Spaces:
Sleeping
Sleeping
File size: 136,393 Bytes
a72140d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 | {
"paper_id": "Transformer",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2025-04-11T13:00:43.870321Z"
},
"title": "Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works. Attention Is All You Need",
"authors": [
{
"first": "Ashish",
"middle": [],
"last": "Vaswani",
"suffix": "",
"affiliation": {},
"email": "avaswani@google.com"
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Google",
"middle": [],
"last": "Brain",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Niki",
"middle": [],
"last": "Parmar",
"suffix": "",
"affiliation": {},
"email": "nikip@google.com"
},
{
"first": "Jakob",
"middle": [],
"last": "Uszkoreit",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Llion",
"middle": [],
"last": "Jones",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Aidan",
"middle": [
"N"
],
"last": "Gomez",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "\u0141ukasz",
"middle": [],
"last": "Kaiser",
"suffix": "",
"affiliation": {},
"email": "lukaszkaiser@google.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 Englishto-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data. * Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.\n\u2020 Work performed while at Google Brain.\n\u2021 Work performed while at Google Research.",
"pdf_parse": {
"paper_id": "Transformer",
"_pdf_hash": "",
"abstract": [
{
"text": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 Englishto-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data. * Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
},
{
"text": "\u2020 Work performed while at Google Brain.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
},
{
"text": "\u2021 Work performed while at Google Research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5] . Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15] .",
"cite_spans": [
{
"start": 50,
"end": 54,
"text": "[13]",
"ref_id": "BIBREF12"
},
{
"start": 75,
"end": 78,
"text": "[7]",
"ref_id": "BIBREF6"
},
{
"start": 267,
"end": 271,
"text": "[35,",
"ref_id": "BIBREF34"
},
{
"start": 272,
"end": 274,
"text": "2,",
"ref_id": "BIBREF1"
},
{
"start": 275,
"end": 277,
"text": "5]",
"ref_id": "BIBREF4"
},
{
"start": 404,
"end": 408,
"text": "[38,",
"ref_id": "BIBREF37"
},
{
"start": 409,
"end": 412,
"text": "24,",
"ref_id": "BIBREF23"
},
{
"start": 413,
"end": 416,
"text": "15]",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states h t , as a function of the previous hidden state h t-1 and the input for position t. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional computation [32] , while also improving model performance in case of the latter. The fundamental constraint of sequential computation, however, remains.",
"cite_spans": [
{
"start": 586,
"end": 590,
"text": "[21]",
"ref_id": "BIBREF20"
},
{
"start": 619,
"end": 623,
"text": "[32]",
"ref_id": "BIBREF31"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Attention mechanisms have become an integral part of compelling sequence modeling and transduction models in various tasks, allowing modeling of dependencies without regard to their distance in the input or output sequences [2, 19] . In all but a few cases [27] , however, such attention mechanisms are used in conjunction with a recurrent network.",
"cite_spans": [
{
"start": 224,
"end": 227,
"text": "[2,",
"ref_id": "BIBREF1"
},
{
"start": 228,
"end": 231,
"text": "19]",
"ref_id": "BIBREF18"
},
{
"start": 257,
"end": 261,
"text": "[27]",
"ref_id": "BIBREF26"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output. The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as little as twelve hours on eight P100 GPUs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16] , ByteNet [18] and ConvS2S [9] , all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions. In these models, the number of operations required to relate signals from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes it more difficult to learn dependencies between distant positions [12] . In the Transformer this is reduced to a constant number of operations, albeit at the cost of reduced effective resolution due to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as described in section 3.2.",
"cite_spans": [
{
"start": 97,
"end": 101,
"text": "[16]",
"ref_id": "BIBREF15"
},
{
"start": 112,
"end": 116,
"text": "[18]",
"ref_id": "BIBREF17"
},
{
"start": 129,
"end": 132,
"text": "[9]",
"ref_id": "BIBREF8"
},
{
"start": 574,
"end": 578,
"text": "[12]",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Self-attention, sometimes called intra-attention is an attention mechanism relating different positions of a single sequence in order to compute a representation of the sequence. Self-attention has been used successfully in a variety of tasks including reading comprehension, abstractive summarization, textual entailment and learning task-independent sentence representations [4, 27, 28, 22] .",
"cite_spans": [
{
"start": 377,
"end": 380,
"text": "[4,",
"ref_id": "BIBREF3"
},
{
"start": 381,
"end": 384,
"text": "27,",
"ref_id": "BIBREF26"
},
{
"start": 385,
"end": 388,
"text": "28,",
"ref_id": "BIBREF27"
},
{
"start": 389,
"end": 392,
"text": "22]",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "End-to-end memory networks are based on a recurrent attention mechanism instead of sequencealigned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [34] .",
"cite_spans": [
{
"start": 212,
"end": 216,
"text": "[34]",
"ref_id": "BIBREF33"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using sequencealigned RNNs or convolution. In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9] .",
"cite_spans": [
{
"start": 356,
"end": 360,
"text": "[17,",
"ref_id": "BIBREF16"
},
{
"start": 361,
"end": 364,
"text": "18]",
"ref_id": "BIBREF17"
},
{
"start": 369,
"end": 372,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35] . Here, the encoder maps an input sequence of symbol representations (x 1 , ..., x n ) to a sequence of continuous representations z = (z 1 , ..., z n ). Given z, the decoder then generates an output sequence (y 1 , ..., y m ) of symbols one element at a time. At each step the model is auto-regressive [10] , consuming the previously generated symbols as additional input when generating the next. The Transformer follows this overall architecture using stacked self-attention and point-wise, fully connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1 , respectively.",
"cite_spans": [
{
"start": 87,
"end": 90,
"text": "[5,",
"ref_id": "BIBREF4"
},
{
"start": 91,
"end": 93,
"text": "2,",
"ref_id": "BIBREF1"
},
{
"start": 94,
"end": 97,
"text": "35]",
"ref_id": "BIBREF34"
},
{
"start": 401,
"end": 405,
"text": "[10]",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 694,
"end": 695,
"text": "1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Model Architecture",
