File size: 106,481 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 |
{
"paper_id": "I17-1044",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:38:16.143855Z"
},
"title": "Local Monotonic Attention Mechanism for End-to-End Speech and Language Processing",
"authors": [
{
"first": "Andros",
"middle": [],
"last": "Tjandra",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Nara Institute of Science and Technology",
"location": {
"country": "Japan"
}
},
"email": "andros.tjandra.ai6@is.naist.jp"
},
{
"first": "Sakriani",
"middle": [],
"last": "Sakti",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Nara Institute of Science and Technology",
"location": {
"country": "Japan"
}
},
"email": "ssakti@is.naist.jp"
},
{
"first": "Satoshi",
"middle": [],
"last": "Nakamura",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Nara Institute of Science and Technology",
"location": {
"country": "Japan"
}
},
"email": "s-nakamura@is.naist.jp"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Recently, encoder-decoder neural networks have shown impressive performance on many sequence-related tasks. The architecture commonly uses an attentional mechanism which allows the model to learn alignments between the source and the target sequence. Most attentional mechanisms used today is based on a global attention property which requires a computation of a weighted summarization of the whole input sequence generated by encoder states. However, it is computationally expensive and often produces misalignment on the longer input sequence. Furthermore, it does not fit with monotonous or left-to-right nature in several tasks, such as automatic speech recognition (ASR), grapheme-to-phoneme (G2P), etc. In this paper, we propose a novel attention mechanism that has local and monotonic properties. Various ways to control those properties are also explored. Experimental results on ASR, G2P and machine translation between two languages with similar sentence structures, demonstrate that the proposed encoderdecoder model with local monotonic attention could achieve significant performance improvements and reduce the computational complexity in comparison with the one that used the standard global attention architecture.",
"pdf_parse": {
"paper_id": "I17-1044",
"_pdf_hash": "",
"abstract": [
{
"text": "Recently, encoder-decoder neural networks have shown impressive performance on many sequence-related tasks. The architecture commonly uses an attentional mechanism which allows the model to learn alignments between the source and the target sequence. Most attentional mechanisms used today is based on a global attention property which requires a computation of a weighted summarization of the whole input sequence generated by encoder states. However, it is computationally expensive and often produces misalignment on the longer input sequence. Furthermore, it does not fit with monotonous or left-to-right nature in several tasks, such as automatic speech recognition (ASR), grapheme-to-phoneme (G2P), etc. In this paper, we propose a novel attention mechanism that has local and monotonic properties. Various ways to control those properties are also explored. Experimental results on ASR, G2P and machine translation between two languages with similar sentence structures, demonstrate that the proposed encoderdecoder model with local monotonic attention could achieve significant performance improvements and reduce the computational complexity in comparison with the one that used the standard global attention architecture.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "End-to-end training is a newly emerging approach to sequence-to-sequence mapping tasks, that allows the model to directly learn the mapping between variable-length representation of different modalities (i.e., text-to-text sequence Sutskever et al., 2014) , speech-totext sequence (Chorowski et al., 2014; Chan et al., 2016) , image-to-text sequence (Xu et al., 2015) , etc).",
"cite_spans": [
{
"start": 232,
"end": 255,
"text": "Sutskever et al., 2014)",
"ref_id": "BIBREF20"
},
{
"start": 281,
"end": 305,
"text": "(Chorowski et al., 2014;",
"ref_id": "BIBREF6"
},
{
"start": 306,
"end": 324,
"text": "Chan et al., 2016)",
"ref_id": "BIBREF2"
},
{
"start": 350,
"end": 367,
"text": "(Xu et al., 2015)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One popular approaches in the end-to-end mapping tasks of different modalities is based on encoder-decoder architecture. The earlier version of an encoder-decoder model is built with only two different components (Sutskever et al., 2014; Cho et al., 2014b) : (1) an encoder that processes the source sequence and encodes them into a fixedlength vector; and (2) a decoder that produces the target sequence based on information from fixedlength vector given by encoder. Both the encoder and decoder are jointly trained to maximize the probability of a correct target sequence given a source sequence. This architecture has been applied in many applications such as machine translation (Sutskever et al., 2014; Cho et al., 2014b) , image captioning (Karpathy and Fei-Fei, 2015) , and so on.",
"cite_spans": [
{
"start": 213,
"end": 237,
"text": "(Sutskever et al., 2014;",
"ref_id": "BIBREF20"
},
{
"start": 238,
"end": 256,
"text": "Cho et al., 2014b)",
"ref_id": "BIBREF5"
},
{
"start": 683,
"end": 707,
"text": "(Sutskever et al., 2014;",
"ref_id": "BIBREF20"
},
{
"start": 708,
"end": 726,
"text": "Cho et al., 2014b)",
"ref_id": "BIBREF5"
},
{
"start": 746,
"end": 774,
"text": "(Karpathy and Fei-Fei, 2015)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "However, such architecture encounters difficulties, especially for coping with long sequences. Because in order to generate the correct target sequence, the decoder solely depends only on the last hidden state of the encoder. In other words, the network needs to compress all of the information contained in the source sequence into a single fixed-length vector. (Cho et al., 2014a) demonstrated a decrease in the performance of the encoder-decoder model associated with an increase in the length of the input sentence sequence. Therefore, introduced attention mechanism to address these issues. Instead of relying on a fixed-length vector, the decoder is assisted by the attention module to get the related context from the encoder sides, depends on the current decoder states.",
"cite_spans": [
{
"start": 363,
"end": 382,
"text": "(Cho et al., 2014a)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Most attention-based encoder-decoder model used today has a \"global\" property Luong et al., 2015) . Every time the decoder needs to predict the output given the previous output, it must compute a weighted summarization of the whole input sequence generated by the encoder states. This global property allows the decoder to address any parts of the source sequence at each step of the output generation and provides advantages in some cases like machine translation tasks. Specifically, when the source and the target languages have different sentence structures and the last part of the target sequence may depend on the first part of the source sequence. However, although the global attention mechanism has often improved performance in some tasks, it is very computationally expensive. For a case that requires mapping between long sequences, misalignments might happen in standard attention mechanism (Kim et al., 2017) . Furthermore, it does not fit with monotonous or left-toright natures in several tasks, such as ASR, G2P, etc.",
"cite_spans": [
{
"start": 78,
"end": 97,
"text": "Luong et al., 2015)",
"ref_id": "BIBREF15"
},
{
"start": 905,
"end": 923,
