File size: 99,659 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 |
{
"paper_id": "I11-1014",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:30:51.702490Z"
},
"title": "Japanese Pronunciation Prediction as Phrasal Statistical Machine Translation",
"authors": [
{
"first": "Jun",
"middle": [],
"last": "Hatori",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tokyo",
"location": {
"addrLine": "7-3-1 Hongo",
"postCode": "113-0033",
"settlement": "Bunkyo",
"region": "Tokyo",
"country": "Japan"
}
},
"email": "hatori@is.s.u-tokyo.ac.jp"
},
{
"first": "Hisami",
"middle": [],
"last": "Suzuki",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Microsoft Research / One Microsoft Way",
"location": {
"postCode": "98052",
"settlement": "Redmond",
"region": "WA",
"country": "USA"
}
},
"email": "hisamis@microsoft.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper addresses the problem of predicting the pronunciation of Japanese text. The difficulty of this task lies in the high degree of ambiguity in the pronunciation of Japanese characters and words. Previous approaches have either considered the task as a word-level classification problem based on a dictionary, which does not fare well in handling out-of-vocabulary (OOV) words; or solely focused on the pronunciation prediction of OOV words without considering the contextual disambiguation of word pronunciations in text. In this paper, we propose a unified approach within the framework of phrasal statistical machine translation (SMT) that combines the strengths of the dictionary-based and substring-based approaches. Our approach is novel in that we combine wordand character-based pronunciations from a dictionary within an SMT framework: the former captures the idiosyncratic properties of word pronunciation, while the latter provides the flexibility to predict the pronunciation of OOV words. We show that based on an extensive evaluation on various test sets, our model significantly outperforms the previous state-of-the-art systems, achieving around 90% accuracy in most domains.",
"pdf_parse": {
"paper_id": "I11-1014",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper addresses the problem of predicting the pronunciation of Japanese text. The difficulty of this task lies in the high degree of ambiguity in the pronunciation of Japanese characters and words. Previous approaches have either considered the task as a word-level classification problem based on a dictionary, which does not fare well in handling out-of-vocabulary (OOV) words; or solely focused on the pronunciation prediction of OOV words without considering the contextual disambiguation of word pronunciations in text. In this paper, we propose a unified approach within the framework of phrasal statistical machine translation (SMT) that combines the strengths of the dictionary-based and substring-based approaches. Our approach is novel in that we combine wordand character-based pronunciations from a dictionary within an SMT framework: the former captures the idiosyncratic properties of word pronunciation, while the latter provides the flexibility to predict the pronunciation of OOV words. We show that based on an extensive evaluation on various test sets, our model significantly outperforms the previous state-of-the-art systems, achieving around 90% accuracy in most domains.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "This paper 1 explores the problem of assigning pronunciation to Japanese text, which consists of a mixture of ideographic and phonetic characters. The task is naturally important for the text-tospeech application (Schroeter et al., 2002) , and has been researched in that context as letter-tophoneme conversion, which converts an ortho-graphic character sequence into phonemes. In addition to speech applications, the task is also crucial for those languages such as Chinese and Japanese, where users generally type in the pronunciations of words, which are then converted into the desired character string via the software application called input methods (e.g. Gao et al. (2002a) ; Gao et al. (2002b) ).",
"cite_spans": [
{
"start": 213,
"end": 237,
"text": "(Schroeter et al., 2002)",
"ref_id": "BIBREF19"
},
{
"start": 663,
"end": 681,
"text": "Gao et al. (2002a)",
"ref_id": "BIBREF6"
},
{
"start": 684,
"end": 702,
"text": "Gao et al. (2002b)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Predicting the pronunciation of Japanese text is particularly challenging because the word and character pronunciations are highly ambiguous. Japanese orthography employs four sets of characters: hiragana and katakana (called generally as kana), which are syllabary systems and thus phonemic; kanji, which is ideographic and consists of several thousand characters; and Roman alphabet. Out of these, kanji characters typically have multiple possible pronunciations 2 ; especially those in frequent use tend to have many -between 5 and 10, sometimes as many as 20. This yields an exponential number of pronunciation possibilities when multiple kanji characters are combined in a word. Also, the pronunciation of a word is frequently idiosyncratic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This idiosyncratic property of the word pronunciation naturally motivates us to take a dictionarybased approach. Traditionally, most approaches to Japanese pronunciation prediction have regarded the problem as a word pronunciation disambiguation task. Since there are no white spaces between words in Japanese text, these approaches first segment an input sentence/phrase into words, and then select a word-level pronunciation among those defined in a dictionary (Nagano et al., 2006; Neubig and Mori, 2010) . For example, given a word \" \", these methods try to select the most appropriate pronunciation out of the three dictionary entries: ninki (popularity), hitoke (sign of life) and jinki (people's atmosphere), depending on the context. However, in these approaches, seg-mentation errors tend to result in the failure of the following step of pronunciation prediction. Moreover, since the dictionary-based approach is inapplicable to those words that are not in the dictionary, there needs to be a separate mechanism for handling out-of-vocabulary (OOV) words.",
"cite_spans": [
{
"start": 463,
"end": 484,
"text": "(Nagano et al., 2006;",
"ref_id": null
},
{
"start": 485,
"end": 507,
"text": "Neubig and Mori, 2010)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Nonetheless, the problem of OOV words has received little attention to date. Traditional systems either bypass this problem completely and assign no pronunciation to OOV words, as Mecab (Kudo et al., 2004) , a Japanese morphological analyzer, does; or use a simple model to cover them (e.g. Neubig and Mori (2010) uses a noisychannel model with a character bigram language model). Our previous work (Hatori and Suzuki, 2011) explicitly addresses the problem of predicting the pronunciation of OOV words, but focuses solely on predicting the pronunciation of nouns that are found in Wikipedia in isolation, and does not address the contextual disambiguation of pronunciation at the sentence level.",
"cite_spans": [
{
"start": 186,
"end": 205,
"text": "(Kudo et al., 2004)",
"ref_id": "BIBREF12"
},
{
"start": 291,
"end": 313,
"text": "Neubig and Mori (2010)",
"ref_id": "BIBREF16"
},
{
"start": 399,
"end": 424,
