File size: 87,578 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 |
{
"paper_id": "O04-3003",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:00:48.431400Z"
},
"title": "Latent Semantic Language Modeling and Smoothing",
"authors": [
{
"first": "Jen-Tzung",
"middle": [],
"last": "Chien",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Cheng Kung University",
"location": {
"settlement": "Tainan",
"country": "Taiwan, ROC"
}
},
"email": "jtchien@mail.ncku.edu.tw"
},
{
"first": "Meng-Sung",
"middle": [],
"last": "Wu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Cheng Kung University",
"location": {
"settlement": "Tainan",
"country": "Taiwan, ROC"
}
},
"email": ""
},
{
"first": "Hua-Jui",
"middle": [],
"last": "Peng",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Cheng Kung University",
"location": {
"settlement": "Tainan",
"country": "Taiwan, ROC"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Language modeling plays a critical role for automatic speech recognition. Typically, the n-gram language models suffer from the lack of a good representation of historical words and an inability to estimate unseen parameters due to insufficient training data. In this study, we explore the application of latent semantic information (LSI) to language modeling and parameter smoothing. Our approach adopts latent semantic analysis to transform all words and documents into a common semantic space. The word-to-word, word-to-document and document-to-document relations are, accordingly, exploited for language modeling and smoothing. For language modeling, we present a new representation of historical words based on retrieval of the most relevant document. We also develop a novel parameter smoothing method, where the language models of seen and unseen words are estimated by interpolating the k nearest seen words in the training corpus. The interpolation coefficients are determined according to the closeness of words in the semantic space. As shown by experiments, the proposed modeling and smoothing methods can significantly reduce the perplexity of language models with moderate computational cost.",
"pdf_parse": {
"paper_id": "O04-3003",
"_pdf_hash": "",
"abstract": [
{
"text": "Language modeling plays a critical role for automatic speech recognition. Typically, the n-gram language models suffer from the lack of a good representation of historical words and an inability to estimate unseen parameters due to insufficient training data. In this study, we explore the application of latent semantic information (LSI) to language modeling and parameter smoothing. Our approach adopts latent semantic analysis to transform all words and documents into a common semantic space. The word-to-word, word-to-document and document-to-document relations are, accordingly, exploited for language modeling and smoothing. For language modeling, we present a new representation of historical words based on retrieval of the most relevant document. We also develop a novel parameter smoothing method, where the language models of seen and unseen words are estimated by interpolating the k nearest seen words in the training corpus. The interpolation coefficients are determined according to the closeness of words in the semantic space. As shown by experiments, the proposed modeling and smoothing methods can significantly reduce the perplexity of language models with moderate computational cost.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Language models have been successfully developed for speech recognition, optical character recognition, machine translation, information retrieval, etc. Many studies in the field of speech recognition have focused on this topic [Jelinek 1990 , Jelinek 1991 . As shown in Figure 1 , a speech recognition system is composed of syllable-level and word-level matching processes, in which the acoustic model \u03bb and language model \u03c4 are applied, respectively. In theory, the speech recognition procedure combines the acoustic model and language model according to the Bayes rule. Let O denote the acoustic data, and let words. The speech recognition task aims to find the most likely word string \u0174 by maximizing the a posteriori probability given the observed acoustic data O:",
"cite_spans": [
{
"start": 228,
"end": 241,
"text": "[Jelinek 1990",
"ref_id": "BIBREF11"
},
{
"start": 242,
"end": 256,
"text": ", Jelinek 1991",
"ref_id": "BIBREF12"
}
],
"ref_spans": [
{
"start": 271,
"end": 279,
"text": "Figure 1",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": ")()(maxarg)(maxar\u011d WPWOPOWPW WW \u03c4\u03bb == ,",
"eq_num": "(1)"
}