"text": "(Kim et al., 2017)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we propose a novel attention module that has two important characteristics to address those problems: local and monotonicity properties. The local property helps our attention module focus on certain parts from the source sequence that the decoder wants to transcribe, and the monotonicity property strictly generates alignment left-to-right from beginning to the end of the source sequence. In case of speech recognition task that need to produces a transcription given the speech signal, the attention module is now able to focus on the audio's specific timing and always move in one direction from the start to the end of the audio. Similar way can be applied also for G2P or machine translation (MT) between two languages with similar sentences structure, i.e., Subject-Verb-Object (SVO) word order in English and French languages. Experimental results demonstrate that the proposed encoder-decoder model with local monotonic attention could achieve significant performance improvements and reduce the computational complexity in comparison with the one that used the standard global attention architecture.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The encoder-decoder model is a neural network that directly models conditional probability",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "Figure 1: Attention-based encoder-decoder archi- tecture. p(y|x), where x = [x 1 , ..., x S ]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "is the source sequence with length S and y = [y 1 , ..., y T ] is the target sequence with length T . Figure 1 shows the overall structure of the attention-based encoderdecoder model that consists of encoder, decoder and attention modules.",
"cite_spans": [],
"ref_spans": [
{
"start": 102,
"end": 110,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "The encoder task processes input sequence x and outputs representative information h e = [h e 1 , ..., h e S ] for the decoder. The attention module is an extension scheme for assisting the decoder to find relevant information on the encoder side based on the current decoder hidden states Luong et al., 2015) . Usually, attention modules produces context information c t at the time t based on the encoder and decoder hidden states:",
"cite_spans": [
{
"start": 290,
"end": 309,
"text": "Luong et al., 2015)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "c t = S s=1 a t (s) * h e s (1) a t (s) = Align(h e s , h d t ) = exp(Score(h e s , h d t )) S s=1 exp(Score(h e s , h d t ))",
"eq_num": "(2)"
}
],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "There are several variations for score functions:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Score(h e s , h d t ) = \uf8f1 \uf8f4 \uf8f2 \uf8f4 \uf8f3 h e s , h d t , dot product h e s W s h d t , bilinear V s tanh(W s [h e s , h d t ]), MLP",
"eq_num": "(3)"
}
],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "where Score :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "(R M \u00d7 R N ) \u2192 R, M",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "is the number of hidden units for encoder and N is the number of hidden units for decoder. Finally, the decoder task, which predicts the target sequence probability at time t based on previous output and context information c t can be formulated:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "log p(y|x) = T t=1 log p(y t |y <t , c t )",
"eq_num": "(4)"
}
],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "For speech recognition task, most common input x is a sequence of feature vectors like Mel-spectral filterbank and/or MFCC. Therefore, x \u2208 R S\u00d7D where D is the number of features and S is the total frame length for an utterance. Output y, which is a speech transcription sequence, can be either phoneme or grapheme (character) sequence. In text-related task such as machine translation, x and y are a sequence of word or character indexes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Attention-based Encoder Decoder Neural Network",
"sec_num": "2"
},
{
"text": "In the previous section, we explained the standard global attention-based encoder-decoder model. However, in order to control the area and focus attention given previous information, such mechanism requires to apply the scoring function into all the encoder states and normalizes them with a softmax function. Another problem is we cannot explicitly enforce the probability mass generated by the current attention modules that are always moving incrementally to the end of the source sequence. In this section, we discuss and explain how to model the locality and monotonicity properties on the attention module. This way, we could improve the sensitivity of capturing regularities and ensure to focus only an important subset instead of whole sequence. Figure 2 illustrates the overall mechanism of our proposed local monotonic attention, and details are described blow.",
"cite_spans": [],
"ref_spans": [
{
"start": 754,
"end": 762,
"text": "Figure 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Locality and Monotonicity Properties",
"sec_num": "3"
},
{
"text": "Position First, we define how to predict the next central position of the alignment illustrated in . At time t, we want to decode the t-th target output given the source sequence, previous output y t\u22121 , and current decoder hidden states h d t \u2208 R N . In standard approaches, we use hidden states h d t to predict the position difference \u2206p t with a multilayer perceptron (MLP). We use variable \u2206p t to determine how far we should move the center of the alignment compared to previous center p t\u22121 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "In this paper, we propose two different formulations for estimating \u2206p t to ensure a forward or monotonicity movement:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "\u2022 Constrained position prediction:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "We limit maximum range from \u2206p t with hyperparameter C max with the following equation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "\u2206p t = C max * sigmoid(V p tanh(W p h d t )) (5)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "Here we can control how far our next center of alignment position p t relies on our datasets and guarantee 0 \u2264 \u2206p t \u2264 C max . However, it requires us to handle hyperparameter C max .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "\u2022 Unconstrained position prediction:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "Compared to a previous formulation, since we do not limit the maximum range of \u2206p t , here we can ignore hyperparameter C max and use exponential (exp) function instead of sigmoid. We can also use another function (e.g softplus) as long as the function satisfy f : R \u2192 R + 0 and the result of \u2206p t \u2265 0. We formulate unconstrained position prediction with following equation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "\u2206p t = exp(V p tanh(W p h d t )) (6) Here V p \u2208 R K\u00d71 , W p \u2208 R K\u00d7N , N",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "is the number of decoder hidden units and K is the number of hidden projection layer units. We omit the bias for simplicity. Both equations guarantee monotonicity properties since \u2200t \u2208",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "[1..T ], p t \u2265 (p t\u22121 + \u2206p t ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "Additionally, we also used scaling variable \u03bb t to scale the unnormalized Gaussian distribution that depends on h t . We calculated \u03bb t with following equation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03bb t = exp(V \u03bb tanh(W p h d t ))",
"eq_num": "(7)"
}
],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "where V \u03bb \u2208 R K\u00d71 . In our initial experiments, we discovered that we improved our model performance by scaling with \u03bb t for each time-step. The main objective of this step is to generate a scaled Gaussian distribution a N t :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "a N t (s) = \u03bb t * exp \u2212 (s \u2212 p t ) 2 2\u03c3 2 .",
"eq_num": "(8)"
}
],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "where p t is the mean and \u03c3 is the standard deviation, both of which are used to calculate the weighted sum from the encoder states to generate context vector c t later. In this paper, we treat \u03c3 as a hyperparameter.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Monotonicity-based Prediction of Central",
"sec_num": "1."