"text": "(Hatori and Suzuki, 2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we propose a unified approach based on the framework of phrasal statistical machine translation (SMT), addressing the whole sentence pronunciation assignment while integrating the OOV pronunciation prediction as part of the whole task. The novelty of our approach lies in using word and single-character pronunciations from a dictionary within the SMT framework: the former captures the idiosyncratic properties of word pronunciation, while the latter provides the flexibility to predict the pronunciation of OOV words based on the sequence of pronunciations at the substring level.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In addressing the pronunciation disambiguation problem within the framework of phrasal SMT, we extend the use of composed operations, which were applied in a limited manner in Hatori and Suzuki (2011) . Within our dictionarybased model, the composed operations are able to incorporate the composition of dictionary words (i.e. phrases) as well as substrings of the character sequence (i.e. (partial) words). In this sense, our approach is more like a standard monotone phrasal SMT, rather than the substring-based string transduction. We also propose to use the joint n-gram model as a feature function, which has been proven to be effective in the letter-tophoneme conversion task (Bisani and Ney, 2008; Jiampojamarn et al., 2010) . In the context of our current task, this feature not only incorporates smoothed contextual information for the purpose of pronunciation disambiguation, but also captures the dependency between single-kanji pronuncia-tions, which is effective for predicting the pronunciation of OOV words.",
"cite_spans": [
{
"start": 187,
"end": 200,
"text": "Suzuki (2011)",
"ref_id": "BIBREF8"
},
{
"start": 682,
"end": 704,
"text": "(Bisani and Ney, 2008;",
"ref_id": "BIBREF1"
},
{
"start": 705,
"end": 731,
"text": "Jiampojamarn et al., 2010)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We collected an extensive evaluation set for the task, including newswire articles, search query logs, person names, and Wikipedia-derived instances. Using these test sets, we show that our model significantly outperforms the previous state-of-the-art systems, achieving around 90% accuracy in most test domains, which is the best known result on the task of Japanese pronunciation prediction to date. We also give a detailed analysis of the comparison of the proposed model with an SVM-based model, KyTea (Neubig and Mori, 2010) , through which we hope to shed light on the remaining issues in solving this task.",
"cite_spans": [
{
"start": 506,
"end": 529,
"text": "(Neubig and Mori, 2010)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We define the task of pronunciation prediction as converting a string of orthographic characters representing a sentence (or a word or phrase) into a sequence of hiragana, which corresponds to how the string is pronounced. For example, given a Japanese sentence \" \" (\"I went to the Exhibition of Tanyu Kano at the Tokyo Metropolitan Art Museum.\"), the system is expected to output a sequence of hiragana, \" \", pronounced as tookyoo to bijutsukan no kanoo tanyuu ten ni itta. The task involves two sub-problems: (a) contextual disambiguation of a word pronunciation, e.g., can be pronounced either as itta \"went\" or okonatta \"did\" depending on the context; (b) pronunciation prediction of OOV words, e.g., in the above example,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background 2.1 Pronunciation Prediction: Task Setting",
"sec_num": "2"
},
{
"text": "(\"the Exhibition of Tanyu Kano\") is not likely to be in the dictionary, so the pronunciation must be reasonably guessed based on the possible pronunciations of individual characters.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background 2.1 Pronunciation Prediction: Task Setting",
"sec_num": "2"
},
{
"text": "Our research on pronunciation prediction is inspired by previous research on string transduction. The most directly relevant is the work on letter-tophoneme conversion. Previous approaches to this task include joint n-gram models (e.g., Bisani and Ney (2002) ; Chen (2003); Bisani and Ney (2008) ) and discriminatively trained substring-based models (e.g., Jiampojamarn et al. (2007) ; Jiampojamarn et al. (2008) ). This task is typically evaluated at the word level, and therefore does not include contextual disambiguation.",
"cite_spans": [
{
"start": 237,
"end": 258,
"text": "Bisani and Ney (2002)",
"ref_id": "BIBREF0"
},
{
"start": 274,
"end": 295,
"text": "Bisani and Ney (2008)",
"ref_id": "BIBREF1"
},
{
"start": 357,
"end": 383,
"text": "Jiampojamarn et al. (2007)",
"ref_id": "BIBREF9"
},
{
"start": 386,
"end": 412,
"text": "Jiampojamarn et al. (2008)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2.2"
},
{
"text": "Similar techniques to the letter-to-phoneme task have also been widely applied to the transliteration task (Knight and Graehl (1998) ). The most relevant to the current task include an approach based on substring operations in the SMT framework (e.g., , Cherry and Suzuki (2009) ), and those that use joint n-gram estimation method for the task of transliteration (e.g., Li et al. (2004) ; Jiampojamarn et al. (2010) ). However, similarly to the letter-to-phoneme task, the contextual disambiguation of the words has not received much attention. The task of Japanese pronunciation prediction itself has been a topic of investigation. Sumita and Sugaya (2006) proposed a method to use the web for assigning word pronunciation, but their focus is limited to the pronunciation disambiguation of known proper nouns. Kurata et al. (2007) and Sasada et al. (2009) discuss the methods of disambiguating new word pronunciation candidates using speech data. Nagano et al. (2006) and Mori et al. (2010b) investigated the use of the joint ngram estimation to this task.",
"cite_spans": [
{
"start": 107,
"end": 132,
"text": "(Knight and Graehl (1998)",
"ref_id": "BIBREF12"
},
{
"start": 254,
"end": 278,
"text": "Cherry and Suzuki (2009)",
"ref_id": "BIBREF3"
},
{
"start": 371,
"end": 387,
"text": "Li et al. (2004)",
"ref_id": "BIBREF14"
},
{
"start": 390,
"end": 416,
"text": "Jiampojamarn et al. (2010)",
"ref_id": "BIBREF11"
},
{
"start": 634,
"end": 658,
"text": "Sumita and Sugaya (2006)",
"ref_id": "BIBREF21"
},
{
"start": 812,
"end": 832,
"text": "Kurata et al. (2007)",
"ref_id": "BIBREF13"
},
{
"start": 837,
"end": 857,
"text": "Sasada et al. (2009)",
"ref_id": "BIBREF18"
},
{
"start": 949,
"end": 969,
"text": "Nagano et al. (2006)",
"ref_id": null
},
{
"start": 974,
"end": 993,
"text": "Mori et al. (2010b)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2.2"
},
{
"text": "More recently, Neubig and Mori (2010) proposed a classifier-based system called KyTea, which is one of the current state-of-the-art systems for the task of Japanese pronunciation prediction. As we use this system as one of our baseline systems, we describe this work in some detail here. KyTea exploits an SVM-based two-step approach, which performs a word segmentation step, followed by a pronunciation disambiguation step for each word segment. In the pronunciation prediction step, if the word in question exists in the dictionary, KyTea uses character and character-type n-grams within a window as features for the SVM classifier. For OOV words, a simple OOV model based on a noisy channel model with a character bigram language model is used. While KyTea uses the discriminative indicator features, our model instead uses character/joint n-gram language models and composed operations (to be explained in Section 3.3.2) to capture the context for the purpose of pronunciation disambiguation. The use of the indicator features essentially requires probabilistic optimization of a large number of weights, making the training less scalable than our model, which only requires frequencies of operations and phrases in the training data.",
"cite_spans": [
{
"start": 15,
"end": 37,
"text": "Neubig and Mori (2010)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2.2"