],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "where )( WP \u03c4 is the a priori probability of the occurring word string W, and )( WOP \u03bb is the probability of observing data O given the word string W. The parameters \u03c4 and \u03bb are the language model and speech hidden Markov models (HMM's), respectively. Hereafter, we will neglect the notation \u03c4 in )( WP \u03c4 . The language model )Pr(W aims to measure the probability of word occurrence. This model is employed to predict the word occurrence given the history words. In an n-gram model, we assume that the probability of a word depends only on the preceding n-1 words. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "This probability estimation is called the maximum likelihood estimation (MLE) could not be estimated. We may apply parameter smoothing to find the unseen trigram model. In the literature, several smoothing methods have been proposed to deal with the data sparseness problem [Katz 1987 , Kawabata and Tamoto 1996 , Lau et al. 1993 , Zhai and Lafferty 2001 . Also, maximum a posteriori adaptation of the language model has been presented to resolve the problem of domain mismatch between training and test corpora [Bellegarda 2000a , Federico 1996 , Masataki et al. 1997 . Besides the problems of data sparseness and domain mismatch, the n-gram model is inferior in terms of characterizing long-distance word relationships. For example, the trigram model is unable to characterize word dependence beyond the span of three successive words. In [Lau et al. 1993, Zhou and Lua 1999 ], the trigram model was improved by extracting word relationships from the document history. This approach was exploited to search the trigger pair, In this paper, a new language modeling and smoothing method is proposed based on the framework of latent semantic analysis (LSA). The traditional n-gram model is weak in terms of characterizing the information in historical words. This weakness is compensated for herein by using the LSA framework, where word-to-word, word-to-document and document-to-document similarities are found in the semantic space. With the use of LSA, all the words are mapped to a common semantic space, which is constructed via the singular value decomposition (SVD) of a word-by-document matrix. Bellegarda [1998 Bellegarda [ , 2000a Bellegarda [ , 2000b applied the LSA framework to the n-gram model such that the resulting word error rate and perplexity were substantially reduced.",
"cite_spans": [
{
"start": 72,
"end": 77,
"text": "(MLE)",
"ref_id": null
},
{
"start": 274,
"end": 284,
"text": "[Katz 1987",
"ref_id": "BIBREF14"
},
{
"start": 285,
"end": 311,
"text": ", Kawabata and Tamoto 1996",
"ref_id": "BIBREF15"
},
{
"start": 312,
"end": 329,
"text": ", Lau et al. 1993",
"ref_id": "BIBREF16"
},
{
"start": 330,
"end": 354,
"text": ", Zhai and Lafferty 2001",
"ref_id": "BIBREF21"
},
{
"start": 512,
"end": 529,
"text": "[Bellegarda 2000a",
"ref_id": "BIBREF2"
},
{
"start": 530,
"end": 545,
"text": ", Federico 1996",
"ref_id": "BIBREF9"
},
{
"start": 546,
"end": 568,
"text": ", Masataki et al. 1997",
"ref_id": "BIBREF17"
},
{
"start": 841,
"end": 867,
"text": "[Lau et al. 1993, Zhou and",
"ref_id": null
},
{
"start": 868,
"end": 876,
"text": "Lua 1999",
"ref_id": "BIBREF22"
},
{
"start": 1602,
"end": 1618,
"text": "Bellegarda [1998",
"ref_id": "BIBREF0"
},
{
"start": 1619,
"end": 1639,
"text": "Bellegarda [ , 2000a",
"ref_id": "BIBREF2"
},
{
"start": 1640,
"end": 1660,
"text": "Bellegarda [ , 2000b",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": ".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The LSA representation of the history suffers from a drawback in that the representation of the history carries insufficient information at the beginning of a text document. To overcome this problem, we propose a relevance retrieval framework to represent the history. For language model smoothing, we estimate unseen language models by using the seen models corresponding to the k nearest neighbor words. Because this smoothing method extracts synonym and semantic information, it can be also referred to as \"semantic smoothing.\" In the following section, we briefly introduce the framework of LSA. Section 3 addresses the proposed language modeling and smoothing approaches. The LSA framework is applied to relevance feedback language modeling and k nearest neighbor language smoothing. Section 4 describes the experimental setup and reports the results for the perplexity and computational cost. Finally, we draw conclusions in Section 5.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "In the literature [Berry et al. 1995 , Deerwester et al. 1990 , Ricardo and Berthier 2000 , latent semantic analysis (LSA) has been widely applied to vector space based information retrieval. During the past few years, LSA has also been applied to language model adaptation [Bellegarda 1998 , Bellegarda 2000a , Novak and Mammone 2001 . Latent semantic analysis is a dimension reduction technique that projects the query and document into a common semantic space [Deerwester et al. 1990 , Ding 1999 . This projection reduces the document vector from a high dimensional space to a low dimensional space, which is referred as the latent semantic space.",
"cite_spans": [
{
"start": 18,
"end": 36,
"text": "[Berry et al. 1995",
"ref_id": "BIBREF4"
},
{
"start": 37,
"end": 61,
"text": ", Deerwester et al. 1990",
"ref_id": "BIBREF7"
},
{
"start": 62,
"end": 89,
"text": ", Ricardo and Berthier 2000",
"ref_id": "BIBREF19"
},
{
"start": 274,
"end": 290,
"text": "[Bellegarda 1998",
"ref_id": "BIBREF0"
},
{
"start": 291,
"end": 309,
"text": ", Bellegarda 2000a",
"ref_id": "BIBREF2"
},
{
"start": 310,
"end": 334,
"text": ", Novak and Mammone 2001",
"ref_id": "BIBREF18"
},
{
"start": 463,
"end": 486,
"text": "[Deerwester et al. 1990",
"ref_id": "BIBREF7"
},
{
"start": 487,
"end": 498,
"text": ", Ding 1999",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Latent semantic analysis",
"sec_num": "2."
