/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://www.tensorflow.org/get_started/ embedding_viz",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The Korean translation does not contain spaces and uses '\u3002' as punctuation symbol, and these are all artifacts of applying a Japanese postprocessor.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We would like to thank the entire Google Brain Team and Google Translate Team for their foundational contributions to this project. In particular, we thank Junyoung Chung for his insights on the topic and Alex Rudnick and Otavio Good for helpful suggestions. We would also like to thank the TACL Action Editor and the reviewers for their feedback.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Tensorflow: A system for large-scale machine learning",
"authors": [
{
"first": "Martin",
"middle": [],
"last": "Abadi",
"suffix": ""
},
{
"first": "Paul",
"middle": [],
"last": "Barham",
"suffix": ""
}
],
"year": 2016,
"venue": "12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16)",
"volume": "",
"issue": "",
"pages": "265--283",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Martin Abadi and Paul Barham et al. 2016. Tensorflow: A system for large-scale machine learning. In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16), pages 265-283.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Neural machine translation by jointly learning to align and translate",
"authors": [
{
"first": "Dzmitry",
"middle": [],
"last": "Bahdanau",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2015,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Does multimodality help human and machine for translation and image captioning?",
"authors": [
{
"first": "Ozan",
"middle": [],
"last": "Caglayan",
"suffix": ""
},
{
"first": "Walid",
"middle": [],
"last": "Aransa",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the First Conference on Machine Translation",
"volume": "",
"issue": "",
"pages": "627--633",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ozan Caglayan and Walid Aransa et al. 2016. Does mul- timodality help human and machine for translation and image captioning? In Proceedings of the First Confer- ence on Machine Translation, pages 627-633, Berlin, Germany, August. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Multitask learning",
"authors": [
{
"first": "Rich",
"middle": [],
"last": "Caruana",
"suffix": ""
}
],
"year": 1998,
"venue": "Learning to learn",
"volume": "",
"issue": "",
"pages": "95--133",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rich Caruana. 1998. Multitask learning. In Learning to learn, pages 95-133. Springer.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Learning phrase representations using RNN encoder-decoder for statistical machine translation",
"authors": [
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Bart",
"middle": [],
"last": "Van Merrienboer",
"suffix": ""
},
{
"first": "\u00c7aglar",
"middle": [],
"last": "G\u00fcl\u00e7ehre",
"suffix": ""
},
{
"first": "Fethi",
"middle": [],
"last": "Bougares",
"suffix": ""
},
{
"first": "Holger",
"middle": [],
"last": "Schwenk",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2014,
"venue": "Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kyunghyun Cho, Bart van Merrienboer, \u00c7aglar G\u00fcl\u00e7ehre, Fethi Bougares, Holger Schwenk, and Yoshua Ben- gio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Conference on Empirical Methods in Natural Language Processing.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Systran's pure neural machine translation systems",
"authors": [
{
"first": "Josep",
"middle": [],
"last": "Crego",
"suffix": ""
},
{
"first": "Jungi",
"middle": [],
"last": "Kim",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1610.05540"
]
},
"num": null,
"urls": [],
"raw_text": "Josep Crego and Jungi Kim et al. 2016. Systran's pure neural machine translation systems. arXiv preprint arXiv:1610.05540.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Multi-task learning for multiple language translation",
"authors": [
{
"first": "Daxiang",
"middle": [],
"last": "Dong",
"suffix": ""
},
{