},
{
"text": "In our previous work (Hatori and Suzuki, 2011) , we addressed the pronunciation prediction of Japanese words in a semi-supervised, substringbased framework, using word-pronunciation pairs automatically extracted from Wikipedia. Though we obtained more than 70% accuracy on Wikipedia data, the model is quite specific to handling the noun phrases in Wikipedia, and it is not clear if the approach can handle the pronunciation assignment of a general text, which includes the pronunciation prediction and disambiguation of the words of all types at the sentence level.",
"cite_spans": [
{
"start": 21,
"end": 46,
"text": "(Hatori and Suzuki, 2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2.2"
},
{
"text": "Since our current work is an extension of this approach, we also adopt our previous work as one of our baseline models in Section 4.4.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2.2"
},
{
"text": "This section describes our phrasal SMT-based approach to pronunciation prediction, which is an extension of our previous work (Hatori and Suzuki, 2011) . We assume that the task of translating a Japanese orthography string to a hiragana string is basically monotone and without insertion or deletion. The overview of our model is given in Figure 1 . The components of the model will be explained below.",
"cite_spans": [
{
"start": 126,
"end": 151,
"text": "(Hatori and Suzuki, 2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 339,
"end": 347,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Pronunciation Prediction Model",
"sec_num": "3"
},
{
"text": "As is widely used in SMT research (Och, 2003) , we adopt a discriminative learning framework that uses component generative models as real-valued features (Cherry and Suzuki, 2009) . Given the source sequence s and the target character sequence t, we define real-valued features over s and",
"cite_spans": [
{
"start": 34,
"end": 45,
"text": "(Och, 2003)",
"ref_id": "BIBREF17"
},
{
"start": 155,
"end": 180,
"text": "(Cherry and Suzuki, 2009)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training and Decoding",
"sec_num": "3.1"
},
{
"text": "t, f i (s, t) for i \u2208 {1, . . . , n}.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training and Decoding",
"sec_num": "3.1"
},
{
"text": "The score of a sequence pair s, t is given by the inner product of the weight vector \u03bb = (\u03bb 1 , . . . , \u03bb n ) and the feature vector f (s, t).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training and Decoding",
"sec_num": "3.1"
},
{
"text": "For the training of model parameters, we use the averaged perceptron (Collins and Roark, 2004) : given a training corpus of transduction derivations, each of which describes a word/substring operation sequence converting s into t, the perceptron iteratively updates the weight vector every time it encounters an instance for which the model outputs a wrong sequence. For decoding, we use a stack decoder (Zens and Ney, 2004) .",
"cite_spans": [
{
"start": 69,
"end": 94,
"text": "(Collins and Roark, 2004)",
"ref_id": "BIBREF4"
},
{
"start": 404,
"end": 424,
"text": "(Zens and Ney, 2004)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training and Decoding",
"sec_num": "3.1"
},
{
"text": "For our baseline model features, we first use those from Hatori and Suzuki (2011) : the bidirectional translation probabilities, P (t|s) and P (s|t), the target character n-gram probability, P (t), the target character count, and the phrase count. In addition, we incorporate the joint n-gram probability, P (s, t), as a feature (described in Section 3.2.1). The estimation of the translation and joint/character n-gram probabilities requires a set of training corpus with source and target alignment at the word/substring level. Once these probabilities have been estimated by using the frequency of (the sequences of) operations in the training set, we only need a small tuning set to adjust the feature weights of the model. This makes online training and domain adaptation easy, and makes our model more scalable compared to fully discriminative systems with indicator features, such as KyTea.",
"cite_spans": [
{
"start": 57,
"end": 81,
"text": "Hatori and Suzuki (2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Features",
"sec_num": "3.2"
},
{
"text": "Motivated by the success in the transliteration task (Jiampojamarn et al., 2010) , we incorporate the joint n-gram language model into our SMT-based framework. The joint n-gram sequence is the sequence of operations used in the transduction: for example, when a paired sentence \" \" is decomposed into three operations \" , , \", the corresponding joint n-gram sequence is \" , , ,",
"cite_spans": [
{
"start": 53,
"end": 80,
"text": "(Jiampojamarn et al., 2010)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Joint n-gram Language Model Feature",
"sec_num": "3.2.1"
},
{
"text": "\". The effectiveness of this feature is confirmed in our experiments in Section 5.2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint n-gram Language Model Feature",
"sec_num": "3.2.1"
},
{
"text": "The corpora we use are a collection of pairs of a Japanese sentence and its hiragana sequence, as described as \"paired corpus\" in Figure 2 . These are just like bilingual corpora if we regard the hiragana sequence as monotonically translated from Japanese text. Since the original corpora do not have any word segmentation or word/substring alignments, we first need to obtain them to construct the translation table for the decoder. In previous work, KyTea used a corpus that is manually aligned using words as a unit of alignment, while Hatori and Suzuki (2011) used an unsupervised substring-based alignment. The former is not scalable easily, while the latter cannot take advantage of existing dictionaries. In this work, we use a novel application of dictionary-based phrasal decoder in order to create an aligned corpus, which allows us to use dictionary information while learning substring-based alignments for handling OOV pronunciation prediction.",
"cite_spans": [
{
"start": 539,
"end": 563,
"text": "Hatori and Suzuki (2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 130,
"end": 138,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Translation Table",
"sec_num": "3.3"
},
{
"text": "In the dictionary-based model we propose, alignments are obtained using a phrasal decoder which is based on a dictionary. This essentially treats the dictionary entries as the minimal unit of substring operations, instead of using single-kanji pronunciations estimated from training corpora as in the case of the substring-based model (Hatori and Suzuki, 2011) . We first build a simple dictionarybased decoder with only two features: the forward translation probability and the phrase count; and then use it to decode a paired corpus to obtain the alignments between the source and target strings. In this process, instances including any operation that is not defined in the dictionary are discarded; this is a major difference with the substring-based model of Hatori and Suzuki (2011) , which uses all instances of training data.",
"cite_spans": [
{
"start": 335,
"end": 360,
"text": "(Hatori and Suzuki, 2011)",
"ref_id": "BIBREF8"
},
{
"start": 764,
"end": 788,
"text": "Hatori and Suzuki (2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dictionary-based model",
"sec_num": "3.3.1"
},
{
"text": "Since Japanese dictionaries typically include single-kanji entries as well as word entries 3 , dictionary-based substring operations actually consist of both single-kanji (that is not a word per se) and word pronunciations. This is why our dictionary-based model is still able to handle OOV words. We show in Section 5 that the benefit of removing noisy training samples by this process outweighs the risk of discarding infrequent or nonstandard pronunciations that do not exist in the dictionary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dictionary-based model",