},
{
"text": "After calculating new position p t , we generate locality-based alignment, as shown in Part (2) of Figure 2 . Based on predicted position p t , we follow (Luong et al., 2015) to generate alignment a S t only within [p t \u2212 2\u03c3, p t + 2\u03c3]:",
"cite_spans": [
{
"start": 154,
"end": 174,
"text": "(Luong et al., 2015)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 99,
"end": 107,
"text": "Figure 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Locality-based Alignment Generation",
"sec_num": "2."
},
{
"text": "a S t (s) = Align(h e s , h d t ), (9) \u2200s \u2208 [p t \u2212 2\u03c3, p t + 2\u03c3].",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Locality-based Alignment Generation",
"sec_num": "2."
},
{
"text": "Since p t is a real number and the indexes for the encoder states are integers, we convert p t into an integer with floor operation. After we know the center of the position p t , we only need to calculate the scores (Eq. 3) for each encoder states in [p t \u22122\u03c3, .., p t +2\u03c3] then calculate the context alignment scores (Eq. 2).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Locality-based Alignment Generation",
"sec_num": "2."
},
{
"text": "Compared to the standard global attention, we can reduce the decoding computational complexity O(T * S) into O(T * \u03c3) where \u03c3 S and \u03c3 is constant, T is total decoding step, S is the length of the encoder states.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Locality-based Alignment Generation",
"sec_num": "2."
},
{
"text": "In the last step, we calculate context c t with alignments a N t and a S t , as shown in Part (3) of Figure 2 :",
"cite_spans": [],
"ref_spans": [
{
"start": 101,
"end": 109,
"text": "Figure 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Context Calculation",
"sec_num": "3."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "c t = (pt+2\u03c3) s=(pt\u22122\u03c3) a N t (s) * a S t (s) * h e s",
"eq_num": "(10)"
}
],
"section": "Context Calculation",
"sec_num": "3."
},
{
"text": "Context c t and current hidden state h d t will later be utilized for calculating current output y t .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Calculation",
"sec_num": "3."
},
{
"text": "Overall, we can rephrase the first step as generating \"prior\" probabilities a N t based on the previous p t\u22121 position and the current decoder states. Then the second step task generates \"likelihood\" probabilities a S t by measuring the relevance of our encoder states with the current decoder states. In the third step, we combine our \"prior\" and \"likelihood\" probability into an unnormalized \"posterior\" probability a t and calculate expected context c t .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Calculation",
"sec_num": "3."
},
{
"text": "We applied our proposed architecture on ASR task. The local property helps our attention module focus on certain parts from the speech that the decoder wants to transcribe, and the monotonicity property strictly generates alignment left-to-right from beginning to the end of the speech.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment on Speech Recognition",
"sec_num": "4"
},
{
"text": "We conducted our experiments on the TIMIT 1 (Garofolo et al., 1993) dataset with the same set-up for training, development, and test sets as defined in the Kaldi s5 recipe (Povey et al., 2011) . The training set contains 3696 sentences from 462 speakers. We also used another sets of 50 speakers for the development set and the test set contains 192 utterances, 8 each from 24 speakers. For every experiment, we used 40-dimensional fbank with delta and acceleration (total 120-dimension feature vector) extracted from the Kaldi toolkit. The input features were normalized by subtracting the mean and divided by the standard deviation from the training set. For our decoder target, we re-mapped the original target phoneme set from 61 into 39 phoneme class plus the end of sequence mark (eos).",
"cite_spans": [
{
"start": 172,
"end": 192,
"text": "(Povey et al., 2011)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Speech Data",
"sec_num": "4.1"
},
{
"text": "On the encoder sides, we projected our input features with a linear layer with 512 hidden units followed by tanh activation function. We used three bidirectional LSTMs (Bi-LSTM) for our encoder with 256 hidden units for each LSTM (total 512 hidden units for Bi-LSTM). To reduce the computational time, we used hierarchical subsampling (Graves, 2012; Bahdanau et al., 2016) , applied it to the top two Bi-LSTM layers, and reduced their length by a factor of 4.",
"cite_spans": [
{
"start": 335,
"end": 349,
"text": "(Graves, 2012;",
"ref_id": "BIBREF8"
},
{
"start": 350,
"end": 372,
"text": "Bahdanau et al., 2016)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architectures",
"sec_num": "4.2"
},
{
"text": "On the decoder sides, we used a 64dimensional embedding matrix to transform the input phonemes into a continuous vector, followed by two unidirectional LSTMs with 512 hidden units. For every local monotonic model, we used an MLP with 256 hidden units to generate \u2206p t and \u03bb t . Hyperparameter 2\u03c3 was set to 3, and C max for constrained position prediction (see Eq. 5) was set to 5. Both hyperparameters were empirically selected and generally gave consistent results across various settings in our proposed model. For our scorer module, we used bilinear and MLP scorers (see Eq 3) with 256 hidden units. We used an Adam (Kingma and Ba, 2014) optimizer with a learning rate of 5e \u2212 4.",
"cite_spans": [
{
"start": 620,
"end": 641,
"text": "(Kingma and Ba, 2014)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architectures",
"sec_num": "4.2"
},
{
"text": "In the recognition phase, we generated transcriptions with best-1 (greedy) search from the decoder. We did not use any language model in this work. All of our models were implemented on the Chainer framework (Tokui et al., 2015) .",
"cite_spans": [
{
"start": 208,
"end": 228,
"text": "(Tokui et al., 2015)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architectures",
"sec_num": "4.2"
},
{
"text": "For comparison, we evaluated our proposed model with the standard global attention-based encoder-decoder model and local-m attention (Luong et al., 2015) as the baseline. Most of the con-figurations follow the above descriptions, except the baseline model that does not have an MLP for generating \u2206p t and \u03bb t . Table 1 summarizes our experiments on our proposed local attention models and compares them to the baseline model using several possible scenarios.",
"cite_spans": [],
"ref_spans": [
{
"start": 312,
"end": 319,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Model Architectures",
"sec_num": "4.2"
},
{