"first": "Hua",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Dianhai",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "Haifeng",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1723--1732",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daxiang Dong, Hua Wu, Wei He, Dianhai Yu, and Haifeng Wang. 2015. Multi-task learning for multiple language translation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics, pages 1723-1732.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Multi-way, multilingual neural machine translation with a shared attention mechanism",
"authors": [
{
"first": "Orhan",
"middle": [],
"last": "Firat",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2016,
"venue": "The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "866--875",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Orhan Firat, Kyunghyun Cho, and Yoshua Bengio. 2016a. Multi-way, multilingual neural machine translation with a shared attention mechanism. In The 2016 Con- ference of the North American Chapter of the Associa- tion for Computational Linguistics: Human Language Technologies, San Diego California, USA, June 12-17, 2016, pages 866-875.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Multi-way, multilingual neural machine translation",
"authors": [
{
"first": "Orhan",
"middle": [],
"last": "Firat",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Baskaran",
"middle": [],
"last": "Sankaran",
"suffix": ""
},
{
"first": "T",
"middle": [
"Yarman"
],
"last": "Fatos",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Vural",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2016,
"venue": "Computer Speech and Language",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Orhan Firat, Kyunghyun Cho, Baskaran Sankaran, Fatos T. Yarman Vural, and Yoshua Bengio. 2016b. Multi-way, multilingual neural machine translation. Computer Speech and Language.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Zero-resource translation with multi-lingual neural machine translation",
"authors": [
{
"first": "Orhan",
"middle": [],
"last": "Firat",
"suffix": ""
},
{
"first": "Baskaran",
"middle": [],
"last": "Sankaran",
"suffix": ""
},
{
"first": "Yaser",
"middle": [],
"last": "Al-Onaizan",
"suffix": ""
},
{
"first": "Fatos",
"middle": [
"T Yarman"
],
"last": "Vural",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "268--277",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Orhan Firat, Baskaran Sankaran, Yaser Al-Onaizan, Fatos T. Yarman Vural, and Kyunghyun Cho. 2016c. Zero-resource translation with multi-lingual neural ma- chine translation. In Proceedings of the 2016 Con- ference on Empirical Methods in Natural Language Processing, pages 268-277, Austin, Texas, November. Association for Computational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Catastrophic forgetting in connectionist networks",
"authors": [
{
"first": "M",
"middle": [],
"last": "Robert",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "French",
"suffix": ""
}
],
"year": 1999,
"venue": "Trends in cognitive sciences",
"volume": "3",
"issue": "4",
"pages": "128--135",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Robert M French. 1999. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A new algorithm for data compression",
"authors": [
{
"first": "Philip",
"middle": [],
"last": "Gage",
"suffix": ""
}
],
"year": 1994,
"venue": "C Users Journal",
"volume": "12",
"issue": "2",
"pages": "23--38",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philip Gage. 1994. A new algorithm for data compression. C Users Journal, 12(2):23-38, February.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Multilingual language processing from bytes",
"authors": [
{
"first": "Dan",
"middle": [],
"last": "Gillick",
"suffix": ""
},
{
"first": "Cliff",
"middle": [],
"last": "Brunk",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "1296--1306",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dan Gillick, Cliff Brunk, Oriol Vinyals, and Amarnag Sub- ramanya. 2016. Multilingual language processing from bytes. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Compu- tational Linguistics: Human Language Technologies, pages 1296-1306, San Diego, California, June. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "An introduction to machine translation",
"authors": [
{