"sec_num": "3.3.1"
},
{
"text": "Our previous work (Hatori and Suzuki, 2011) exploits composed operations in order to include local contextual information in the substring-based model. Given a paired corpus, they use an aligner to obtain single-character alignments, which maps one kanji to one or more kana characters, which are then composed into larger operations. This procedure makes it possible to obtain longer alignments with limited memory, rather than using the source phrase length larger than one. In the current work, we extend the use of composed operations so that they work properly with the joint n-gram estimation.",
"cite_spans": [
{
"start": 18,
"end": 43,
"text": "(Hatori and Suzuki, 2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Composed operations",
"sec_num": "3.3.2"
},
{
"text": "The composed operations are beneficial for capturing contextual information. For example, the phrase \" \" can be pronounced in two ways: itta \"went\" and okonatta \"did\", which cannot be distinguished without any context. However, if this phrase is preceded by a hiragana particle ni \"to\", we can assume that the correct pronunciation is most likely itta, because the pronunciation ni okonatta is unusual ( okonatta is seldom preceded by ni). The composed operations are also useful in capturing the pronunciation of compound nouns: for example, due to the phonological process called rendaku (sequential voicing) (Vance, 1987) ,",
"cite_spans": [
{
"start": 611,
"end": 624,
"text": "(Vance, 1987)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Composed operations",
"sec_num": "3.3.2"
},
{
"text": "-\"plate rack\" is pronounced as shokki-dana, while the components of this word are individually pronounced as shokki (\"plate\") and tana (\"rack\"). By considering the compositions of operations, we can capture the pronunciation in the context of a compound word. Our phrasal decoder considers all (i.e. composed and non-composed) operations during the decoding, but longer (composed) operations are generally preferred when available because the phrase count feature usually receives a negative weight.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Composed operations",
"sec_num": "3.3.2"
},
{
"text": "However, the simultaneous use of these operations of different size may cause a problem when the joint n-gram estimation is applied: because composed operations include multiple noncomposed operations, they break the independence assumption of n-gram occurrences in the language model. For example, given a parallel phrase \" \" (went to an exhibition), which is decomposed into \" , , \" by dictionary-based alignments, the joint n-gram language model expects that the occurrence of \" \" (non-composed operation) is independent of that of \" --\" (composed operation), but this is not the case. To avoid this, we let the model retain the original operations even after they are composed. As shown in Figure 1 , even after the two operations \" \" and \" \" are merged into a composed operation \" --\", the joint n-gram probability is still estimated based on the original (non-composed) operations. For efficiency purposes, we only retain the decomposition of the first appearance of each composed operation even if multiple different decompositions are possible.",
"cite_spans": [],
"ref_spans": [
{
"start": 694,
"end": 702,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Composed operations",
"sec_num": "3.3.2"
},
{
"text": "In the dictionary-based framework, we need a dictionary based on which we obtain the alignments. We use a combination of three dictionaries: Uni-Dic (Den et al., 2007) , Iwanami Dictionary, and an in-house dictionary that was available to us of unknown origin. UniDic is a dictionary resource available for research purposes, which is updated on a regular basis and includes 625k word forms as of the version 1.3.12 release (July 2009). Iwanami Dictionary consists of 107k words, which expands into 325k surface forms after considering okurigana (verb inflectional ending) variants. The inhouse dictionary consists of a total of 226k words and single-kanji pronunciations. After removing duplicates, the combined dictionary consists of 770k entries. Note that these dictionaries are also used as part of training data.",
"cite_spans": [
{
"start": 149,
"end": 167,
"text": "(Den et al., 2007)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dictionary",
"sec_num": "4.1"
},
{
"text": "As described in Section 3, we need word/substring-aligned parallel corpora to train the models. We used three different sources of training data in our experiments. First, following Hatori and Suzuki (2011), we used Wikipedia: following the heuristics described in the paper, we extracted about 460k noisy word-pronunciation pairs from Japanese Wikipedia articles as of January 24, 2010. Of these pairs, we set aside 3k instances for use in development and evaluation, and used the rest for training (referred to as \"Wiki-Train\"). Secondly, since word-pronunciation pairs extracted from Wikipedia are noisy 4 and mostly consist of noun phrases, we also used a newspaper corpus, which is comprised of 1.4m sentence pairs, referred to as \"News-Train\". Finally, for the comparison with KyTea, we use a publicly available corpus, the Balanced Corpus of Contemporary Written Japanese (Maekawa (2008) ). Specifically, we use the 2009 Core Data of this corpus, which consists of 37k sentences annotated with pronunciations (referred to as \"BCCWJ\").",
"cite_spans": [],
"ref_spans": [
{
"start": 879,
"end": 894,
"text": "(Maekawa (2008)",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Training and Test Data",
"sec_num": "4.2"
},
{
"text": "Our test data consist of six datasets from various domains. Table 1 shows the statistics of these corpora, with the OOV rate estimated using KyTea 5 Table 1 : Statistics of test sets, where \"Avg. len.\" is the average length of an instance in the number of characters.",
"cite_spans": [],
"ref_spans": [
{
"start": 60,
"end": 67,
"text": "Table 1",
"ref_id": null
},
{
"start": 149,
"end": 156,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Training and Test Data",
"sec_num": "4.2"
},
{
"text": "\u2022 News-1(N1) and News-2(N2): collections of newswire articles available as Microsoft Research IME Corpus (Suzuki and Gao, 2005) . These articles are from different newspapers from the news corpus we used in training. In preparing these test sets, instances including Arabic and kanji numerals (0,1, ,9, , , , ) , or Roman alphabets are excluded 6 . \u2022 Query-1(Q1) and Query-2(Q2): query logs from a search engine (source undisclosed for blind reviewing). These sets consist of various instances ranging from general noun phrases to relatively new proper nouns. For the tuning of the weights of the model, we used 200 held-out instances for each test domain, except that the development set of Query-1 is also used for the tuning for Query-2, and the set of Wiki is used for the tuning for Name.",
"cite_spans": [
{
"start": 105,
"end": 127,
"text": "(Suzuki and Gao, 2005)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [
{
"start": 293,
"end": 311,
"text": "(0,1, ,9, , , , )",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Training and Test Data",
"sec_num": "4.2"
},
{