},
{
"text": "The goal is to represent similar documents as close points in the latent semantic space, based on an appropriate metric. This metric can capture the significant associations between words and documents. Given an M \u00d7 N matrix A, with M terms and N documents, NM \u2265 and rank (A) =",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Latent semantic analysis",
"sec_num": "2."
},
{
"text": "The weighted count ji a , of matrix A is the number of occurrences of each word i w in a",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "document j d , calculated as follows: j ji iji n c a , , )1( \u03b5 \u2212= .",
"eq_num": "(4)"
}
],
"section": "R.",
"sec_num": null
},
{
"text": "Here, ji c , is the number of terms i w occurring in document j d , j n is the total number of words in j d , and i \u03b5 is the normalized",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "entropy of i w in the collection of data consisting of N documents, i.e., i ji N j i ji i t c t c N , 1 , log log 1 \u2211 = \u2212=\u03b5 ,",
"eq_num": "(5)"
}
],
"section": "R.",
"sec_num": null
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "\u2211 = j jii ct .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "is the total number of times i w occurs in the collection of data. A value of i \u03b5 that is close to one occurs in case of Ntc iji = , . This means that the word i w is distributed across many documents throughout the corpus. A value of i \u03b5 that is close to zero, i.e., the case in which iji tc = , , indicates that the word i w is present in only a few documents.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "Hence, in (4), i \u03b5 \u2212 1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "represents a global indexing weight for the word i w , and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "jji nc , indicates that the word i w occurs in frequently in document j d .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "Latent semantic analysis is a conceptual-indexing method, which uses singular value decomposition (SVD) [Berry et al. 1995, Golub and Van Loan 1989 ] to find the latent semantic structure of word to document association. SVD decomposes the matrix A into three sub-matrices:",
"cite_spans": [
{
"start": 104,
"end": 133,
"text": "[Berry et al. 1995, Golub and",
"ref_id": null
},
{
"start": 134,
"end": 147,
"text": "Van Loan 1989",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "T VUA \u03a3= ,",
"eq_num": "(6)"
}
],
"section": "R.",
"sec_num": null
},
{
"text": "where U and V are orthogonal matrices,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "R TT IVVUU ==",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": ", and \u03a3 is a diagonal matrix. As shown in Figure 2 , the first R columns of U and V, and the first R diagonal elements of \u03a3 can be used to approach A with R = )rank(A by means of",
"cite_spans": [],
"ref_spans": [
{
"start": 42,
"end": 50,
"text": "Figure 2",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "T RRRR VUA \u03a3= , where R",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "A is a representative matrix A. The result of SVD is a set of vectors representing the location of each term and document in the reduced R-dimensional LSA space [Berry 1992] . For a given training corpus, T AA characterizes all the co-occurrences between words, and A A T characterizes all the co-occurrences between documents. That is, a similar pattern of occurring words ji ww and can be inferred from the ),( ji cell of T AA , and a similar pattern of words contained in documents ji dd and can be inferred from the ),( ji cell of A A T [Bellegarda 1998 , Bellegarda 1997 , Bellegarda 2000a , Chen and Goodman 1999 . This LSA approach performs well when a major portion of the meaningful semantic structure [Deerwester et al. 1990 ] is captured. ww , which characterizes the linguistic regularity in a span of n words. When the window size n is limited, the n-gram is weak in terms of capturing long distance dependencies. Long distance correlation between words is commonly found in language and is caused by closeness in meaning; e.g., the words \"stock\" and \"fund\" are both likely to occur in financial news. To deal with long distance modeling, the LSA approach can be applied to extract large span semantic knowledge. Our motivation lies in the fact that there exists some latent structure in the occurrence patterns of words across documents. Hence, the n-gram language model can be improved by employing LSA to perform large span prediction of word occurrence. ",
"cite_spans": [
{
"start": 161,
"end": 173,
"text": "[Berry 1992]",
"ref_id": "BIBREF5"
},
{
"start": 541,
"end": 557,
"text": "[Bellegarda 1998",