"sec_num": "3"
},
{
"text": "Encoder: The encoder is composed of a stack of N = 6 identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, positionwise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1] . That is, the output of each sub-layer is LayerNorm(x + Sublayer(x)), where Sublayer(x) is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension d model = 512.",
"cite_spans": [
{
"start": 265,
"end": 269,
"text": "[11]",
"ref_id": "BIBREF10"
},
{
"start": 337,
"end": 340,
"text": "[1]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Encoder and Decoder Stacks",
"sec_num": "3.1"
},
{
"text": "Decoder: The decoder is also composed of a stack of N = 6 identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position i can depend only on the known outputs at positions less than i.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Encoder and Decoder Stacks",
"sec_num": "3.1"
},
{
"text": "An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum Scaled Dot-Product Attention Multi-Head Attention of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention",
"sec_num": "3.2"
},
{
"text": "We call our particular attention \"Scaled Dot-Product Attention\" (Figure 2 ). The input consists of queries and keys of dimension d k , and values of dimension d v . We compute the dot products of the query with all keys, divide each by \u221a d k , and apply a softmax function to obtain the weights on the values.",
"cite_spans": [],
"ref_spans": [
{
"start": 72,
"end": 73,
"text": "2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": "In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix Q. The keys and values are also packed together into matrices K and V . We compute the matrix of outputs as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Attention(Q, K, V ) = softmax( QK T \u221a d k )V",
"eq_num": "(1)"
}
],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": "The two most commonly used attention functions are additive attention [2] , and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor of 1",
"cite_spans": [
{
"start": 70,
"end": 73,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": "\u221a d k",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": ". Additive attention computes the compatibility function using a feed-forward network with a single hidden layer. While the two are similar in theoretical complexity, dot-product attention is much faster and more space-efficient in practice, since it can be implemented using highly optimized matrix multiplication code.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": "While for small values of d k the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of d k [3] . We suspect that for large values of d k , the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients4 . To counteract this effect, we scale the dot products by 1",
"cite_spans": [
{
"start": 162,
"end": 165,
"text": "[3]",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": "\u221a d k .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scaled Dot-Product Attention",
"sec_num": "3.2.1"
},
{
"text": "Instead of performing a single attention function with d model -dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values h times with different, learned linear projections to d k , d k and d v dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding d v -dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 531,
"end": 532,
"text": "2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "MultiHead(Q, K, V ) = Concat(head 1 , ..., head h )W O",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "head i = Attention(QW Q i , KW K i , V W V i )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "Where the projections are parameter matrices",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "W Q i \u2208 R dmodel\u00d7d k , W K i \u2208 R dmodel\u00d7d k , W V i \u2208 R dmodel\u00d7dv and W O \u2208 R hdv\u00d7dmodel .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "In this work we employ h = 8 parallel attention layers, or heads. For each of these we use",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "d k = d v = d model /h = 64.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-Head Attention",
"sec_num": "3.2.2"
},
{
"text": "The Transformer uses multi-head attention in three different ways:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Applications of Attention in our Model",
"sec_num": "3.2.3"
},
{
"text": "\u2022 In \"encoder-decoder attention\" layers, the queries come from the previous decoder layer, and the memory keys and values come from the output of the encoder. This allows every position in the decoder to attend over all positions in the input sequence. This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9] .",
"cite_spans": [
{
"start": 353,
"end": 357,
"text": "[38,",
"ref_id": "BIBREF37"
},
{
"start": 358,
"end": 360,
"text": "2,",
"ref_id": "BIBREF1"
},
{
"start": 361,
"end": 363,
"text": "9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Applications of Attention in our Model",
"sec_num": "3.2.3"
},
{
"text": "\u2022 The encoder contains self-attention layers. In a self-attention layer all of the keys, values and queries come from the same place, in this case, the output of the previous layer in the encoder. Each position in the encoder can attend to all positions in the previous layer of the encoder.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Applications of Attention in our Model",
"sec_num": "3.2.3"
},
{
"text": "\u2022 Similarly, self-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to -\u221e) all values in the input of the softmax which correspond to illegal connections. See Figure 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 443,
"end": 444,
"text": "2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Applications of Attention in our Model",
"sec_num": "3.2.3"
},
{
"text": "In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Position-wise Feed-Forward Networks",
"sec_num": "3.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "FFN(x) = max(0, xW 1 + b 1 )W 2 + b 2",
"eq_num": "(2)"
}
],
"section": "Position-wise Feed-Forward Networks",
"sec_num": "3.3"
},
{
"text": "While the linear transformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kernel size 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Position-wise Feed-Forward Networks",
"sec_num": "3.3"
},
{
"text": "The dimensionality of input and output is d model = 512, and the inner-layer has dimensionality d f f = 2048.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Position-wise Feed-Forward Networks",
"sec_num": "3.3"
},
{
"text": "Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension d model . We also use the usual learned linear transformation and softmax function to convert the decoder output to predicted next-token probabilities. In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [30] . In the embedding layers, we multiply those weights by \u221a d model . ",