"first": "William John",
"middle": [],
"last": "Hutchins",
"suffix": ""
},
{
"first": "Harold",
"middle": [
"L"
],
"last": "Somers",
"suffix": ""
}
],
"year": 1992,
"venue": "",
"volume": "362",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "William John Hutchins and Harold L. Somers. 1992. An introduction to machine translation, volume 362. Academic Press London.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Recurrent continuous translation models",
"authors": [
{
"first": "Nal",
"middle": [],
"last": "Kalchbrenner",
"suffix": ""
},
{
"first": "Phil",
"middle": [],
"last": "Blunsom",
"suffix": ""
}
],
"year": 2013,
"venue": "Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent continuous translation models. In Conference on Em- pirical Methods in Natural Language Processing.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Fully character-level neural machine translation without explicit segmentation",
"authors": [
{
"first": "Jason",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Hofmann",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1610.03017"
]
},
"num": null,
"urls": [],
"raw_text": "Jason Lee, Kyunghyun Cho, and Thomas Hofmann. 2016. Fully character-level neural machine transla- tion without explicit segmentation. arXiv preprint arXiv:1610.03017.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Multi-task sequence to sequence learning",
"authors": [
{
"first": "Minh-Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Quoc",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Lukasz",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kaiser",
"suffix": ""
}
],
"year": 2015,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Minh-Thang Luong, Quoc V Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. 2015a. Multi-task se- quence to sequence learning. In International Confer- ence on Learning Representations.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Effective approaches to attention-based neural machine translation",
"authors": [
{
"first": "Minh-Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "Hieu",
"middle": [],
"last": "Pham",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2015,
"venue": "Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Minh-Thang Luong, Hieu Pham, and Christopher D. Man- ning. 2015b. Effective approaches to attention-based neural machine translation. In Conference on Empiri- cal Methods in Natural Language Processing.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Addressing the rare word problem in neural machine translation",
"authors": [
{
"first": "Minh-Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Quoc",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Wojciech",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Zaremba",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Minh-Thang Luong, Ilya Sutskever, Quoc V Le, Oriol Vinyals, and Wojciech Zaremba. 2015c. Addressing the rare word problem in neural machine translation. In Proceedings of the 53rd Annual Meeting of the As- sociation for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Visualizing Data using t-SNE",
"authors": [
{
"first": "Laurens",
"middle": [],
"last": "Van Der Maaten",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [],
"last": "Hinton",
"suffix": ""
}
],
"year": 2008,
"venue": "Journal of Machine Learning Research",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Laurens Van Der Maaten and Geoffrey Hinton. 2008. Visualizing Data using t-SNE. Journal of Machine Learning Research, 9.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Interlingual machine translation",
"authors": [
{
"first": "H",
"middle": [],
"last": "Richard",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Richens",
"suffix": ""
}
],
"year": 1958,
"venue": "The Computer Journal",
"volume": "1",
"issue": "3",
"pages": "144--147",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Richard H Richens. 1958. Interlingual machine transla- tion. The Computer Journal, 1(3):144-147.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Multilingual speech processing",
"authors": [
{
"first": "Tanja",
"middle": [],
"last": "Schultz",
"suffix": ""
},
{
"first": "Katrin",
"middle": [],