"ref_id": "BIBREF0"
},
{
"start": 558,
"end": 575,
"text": ", Bellegarda 1997",
"ref_id": "BIBREF1"
},
{
"start": 576,
"end": 594,
"text": ", Bellegarda 2000a",
"ref_id": "BIBREF2"
},
{
"start": 595,
"end": 618,
"text": ", Chen and Goodman 1999",
"ref_id": "BIBREF6"
},
{
"start": 711,
"end": 734,
"text": "[Deerwester et al. 1990",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "A R U R \u03a3 T R V 1 w M w 1 d N d 1 u M u 1 v N v word vectors words documents document vectors 0 0 ( M x N ) ( R x R ) M x R ( R x N ) \u2245 \u00d7 \u00d7",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "where the conditioning on S reflects the fact that the probability depends on the particular vector space arising from the SVD representation, and where )Pr( [Bellegarda 1998 , Bellegarda 2000a , Bellegarda 2000b . The representation",
"cite_spans": [
{
"start": 158,
"end": 174,
"text": "[Bellegarda 1998",
"ref_id": "BIBREF0"
},
{
"start": 175,
"end": 193,
"text": ", Bellegarda 2000a",
"ref_id": "BIBREF2"
},
{
"start": 194,
"end": 212,
"text": ", Bellegarda 2000b",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "1 \u2212 q v",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "for the pseudodocument vector",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "1 \u2212 q d in the space S is given by 1 11 \u2212 \u2212\u2212 \u03a3= Udv T qq .",
"eq_num": "(8)"
}
],
"section": "R.",
"sec_num": null
},
{
"text": "By referring to (4), we can obtain the pseudodocument vector q d recursively in the LSA space via [Bellegarda 2000a , Bellegarda 2000b ",
"cite_spans": [
{
"start": 98,
"end": 115,
"text": "[Bellegarda 2000a",
"ref_id": "BIBREF2"
},
{
"start": 116,
"end": 134,
"text": ", Bellegarda 2000b",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "] T q q q q q q nn n ]00 1 00[ 1 1 LL \u03b5 \u2212 + \u2212 = \u2212 dd .",
"eq_num": "(9)"
}
],
"section": "R.",
"sec_num": null
},
{
"text": "To clarify (8) and (9), we provide their derivations in the Appendix.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "However, at the beginning of a text document, it is difficult to capture long distance word dependencies for calculating )Pr( ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "1 \u2212 q d . The LSA probability )Pr( 1 \u2212 qq w d is replaced by ) Pr( 1 \u2212 qq w d",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": ". Accordingly, the pseudodocument",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "1 \u2212 q d is estimated by Ni qiq i ,,1 ),Pr(maxar\u011d 11 L == \u2212\u2212 ddd d .",
"eq_num": "(10)"
}
],
"section": "R.",
"sec_num": null
},
{
"text": "Here,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "1 \u2212 q d",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "is obtained recursively from (9). The probability )Pr( i.e., by using the vectors \u03a3",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "i v and \u03a3 \u22121 q v in \u03a3\u03a3 \u03a3 =\u03a3\u03a3= \u2212 \u2212 \u2212\u2212 1 1 2 11 ),cos()Pr( qi T qi qiqi vv vv vvdd . (11)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "When q is increased, the most likely document vector",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "1 \u2212 q d",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "moves around in the LSA space. Assuming that",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "1 \u2212 q d",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "is semantically homogeneous, we can expect the resulting trajectory to eventually settle down in the vicinity of the document cluster corresponding to the closest semantic content.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "In this study, the LSA language model is exploited by integrating the effects of histories ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": ")( 1 )( 1 )( 1 )( 1 )( 1 )( 1 )( 1 )( 1 )( 1 )( 1 )",
"eq_num": "("
}
],
"section": "R.",
"sec_num": null
},
{
"text": "\u03a3 \u2212q v",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": ", respectively. The LSA probability is calculated as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "R.",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "2 1 2 1 2 1 2 1 1 1 11\u02c6 \u02c6)\u02c6, cos()\u02c6Pr( \u03a3\u03a3 \u03a3 =\u03a3\u03a3= \u2212 \u2212 \u2212\u2212 qq T qq qqqq w vu vu vud .",
"eq_num": "(13)"
}
],
"section": "R.",
"sec_num": null
},
{
"text": "In the real world, a training corpus is not sufficient to estimate the n-gram model for all word",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "occurrences },,,{ 11 qqnq www \u2212+\u2212 L .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "To overcome the problem of insufficient data, the parameter smoothing method can be used to estimate the joint probabilities of unseen word occurrences and, simultaneously, smooth those of seen word occurrences in the training corpus. It is common to interpolate the n-gram and (n-1)-gram for the purpose of language model smoothing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "Jelinek-Mercer smoothing [Jelinek and Mercer 1980] is represented as follows:",