"cite_spans": [
{
"start": 430,
"end": 434,
"text": "[30]",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Embeddings and Softmax",
"sec_num": "3.4"
},
{
"text": "(n 2 \u2022 d) O(1) O(1) Recurrent O(n \u2022 d 2 ) O(n) O(n) Convolutional O(k \u2022 n \u2022 d 2 ) O(1) O(log k (n)) Self-Attention (restricted) O(r \u2022 n \u2022 d) O(1) O(n/r)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Embeddings and Softmax",
"sec_num": "3.4"
},
{
"text": "Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add \"positional encodings\" to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension d model as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9] .",
"cite_spans": [
{
"start": 517,
"end": 520,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Positional Encoding",
"sec_num": "3.5"
},
{
"text": "In this work, we use sine and cosine functions of different frequencies:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Positional Encoding",
"sec_num": "3.5"
},
{
"text": "P E (pos,2i) = sin(pos/10000 2i/dmodel ) P E (pos,2i+1) = cos(pos/10000 2i/dmodel )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Positional Encoding",
"sec_num": "3.5"
},
{
"text": "where pos is the position and i is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from 2\u03c0 to 10000 \u2022 2\u03c0. We chose this function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset k, P E pos+k can be represented as a linear function of P E pos .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Positional Encoding",
"sec_num": "3.5"
},
{
"text": "We also experimented with using learned positional embeddings [9] instead, and found that the two versions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training.",
"cite_spans": [
{
"start": 62,
"end": 65,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 152,
"end": 153,
"text": "3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Positional Encoding",
"sec_num": "3.5"
},
{
"text": "In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations (x 1 , ..., x n ) to another sequence of equal length (z 1 , ..., z n ), with x i , z i \u2208 R d , such as a hidden layer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we consider three desiderata.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "The third is the path length between long-range dependencies in the network. Learning long-range dependencies is a key challenge in many sequence transduction tasks. One key factor affecting the ability to learn such dependencies is the length of the paths forward and backward signals have to traverse in the network. The shorter these paths between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies [12] . Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.",
"cite_spans": [
{
"start": 465,
"end": 469,
"text": "[12]",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "As noted in Table 1 , a self-attention layer connects all positions with a constant number of sequentially executed operations, whereas a recurrent layer requires O(n) sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence length n is smaller than the representation dimensionality d, which is most often the case with sentence representations used by state-of-the-art models in machine translations, such as word-piece [38] and byte-pair [31] representations. To improve computational performance for tasks involving very long sequences, self-attention could be restricted to considering only a neighborhood of size r in the input sequence centered around the respective output position. This would increase the maximum path length to O(n/r). We plan to investigate this approach further in future work.",
"cite_spans": [
{
"start": 499,
"end": 503,
"text": "[38]",
"ref_id": "BIBREF37"
},
{
"start": 518,
"end": 522,
"text": "[31]",
"ref_id": "BIBREF30"
}
],
"ref_spans": [
{
"start": 18,
"end": 19,
"text": "1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "A single convolutional layer with kernel width k < n does not connect all pairs of input and output positions. Doing so requires a stack of O(n/k) convolutional layers in the case of contiguous kernels, or O(log k (n)) in the case of dilated convolutions [18] , increasing the length of the longest paths between any two positions in the network. Convolutional layers are generally more expensive than recurrent layers, by a factor of k. Separable convolutions [6] , however, decrease the complexity considerably, to O(k",
"cite_spans": [
{
"start": 255,
"end": 259,
"text": "[18]",
"ref_id": "BIBREF17"
},
{
"start": 461,
"end": 464,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "\u2022 n \u2022 d + n \u2022 d 2 ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "Even with k = n, however, the complexity of a separable convolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer, the approach we take in our model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "As side benefit, self-attention could yield more interpretable models. We inspect attention distributions from our models and present and discuss examples in the appendix. Not only do individual attention heads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic and semantic structure of the sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Why Self-Attention",
"sec_num": "4"
},
{
"text": "This section describes the training regime for our models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "5"
},
{
"text": "We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs. Sentences were encoded using byte-pair encoding [3] , which has a shared sourcetarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT 2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece vocabulary [38] . Sentence pairs were batched together by approximate sequence length. Each training batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000 target tokens.",
"cite_spans": [
{
"start": 155,
"end": 158,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 392,
"end": 396,
"text": "[38]",
"ref_id": "BIBREF37"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training Data and Batching",
"sec_num": "5.1"
},
{
"text": "We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using the hyperparameters described throughout the paper, each training step took about 0.4 seconds. We trained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the bottom line of table 3 ), step time was 1.0 seconds. The big models were trained for 300,000 steps (3.5 days).",
"cite_spans": [],
"ref_spans": [
{
"start": 310,
"end": 311,
"text": "3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Hardware and Schedule",
"sec_num": "5.2"
},
{
"text": "We used the Adam optimizer [20] with \u03b2 1 = 0.9, \u03b2 2 = 0.98 and \u03f5 = 10 -9 . We varied the learning rate over the course of training, according to the formula:",