"last": "Kirchhoff",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tanja Schultz and Katrin Kirchhoff. 2006. Multilingual speech processing. Elsevier Academic Press, Amster- dam, Boston, Paris.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Japanese and Korean voice search",
"authors": [
{
"first": "Mike",
"middle": [],
"last": "Schuster",
"suffix": ""
},
{
"first": "Kaisuke",
"middle": [],
"last": "Nakajima",
"suffix": ""
}
],
"year": 2012,
"venue": "IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mike Schuster and Kaisuke Nakajima. 2012. Japanese and Korean voice search. 2012 IEEE International Conference on Acoustics, Speech and Signal Process- ing.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "On using very large target vocabulary for neural machine translation",
"authors": [
{
"first": "Jean",
"middle": [],
"last": "S\u00e9bastien",
"suffix": ""
},
{
"first": "Cho",
"middle": [],
"last": "Kyunghyun",
"suffix": ""
},
{
"first": "Roland",
"middle": [],
"last": "Memisevic",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jean S\u00e9bastien, Cho Kyunghyun, Roland Memisevic, and Yoshua Bengio. 2015. On using very large target vo- cabulary for neural machine translation. In Proceedings of the 53rd Annual Meeting of the Association for Com- putational Linguistics and the 7th International Joint Conference on Natural Language Processing.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Controlling politeness in neural machine translation via side constraints",
"authors": [
{
"first": "Rico",
"middle": [],
"last": "Sennrich",
"suffix": ""
},
{
"first": "Barry",
"middle": [],
"last": "Haddow",
"suffix": ""
},
{
"first": "Alexandra",
"middle": [],
"last": "Birch",
"suffix": ""
}
],
"year": 2016,
"venue": "The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "35--40",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016a. Controlling politeness in neural machine trans- lation via side constraints. In The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, San Diego California, USA, June 12-17, 2016, pages 35-40.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Neural machine translation of rare words with subword units",
"authors": [
{
"first": "Rico",
"middle": [],
"last": "Sennrich",
"suffix": ""
},
{
"first": "Barry",
"middle": [],
"last": "Haddow",
"suffix": ""
},
{
"first": "Alexandra",
"middle": [],
"last": "Birch",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016b. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meet- ing of the Association for Computational Linguistics.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Sequence to sequence learning with neural networks",
"authors": [
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Quoc",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Le",
"suffix": ""
}
],
"year": 2014,
"venue": "Advances in Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "3104--3112",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pages 3104-3112.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Polyglot neural language models: A case study in crosslingual phonetic representation learning",
"authors": [
{
"first": "Yulia",
"middle": [],
"last": "Tsvetkov",
"suffix": ""
},
{
"first": "Sunayana",
"middle": [],
"last": "Sitaram",
"suffix": ""
},
{
"first": "Manaal",
"middle": [],
"last": "Faruqui",
"suffix": ""
},
{
"first": "Guillaume",
"middle": [],
"last": "Lample",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Littell",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Mortensen",
"suffix": ""
},
{
"first": "Alan",
"middle": [
"W"
],
"last": "Black",
"suffix": ""
},
{
"first": "Lori",
"middle": [],
"last": "Levin",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Dyer",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "1357--1366",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yulia Tsvetkov, Sunayana Sitaram, Manaal Faruqui, Guillaume Lample, Patrick Littell, David Mortensen, Alan W Black, Lori Levin, and Chris Dyer. 2016. Poly- glot neural language models: A case study in cross- lingual phonetic representation learning. In Proceed- ings of the 2016 Conference of the North American Chapter of the Association for Computational Linguis- tics: Human Language Technologies, pages 1357-1366, San Diego, California, June. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Google's neural machine translation system: Bridging the gap between human and machine translation",