"cite_spans": [
{
"start": 25,
"end": 50,
"text": "[Jelinek and Mercer 1980]",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": ")(Pr)1()Pr()(Pr 1 2 1 1 1 1 \u2212 +\u2212 \u2212 +\u2212 \u2212 +\u2212 \u2212+= q nqqq q nqqq q nqq wwwwww JMJM \u03bb\u03bb .",
"eq_num": "(14)"
}
],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "The smoothed n-gram model )( ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "{ } 0)(: 1 1 >= \u2212 +\u2212 q q nq qq wwcwN .",
"eq_num": "(15)"
}
],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "The well-known Witten-Bell smoothing approach [Written and Bell 1991] incorporates the interpolation coefficient",
"cite_spans": [
{
"start": 46,
"end": 69,
"text": "[Written and Bell 1991]",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u2211 +\u2212 + =\u2212 q w q nq q q q wcN N )( 1 1 \u03bb ,",
"eq_num": "(16)"
}
],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "into 14 ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "In this paper, we will present a novel smoothing method in which the language models of seen and unseen word occurrences are estimated by interpolating the LSA language model of a word occurrence and of the k nearest word occurrences. Let us consider the words \"car,\" \"automobile,\" \"driver,\" and \"elephant\". \"Car\" and \"automobile\" are synonyms. \"Driver\" is related and \"elephant\" is unrelated to \"car\" and \"automobile.\" If the words \"car\" and \"automobile\" do not appear in the given documents, we may collect many documents containing related words, e.g., the motor, vehicle, engine, etc. The statistics of these nearest seen words can be used to estimate the language model of the unseen words. When the bigram model is used, the smoothed model )r(P~1 ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "The interpolation is performed as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u2211 = \u2212\u2212 \u2212+= k j q jq q jqqqqqq wwwwww 1 11 )\u02c6Pr()1()Pr()r(P~\u03b2\u03b1\u03b1 ,",
"eq_num": "(19)"
}
],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "where the weighting coefficients }1 ,{ kj ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "respectively. As seen in (20), the weighting coefficient ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LSA Parameter Smoothing",
"sec_num": "3.2"
},
{
"text": "We evaluated the performance of the proposed language model through experiments. Two databases were employed. The first database was the CKIP balanced corpus of Modern Chinese (http://godel.iis.sinica.edu.tw), which was collected by Academia Sinica in Taiwan, ROC.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "Totally, this database has twenty-five million Chinese characters and a vocabulary size of 80,000 words. In addition, we collected 9,372 news documents during 2001 and 2002 from the news websites of CNA (http://www.cna.com.tw), ChinaTimes (http://news.chinatimes. com) and UDNnews ( http://www.udnnews.com.tw). We randomly sampled 9,148 documents for training and the remaining 224 documents for testing. The news documents were divided into eight categories, including technology, society, international, leisure, politics, finance, entertainment, and sports news. The numbers of training and testing documents in the eight news categories are listed in Table 1 . We chose the most frequent 32,941 words to construct our dictionary. Using the LSA procedure, we built a 32,941*9,148 word by document matrix A using training data. The SVD algorithm was applied with different numbers of singular values. In this study, we used MATLAB for the SVD operation and compared the performance of LSA language modeling, with the number of singular values R set at 25, 50, 75, and 100.",
"cite_spans": [],
"ref_spans": [
{
"start": 655,
"end": 662,
"text": "Table 1",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "The measure of perplexity was adopted to evaluate the different language models. The computational costs were reported for comparison. Here, the computation time was measured in minutes by testing 224 documents using a personal computer with a Pentium IV-1.6GHz",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "processor and 256 MB RAM. The bigram model was employed in the experiments. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "Given testing documents with l words, the perplexity is calculated as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "E Perplexity~2 = .",