"cite_spans": [
{
"start": 27,
"end": 31,
"text": "[20]",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Optimizer",
"sec_num": "5.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "lrate = d -0.5 model \u2022 min(step_num -0.5 , step_num \u2022 warmup_steps -1.5 )",
"eq_num": "(3)"
}
],
"section": "Optimizer",
"sec_num": "5.3"
},
{
"text": "This corresponds to increasing the learning rate linearly for the first warmup_steps training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used warmup_steps = 4000.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Optimizer",
"sec_num": "5.3"
},
{
"text": "We employ three types of regularization during training: Residual Dropout We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of P drop = 0.1.",
"cite_spans": [
{
"start": 91,
"end": 95,
"text": "[33]",
"ref_id": "BIBREF32"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Regularization",
"sec_num": "5.4"
},
{
"text": "During training, we employed label smoothing of value \u03f5 ls = 0.1 [36] . This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.",
"cite_spans": [
{
"start": 65,
"end": 69,
"text": "[36]",
"ref_id": "BIBREF35"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Label Smoothing",
"sec_num": null
},
{
"text": "6 Results",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Label Smoothing",
"sec_num": null
},
{
"text": "On the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big) in Table 2 ) outperforms the best previously reported models (including ensembles) by more than 2.0 BLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is listed in the bottom line of Table 3 . Training took 3.5 days on 8 P100 GPUs. Even our base model surpasses all previously published models and ensembles, at a fraction of the training cost of any of the competitive models.",
"cite_spans": [],
"ref_spans": [
{
"start": 106,
"end": 107,
"text": "2",
"ref_id": "TABREF1"
},
{
"start": 329,
"end": 330,
"text": "3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Machine Translation",
"sec_num": "6.1"
},
{
"text": "On the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0, outperforming all of the previously published single models, at less than 1/4 the training cost of the previous state-of-the-art model. The Transformer (big) model trained for English-to-French used dropout rate P drop = 0.1, instead of 0.3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Translation",
"sec_num": "6.1"
},
{
"text": "For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We used beam search with a beam size of 4 and length penalty \u03b1 = 0.6 [38] . These hyperparameters were chosen after experimentation on the development set. We set the maximum output length during inference to input length + 50, but terminate early when possible [38] .",
"cite_spans": [
{
"start": 259,
"end": 263,
"text": "[38]",
"ref_id": "BIBREF37"
},
{
"start": 452,
"end": 456,
"text": "[38]",
"ref_id": "BIBREF37"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Translation",
"sec_num": "6.1"
},
{
"text": "Table 2 summarizes our results and compares our translation quality and training costs to other model architectures from the literature. We estimate the number of floating point operations used to train a model by multiplying the training time, the number of GPUs used, and an estimate of the sustained single-precision floating-point capacity of each GPU5 .",
"cite_spans": [],
"ref_spans": [
{
"start": 6,
"end": 7,
"text": "2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Machine Translation",
"sec_num": "6.1"
},
{
"text": "To evaluate the importance of different components of the Transformer, we varied our base model in different ways, measuring the change in performance on English-to-German translation on the . We used beam search as described in the previous section, but no checkpoint averaging. We present these results in Table 3 .",
"cite_spans": [],
"ref_spans": [
{
"start": 314,
"end": 315,
"text": "3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Model Variations",
"sec_num": "6.2"
},
{
"text": "In Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions, keeping the amount of computation constant, as described in Section 3.2.2. While single-head attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.",
"cite_spans": [],
"ref_spans": [
{
"start": 9,
"end": 10,
"text": "3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Model Variations",
"sec_num": "6.2"
},
{
"text": "In Table 3 rows (B), we observe that reducing the attention key size d k hurts model quality. This suggests that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected, bigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our sinusoidal positional encoding with learned positional embeddings [9] , and observe nearly identical results to the base model.",
"cite_spans": [
{
"start": 470,
"end": 473,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 9,
"end": 10,
"text": "3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Model Variations",
"sec_num": "6.2"
},
{
"text": "To evaluate if the Transformer can generalize to other tasks we performed experiments on English constituency parsing. This task presents specific challenges: the output is subject to strong structural constraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence models have not been able to attain state-of-the-art results in small-data regimes [37] .",
"cite_spans": [
{
"start": 379,
"end": 383,
"text": "[37]",
"ref_id": "BIBREF36"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "English Constituency Parsing",
"sec_num": "6.3"
},
{
"text": "We trained a 4-layer transformer with d model = 1024 on the Wall Street Journal (WSJ) portion of the Penn Treebank [25] , about 40K training sentences. We also trained it in a semi-supervised setting, using the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences [37] . We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens for the semi-supervised setting.",
"cite_spans": [
{
"start": 115,
"end": 119,
"text": "[25]",
"ref_id": "BIBREF24"
},
{
"start": 298,
"end": 302,
"text": "[37]",
"ref_id": "BIBREF36"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "English Constituency Parsing",
"sec_num": "6.3"
},
{
"text": "We performed only a small number of experiments to select the dropout, both attention and residual (section 5.4), learning rates and beam size on the Section 22 development set, all other parameters remained unchanged from the English-to-German base translation model. During inference, we increased the maximum output length to input length + 300. We used a beam size of 21 and \u03b1 = 0.3 for both WSJ only and the semi-supervised setting.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "English Constituency Parsing",
"sec_num": "6.3"
},
{
"text": "Our results in Table 4 show that despite the lack of task-specific tuning our model performs surprisingly well, yielding better results than all previously reported models with the exception of the Recurrent Neural Network Grammar [8] .",