"text": "Considering the use of constrained and unconstrained position prediction \u2206p t , our results show that the model with the unconstrained position prediction (exp) model gives better results than one based on the constrained position prediction (sigmoid) model on both MLP and bilinear scorers. We conclude that it is more beneficial to use the unconstrained position prediction formulation since it gives better performance and we do not need to handle the additional hyperparameter C max .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constrained vs Unconstrained Position Prediction",
"sec_num": "5.1"
},
{
"text": "Next we investigate the importance of the scorer module by comparing the results between a model with and without it. Our results reveal that, by only relying on Gaussian alignment a N t and set a S t = 1, our model performance's was worse than one that used both the scorer and Gaussian alignment. This might be because the scorer modules are able to correct the details from the Gaussian alignment based on the relevance of the encoder states in the current decoder states. Thus, we conclude that alignment with the scorer is essential for our proposed models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Alignment Scorer vs Non-Scorer",
"sec_num": "5.2"
},
{
"text": "Overall, our proposed encoder-decoder model with local monotonic attention significantly improved the performance and reduced the computational complexity in comparison with one that used standard global attention mechanism (we cannot compare directly with (Chorowski et al., 2014) since its pretrained with HMM state alignment). We also tried local-m attention from (Luong et al., 2015), however our model cannot converge and we hypothesize the reason is because ratio length between the speech and their corresponding text is larger than 1, therefore the Table 1 : Results from baseline and proposed models on ASR task with TIMIT test set.",
"cite_spans": [
{
"start": 257,
"end": 281,
"text": "(Chorowski et al., 2014)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [
{
"start": 557,
"end": 564,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Overall comparison to the baseline",
"sec_num": "5.3"
},
{
"text": "Test PER (%) Global Attention Model (Baseline) Att Enc-Dec (pretrained with HMM align) (Chorowski et al., 2014) 18.6",
"cite_spans": [
{
"start": 87,
"end": 111,
"text": "(Chorowski et al., 2014)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "Att Enc-Dec (Pereyra et al., 2017) 23.2 Att Enc-Dec (Luo et al., 2016) 24.5 Att Enc-Dec with MLP Scorer (ours) 23.8 Att Enc-Dec with local-m (ours) (Luong et al., 2015 \u2206p t cannot be represented by fixed value. The best performance achieved by our proposed model with unconstrained position prediction and bilinear scorer, and provided 12.2% relative error rate reduction to our baseline.",
"cite_spans": [
{
"start": 12,
"end": 34,
"text": "(Pereyra et al., 2017)",
"ref_id": "BIBREF16"
},
{
"start": 52,
"end": 70,
"text": "(Luo et al., 2016)",
"ref_id": null
},
{
"start": 104,
"end": 110,
"text": "(ours)",
"ref_id": null
},
{
"start": 148,
"end": 167,
"text": "(Luong et al., 2015",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "We also investigated our proposed architecture on G2P conversion task. Here, the model need to generate corresponding phoneme given small segment of characters and its always moving from left to right. The local property helps our attention module focus on certain parts from the grapheme source sequence that the decoder wants to convert into phoneme, and the monotonicity property strictly generates alignment left-to-right from beginning to the end of the grapheme source sequence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment on Grapheme-to-Phoneme",
"sec_num": "6"
},
{
"text": "Here, we used the CMUDict dataset 2 . It contains 113438 words for training and 12753 for testing (12000 unique words). For validation, we randomly select 3000 sentences from the training set. The evaluation metrics for this task are phoneme error rate (PER) and word error rate (WER). In the evaluation process, there are some words has multiple references (pronunciations). Therefore, we select one of the references that has lowest PER between compared to our hypothesis, and if the hypothesis completely match with one of those references, then the WER is not increasing. For our encoder input, we used 26 letter (A-Z) + single quotes ('). For our decoder target, we used 39 phonemes plus the end of sequence mark (eos).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "6.1"
},
{
"text": "On the encoder sides, the characters input were projected into 256 dims using embedding matrix. We used two bidirectional LSTMs (Bi-LSTM) for our encoder with 512 hidden units for each LSTM (total 1024 hidden units for Bi-LSTM). On the decoder sides, the phonemes input were projected into 256 dims using embedding matrix, followed by two unidirectional LSTMs with 512 hidden units. For local monotonic model, we used an MLP with 256 hidden units to generate \u2206p t and \u03bb t . For this task, we only used the unconstrained formulation because based on previous sections, we able to achieved better performance and we didn't need to find optimal hyperparameter for C max . For our scorer module, we used MLP scorer with 256 hidden units.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architectures",
"sec_num": "6.2"
},
{
"text": "In the decoding phase, we used beam search strategy with beam size 3 to generate the phonemes given the character sequences. For comparison, we evaluated our model with standard global attention and local-m attention model (Luong et al., 2015) as the baseline. Table 2 summarizes our experiment on proposed local attention models. We compared our proposed models with several baselines from other algorithm as well. Our model significantly improving the PER and WER compared to encoderdecoder, attention-based global softmax and localm attention (fixed-step size). Compared to Bi-LSTM model which was trained with explicit alignment, we achieve slightly better PER and WER with larger window size (2\u03c3 = 3).",
"cite_spans": [
{
"start": 223,
"end": 243,
"text": "(Luong et al., 2015)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 261,
"end": 268,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Model Architectures",
"sec_num": "6.2"
},
{
"text": "We also conducted experiment on machine translation task, specifically between two languages with similar sentences structure. By using our proposed method, we able to focus only to a small related segment on the source side and the target generation process usually follows the source sentence structure without many reordering process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment on Machine Translation",
"sec_num": "7"
},
{