"text": "We use our original implementation of the phrasal aligner and decoder, which is also used as our implementation of the substring-based model of Hatori and Suzuki (2011) . An ITG-based aligner with EM algorithm is used with monotonic setting; we set the source (kanji) and target (kana) phrase length limits to 1 and 4, and prohibit alignments to a null symbol in either source or target side. The decoder runs with the beam size of 20. The maximum number of composed operations is 4 for the substringbased model of Hatori and Suzuki (2011) , and 3 for the proposed dictionary-based model. In the substring-based model, character 5-gram and joint 4-gram language models with Kneser-Ney smoothing and the BoS (beginning-of-string) and EoS (end-of-string) symbols are used; in the dictionary-based model, character 5-gram and joint 3-gram models with the same settings are used. We did not use the infrequent operation cutoff. All of these parameters and settings are set based on the preliminary experiments. As the evaluation measure, we use instance-level accuracy, which is calculated based on the percentage of the outputs that exactly match the gold standard: instances correspond to sentences in News-1/2, and to words or phrases in all other test domains. The statistical significance of the results is given using McNemar's test.",
"cite_spans": [
{
"start": 155,
"end": 168,
"text": "Suzuki (2011)",
"ref_id": "BIBREF8"
},
{
"start": 515,
"end": 539,
"text": "Hatori and Suzuki (2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental settings",
"sec_num": "4.3"
},
{
"text": "We describe three baseline models that we use as reference in our experiment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline Models",
"sec_num": "4.4"
},
{
"text": "\u2022 Mecab: Mecab version 0.98 7 , which is the state-of-the-art morphological analyzer for Japanese that also outputs pronunciations of words (Kudo et al., 2004) , with the off-the-shelf IPA Dictionary containing 392k word entries provided at the author's page. \u2022 KyTea: KyTea version 0.13 8 , which is described in Section 2.2. In our comparison experiment, we run KyTea version 0.13 both as is (using their pre-trained model), and as trained by us to allow the comparison of the framework using the same publicly available training data. \u2022 HS11: HS11 is our reimplementation of the substring-based model by Hatori and Suzuki (2011) , which was shown to outperform the substring-based joint trigram model on a Wikipedia test set. Table 2 : Instance-level accuracy (in %) of pronunciation prediction models. The upper two models use the off-the-shelf models; the lower three models are trained using the same resources: Wiki-Train, News-Train, and the combined dictionary. the system does not have a mechanism to handle OOV words. The second row shows the result of KyTea using the off-the-shelf \"full SVM model\" 9 , which is trained on several resources including BCCWJ and UniDic. It generally does better than Mecab, but the accuracies on the high OOV rate domains (i.e. Name and Wiki) are still quite low.",
"cite_spans": [
{
"start": 140,
"end": 159,
"text": "(Kudo et al., 2004)",
"ref_id": "BIBREF12"
},
{
"start": 607,
"end": 631,
"text": "Hatori and Suzuki (2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 729,
"end": 736,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Baseline Models",
"sec_num": "4.4"
},
{
"text": "The bottom three models are all trained with the same resources: Wiki-Train and News-Train with all the three dictionaries. \"HS11\" is the substring-based model proposed by Hatori and Suzuki (2011) , while \"HS11+\" is the model enhanced with two additional features: the joint ngram feature (as described in Section 3.2), and the dictionary feature, whose value is the total length (in souce characters) of words matching any dictionary entry. 10 By comparing these two models, the effectiveness of these features over the model \"HS11\" is quite clear. However, the accuracy is below 40% on newswire test sets, where each instance is a full sentence. We assume that this is because the substring-based model cannot capture the contextual information that is broad enough, and also is easily affected by noise in the training data. Our proposed model, corresponding to the last line in the table, overcomes this problem and achives the best accuracy in all but one test domain (Wiki), showing the effectiveness and robustness of the dictionary-based approach. We lags behind \"HS11+\" on Wiki, probably because the dictionary-based model discards many operations that are uncommon, but are still useful for the pronunciation of OOV words in Wikipedia. Table 3 shows the direct comparison between KyTea and the proposed model trained 11 with exactly the same datasets: BCCWJ, Wiki-Train, 9 We could not train KyTea with the same dataset as the proposed model uses due to memory limitation.",
"cite_spans": [
{
"start": 172,
"end": 196,
"text": "Hatori and Suzuki (2011)",
"ref_id": "BIBREF8"
},
{
"start": 1381,
"end": 1382,
"text": "9",
"ref_id": null
}
],
"ref_spans": [
{
"start": 1246,
"end": 1253,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Main Results",
"sec_num": "5.1"
},
{
"text": "10 The dictionary is also used as the training data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Main Results",
"sec_num": "5.1"
},
{
"text": "11 Our training of KyTea is performed as follows: we first train a segmentation model for KyTea using BCCWJ and UniDic, and use this model to segment the substring-aligned Wiki-Train instances to obtain a corpus with consistent segmentation, which is then used to train the final model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Main Results",
"sec_num": "5.1"
},
{
"text": "N1 N2 Q1 Q2 PN WP KyTea (w/noise) 68.5 65.3 88.0 79.5 67.9 65.8 KyTea (wo/noise) 75.3 75.5 91.5 83.4 61.7 64.1 Proposed 73.8 75.4 92.8 \u2020 84.9 \u2020 62.8 64.3 Table 3 : Instance-level accuracy (in %) of the models trained on Wiki-Train and BC-CWJ with UniDic. \" \u2020\" denotes a statisticallysignificant (p < 0.01) difference between \"KyTea (wo/noise)\" and \"Proposed\".",
"cite_spans": [],
"ref_spans": [
{
"start": 154,
"end": 161,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "and UniDic, all of which are from publicly available resources. Whereas \"KyTea (w/noise)\" uses all the instances for training, \"KyTea (wo/noise)\" uses only the instances that are filtered using dictionary-based operations 12 . Note that this cleaning process is also a novel contribution of our work. As is observed from Table 3 , this cleaning process resulted in a large improvement in accuracy, with the exception of the Name and Wiki sets. After inspecting the errors manually, we have found that this is because the UniDicbased operations do not include many single-kanji pronunciations that are commonly used in person's names, such as \" mi\" and \" to\". However, this problems seems negligible when a larger dictionary including common pronunciations for person's names is available. In the comparison in Table 2 , where the models use a combination of three dictionaries, the dictionary-based model \"Proposed\" performs better than the substringbased model \"HS11+\" even on the Name set. Overall, the proposed model outperforms \"KyTea (wo/noise)\" in four out of six test sets, and the differences in the remaining two sets (News-1/2) are not statistically significant. Considering also that the training data is relatively small in this comparison experiment 13 , we can conclude that our model has at least a comparable performance to KyTea for the task of pronunciation disambiguation, while achieving a superior performance on the task of pronunciation prediction for OOV words. A manual analysis of the results also showed that our model indeed has an advantage in outputting phonetically natural pronunciation sequences, partially resolving problems related to on/kun 14 and rendaku, as in keiyaku-12 27.6% of the instances in Wiki-Train is filtered out. This percentage is larger than the noise rate of 10% in this corpus, which Hatori and Suzuki (2011) reported, because the sole use of UniDic does not cover many single-kanji pronunciations, as mentioned later in this paragraph.",