"cite_spans": [
{
"start": 231,
"end": 234,
"text": "[8]",
"ref_id": "BIBREF7"
}
],
"ref_spans": [
{
"start": 21,
"end": 22,
"text": "4",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "English Constituency Parsing",
"sec_num": "6.3"
},
{
"text": "In contrast to RNN sequence-to-sequence models [37] , the Transformer outperforms the Berkeley-Parser [29] even when training only on the WSJ training set of 40K sentences.",
"cite_spans": [
{
"start": 47,
"end": 51,
"text": "[37]",
"ref_id": "BIBREF36"
},
{
"start": 102,
"end": 106,
"text": "[29]",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "English Constituency Parsing",
"sec_num": "6.3"
},
{
"text": "In this work, we presented the Transformer, the first sequence transduction model based entirely on attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art. In the former task our best model outperforms even all previously reported ensembles.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "We are excited about the future of attention-based models and plan to apply them to other tasks. We plan to extend the Transformer to problems involving input and output modalities other than text and to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs such as images, audio and video. Making generation less sequential is another research goals of ours.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "The code we used to train and evaluate our models is available at https://github.com/ tensorflow/tensor2tensor. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "The Full attentions for head 5. Bottom: Isolated attentions from just the word 'its' for attention heads 5 and 6. Note that the attentions are very sharp for this word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input-Input Layer5",
"sec_num": null
},
{
"text": "The ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input-Input Layer5",
"sec_num": null
},
{
"text": "To illustrate why the dot products get large, assume that the components of q and k are independent random variables with mean 0 and variance 1. Then their dot product, q \u2022 k = d k i=1 qiki, has mean 0 and variance d k .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We used values of 2.8, 3.7,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Acknowledgements We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "acknowledgement",
"sec_num": null
}
],
"bib_entries": {
"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": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Massive exploration of neural machine translation architectures",
"authors": [
{
"first": "Denny",
"middle": [],
"last": "Britz",
"suffix": ""
},
{
"first": "Anna",
"middle": [],
"last": "Goldie",
"suffix": ""
},
{
"first": "Minh-Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Quoc",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Le",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural machine translation architectures. CoRR, abs/1703.03906, 2017.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Long short-term memory-networks for machine reading",
"authors": [
{
"first": "Jianpeng",
"middle": [],
"last": "Cheng",
"suffix": ""
},
{
"first": "Li",
"middle": [],
"last": "Dong",
"suffix": ""
},
{
"first": "Mirella",
"middle": [],
"last": "Lapata",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1601.06733"
]
},
"num": null,
"urls": [],
"raw_text": "Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine reading. arXiv preprint arXiv:1601.06733, 2016.",
"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": "Caglar",
"middle": [],
"last": "Gulcehre",
"suffix": ""
},
{
"first": "Fethi",
"middle": [],
"last": "Bougares",
"suffix": ""
},
{
"first": "Holger",
"middle": [],
"last": "Schwenk",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. CoRR, abs/1406.1078, 2014.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Xception: Deep learning with depthwise separable convolutions",
"authors": [
{
"first": "Francois",
"middle": [],
"last": "Chollet",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1610.02357"
]
},
"num": null,
"urls": [],
"raw_text": "Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv preprint arXiv:1610.02357, 2016.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Empirical evaluation of gated recurrent neural networks on sequence modeling",
"authors": [
{
"first": "Junyoung",
"middle": [],
"last": "Chung",
"suffix": ""
},
{
"first": "\u00c7aglar",
"middle": [],
"last": "G\u00fcl\u00e7ehre",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Junyoung Chung, \u00c7aglar G\u00fcl\u00e7ehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. CoRR, abs/1412.3555, 2014.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Recurrent neural network grammars",
"authors": [
{
"first": "Chris",
"middle": [],
"last": "Dyer",
"suffix": ""
},
{
"first": "Adhiguna",
"middle": [],
"last": "Kuncoro",
"suffix": ""
},
{
"first": "Miguel",
"middle": [],
"last": "Ballesteros",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
}
],
"year": 2016,
"venue": "Proc. of NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural network grammars. In Proc. of NAACL, 2016.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Convolutional sequence to sequence learning",
"authors": [
{
"first": "Jonas",
"middle": [],
"last": "Gehring",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Auli",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Grangier",
"suffix": ""
},
{
"first": "Denis",
"middle": [],
"last": "Yarats",
"suffix": ""
},
{
"first": "Yann",
"middle": [
"N"
],
"last": "Dauphin",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1705.03122v2"
]
},
"num": null,
"urls": [],
"raw_text": "Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolu- tional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2, 2017.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Generating sequences with recurrent neural networks",
"authors": [
{
"first": "Alex",
"middle": [],
"last": "Graves",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1308.0850"
]
},
"num": null,
"urls": [],
"raw_text": "Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Deep residual learning for image recognition",
"authors": [
{
"first": "Kaiming",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Xiangyu",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Shaoqing",
"middle": [],
"last": "Ren",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Sun",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "770--778",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im- age recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Gradient flow in recurrent nets: the difficulty of learning long-term dependencies",
"authors": [
{
"first": "Sepp",
"middle": [],
"last": "Hochreiter",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