"text": "We used BTEC dataset (Kikui et al., 2003) and chose English-to-France and Indonesian-to-English parallel corpus. From BTEC dataset, we extracted 162318 sentences for training and 510 sentences for test data. Because there are no default development set, we randomly sampled 1000 sentences from training data for validation set. For all language pairs, we preprocessed our dataset using Moses (Koehn et al., 2007) tokenizer. For training, we replaced any word that appear less then twice with unknown (unk) symbol. In details, we keep 10105 words for French corpus, 8265 words for English corpus and 9577 words for Indonesian corpus. We only used sentence pairs where the source is no longer than 60 words in training phase.",
"cite_spans": [
{
"start": 21,
"end": 41,
"text": "(Kikui et al., 2003)",
"ref_id": "BIBREF10"
},
{
"start": 392,
"end": 412,
"text": "(Koehn et al., 2007)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "7.1"
},
{
"text": "On both encoder and decoder sides, the input words were projected into 256 dims using embedding matrix. We used three Bi-LSTM for our encoder with 512 hidden units for each LSTM (total 1024 hidden unit for Bi-LSTM). For our decoder, we used three LSTM with 512 hidden units. For local monotonic model, we used an MLP with 256 hidden units to generate \u2206p t and \u03bb t . Same as previous section, we only used the unconstrained for-mulation for local monotonic experiment. For our scorer module, we used MLP scorer with 256 hidden units. In the decoding phase, we used beam search strategy with beam size 5 and normalized length penalty with \u03b1 = 1 (Wu et al., 2016) . For comparison, we evaluate our model with standard global attention and local-m attention model (Luo et al., 2016) as the baseline. Table 3 summarizes our experiment on proposed local attention models compared to baseline global attention model and local-m attention model (Luong et al., 2015) . Generally, local monotonic attention had better result compared to global attention on both English-to-France and Indonesianto-English translation task. Our proposed model were able to improve the BLEU up to 2.2 points on English-to-France and 3.6 points on Indonesianto-English translation task compared to standard global attention. Compared to local-m attention with fixed step size, our proposed model able to improve the performance up to 0.8 BLEU on English-to-France and 2.0 BLEU on Indonesianto-English translation task.",
"cite_spans": [
{
"start": 643,
"end": 660,
"text": "(Wu et al., 2016)",
"ref_id": "BIBREF22"
},
{
"start": 760,
"end": 778,
"text": "(Luo et al., 2016)",
"ref_id": null
},
{
"start": 937,
"end": 957,
"text": "(Luong et al., 2015)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 796,
"end": 803,
"text": "Table 3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Model Architecture",
"sec_num": "7.2"
},
{
"text": "Humans do not generally process all of the information that they encounter at once. Selective attention, which is a critical property in human perception, allows attention to be focused on particular information while filtering out a range of other information. The biological structure of the eye and the eye movement mechanism is one part of visual selective attention that provides the ability to focus attention selectively on parts of the visual space to acquire information when and where it is needed (Rensink, 2000) . In the case of the cocktail party effect, humans can selectively focus their attentive hearing on a single speaker among various conversation and background noise sources (Cherry, 1953) . The attention mechanism in deep learning has been studied for many years. But, only recently have attention mechanisms made their way into the sequence-to-sequence deep learning architectures that were proposed to solve machine translation tasks. Such mechanisms provide a model with the ability to jointly align and translate . With the attention-based model, the encoder-decoder model significantly improved the performance on machine translation Luong et al., 2015) and has successfully been applied to ASR tasks (Chorowski et al., 2014; Chan et al., 2016) .",
"cite_spans": [
{
"start": 508,
"end": 523,
"text": "(Rensink, 2000)",
"ref_id": "BIBREF19"
},
{
"start": 697,
"end": 711,
"text": "(Cherry, 1953)",
"ref_id": "BIBREF3"
},
{
"start": 1163,
"end": 1182,
"text": "Luong et al., 2015)",
"ref_id": "BIBREF15"
},
{
"start": 1230,
"end": 1254,
"text": "(Chorowski et al., 2014;",
"ref_id": "BIBREF6"
},
{
"start": 1255,
"end": 1273,
"text": "Chan et al., 2016)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "8"
},
{
"text": "However, as we mentioned earlier, most of those attention mechanism are based on \"global\" property, where the attention module tries to match the current hidden states with all the states from the encoder sides. This approach is inefficient and computationally expensive on longer source sequences. A \"local attention\" was recently introduced by (Luong et al., 2015) which provided the capability to only focus small subset of the encoder sides. They also proposed monotonic attention but limited to fixed step-size and not suitable for a task where the length ratio between source and target sequence is vastly different. Our proposed method are able to elevated this problem by predicting the step size dynamically instead of using fixed step size. After we constructed our proposed framework, we found work by (Raffel et al., 2017) recently that also proposed a method for producing monotonic alignment by using Bernoulli random variable to control when the alignment should stop and generate output. However, it cannot attend the source sequence outside the range between previous and current position. In contrast with our approach, we are able to control how large the area we want to attend based on the window size. (Chorowski et al., 2014) also proposed a soft constraint to encourage monotonicity by invoking a penalty based on the current alignment and previous alignments. However, the methods still did not guarantee a monotonicity movement of the attention.",
"cite_spans": [
{
"start": 346,
"end": 366,
"text": "(Luong et al., 2015)",
"ref_id": "BIBREF15"
},
{
"start": 813,
"end": 834,
"text": "(Raffel et al., 2017)",
"ref_id": "BIBREF18"
},
{
"start": 1224,
"end": 1248,
"text": "(Chorowski et al., 2014)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "8"
},
{
"text": "To the best of our knowledge, only few studies have explored about local and monotonicity properties on an attention-based model. This work presents a novel attention module with locality and monotonicity properties. Our proposed mechanism strictly enforces monotonicity and locality properties in their alignment by explicitly modeling them in mathematical equations. The observation on our proposed model can also possibly act as regularizer by only observed a subset of encoder states. Here, we also explore various ways to control both properties and evaluate the impact of each variations on our proposed model. Experimental results also demonstrate that the proposed encoder-decoder model with local monotonic attention could provide a better performances in comparison with the standard global attention architecture and local-m attention model (Luong et al., 2015) .",