"cite_spans": [
{
"start": 1839,
"end": 1863,
"text": "Hatori and Suzuki (2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 321,
"end": 328,
"text": "Table 3",
"ref_id": null
},
{
"start": 810,
"end": 817,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "13 Since the translation probabilities in our model are based on unregularized frequency, our model is less powerful with small training data, while it is more scalable.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "14 Pronunciations of kanji are classified into on and kun pronunciations (corresponding to their origin, Chinese and Model N1 N2 Q1 Q2 PN WP Proposed (D) 89.7 88.6 95.5 87.8 92.9 70.2 -wo/joint n-gram -5.5 -3.3 -1.5 -3.8 -4.4 -4.2 -wo/composed op. -3.9 -4.0 -2.6 -1.2 -1.8 -2.9 Table 4 : Feature ablation results for the dictionarybased model trained with Wiki-Train, News-Train and the combined dictionary. All the losses in accuracy were statistically significant (p < 0.01).",
"cite_spans": [
{
"start": 135,
"end": 153,
"text": "PN WP Proposed (D)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 278,
"end": 285,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "gire (individually pronounced as keiyaku and kire; \"contract expiration\"). Although KyTea wrongly output keiyaku-kire to this instance, the proposed model was able to output the correct pronunciation by learning that the pronunciation of tends to be gire after the pronunciation ku, from other instances such asku-gire (segments in haiku). On the other hand, KyTea is better at capturing generalized context by using a charactertype feature, resolving instances such as \"",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "-\" (katakana + mai; \"brand rice\"), while the proposed model wrongly output the most frequent pronunciation bei for . Table 4 shows the results of the feature ablation experiment of the proposed model. As we mentioned in Section 3.2.1, the advantage of the joint n-gram language model is twofold: incorporating smoothed context into word pronunciation disambiguation (which is the dominant problem in News-1/2), as well as incorporating singlekanji pronunciation dependencies into pronunciation prediction for OOV words (considered to be common in Name and Wiki). The improvement observed in these domains suggests that the joint n-gram probability successfully captured these two aspects. The use of composed operations showed large improvement particularly on News-1/2, proving its utility for the pronunciation disambiguation aspect of this task. Figure 3 shows the performance of the proposed model with respect to the number of News-Train sentences used for training. In this experiment, the model is first trained only with Wiki-Train; then, sentences from News-Train are incrementally added. This can be seen as a process for adapting a word-based model to a fully sentential, disambiguation-capable model. As expected, the accuracy is consistently improved in the news domain as more sentences are added, while the accuracy remains almost unchanged in the rest of the Japanese), each of which tends to be used consecutively. domains, without showing any negative effect by the additional out-of-domain training data. These results suggest that our model is robust and can adapt to new domains with a simple addition of training data.",
"cite_spans": [],
"ref_spans": [
{
"start": 117,
"end": 124,
"text": "Table 4",
"ref_id": null
},
{
"start": 849,
"end": 857,
"text": "Figure 3",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "We have presented a unified approach to the task of Japanese pronunciation prediction. Based on the framework of phrasal SMT, our model seamlessly and robustly integrates the task of word pronunciation disambiguation and pronunciation prediction for OOV words. Its basic components are trained in an unsupervised manner, and work in the presence of noise in training data. The model also has potential to adapt to a new domain when additional training data is available. We have performed an extensive evaluation on various test sets, and showed that our model achieves the new state-of-the-art accuracy on the task of Japanese pronunciation prediction. Looking into the future, we would like to see if the proposed model is effective in a general task of transliteration within a sentential context, which is conceivable as an application of phonetic input (e.g., inputting Arabic using Roman text and converting it automatically into Arabic scripts). On the task of Japanese pronunciation prediction, we are also interested in incorporating class-based features, such as character type information and on/kun dependencies, by using both existing resources and clustering methods.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "This work was conducted during the first author's internship at Microsoft Research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "In UniDic(Den et al., 2007), the average number of pronunciations per kanji character is 2.3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "This is because each kanji character is a morpheme representing a meaning, and is worth an entry in dictionaries.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We have found that roughly 10% of these instances are invalid word-pronunciation pairs.5 We ran KyTea 0.13 with the built-in default model. For",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "News-1/2, the OOV rate in the table is the OOV word rate based on the KyTea's output. For the other test sets, the figures show the rate of the instances (words or phrases) that contain any OOV word, again based on the KyTea's output 6 This is because there exist different standards in how to pronounce them. For example, the literal pronunciation is preferred for text-to-speech applications, whereas just outputting numerals as such suits better for the training of Japanese input methods.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://mecab.sourceforge.net/ 8 http://www.phontron.com/kytea/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We are grateful to Graham Neubig for providing us with detailed information on KyTea, and to anonymous reviewers for useful comments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Investigations on joint-multigram models for grapheme-tophoneme conversion",
"authors": [
{
"first": "Maximilian",
"middle": [],
"last": "Bisani",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the International Conference on Spoken Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maximilian Bisani and Hermann Ney. 2002. Investi- gations on joint-multigram models for grapheme-to- phoneme conversion. In Proceedings of the Interna- tional Conference on Spoken Language Processing.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Jointsequence models for grapheme-to-phoneme conversion",
"authors": [
{
"first": "Maximilian",
"middle": [],
"last": "Bisani",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2008,
"venue": "Speech Communication",
"volume": "50",
"issue": "",
"pages": "434--451",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maximilian Bisani and Hermann Ney. 2008. Joint- sequence models for grapheme-to-phoneme conver- sion. Speech Communication, 50:434-451.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Conditional and joint models for grapheme-to-phoneme conversion",
"authors": [
{
"first": "F",
"middle": [],