},
{
"first": "Paolo",
"middle": [],
"last": "Frasconi",
"suffix": ""
},
{
"first": "J\u00fcrgen",
"middle": [],
"last": "Schmidhuber",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and J\u00fcrgen Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Long short-term memory",
"authors": [
{
"first": "Sepp",
"middle": [],
"last": "Hochreiter",
"suffix": ""
},
{
"first": "J\u00fcrgen",
"middle": [],
"last": "Schmidhuber",
"suffix": ""
}
],
"year": 1997,
"venue": "Neural computation",
"volume": "9",
"issue": "8",
"pages": "1735--1780",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Self-training PCFG grammars with latent annotations across languages",
"authors": [
{
"first": "Zhongqiang",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Mary",
"middle": [],
"last": "Harper",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "832--841",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations across languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 832-841. ACL, August 2009.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Exploring the limits of language modeling",
"authors": [
{
"first": "Rafal",
"middle": [],
"last": "Jozefowicz",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Schuster",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Yonghui",
"middle": [],
"last": "Wu",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1602.02410"
]
},
"num": null,
"urls": [],
"raw_text": "Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Can active memory replace attention?",
"authors": [
{
"first": "\u0141ukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "Samy",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2016,
"venue": "Advances in Neural Information Processing Systems, (NIPS)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "\u0141ukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural Information Processing Systems, (NIPS), 2016.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Neural GPUs learn algorithms",
"authors": [
{
"first": "\u0141ukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
}
],
"year": 2016,
"venue": "International Conference on Learning Representations (ICLR)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "\u0141ukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference on Learning Representations (ICLR), 2016.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Neural machine translation in linear time",
"authors": [
{
"first": "Nal",
"middle": [],
"last": "Kalchbrenner",
"suffix": ""
},
{
"first": "Lasse",
"middle": [],
"last": "Espeholt",
"suffix": ""
},
{
"first": "Karen",
"middle": [],
"last": "Simonyan",
"suffix": ""
},
{
"first": "Aaron",
"middle": [],
"last": "Van Den Oord",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Graves",
"suffix": ""
},
{
"first": "Koray",
"middle": [],
"last": "Kavukcuoglu",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1610.10099v2"
]
},
"num": null,
"urls": [],
"raw_text": "Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Ko- ray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2, 2017.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Structured attention networks",
"authors": [
{
"first": "Yoon",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Carl",
"middle": [],
"last": "Denton",
"suffix": ""
},
{
"first": "Luong",
"middle": [],
"last": "Hoang",
"suffix": ""
},
{
"first": "Alexander",
"middle": [
"M"
],
"last": "Rush",
"suffix": ""
}
],
"year": 2017,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In International Conference on Learning Representations, 2017.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Adam: A method for stochastic optimization",
"authors": [
{
"first": "Diederik",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Ba",
"suffix": ""
}
],
"year": 2015,
"venue": "ICLR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Factorization tricks for LSTM networks",
"authors": [
{
"first": "Oleksii",
"middle": [],
"last": "Kuchaiev",
"suffix": ""
},
{
"first": "Boris",
"middle": [],
"last": "Ginsburg",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1703.10722"
]
},
"num": null,
"urls": [],
"raw_text": "Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint arXiv:1703.10722, 2017.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "A structured self-attentive sentence embedding",
"authors": [
{
"first": "Zhouhan",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Minwei",
"middle": [],
"last": "Feng",
"suffix": ""
},
{
"first": "Cicero",
"middle": [],
"last": "Nogueira Dos Santos",
"suffix": ""
},
{
"first": "Mo",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "Bing",
"middle": [],
"last": "Xiang",
"suffix": ""
},
{
"first": "Bowen",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1703.03130"
]
},
"num": null,
"urls": [],
"raw_text": "Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint arXiv:1703.03130, 2017.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Multi-task sequence to sequence learning",
"authors": [
{
"first": "Minh-Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "Quoc",
"middle": [
"V"
],
"last": "Le",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "Lukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1511.06114"
]
},
"num": null,
"urls": [],
"raw_text": "Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task sequence to sequence learning. arXiv preprint arXiv:1511.06114, 2015.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Effective approaches to attentionbased 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": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1508.04025"
]
},
"num": null,
"urls": [],
"raw_text": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention- based neural machine translation. arXiv preprint arXiv:1508.04025, 2015.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Building a large annotated corpus of english: The penn treebank",
"authors": [
{
"first": "Mary",
"middle": [],
"last": "Mitchell P Marcus",
"suffix": ""
},
{
"first": "Ann",
"middle": [],
"last": "Marcinkiewicz",
"suffix": ""
},
{
"first": "Beatrice",
"middle": [],
"last": "Santorini",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational linguistics",
"volume": "19",
"issue": "2",
"pages": "313--330",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313-330, 1993.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Effective self-training for parsing",
"authors": [
{
"first": "David",
"middle": [],
"last": "Mcclosky",
"suffix": ""
},
{
"first": "Eugene",
"middle": [],
"last": "Charniak",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Johnson",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Human Language Technology Conference of the NAACL, Main Conference",
"volume": "",
"issue": "",
"pages": "152--159",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152-159. ACL, June 2006.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "A decomposable attention model",