"cite_spans": [
{
"start": 852,
"end": 872,
"text": "(Luong et al., 2015)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "8"
},
{
"text": "This paper demonstrated a novel attention mechanism for encoder decoder model that ensures monotonicity and locality properties. We explored various ways to control these properties, including dynamic monotonicity-based position prediction and locality-based alignment generation. The results reveal our proposed encoder-decoder model with local monotonic attention significantly improved the performance on three different tasks and able to reduced the computational complexity more than one that used standard global attention architecture.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "9"
},
{
"text": "Part of this work was supported by JSPS KAKENHI Grant Numbers JP17H06101 and JP17K00237.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": "10"
},
{
"text": "https://catalog.ldc.upenn.edu/ldc93s1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "CMUdict: https://sourceforge.net/ projects/cmusphinx/files/G2P%20Models/ phonetisaurus-cmudict-split.tar.gz",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"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": {
"arXiv": [
"arXiv:1409.0473"
]
},
"num": null,
"urls": [],
"raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- gio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Endto-end attention-based large vocabulary speech recognition",
"authors": [
{
"first": "Dzmitry",
"middle": [],
"last": "Bahdanau",
"suffix": ""
},
{
"first": "Jan",
"middle": [],
"last": "Chorowski",
"suffix": ""
},
{
"first": "Dmitriy",
"middle": [],
"last": "Serdyuk",
"suffix": ""
}
],
"year": 2016,
"venue": "Acoustics, Speech and Signal Processing (ICASSP)",
"volume": "",
"issue": "",
"pages": "4945--4949",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dzmitry Bahdanau, Jan Chorowski, Dmitriy Serdyuk, Philemon Brakel, and Yoshua Bengio. 2016. End- to-end attention-based large vocabulary speech recognition. In Acoustics, Speech and Signal Pro- cessing (ICASSP), 2016 IEEE International Confer- ence on, pages 4945-4949. IEEE.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Listen, attend and spell: A neural network for large vocabulary conversational speech recognition",
"authors": [
{
"first": "William",
"middle": [],
"last": "Chan",
"suffix": ""
},
{
"first": "Navdeep",
"middle": [],
"last": "Jaitly",
"suffix": ""
},
{
"first": "Quoc",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
}
],
"year": 2016,
"venue": "Acoustics, Speech and Signal Processing (ICASSP)",
"volume": "",
"issue": "",
"pages": "4960--4964",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. 2016. Listen, attend and spell: A neural network for large vocabulary conversational speech recognition. In Acoustics, Speech and Signal Pro- cessing (ICASSP), 2016 IEEE International Confer- ence on, pages 4960-4964. IEEE.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Some experiments on the recognition of speech, with one and with two ears",
"authors": [
{
"first": "Cherry",
"middle": [],
"last": "E Colin",
"suffix": ""
}
],
"year": 1953,
"venue": "The Journal of the acoustical society of America",
"volume": "25",
"issue": "5",
"pages": "975--979",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "E Colin Cherry. 1953. Some experiments on the recognition of speech, with one and with two ears. The Journal of the acoustical society of America, 25(5):975-979.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "On the properties of neural machine translation: Encoder-decoder approaches. Syntax, Semantics and Structure in Statistical Translation",
"authors": [
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Bart",
"middle": [],
"last": "Van Merri\u00ebnboer",
"suffix": ""
},
{
"first": "Dzmitry",
"middle": [],
"last": "Bahdanau",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kyunghyun Cho, Bart van Merri\u00ebnboer, Dzmitry Bah- danau, and Yoshua Bengio. 2014a. On the proper- ties of neural machine translation: Encoder-decoder approaches. Syntax, Semantics and Structure in Sta- tistical Translation, page 103.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Learning phrase representations using RNN encoder-decoder for statistical machine translation",
"authors": [
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Bart",
"middle": [],
"last": "Van Merri\u00ebnboer",
"suffix": ""
},
{
"first": "Caglar",
"middle": [],
"last": "Gulcehre",
"suffix": ""
},
{
"first": "Dzmitry",
"middle": [],
"last": "Bahdanau",
"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": {
"arXiv": [
"arXiv:1406.1078"
]
},
"num": null,
"urls": [],
"raw_text": "Kyunghyun Cho, Bart Van Merri\u00ebnboer, Caglar Gul- cehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014b. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "End-to-end continuous speech recognition using attention-based recurrent NN: First results",
"authors": [
{
"first": "Jan",
"middle": [],
"last": "Chorowski",
"suffix": ""
},
{
"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": {
"arXiv": [
"arXiv:1412.1602"
]
},
"num": null,
"urls": [],
"raw_text": "Jan Chorowski, Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. End-to-end continuous speech recognition using attention-based recurrent NN: First results. arXiv preprint arXiv:1412.1602.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Darpa TIMIT acoustic-phonetic continous speech corpus cd-rom",
"authors": [
{
"first": "Lori",
"middle": [
"F"
],
"last": "John S Garofolo",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lamel",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "William",
"suffix": ""
},
{
"first": "Jonathon",
"middle": [
"G"
],
"last": "Fisher",
"suffix": ""
},
{
"first": "David",
"middle": [
"S"
],
"last": "Fiscus",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Pallett",
"suffix": ""
}
],
"year": 1993,
"venue": "NIST speech disc 1-1.1. NASA STI/Recon technical report n",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John S Garofolo, Lori F Lamel, William M Fisher, Jonathon G Fiscus, and David S Pallett. 1993. Darpa TIMIT acoustic-phonetic continous speech corpus cd-rom. NIST speech disc 1-1.1. NASA STI/Recon technical report n, 93.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Supervised sequence labelling",
"authors": [
{
"first": "Alex",
"middle": [],
"last": "Graves",
"suffix": ""
}
],
"year": 2012,