"eq_num": "(25)"
}
],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "In general, the entropy E~ is the average difficulty or uncertainty of each word using the language model. The lower measured the perplexity, the better the speech recognition accuracy that can be achieved.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "In the experiments, we evaluated different language modeling and smoothing methods in terms of perplexity and computation time. First of all, we investigated the effect of the SVD dimension in the proposed LSA bigram model. No parameter smoothing was performed. In Figures 3 and 4 , we compare the perplexity and computation time for different SVD dimensions.",
"cite_spans": [],
"ref_spans": [
{
"start": 265,
"end": 280,
"text": "Figures 3 and 4",
"ref_id": "FIGREF12"
}
],
"eq_spans": [],
"section": "Evaluation of Different Language Modeling and Smoothing Methods",
"sec_num": "4.2"
},
{
"text": "Here, the computation time was a measure of the SVD operation of a 32,941*9,148 word by document matrix A . We found that an SVD dimension of 25 was appropriate for constructing the semantic space. In the subsequent evaluation, the SVD dimension was fixed at 25 for the proposed LSA bigram and LSA smoothing. Next, we examined the effect of the parameter k in the proposed LSA smoothing method. LSA smoothing of seen and unseen bigrams was performed by combining the bigrams corresponding to the k nearest words. In Figure 5 , we show the results for perplexity versus the k nearest neighbor words when LSA smoothing was applied to the standard bigram and proposed LSA bigram. The values 5 = k , 10, 30 and 50 were examined. When proposed LSA modeling and smoothing was used, the lowest perplexity The proposed LSA smoothing combines the bigrams corresponding to the k nearest seen words in the training corpus. We can see that the baseline bigram model has a perplexity of 158.3. The perplexity was reduced to 128.7 and 124.4 by applying Bellegarda's LSA bigram and proposed LSA bigram, respectively. However, when Witten-Bell smoothing was incorporated, the perplexity is greatly reduced from 158.3 without smoothing to 122.6 with smoothing. When the proposed LSA bigram with Witten-Bell smoothing were used, the perplexity could be improved to 108.7. This indicates the importance of adopting a smoothing algorithm in the language model. Furthermore, when the proposed LSA smoothing was used, the perplexity was reduced to 102, which is better than the perplexity of 122.6 obtained using Witten-Bell smoothing. This is because the Witten-Bell smoothing method estimates the n-gram model by using the (n-1)-gram, while the proposed LSA smoothing approach always adopts nearest n-gram models without using the (n-1)-gram. Among the different combinations, the lowest perplexity of 81 was achieved by applying the proposed LSA bigram with LSA smoothing. Compared to baseline system, the perplexity could be improved by up to 48.8%. The computation times of the different methods were also compared. The results show that the computation overhead of using a smoothing algorithm is slight. The computation load of the LSA bigram is much higher than that of the standard bigram. This result indicates that the smoothing algorithm can lead to greater improvement in perplexity with a lower computation cost than can be achieved by modifying the language model. ",
"cite_spans": [],
"ref_spans": [
{
"start": 516,
"end": 524,
"text": "Figure 5",
"ref_id": "FIGREF13"
}
],
"eq_spans": [],
"section": "Evaluation of Different Language Modeling and Smoothing Methods",
"sec_num": "4.2"
},
{
"text": "Statistical n-gram modeling is limited in terms of its ability to represent the historical words and estimate the unseen parameters of an inadequate training corpus. In this paper, we have presented new language modeling and smoothing methods that are based on the framework of latent semantic analysis. The concept of relevance retrieval has been adopted in order to exploit a new language modeling approach, where the most likely pseudodocument is retrieved to represent the historical words. The language model is estimated according to the closeness of the current word vector and the historical pseudodocument vector in the common LSA space.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5."
},
{
"text": "To overcome the problem of insufficient training data, we perform LSA smoothing, where the bigram of the current word is computed by interpolating with the bigrams corresponding to the k nearest words. The weighting coefficients of the k nearest words are proportional to the",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5."