"authors": [
{
"first": "Ankur",
"middle": [],
"last": "Parikh",
"suffix": ""
},
{
"first": "Oscar",
"middle": [],
"last": "T\u00e4ckstr\u00f6m",
"suffix": ""
},
{
"first": "Dipanjan",
"middle": [],
"last": "Das",
"suffix": ""
},
{
"first": "Jakob",
"middle": [],
"last": "Uszkoreit",
"suffix": ""
}
],
"year": 2016,
"venue": "Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ankur Parikh, Oscar T\u00e4ckstr\u00f6m, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model. In Empirical Methods in Natural Language Processing, 2016.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "A deep reinforced model for abstractive summarization",
"authors": [
{
"first": "Romain",
"middle": [],
"last": "Paulus",
"suffix": ""
},
{
"first": "Caiming",
"middle": [],
"last": "Xiong",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1705.04304"
]
},
"num": null,
"urls": [],
"raw_text": "Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. arXiv preprint arXiv:1705.04304, 2017.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Learning accurate, compact, and interpretable tree annotation",
"authors": [
{
"first": "Slav",
"middle": [],
"last": "Petrov",
"suffix": ""
},
{
"first": "Leon",
"middle": [],
"last": "Barrett",
"suffix": ""
},
{
"first": "Romain",
"middle": [],
"last": "Thibaux",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "433--440",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact, and interpretable tree annotation. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 433-440. ACL, July 2006.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Using the output embedding to improve language models",
"authors": [],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1608.05859"
]
},
"num": null,
"urls": [],
"raw_text": "Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint arXiv:1608.05859, 2016.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"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": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1508.07909"
]
},
"num": null,
"urls": [],
"raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer",
"authors": [
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Azalia",
"middle": [],
"last": "Mirhoseini",
"suffix": ""
},
{
"first": "Krzysztof",
"middle": [],
"last": "Maziarz",
"suffix": ""
},
{
"first": "Andy",
"middle": [],
"last": "Davis",
"suffix": ""
},
{
"first": "Quoc",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [],
"last": "Hinton",
"suffix": ""
},
{
"first": "Jeff",
"middle": [],
"last": "Dean",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1701.06538"
]
},
"num": null,
"urls": [],
"raw_text": "Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Dropout: a simple way to prevent neural networks from overfitting",
"authors": [
{
"first": "Nitish",
"middle": [],
"last": "Srivastava",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Krizhevsky",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
}
],
"year": 2014,
"venue": "Journal of Machine Learning Research",
"volume": "15",
"issue": "1",
"pages": "1929--1958",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi- nov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929-1958, 2014.",
"links": null
},
"BIBREF33": {
"ref_id": "b33",
"title": "End-to-end memory networks",
"authors": [
{
"first": "Sainbayar",
"middle": [],
"last": "Sukhbaatar",
"suffix": ""
},
{
"first": "Arthur",
"middle": [],
"last": "Szlam",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Weston",
"suffix": ""
},
{
"first": "Rob",
"middle": [],
"last": "Fergus",
"suffix": ""
}
],
"year": 2015,
"venue": "Advances in Neural Information Processing Systems",
"volume": "28",
"issue": "",
"pages": "2440--2448",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2440-2448. Curran Associates, Inc., 2015.",
"links": null
},
"BIBREF34": {
"ref_id": "b34",
"title": "Sequence to sequence learning with neural networks",
"authors": [
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "Quoc Vv",
"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 VV Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pages 3104-3112, 2014.",
"links": null
},
"BIBREF35": {
"ref_id": "b35",
"title": "Rethinking the inception architecture for computer vision",
"authors": [
{
"first": "Christian",
"middle": [],
"last": "Szegedy",
"suffix": ""
},
{
"first": "Vincent",
"middle": [],
"last": "Vanhoucke",
"suffix": ""
},
{
"first": "Sergey",
"middle": [],
"last": "Ioffe",
"suffix": ""
},
{
"first": "Jonathon",
"middle": [],
"last": "Shlens",
"suffix": ""
},
{
"first": "Zbigniew",
"middle": [],
"last": "Wojna",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. CoRR, abs/1512.00567, 2015.",
"links": null
},
"BIBREF36": {
"ref_id": "b36",
"title": "Grammar as a foreign language",
"authors": [
{
"first": "",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "Koo",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Petrov",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hinton",
"suffix": ""
}
],
"year": 2015,
"venue": "Advances in Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In Advances in Neural Information Processing Systems, 2015.",
"links": null
},
"BIBREF37": {
"ref_id": "b37",
"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": ""
},
{
"first": "V",
"middle": [],
"last": "Quoc",
"suffix": ""
},
{
"first": "Mohammad",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Wolfgang",
"middle": [],
"last": "Norouzi",
"suffix": ""
},
{
"first": "Maxim",
"middle": [],
"last": "Macherey",
"suffix": ""
},
{
"first": "Yuan",
"middle": [],
"last": "Krikun",
"suffix": ""
},
{
"first": "Qin",
"middle": [],
"last": "Cao",
"suffix": ""
},
{
"first": "Klaus",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Macherey",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1609.08144"
]
},
"num": null,
"urls": [],
"raw_text": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016.",
"links": null
},
"BIBREF38": {
"ref_id": "b38",
"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": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with fast-forward connections for neural machine translation. CoRR, abs/1606.04199, 2016.",
"links": null
},
"BIBREF39": {
"ref_id": "b39",
"title": "Fast and accurate shift-reduce constituent parsing",
"authors": [
{
"first": "Muhua",
"middle": [],
"last": "Zhu",
"suffix": ""
},
{
"first": "Yue",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Wenliang",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Min",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jingbo",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the 51st Annual Meeting of the ACL",
"volume": "1",
"issue": "",
"pages": "434--443",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate shift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers), pages 434-443. ACL, August 2013.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Figure 1: The Transformer -model architecture.",