"authors": [
{
"first": "Yonghui",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Schuster",
"suffix": ""
},
{
"first": "Zhifeng",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1609.08144v2"
]
},
"num": null,
"urls": [],
"raw_text": "Yonghui Wu, Mike Schuster, and Zhifeng Chen et al. 2016. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144v2.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Controlling the voice of a sentence in Japanese-to-English neural machine translation",
"authors": [
{
"first": "Hayahide",
"middle": [],
"last": "Yamagishi",
"suffix": ""
},
{
"first": "Shin",
"middle": [],
"last": "Kanouchi",
"suffix": ""
},
{
"first": "Mamoru",
"middle": [],
"last": "Komachi",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 3rd Workshop on Asian Translation",
"volume": "",
"issue": "",
"pages": "203--210",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hayahide Yamagishi, Shin Kanouchi, and Mamoru Ko- machi. 2016. Controlling the voice of a sentence in Japanese-to-English neural machine translation. In Proceedings of the 3rd Workshop on Asian Translation, pages 203-210, Osaka, Japan, December.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Deep recurrent models with fast-forward connections for neural machine translation",
"authors": [
{
"first": "Jie",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "Ying",
"middle": [],
"last": "Cao",
"suffix": ""
},
{
"first": "Xuguang",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Peng",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Xu",
"suffix": ""
}
],
"year": 2016,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "4",
"issue": "",
"pages": "371--383",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. 2016. Deep recurrent models with fast-forward connections for neural machine translation. Transac- tions of the Association for Computational Linguistics, 4:371-383.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Multi-source neural translation",
"authors": [
{
"first": "Barret",
"middle": [],
"last": "Zoph",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
}
],
"year": 2016,
"venue": "The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "30--34",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Barret Zoph and Kevin Knight. 2016. Multi-source neural translation. In NAACL HLT 2016, The 2016 Confer- ence of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, San Diego California, USA, June 12-17, 2016, pages 30-34.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"num": null,
"text": "A t-SNE projection of the embedding of 74 semantically identical sentences translated across all 6 possible directions, yielding a total of 9,978 steps (dots in the image), from the model trained on English\u2194Japanese and English\u2194Korean examples. (a) A bird's-eye view of the embedding, coloring by the index of the semantic sentence. Well-defined clusters each having a single color are apparent. (b) A zoomed in view of one of the clusters with the same coloring. All of the sentences within this cluster are translations of \"The stratosphere extends from about 10km to about 50km in altitude.\" (c) The same cluster colored by source language. All three source languages can be seen within this cluster.",
"type_str": "figure"
},
"FIGREF1": {
"uris": null,
"num": null,
"text": "(a) A bird's-eye view of a t-SNE projection of an embedding of the model trained on Portuguese\u2192English (blue) and English\u2192Spanish (yellow) examples with a Portuguese\u2192Spanish zero-shot bridge (red). The large red region on the left primarily contains the zero-shot Portuguese\u2192Spanish translations. (b) A scatter plot of BLEU scores of zero-shot translations versus the average point-wise distance between the zero-shot translation and a non-bridged translation. The Pearson correlation coefficient is \u22120.42.",
"type_str": "figure"
},
"TABREF0": {