}
],
"back_matter": [
{
"text": "The authors thank the anonymous reviewers for providing valuable comments, which considerably improved the quality of this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgment",
"sec_num": null
},
{
"text": "Perplexity is an important parameter used to evaluate the performance of language models. Consider an information source containing of word sequence, l www ,,, 21 K , each of which is chosen from a vocabulary V . The entropy of a source emitting the wordsIf the source is ergodic, the entropy in (22) is equivalent to ),,,Pr(log 1 limSince the n-gram language model is used, E can be estimated as follows:closeness to the current word in the LSA space. From the results of experiments in which Chinese news documents were evaluated, we found that the language modeling performance could be greatly improved by applying the proposed LSA parameter modeling and smoothing algorithms. The proposed methods outperformed Bellegarda's LSA bigram and Witten-Bell smoothing. Compared to the baseline bigram model, the perplexity was reduced by up to 48.8%.Also, the perplexity improvement and computation efficiency that could be achieved through parameter smoothing were better than that which could be achieved through parameter modeling. This approach can be easily extended to the trigram model and other languages. In the future, we will explore theoretical rules for determining the SVD dimension for LSA. We will also investigate the effect of the amount of training data on the LSA framework. We are currently applying the proposed language model to information retrieval and large vocabulary continuous speech recognition.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Perplexity",
"sec_num": "4.1"
},
{
"text": "In 8 . Because U is orthogonal and \u03a3 is diagonal, the representationAlso, from (4), we can derive the recursive formula for By extending this formula using vector representation, we obtain (9) by where the \"1\" appears at coordinate i in the above vector.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Derivations of Equations (8) and (9)",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A Multi-span Language Modeling Framework for Large Vocabulary Speech Recognition",
"authors": [
{
"first": "J",
"middle": [
"R"
],
"last": "Bellegarda",
"suffix": ""
}
],
"year": 1998,
"venue": "IEEE Transactions on Speech and Audio Processing",
"volume": "6",
"issue": "5",
"pages": "456--467",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bellegarda, J. R., \"A Multi-span Language Modeling Framework for Large Vocabulary Speech Recognition,\" IEEE Transactions on Speech and Audio Processing, 6(5) 1998, pp. 456-467.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A statistical language modeling approach integrating local and global constraints",
"authors": [
{
"first": "J",
"middle": [
"R"
],
"last": "Bellegarda",
"suffix": ""
}
],
"year": 1997,
"venue": "Proc. IEEE Workshop on Automatic Speech Recognition and Understanding",
"volume": "",
"issue": "",
"pages": "262--269",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bellegarda, J. R., \"A statistical language modeling approach integrating local and global constraints,\" Proc. IEEE Workshop on Automatic Speech Recognition and Understanding, 1997, pp. 262-269.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Exploiting latent semantic information in statistical language modeling",
"authors": [
{
"first": "J",
"middle": [
"R"
],
"last": "Bellegarda",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceeding of IEEE",
"volume": "88",
"issue": "8",
"pages": "1279--1296",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bellegarda, J. R., \"Exploiting latent semantic information in statistical language modeling, \" Proceeding of IEEE, 88(8) 2000a, pp. 1279-1296.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Large vocabulary speech recognition with multi-span statistical language models",
"authors": [
{
"first": "J",
"middle": [
"R"
],
"last": "Bellegarda",
"suffix": ""
}
],
"year": 2000,
"venue": "IEEE Transactions on Speech and Audio Processing",
"volume": "8",
"issue": "1",
"pages": "76--84",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bellegarda, J. R., \"Large vocabulary speech recognition with multi-span statistical language models,\" IEEE Transactions on Speech and Audio Processing, 8(1) 2000b, pp. 76-84.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Using Linear algebra for Intelligent Information Retrieval",
"authors": [
{
"first": "M",
"middle": [
"W"
],
"last": "Berry",
"suffix": ""
},
{
"first": "S",
"middle": [
"T"
],
"last": "Dumais",
"suffix": ""
},
{
"first": "G",
"middle": [
"W"
],
"last": "O'brien",
"suffix": ""
}
],
"year": 1995,
"venue": "Society for Industrial and Applied Mathematics (SIAM): Review",
"volume": "37",
"issue": "4",
"pages": "573--595",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Berry, M. W., S. T. Dumais and G. W. O'Brien, \"Using Linear algebra for Intelligent Information Retrieval,\" Society for Industrial and Applied Mathematics (SIAM): Review, 37(4) 1995, pp. 573-595.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Large scale singular value computations",