"venue": "Supervised Sequence Labelling with Recurrent Neural Networks",
"volume": "",
"issue": "",
"pages": "5--13",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alex Graves. 2012. Supervised sequence labelling. In Supervised Sequence Labelling with Recurrent Neu- ral Networks, pages 5-13. Springer.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Deep visualsemantic alignments for generating image descriptions",
"authors": [
{
"first": "Andrej",
"middle": [],
"last": "Karpathy",
"suffix": ""
},
{
"first": "Li",
"middle": [],
"last": "Fei-Fei",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "3128--3137",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrej Karpathy and Li Fei-Fei. 2015. Deep visual- semantic alignments for generating image descrip- tions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3128-3137.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Creating corpora for speech-to-speech translation",
"authors": [
{
"first": "Genichiro",
"middle": [],
"last": "Kikui",
"suffix": ""
},
{
"first": "Eiichiro",
"middle": [],
"last": "Sumita",
"suffix": ""
},
{
"first": "Toshiyuki",
"middle": [],
"last": "Takezawa",
"suffix": ""
},
{
"first": "Seiichi",
"middle": [],
"last": "Yamamoto",
"suffix": ""
}
],
"year": 2003,
"venue": "Eighth European Conference on Speech Communication and Technology",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Genichiro Kikui, Eiichiro Sumita, Toshiyuki Takezawa, and Seiichi Yamamoto. 2003. Creating corpora for speech-to-speech translation. In Eighth European Conference on Speech Communication and Technology.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Joint ctc-attention based end-to-end speech recognition using multi-task learning",
"authors": [
{
"first": "Suyoun",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Takaaki",
"middle": [],
"last": "Hori",
"suffix": ""
},
{
"first": "Shinji",
"middle": [],
"last": "Watanabe",
"suffix": ""
}
],
"year": 2017,
"venue": "Acoustics, Speech and Signal Processing",
"volume": "",
"issue": "",
"pages": "4835--4839",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Suyoun Kim, Takaaki Hori, and Shinji Watanabe. 2017. Joint ctc-attention based end-to-end speech recognition using multi-task learning. In Acous- tics, Speech and Signal Processing (ICASSP), 2017 IEEE International Conference on, pages 4835- 4839. IEEE.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Adam: A method for stochastic optimization",
"authors": [
{
"first": "Diederik",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Ba",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1412.6980"
]
},
"num": null,
"urls": [],
"raw_text": "Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Moses: Open source toolkit for statistical machine translation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "Hieu",
"middle": [],
"last": "Hoang",
"suffix": ""
},
{
"first": "Alexandra",
"middle": [],
"last": "Birch",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
},
{
"first": "Marcello",
"middle": [],
"last": "Federico",
"suffix": ""
},
{
"first": "Nicola",
"middle": [],
"last": "Bertoldi",
"suffix": ""
},
{
"first": "Brooke",
"middle": [],
"last": "Cowan",
"suffix": ""
},
{
"first": "Wade",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "Christine",
"middle": [],
"last": "Moran",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Zens",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 45th annual meeting of the ACL on interactive poster and demonstration sessions",
"volume": "",
"issue": "",
"pages": "177--180",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open source toolkit for statistical machine translation. In Pro- ceedings of the 45th annual meeting of the ACL on interactive poster and demonstration sessions, pages 177-180. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Navdeep Jaitly, and Ilya Sutskever. 2016. Learning online alignments with continuous rewards policy gradient",
"authors": [
{
"first": "Yuping",
"middle": [],
"last": "Luo",
"suffix": ""
},
{
"first": "Chung-Cheng",
"middle": [],
"last": "Chiu",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1608.01281"
]
},
"num": null,
"urls": [],
"raw_text": "Yuping Luo, Chung-Cheng Chiu, Navdeep Jaitly, and Ilya Sutskever. 2016. Learning online alignments with continuous rewards policy gradient. arXiv preprint arXiv:1608.01281.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Effective approaches to attentionbased neural machine translation",
"authors": [
{
"first": "Minh-Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "Hieu",
"middle": [],
"last": "Pham",
"suffix": ""
},
{
"first": "Christopher D",
"middle": [],
"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. 2015. Effective approaches to attention- based neural machine translation. arXiv preprint arXiv:1508.04025.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Regularizing neural networks by penalizing confident output distributions",
"authors": [
{
"first": "Gabriel",
"middle": [],
"last": "Pereyra",
"suffix": ""
},
{
"first": "George",
"middle": [],
"last": "Tucker",
"suffix": ""
},
{
"first": "Jan",
"middle": [],
"last": "Chorowski",
"suffix": ""
},
{
"first": "\u0141ukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [],
"last": "Hinton",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1701.06548"
]
},
"num": null,
"urls": [],
"raw_text": "Gabriel Pereyra, George Tucker, Jan Chorowski, \u0141ukasz Kaiser, and Geoffrey Hinton. 2017. Regular- izing neural networks by penalizing confident output distributions. arXiv preprint arXiv:1701.06548.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "The Kaldi speech recognition toolkit",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Povey",
"suffix": ""
},
{
"first": "Arnab",
"middle": [],
"last": "Ghoshal",
"suffix": ""
},
{
"first": "Gilles",
"middle": [],
"last": "Boulianne",
"suffix": ""
},
{
"first": "Lukas",
"middle": [],
"last": "Burget",
"suffix": ""
},
{
"first": "Ondrej",
"middle": [],
"last": "Glembek",
"suffix": ""
},
{
"first": "Nagendra",
"middle": [],
"last": "Goel",
"suffix": ""
},
{
"first": "Mirko",
"middle": [],
"last": "Hannemann",
"suffix": ""
},
{
"first": "Petr",
"middle": [],
"last": "Motlicek",
"suffix": ""
},
{
"first": "Yanmin",
"middle": [],
"last": "Qian",
"suffix": ""
},
{
"first": "Petr",
"middle": [],
"last": "Schwarz",
"suffix": ""
},
{
"first": "Jan",
"middle": [],
"last": "Silovsky",
"suffix": ""
},
{
"first": "Georg",
"middle": [],
"last": "Stemmer",
"suffix": ""
},
{
"first": "Karel",
"middle": [],
"last": "Vesely",
"suffix": ""
}
],
"year": 2011,