"last": "Stanley",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the European Conference on Speech Communication and Technology",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Stanley F. Chen. 2003. Conditional and joint models for grapheme-to-phoneme conversion. In Proceed- ings of the European Conference on Speech Commu- nication and Technology.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Discriminative substring decoding for transliteration",
"authors": [
{
"first": "Colin",
"middle": [],
"last": "Cherry",
"suffix": ""
},
{
"first": "Hisami",
"middle": [],
"last": "Suzuki",
"suffix": ""
}
],
"year": 2009,
"venue": "EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Colin Cherry and Hisami Suzuki. 2009. Discrim- inative substring decoding for transliteration. In EMNLP.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Incremental parsing with the perceptron algorithm",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Collins",
"suffix": ""
},
{
"first": "Brian",
"middle": [],
"last": "Roark",
"suffix": ""
}
],
"year": 2004,
"venue": "ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Collins and Brian Roark. 2004. Incremental parsing with the perceptron algorithm. In ACL.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "The development of an electronic dictionary for morphological analysis and its application to Japanese corpus linguistics",
"authors": [
{
"first": "Yasuharu",
"middle": [],
"last": "Den",
"suffix": ""
},
{
"first": "Toshinobu",
"middle": [],
"last": "Ogiso",
"suffix": ""
},
{
"first": "Hideki",
"middle": [],
"last": "Ogura",
"suffix": ""
},
{
"first": "Atsushi",
"middle": [],
"last": "Yamada",
"suffix": ""
},
{
"first": "Nobuaki",
"middle": [],
"last": "Minematsu",
"suffix": ""
}
],
"year": 2007,
"venue": "Japanese linguistics",
"volume": "22",
"issue": "",
"pages": "101--122",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yasuharu Den, Toshinobu Ogiso, Hideki Ogura, At- sushi Yamada, Nobuaki Minematsu, Kiyotaka Uchi- moto, and Hanae Koiso. 2007. The development of an electronic dictionary for morphological analysis and its application to Japanese corpus linguistics (in Japanese). Japanese linguistics, 22:101-122.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Toward a unified approach to statistical language modeling for chinese",
"authors": [
{
"first": "Jianfeng",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Mingjing",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Joshua",
"middle": [
"T"
],
"last": "Goodman",
"suffix": ""
},
{
"first": "Kai-Fu",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2002,
"venue": "ACM Transactions on Asian Language Information Processing",
"volume": "1",
"issue": "",
"pages": "3--33",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jianfeng Gao, Mingjing Li, Joshua T. Goodman, and Kai-Fu Lee. 2002a. Toward a unified approach to statistical language modeling for chinese. ACM Transactions on Asian Language Information Pro- cessing, 1:3-33.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Exploiting headword dependency and predictive clustering for language modeling",
"authors": [
{
"first": "Jianfeng",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Hisami",
"middle": [],
"last": "Suzuki",
"suffix": ""
},
{
"first": "Yang",
"middle": [],
"last": "Wen",
"suffix": ""
}
],
"year": 2002,
"venue": "EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jianfeng Gao, Hisami Suzuki, and Yang Wen. 2002b. Exploiting headword dependency and predictive clustering for language modeling. In EMNLP.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Predicting word pronunciation in Japanese",
"authors": [
{
"first": "Jun",
"middle": [],
"last": "Hatori",
"suffix": ""
},
{
"first": "Hisami",
"middle": [],
"last": "Suzuki",
"suffix": ""
}
],
"year": 2011,
"venue": "CICLing 2011, Lecture Notes in Computer Science (6609)",
"volume": "",
"issue": "",
"pages": "477--492",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jun Hatori and Hisami Suzuki. 2011. Predicting word pronunciation in Japanese. In CICLing 2011, Lec- ture Notes in Computer Science (6609), pages 477- 492. Springer.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Applying many-to-many alignments and hidden markov models to letter-to-phoneme conversion",
"authors": [
{
"first": "Grzegorz",
"middle": [],
"last": "Sittichai Jiampojamarn",
"suffix": ""
},
{
"first": "Tarek",
"middle": [],
"last": "Kondrak",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Sherif",
"suffix": ""
}
],
"year": 2007,
"venue": "HLT-NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sittichai Jiampojamarn, Grzegorz Kondrak, and Tarek Sherif. 2007. Applying many-to-many alignments and hidden markov models to letter-to-phoneme conversion. In HLT-NAACL.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Joint processing and discriminative training for letter-to-phoneme conversion",
"authors": [
{
"first": "Sittichai",
"middle": [],
"last": "Jiampojamarn",
"suffix": ""
},
{
"first": "Colin",
"middle": [],
"last": "Cherry",
"suffix": ""
},
{
"first": "Grzegorz",
"middle": [],
"last": "Kondrak",
"suffix": ""
}
],
"year": 2008,
"venue": "ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sittichai Jiampojamarn, Colin Cherry, and Grzegorz Kondrak. 2008. Joint processing and discriminative training for letter-to-phoneme conversion. In ACL.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Integrating joint n-gram features into a discriminative training framework",
"authors": [
{
"first": "Sittichai",
"middle": [],
"last": "Jiampojamarn",
"suffix": ""
},
{
"first": "Colin",
"middle": [],
"last": "Cherry",
"suffix": ""
},
{
"first": "Grzegorz",
"middle": [],
"last": "Kondrak",
"suffix": ""
}
],
"year": 2010,
"venue": "NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sittichai Jiampojamarn, Colin Cherry, and Grzegorz Kondrak. 2010. Integrating joint n-gram fea- tures into a discriminative training framework. In NAACL.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Appliying conditional random fields to Japanese morphological analysis",
"authors": [
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "Graehl ; Taku Kudo",
"suffix": ""
},
{
"first": "Kaoru",
"middle": [],
"last": "Yamamoto",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 1998,
"venue": "EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kevin Knight and Jonathan Graehl. 1998. Machine transliteration. Computational Linguistics, 24. Taku Kudo, Kaoru Yamamoto, and Yuji Matsumoto. 2004. Appliying conditional random fields to Japanese morphological analysis. In EMNLP.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Unsupervised lexicon acquisition from speech and text",
"authors": [
{
"first": "Gakuto",
"middle": [],
"last": "Kurata",
"suffix": ""
},
{
"first": "Shinsuke",
"middle": [],
"last": "Mori",
"suffix": ""
},
{
"first": "Nobuyasu",
"middle": [],
"last": "Itoh",
"suffix": ""
},
{
"first": "Masafumi",
"middle": [],
"last": "Nishimura",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of ICASSP-2007",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gakuto Kurata, Shinsuke Mori, Nobuyasu Itoh, and Masafumi Nishimura. 2007. Unsupervised lexicon acquisition from speech and text. In Proceedings of ICASSP-2007.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A joint source-channel model for machine transliteration",
"authors": [
{
"first": "Haizhou",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Min",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Su",
"suffix": ""
}
],
"year": 2004,
"venue": "ACL. Kikuo Maekawa",
"volume": "4",
"issue": "",