"content": "<table><tr><td>WMT De\u2192En</td><td>30.43 30.54 +0.11</td></tr><tr><td>WMT Fr\u2192En</td><td>35.50 36.77 +1.27</td></tr><tr><td colspan=\"2\">Prod Ja\u2192En 23.41 23.87 +0.46</td></tr><tr><td colspan=\"2\">Prod Ko\u2192En 25.42 25.47 +0.05</td></tr><tr><td colspan=\"2\">Prod Es\u2192En 38.00 38.73 +0.73</td></tr><tr><td colspan=\"2\">Prod Pt\u2192En 44.40 45.19 +0.79</td></tr></table>",
"num": null,
"type_str": "table",
"text": "Many to One: BLEU scores on for single language pair and multilingual models. : no oversampling Model Single Multi Diff WMT De\u2192En 30.43 30.59 +0.16 WMT Fr\u2192En 35.50 35.73 +0.23",
"html": null
},
"TABREF1": {
"content": "<table><tr><td colspan=\"2\">Model Single Multi</td><td>Diff</td></tr><tr><td colspan=\"3\">WMT En\u2192De 24.67 24.97 +0.30</td></tr><tr><td colspan=\"3\">WMT En\u2192Fr 38.95 36.84 -2.11</td></tr><tr><td>WMT En\u2192De</td><td colspan=\"2\">24.67 22.61 -2.06</td></tr><tr><td>WMT En\u2192Fr</td><td colspan=\"2\">38.95 38.16 -0.79</td></tr><tr><td colspan=\"3\">Prod En\u2192Ja 23.66 23.73 +0.07</td></tr><tr><td colspan=\"3\">Prod En\u2192Ko 19.75 19.58 -0.17</td></tr><tr><td colspan=\"3\">Prod En\u2192Es 34.50 35.40 +0.90</td></tr><tr><td colspan=\"3\">Prod En\u2192Pt 38.40 38.63 +0.23</td></tr></table>",
"num": null,
"type_str": "table",
"text": "One to Many: BLEU scores for single language pair and multilingual models. : no oversampling",
"html": null
},
"TABREF2": {
"content": "<table><tr><td colspan=\"2\">WMT Fr\u2192En 35.50 34.89 -0.61</td></tr><tr><td>WMT En\u2192De</td><td>24.67 21.92 -2.75</td></tr><tr><td>WMT En\u2192Fr</td><td>38.95 37.45 -1.50</td></tr><tr><td>WMT De\u2192En</td><td>30.43 29.22 -1.21</td></tr><tr><td>WMT Fr\u2192En</td><td>35.50 35.93 +0.43</td></tr><tr><td colspan=\"2\">Prod En\u2192Ja 23.66 23.12 -0.54</td></tr><tr><td colspan=\"2\">Prod En\u2192Ko 19.75 19.73 -0.02</td></tr><tr><td colspan=\"2\">Prod Ja\u2192En 23.41 22.86 -0.55</td></tr><tr><td colspan=\"2\">Prod Ko\u2192En 25.42 24.76 -0.66</td></tr><tr><td colspan=\"2\">Prod En\u2192Es 34.50 34.69 +0.19</td></tr><tr><td colspan=\"2\">Prod En\u2192Pt 38.40 37.25 -1.15</td></tr><tr><td colspan=\"2\">Prod Es\u2192En 38.00 37.65 -0.35</td></tr><tr><td colspan=\"2\">Prod Pt\u2192En 44.40 44.02 -0.38</td></tr></table>",
"num": null,
"type_str": "table",
"text": "Many to Many: BLEU scores for single language pair and multilingual models. : no oversampling Model Single Multi Diff WMT En\u2192De 24.67 24.49 -0.18 WMT En\u2192Fr 38.95 36.23 -2.72 WMT De\u2192En 30.43 29.84 -0.59",
"html": null
},
"TABREF3": {
"content": "<table><tr><td>Model</td><td colspan=\"2\">Single Multi</td><td>Multi</td><td>Multi</td><td>Multi</td></tr><tr><td>#nodes</td><td>1024</td><td>1024</td><td>1280</td><td>1536</td><td>1792</td></tr><tr><td>#params</td><td>3B</td><td colspan=\"4\">255M 367M 499M 650M</td></tr><tr><td>En\u2192Ja</td><td>23.66</td><td colspan=\"4\">21.10 21.17 21.72 21.70</td></tr><tr><td>En\u2192Ko</td><td>19.75</td><td colspan=\"4\">18.41 18.36 18.30 18.28</td></tr><tr><td>Ja\u2192En</td><td>23.41</td><td colspan=\"4\">21.62 22.03 22.51 23.18</td></tr><tr><td>Ko\u2192En</td><td>25.42</td><td colspan=\"4\">22.87 23.46 24.00 24.67</td></tr><tr><td>En\u2192Es</td><td>34.50</td><td colspan=\"4\">34.25 34.40 34.77 34.70</td></tr><tr><td>En\u2192Pt</td><td>38.40</td><td colspan=\"4\">37.35 37.42 37.80 37.92</td></tr><tr><td>Es\u2192En</td><td>38.00</td><td colspan=\"4\">36.04 36.50 37.26 37.45</td></tr><tr><td>Pt\u2192En</td><td>44.40</td><td colspan=\"4\">42.53 42.82 43.64 43.87</td></tr><tr><td>En\u2192De</td><td>26.43</td><td colspan=\"4\">23.15 23.77 23.63 24.01</td></tr><tr><td>En\u2192Fr</td><td>35.37</td><td colspan=\"4\">34.00 34.19 34.91 34.81</td></tr><tr><td>De\u2192En</td><td>31.77</td><td colspan=\"4\">31.17 31.65 32.24 32.32</td></tr><tr><td>Fr\u2192En</td><td>36.47</td><td colspan=\"4\">34.40 34.56 35.35 35.52</td></tr><tr><td>ave diff</td><td>-</td><td>-1.72</td><td>-1.43</td><td>-0.95</td><td>-0.76</td></tr><tr><td>vs single</td><td>-</td><td colspan=\"4\">-5.6% -4.7% -3.1% -2.5%</td></tr></table>",
"num": null,
"type_str": "table",
"text": "Large-scale experiments: BLEU scores for single language pair and multilingual models.",
"html": null
},
"TABREF4": {