"venue": "IEEE 2011 Workshop on Automatic Speech Recognition and Understanding",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukas Burget, Ondrej Glembek, Nagendra Goel, Mirko Hannemann, Petr Motlicek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg Stemmer, and Karel Vesely. 2011. The Kaldi speech recognition toolkit. In IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. IEEE Signal Pro- cessing Society. IEEE Catalog No.: CFP11SRW- USB.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Online and linear-time attention by enforcing monotonic alignments",
"authors": [
{
"first": "Colin",
"middle": [],
"last": "Raffel",
"suffix": ""
},
{
"first": "Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Peter",
"suffix": ""
},
{
"first": "Ron",
"middle": [
"J"
],
"last": "Liu",
"suffix": ""
},
{
"first": "Douglas",
"middle": [],
"last": "Weiss",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Eck",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1704.00784"
]
},
"num": null,
"urls": [],
"raw_text": "Colin Raffel, Thang Luong, Peter J Liu, Ron J Weiss, and Douglas Eck. 2017. Online and linear-time at- tention by enforcing monotonic alignments. arXiv preprint arXiv:1704.00784.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "The dynamic representation of scenes",
"authors": [
{
"first": "A",
"middle": [],
"last": "Ronald",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Rensink",
"suffix": ""
}
],
"year": 2000,
"venue": "Visual cognition",
"volume": "7",
"issue": "1-3",
"pages": "17--42",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ronald A Rensink. 2000. The dynamic representation of scenes. Visual cognition, 7(1-3):17-42.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Sequence-to-Sequence learning with neural networks",
"authors": [
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
},
{
"first": "Oriol",
"middle": [],
"last": "Vinyals",
"suffix": ""
},
{
"first": "Quoc V",
"middle": [],
"last": "Le",
"suffix": ""
}
],
"year": 2014,
"venue": "Advances in neural information processing systems",
"volume": "",
"issue": "",
"pages": "3104--3112",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence-to-Sequence learning with neural net- works. In Advances in neural information process- ing systems, pages 3104-3112.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Chainer: a next-generation open source framework for deep learning",
"authors": [
{
"first": "Seiya",
"middle": [],
"last": "Tokui",
"suffix": ""
},
{
"first": "Kenta",
"middle": [],
"last": "Oono",
"suffix": ""
},
{
"first": "Shohei",
"middle": [],
"last": "Hido",
"suffix": ""
},
{
"first": "Justin",
"middle": [],
"last": "Clayton",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of Workshop on Machine Learning Systems (Learn-ingSys) in The Twenty-ninth Annual Conference on Neural Information Processing Systems (NIPS)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. 2015. Chainer: a next-generation open source framework for deep learning. In Proceedings of Workshop on Machine Learning Systems (Learn- ingSys) in The Twenty-ninth Annual Conference on Neural Information Processing Systems (NIPS).",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"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. 2016. Google's neural ma- chine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Show, attend and tell: Neural image caption generation with visual attention",
"authors": [
{
"first": "Kelvin",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Ba",
"suffix": ""
},
{
"first": "Ryan",
"middle": [],
"last": "Kiros",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Aaron",
"middle": [
"C"
],
"last": "Courville",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "Richard",
"middle": [
"S"
],
"last": "Zemel",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 32nd International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "2048--2057",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. 2015. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the 32nd In- ternational Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 2048- 2057.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Sequenceto-sequence neural net models for grapheme-tophoneme conversion",
"authors": [
{
"first": "Kaisheng",
"middle": [],
"last": "Yao",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [],
"last": "Zweig",
"suffix": ""
}
],
"year": 2015,
"venue": "Sixteenth Annual Conference of the International Speech Communication Association",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kaisheng Yao and Geoffrey Zweig. 2015. Sequence- to-sequence neural net models for grapheme-to- phoneme conversion. In Sixteenth Annual Confer- ence of the International Speech Communication As- sociation.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "Local monotonic attention. Part (1) of Figure 2. Assume we have source sequence with length S, which is encoded by the stack of Bi-LSTM (see Figure 1) into S encoded states h e = [h e 1 , ..., h e S ]",
"num": null,
"type_str": "figure"
},
"TABREF1": {
"text": "Results from baseline and proposed method on G2P task with CMUDict test set",
"content": "<table><tr><td>Model</td><td>PER (%)</td><td>WER (%)</td></tr><tr><td>Baseline</td><td/><td/></tr><tr><td>Enc-Dec LSTM (2 lyr) (Yao and Zweig, 2015)</td><td colspan=\"2\">7.63 28.61</td></tr><tr><td>Bi-LSTM (3 lyr) (Yao and Zweig, 2015)</td><td colspan=\"2\">5.45 23.55</td></tr><tr><td>Att Enc-Dec with Global MLP Scorer (ours)</td><td colspan=\"2\">5.96 25.55</td></tr><tr><td>Att Enc-Dec with local-m (ours) (Luong et al., 2015)</td><td colspan=\"2\">5.64 24.32</td></tr><tr><td>Proposed</td><td/><td/></tr><tr><td>Att Enc-Dec + Unconst (exp) (2\u03c3 = 2)</td><td colspan=\"2\">5.45 23.15</td></tr><tr><td>Att Enc-Dec + Unconst (exp) (2\u03c3 = 3)</td><td colspan=\"2\">5.43 23.19</td></tr></table>",
"num": null,
"type_str": "table",
"html": null
},
"TABREF2": {
"text": "",
"content": "<table><tr><td colspan=\"2\">: Results from baseline and proposed</td></tr><tr><td colspan=\"2\">method on English-to-France and Indonesian-to-</td></tr><tr><td>English translation tasks.</td><td/></tr><tr><td>Model</td><td>BLEU</td></tr><tr><td colspan=\"2\">BTEC English to France</td></tr><tr><td>Baseline</td><td/></tr><tr><td>Att Enc-Dec with Global MLP Scorer</td><td>49.0</td></tr><tr><td>Att Enc-Dec with local-m (ours) (Luong et al., 2015)</td><td>50.4</td></tr><tr><td>Proposed</td><td/></tr><tr><td>Att Enc-Dec + Unconst (exp) (2\u03c3 = 4)</td><td>51.2</td></tr><tr><td>Att Enc-Dec + Unconst (exp) (2\u03c3 = 6)</td><td>51.1</td></tr><tr><td colspan=\"2\">BTEC Indonesian to English</td></tr><tr><td>Baseline</td><td/></tr><tr><td>Att Enc-Dec with Global MLP Scorer</td><td>38.2</td></tr><tr><td>Att Enc-Dec with local-m (ours) (Luong et al., 2015)</td><td>39.8</td></tr><tr><td>Proposed</td><td/></tr><tr><td>Att Enc-Dec + Unconst (exp) (2\u03c3 = 4)</td><td>40.9</td></tr><tr><td>Att Enc-Dec + Unconst (exp) (2\u03c3 = 6)</td><td>41.8</td></tr><tr><td>7.3 Result Discussion</td><td/></tr></table>",
"num": null,
"type_str": "table",
"html": null
}
}
}
} |