"pages": "82--95",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Haizhou Li, Min Zhang, and Jian Su. 2004. A joint source-channel model for machine transliteration. In ACL. Kikuo Maekawa. 2008. Compilation of the KOTONOHA-BCCWJ corpus (in Japanese). Ni- hongo no kenkyu (Studies in Japanese), 4:82-95.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "An n-gram-based approach to phoneme and accent estimation for tts",
"authors": [
{
"first": "Shinsuke",
"middle": [],
"last": "Mori",
"suffix": ""
},
{
"first": "Tetsuro",
"middle": [],
"last": "Sasada",
"suffix": ""
},
{
"first": "Graham",
"middle": [],
"last": "Neubig",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "47",
"issue": "",
"pages": "1793--1801",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shinsuke Mori, Tetsuro Sasada, and Graham Neubig. 2010b. Language model estimation from a stochas- tically tagged corpus (in Japanese). Technical Re- port, SIG, Information Processing Society of Japan. Tohru Nagano, Shinsuke Mori, and Masafumi Nishimura. 2006. An n-gram-based approach to phoneme and accent estimation for tts (in Japanese). Transactions of Information Processing Society of Japan, 47:1793-1801.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Wordbased partial annotation for efficient corpus construction",
"authors": [
{
"first": "Graham",
"middle": [],
"last": "Neubig",
"suffix": ""
},
{
"first": "Shinsuke",
"middle": [],
"last": "Mori",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC 2010)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Graham Neubig and Shinsuke Mori. 2010. Word- based partial annotation for efficient corpus con- struction. In Proceedings of the Seventh Interna- tional Conference on Language Resources and Eval- uation (LREC 2010).",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Minimum error rate training for statistical machine translation",
"authors": [
{
"first": "Franz Josef",
"middle": [],
"last": "Och",
"suffix": ""
}
],
"year": 2003,
"venue": "ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz Josef Och. 2003. Minimum error rate training for statistical machine translation. In ACL.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Domain adaptation of statistical kanakanji conversion system by automatic acquisition of contextual information with unknown words",
"authors": [
{
"first": "Tetsuro",
"middle": [],
"last": "Sasada",
"suffix": ""
},
{
"first": "Shinsuke",
"middle": [],
"last": "Mori",
"suffix": ""
},
{
"first": "Tatsuya",
"middle": [],
"last": "Kawahara",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 15th Annual Meeting of the Association for Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tetsuro Sasada, Shinsuke Mori, and Tatsuya Kawa- hara. 2009. Domain adaptation of statistical kana- kanji conversion system by automatic acquisition of contextual information with unknown words (in Japanese). In Proceedings of the 15th Annual Meet- ing of the Association for Natural Language Pro- cessing.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A perspective on the next challenges for TTS research",
"authors": [
{
"first": "Juergen",
"middle": [],
"last": "Schroeter",
"suffix": ""
},
{
"first": "Alistair",
"middle": [],
"last": "Conkie",
"suffix": ""
},
{
"first": "Ann",
"middle": [],
"last": "Syrdal",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Beutnagel",
"suffix": ""
},
{
"first": "Matthias",
"middle": [],
"last": "Jilka",
"suffix": ""
},
{
"first": "Volker",
"middle": [],
"last": "Strom",
"suffix": ""
},
{
"first": "Yeon-Jun",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Hong-Goo",
"middle": [],
"last": "Kang",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Kapilow",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the IEEE 2002 Workshop on Speech Synthesis",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Juergen Schroeter, Alistair Conkie, Ann Syrdal, Mark Beutnagel, Matthias Jilka, Volker Strom, Yeon-Jun Kim, Hong-Goo Kang, and David Kapilow. 2002. A perspective on the next challenges for TTS re- search. In Proceedings of the IEEE 2002 Workshop on Speech Synthesis.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Substringbased transliteration",
"authors": [
{
"first": "Tarek",
"middle": [],
"last": "Sherif",
"suffix": ""
},
{
"first": "Grzegorz",
"middle": [],
"last": "Kondrak",
"suffix": ""
}
],
"year": 2007,
"venue": "ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tarek Sherif and Grzegorz Kondrak. 2007. Substring- based transliteration. In ACL.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Word pronunciation disambiguation using the web",
"authors": [
{
"first": "Eiichiro",
"middle": [],
"last": "Sumita",
"suffix": ""
},
{
"first": "Fumiaki",
"middle": [],
"last": "Sugaya",
"suffix": ""
}
],
"year": 2006,
"venue": "NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eiichiro Sumita and Fumiaki Sugaya. 2006. Word pronunciation disambiguation using the web. In NAACL.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Microsoft Research IME Corpus",
"authors": [
{
"first": "Hisami",
"middle": [],
"last": "Suzuki",
"suffix": ""
},
{
"first": "Jianfeng",
"middle": [],
"last": "Gao",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hisami Suzuki and Jianfeng Gao. 2005. Microsoft Research IME Corpus. MSR Technical Report No. 2005-168.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "An Introduction to Japanese Phonology",
"authors": [
{
"first": "Timothy",
"middle": [
"J"
],
"last": "Vance",
"suffix": ""
}
],
"year": 1987,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Timothy J. Vance. 1987. An Introduction to Japanese Phonology. State University of New York Press.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Improvements in phrase-based statistical machine translation",
"authors": [
{
"first": "Richard",
"middle": [],
"last": "Zens",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2004,
"venue": "HLT-NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Richard Zens and Hermann Ney. 2004. Improvements in phrase-based statistical machine translation. In HLT-NAACL.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Bayesian learning of noncompositional phrases with synchronous parsing",
"authors": [
{
"first": "Hao",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Quirk",
"suffix": ""
},
{
"first": "Robert",
"middle": [
"C"
],
"last": "Moore",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Gildea",
"suffix": ""
}
],
"year": 2008,
"venue": "ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hao Zhang, Chris Quirk, Robert C. Moore, and Daniel Gildea. 2008. Bayesian learning of non- compositional phrases with synchronous parsing. In ACL.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "Overview of the model.",
"uris": null,
"type_str": "figure"
},
"FIGREF1": {
"num": null,
"text": "Overview of the training.",
"uris": null,
"type_str": "figure"
},
"FIGREF2": {
"num": null,
"text": "Name(PN): a collection of difficult-topronounce words, mostly consisting of person names. \u2022 Wiki(WP):manually-cleaned wordpronunciation pairs from Wikipedia, which consists mostly of proper nouns including names of people and locations as well as terms that are difficult to pronounce.",
"uris": null,
"type_str": "figure"
},
"FIGREF3": {
"num": null,
"text": "Performance (accuracy in %) of the proposed model with respect to the log of the number of additional training sentences from News-Train.",
"uris": null,
"type_str": "figure"
},
"TABREF1": {
"text": "Proposed 89.7 88.6 95.5 87.8 92.9 70.2",
"num": null,
"content": "<table><tr><td>shows the performance of the proposed model along with various baseline models. The first two lines are the result of the off-the-shelf, pre-trained systems. Mecab achieves around or above 80% accuracy on five out of six test sets, although the result on Wiki is below 60% because</td></tr></table>",
"html": null,
"type_str": "table"
}
}
}
} |