"content": "<table><tr><td/><td>Model</td><td colspan=\"2\">Zero-shot BLEU</td></tr><tr><td>(a)</td><td>PBMT bridged</td><td>no</td><td>28.99</td></tr><tr><td>(b)</td><td>NMT bridged</td><td>no</td><td>30.91</td></tr><tr><td>(c)</td><td>NMT Pt\u2192Es</td><td>no</td><td>31.50</td></tr><tr><td>(d)</td><td>Model 1 (Pt\u2192En, En\u2192Es)</td><td>yes</td><td>21.62</td></tr><tr><td>(e)</td><td>Model 2 (En\u2194{Es, Pt})</td><td>yes</td><td>24.75</td></tr><tr><td colspan=\"2\">(f) Model 2 + incremental training</td><td>no</td><td>31.77</td></tr></table>",
"num": null,
"type_str": "table",
"text": "Portuguese\u2192Spanish BLEU scores using various models.",
"html": null
},
"TABREF5": {
"content": "<table><tr><td>Model</td><td>BLEU</td></tr><tr><td>NMT Es\u2192Ja explicitly bridged</td><td>18.00</td></tr><tr><td>NMT Es\u2192Ja implicitly bridged</td><td>9.14</td></tr></table>",
"num": null,
"type_str": "table",
"text": "Spanish\u2192Japanese BLEU scores for explicit and implicit bridging using the 12-language pair large-scale model fromTable 4.",
"html": null
},
"TABREF6": {
"content": "<table><tr><td/><td colspan=\"3\">Zero-Shot From-Scratch Incremental</td></tr><tr><td>En\u2192Be</td><td>16.85</td><td>17.03</td><td>16.99</td></tr><tr><td>En\u2192Ru</td><td>22.21</td><td>22.03</td><td>21.92</td></tr><tr><td>En\u2192Uk</td><td>18.16</td><td>17.75</td><td>18.27</td></tr><tr><td>Be\u2192En</td><td>25.44</td><td>24.72</td><td>25.54</td></tr><tr><td>Ru\u2192En</td><td>28.36</td><td>27.90</td><td>28.46</td></tr><tr><td>Uk\u2192En</td><td>28.60</td><td>28.51</td><td>28.58</td></tr><tr><td>Be\u2192Ru</td><td>56.53</td><td>82.50</td><td>78.63</td></tr><tr><td>Ru\u2192Be</td><td>58.75</td><td>72.06</td><td>70.01</td></tr><tr><td>Ru\u2192Uk</td><td>21.92</td><td>25.75</td><td>25.34</td></tr><tr><td>Uk\u2192Ru</td><td>16.73</td><td>30.53</td><td>29.92</td></tr></table>",
"num": null,
"type_str": "table",
"text": "BLEU scores for English\u2194{Belarusian, Russian, Ukrainian} models.",
"html": null
},
"TABREF7": {
"content": "<table/>",
"num": null,
"type_str": "table",
"text": "Gradually mixing target languages Ja/Ko. w ko I must be getting somewhere near the centre of the earth. 0.00 \u79c1\u306f\u5730\u7403\u306e\u4e2d\u5fc3\u306e\u8fd1\u304f\u306b\u3069\u3053\u304b\u306b\u884c\u3063\u3066\u3044\u308b \u306b\u9055\u3044\u306a\u3044\u3002 0.40 \u79c1\u306f\u5730\u7403\u306e\u4e2d\u5fc3\u8fd1\u304f\u306e\u3069\u3053\u304b\u306b\u7740\u3044\u3066\u3044\u308b\u306b \u9055\u3044\u306a\u3044\u3002 0.56 \u79c1\u306f\u5730\u7403\u306e\u4e2d\u5fc3\u306e\u8fd1\u304f\u306e\u3069\u3053\u304b\u306b\u306a\u3063\u3066\u3044\u308b \u306b\u9055\u3044\u306a\u3044\u3002 0.58 \u79c1\u306f\u110c \u1175\u1100 \u116e\u306e\u4e2d\u5fc3\u110b \u1174\u1100 \u1161\u1101 \u1161\u110b \u1175\u110b \u1166\u110b \u1165\u1103 \u1175 \u11ab\u1100 \u1161\u110b \u1166\u1103 \u1169\u110e \u1161 \u11a8\u1112 \u1161\u1100 \u1169\u110b \u1175 \u11bb \u110b \u1165\u110b \u1163\u1112 \u1161 \u11ab\u1103 \u1161\u3002 0.60 \u1102 \u1161\u1102 \u1173 \u11ab\u110c \u1175\u1100 \u116e\u110b \u1174\u1109 \u1166 \u11ab\u1110 \u1165\u110b \u1174\u1100 \u1161\u1101 \u1161\u110b \u1175\u110b \u1166\u110b \u1165\u1103 \u1175 \u11ab\u1100 \u1161\u110b \u1166\u1103 \u1169\u110e \u1161 \u11a8\u1112 \u1161\u1100 \u1169\u110b \u1175 \u11bb \u110b \u1165\u110b \u1163\u1112 \u1161 \u11ab\u1103 \u1161\u3002 0.70 \u1102 \u1161\u1102 \u1173 \u11ab\u110c \u1175\u1100 \u116e\u110b \u1174\u110c \u116e \u11bc\u1109 \u1175 \u11b7\u1100 \u1173 \u11ab\u110e \u1165\u110b \u1165\u1103 \u1175 \u11ab\u1100 \u1161\u110b \u1166\u1103 \u1169\u110e \u1161 \u11a8\u1112 \u1162\u110b \u1163\u1112 \u1161 \u11b8\u1102 \u1175\u1103 \u1161\u3002 0.90 \u1102 \u1161\u1102 \u1173 \u11ab\u110b \u1165\u1103 \u1175 \u11ab\u1100 \u1161\u110c \u1175\u1100 \u116e\u110b \u1174\u110c \u116e \u11bc\u1109 \u1175 \u11b7\u1100 \u1173 \u11ab\u110e \u1165\u110b \u1166\u1103 \u1169\u110e \u1161 \u11a8\u1112 \u1162\u110b \u1163\u1112 \u1161 \u11b8\u1102 \u1175\u1103 \u1161\u3002 1.00 \u1102 \u1161\u1102 \u1173 \u11ab\u110b \u1165\u1103 \u1175 \u11ab\u1100 \u1161\u110c \u1175\u1100 \u116e\u110b \u1174\u110c \u116e \u11bc\u1109 \u1175 \u11b7\u1100 \u1173 \u11ab\u110e \u1165\u110b \u1166\u1103 \u1169\u110e \u1161 \u11a8\u1112 \u1162\u110b \u1163\u1112 \u1161 \u11b8\u1102 \u1175\u1103 \u1161\u3002",
"html": null
}
}
}
} |