"num": null,
"type_str": "figure",
"fig_num": "1",
"uris": null
},
"FIGREF1": {
"text": "Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel.",
"num": null,
"type_str": "figure",
"fig_num": "2",
"uris": null
},
"FIGREF2": {
"text": "Figure 3: An example of the attention mechanism following long-distance dependencies in the encoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of the verb 'making', completing the phrase 'making...more difficult'. Attentions here shown only for the word 'making'. Different colors represent different heads. Best viewed in color.",
"num": null,
"type_str": "figure",
"fig_num": "3",
"uris": null
},
"FIGREF3": {
"text": "Figure4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top: Full attentions for head 5. Bottom: Isolated attentions from just the word 'its' for attention heads 5 and 6. Note that the attentions are very sharp for this word.",
"num": null,
"type_str": "figure",
"fig_num": "4",
"uris": null
},
"FIGREF4": {
"text": "Figure5: Many of the attention heads exhibit behaviour that seems related to the structure of the sentence. We give two such examples above, from two different heads from the encoder self-attention at layer 5 of 6. The heads clearly learned to perform different tasks.",
"num": null,
"type_str": "figure",
"fig_num": "5",
"uris": null
},
"TABREF0": {
"text": "Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. n is the sequence length, d is the representation dimension, k is the kernel size of convolutions and r the size of the neighborhood in restricted self-attention.",
"num": null,
"type_str": "table",
"html": null,
"content": "<table><tr><td>Layer Type</td><td>Complexity per Layer Sequential Maximum Path Length</td></tr><tr><td/><td>Operations</td></tr><tr><td>Self-Attention</td><td>O</td></tr></table>"
},
"TABREF1": {
"text": "The Transformer achieves better BLEU scores than previous state-of-the-art models on the English-to-German and English-to-French newstest2014 tests at a fraction of the training cost.",
"num": null,
"type_str": "table",
"html": null,
"content": "<table><tr><td>Model</td><td colspan=\"2\">BLEU EN-DE EN-FR</td><td>Training Cost (FLOPs) EN-DE EN-FR</td></tr><tr><td>ByteNet [18]</td><td>23.75</td><td/><td/></tr><tr><td>Deep-Att + PosUnk [39]</td><td/><td>39.2</td><td>1.0 \u2022 10 20</td></tr><tr><td>GNMT + RL [38]</td><td>24.6</td><td>39.92</td><td>2.3 \u2022 10 19 1.4 \u2022 10 20</td></tr><tr><td>ConvS2S [9]</td><td>25.16</td><td>40.46</td><td>9.6 \u2022 10 18 1.5 \u2022 10 20</td></tr><tr><td>MoE [32]</td><td>26.03</td><td>40.56</td><td>2.0 \u2022 10 19 1.2 \u2022 10 20</td></tr><tr><td>Deep-Att + PosUnk Ensemble [39]</td><td/><td>40.4</td><td>8.0 \u2022 10 20</td></tr><tr><td>GNMT + RL Ensemble [38]</td><td>26.30</td><td>41.16</td><td>1.8 \u2022 10 20 1.1 \u2022 10 21</td></tr><tr><td>ConvS2S Ensemble [9]</td><td>26.36</td><td>41.29</td><td>7.7 \u2022 10 19 1.2 \u2022 10 21</td></tr><tr><td>Transformer (base model)</td><td>27.3</td><td>38.1</td><td>3.3 \u2022 10 18</td></tr><tr><td>Transformer (big)</td><td>28.4</td><td>41.8</td><td>2.3 \u2022 10 19</td></tr></table>"
},
"TABREF2": {
"text": "Variations on the Transformer architecture. Unlisted values are identical to those of the base model. All metrics are on the English-to-German translation development set, newstest2013. Listed perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to per-word perplexities.",
"num": null,
"type_str": "table",
"html": null,
"content": "<table><tr><td/><td colspan=\"2\">N d model</td><td>d ff</td><td>h</td><td>d k</td><td>d v</td><td colspan=\"2\">P drop \u03f5 ls</td><td>train steps (dev) (dev) PPL BLEU params \u00d710 6</td></tr><tr><td colspan=\"2\">base 6</td><td>512</td><td colspan=\"2\">2048 8</td><td>64</td><td>64</td><td>0.1</td><td colspan=\"2\">0.1 100K 4.92</td><td>25.8</td><td>65</td></tr><tr><td/><td/><td/><td/><td colspan=\"3\">1 512 512</td><td/><td/><td>5.29</td><td>24.9</td></tr><tr><td>(A)</td><td/><td/><td/><td colspan=\"3\">4 128 128 16 32 32</td><td/><td/><td>5.00 4.91</td><td>25.5 25.8</td></tr><tr><td/><td/><td/><td/><td colspan=\"2\">32 16</td><td>16</td><td/><td/><td>5.01</td><td>25.4</td></tr><tr><td>(B)</td><td/><td/><td/><td/><td>16 32</td><td/><td/><td/><td>5.16 5.01</td><td>25.1 25.4</td><td>58 60</td></tr><tr><td/><td>2</td><td/><td/><td/><td/><td/><td/><td/><td>6.11</td><td>23.7</td><td>36</td></tr><tr><td/><td>4</td><td/><td/><td/><td/><td/><td/><td/><td>5.19</td><td>25.3</td><td>50</td></tr><tr><td/><td>8</td><td/><td/><td/><td/><td/><td/><td/><td>4.88</td><td>25.5</td><td>80</td></tr><tr><td>(C)</td><td/><td>256</td><td/><td/><td>32</td><td>32</td><td/><td/><td>5.75</td><td>24.5</td><td>28</td></tr><tr><td/><td/><td>1024</td><td/><td/><td colspan=\"2\">128 128</td><td/><td/><td>4.66</td><td>26.0</td><td>168</td></tr><tr><td/><td/><td/><td>1024</td><td/><td/><td/><td/><td/><td>5.12</td><td>25.4</td><td>53</td></tr><tr><td/><td/><td/><td>4096</td><td/><td/><td/><td/><td/><td>4.75</td><td>26.2</td><td>90</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td>0.0</td><td/><td>5.77</td><td>24.6</td></tr><tr><td>(D)</td><td/><td/><td/><td/><td/><td/><td>0.2</td><td>0.0</td><td>4.95 4.67</td><td>25.5 25.3</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td>0.2</td><td>5.47</td><td>25.7</td></tr><tr><td>(E)</td><td/><td colspan=\"7\">positional embedding instead of sinusoids</td><td>4.92</td><td>25.7</td></tr><tr><td>big</td><td>6</td><td colspan=\"3\">1024 4096 16</td><td/><td/><td>0.3</td><td/><td>300K 4.33</td><td>26.4</td><td>213</td></tr><tr><td colspan=\"4\">development set, newstest2013</td><td/><td/><td/><td/><td/></tr></table>"
},
"TABREF3": {
"text": "The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)",
"num": null,
"type_str": "table",
"html": null,
"content": "<table><tr><td>Parser</td><td>Training</td><td>WSJ 23 F1</td></tr><tr><td colspan=\"2\">Vinyals & Kaiser el al. (2014) [37] WSJ only, discriminative</td><td>88.3</td></tr><tr><td>Petrov et al. (2006) [29]</td><td>WSJ only, discriminative</td><td>90.4</td></tr><tr><td>Zhu et al. (2013) [40]</td><td>WSJ only, discriminative</td><td>90.4</td></tr><tr><td>Dyer et al. (2016) [8]</td><td>WSJ only, discriminative</td><td>91.7</td></tr><tr><td>Transformer (4 layers)</td><td>WSJ only, discriminative</td><td>91.3</td></tr><tr><td>Zhu et al. (2013) [40]</td><td>semi-supervised</td><td>91.3</td></tr><tr><td>Huang & Harper (2009) [14]</td><td>semi-supervised</td><td>91.3</td></tr><tr><td>McClosky et al. (2006) [26]</td><td>semi-supervised</td><td>92.1</td></tr><tr><td>Vinyals & Kaiser el al. (2014) [37]</td><td>semi-supervised</td><td>92.1</td></tr><tr><td>Transformer (4 layers)</td><td>semi-supervised</td><td>92.7</td></tr><tr><td>Luong et al. (2015) [23]</td><td>multi-task</td><td>93.0</td></tr><tr><td>Dyer et al. (2016) [8]</td><td>generative</td><td>93.3</td></tr></table>"
}
}
}
} |