"authors": [
{
"first": "M",
"middle": [
"W"
],
"last": "Berry",
"suffix": ""
}
],
"year": 1992,
"venue": "International Journal of Supercomputer Applications",
"volume": "6",
"issue": "",
"pages": "13--49",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Berry, M. W., \"Large scale singular value computations,\" International Journal of Supercomputer Applications, vol. 6, 1992, pp. 13-49.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "An Empirical Study of Smoothing Techniques for Language Modeling",
"authors": [
{
"first": "S",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Goodman",
"suffix": ""
}
],
"year": 1999,
"venue": "Computer Speech and Language",
"volume": "13",
"issue": "4",
"pages": "359--394",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chen, S. and J. Goodman, \"An Empirical Study of Smoothing Techniques for Language Modeling,\" Computer Speech and Language, 13(4) 1999, pp. 359-394.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Indexing by Latent Semantic Analysis",
"authors": [
{
"first": "S",
"middle": [],
"last": "Deerwester",
"suffix": ""
},
{
"first": "S",
"middle": [
"T"
],
"last": "Dumais",
"suffix": ""
},
{
"first": "T",
"middle": [
"K"
],
"last": "Landauer",
"suffix": ""
},
{
"first": "G",
"middle": [
"W"
],
"last": "Furnas",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Harshman",
"suffix": ""
}
],
"year": 1990,
"venue": "Journal of the American Society for Information Science",
"volume": "41",
"issue": "6",
"pages": "391--407",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Deerwester, S., S. T. Dumais, T. K. Landauer, G. W. Furnas and R. Harshman, \"Indexing by Latent Semantic Analysis,\" Journal of the American Society for Information Science, 41(6) 1990, pp. 391-407.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "A similarity-based probability model for latent semantic indexing",
"authors": [
{
"first": "C",
"middle": [
"H Q"
],
"last": "Ding",
"suffix": ""
}
],
"year": 1999,
"venue": "Proc. 22 nd Annual International ACM SIGIR Conference",
"volume": "",
"issue": "",
"pages": "58--65",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ding, C. H. Q., \"A similarity-based probability model for latent semantic indexing,\" Proc. 22 nd Annual International ACM SIGIR Conference, 1999, pp. 58-65.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Bayesian estimation methods for n-gram language model adaptation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Federico",
"suffix": ""
}
],
"year": 1996,
"venue": "Proc. of the International Conference on Spoken Language Processing",
"volume": "1",
"issue": "",
"pages": "240--243",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Federico, M., \"Bayesian estimation methods for n-gram language model adaptation,\" Proc. of the International Conference on Spoken Language Processing, vol. 1, 1996, pp. 240-243.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Self-Organized Language Modeling for Speech Recognition",
"authors": [
{
"first": "F",
"middle": [],
"last": "Jelinek",
"suffix": ""
}
],
"year": 1990,
"venue": "Readings in Speech Recognition",
"volume": "",
"issue": "",
"pages": "450--506",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jelinek, F., \"Self-Organized Language Modeling for Speech Recognition,\" Readings in Speech Recognition, Morgan-Kaufmann Publishers, 1990, pp. 450-506.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Up From Trigrams",
"authors": [
{
"first": "F",
"middle": [],
"last": "Jelinek",
"suffix": ""
}
],
"year": 1991,
"venue": "Proc. European Conference on Speech communication and Technology",
"volume": "3",
"issue": "",
"pages": "1037--1040",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jelinek, F., \"Up From Trigrams,\" Proc. European Conference on Speech communication and Technology, vol. 3, 1991, pp. 1037-1040.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Interpolated estimation of Markov source parameters from sparse data",
"authors": [
{
"first": "F",
"middle": [],
"last": "Jelinek",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Mercer",
"suffix": ""
}
],
"year": 1980,
"venue": "Pattern Recognition in Practice",
"volume": "",
"issue": "",
"pages": "381--397",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jelinek, F. and R. Mercer, \"Interpolated estimation of Markov source parameters from sparse data,\" Pattern Recognition in Practice, 1980, pp. 381-397.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Estimation of Probabilities from Sparse Data for the Language Model Component of a Speech Recognizer",
"authors": [
{
"first": "S",
"middle": [
"M"
],
"last": "Katz",
"suffix": ""
}
],
"year": 1987,
"venue": "IEEE Transactions on Acoustics, Speech, and Signal Processing",
"volume": "35",
"issue": "3",
"pages": "400--401",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Katz, S.M., \"Estimation of Probabilities from Sparse Data for the Language Model Component of a Speech Recognizer,\" IEEE Transactions on Acoustics, Speech, and Signal Processing, 35(3) 1987, pp. 400-401.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Back-off Method for N-gram Smoothing based on Binomial Posteriori Distribution",
"authors": [
{
"first": "T",
"middle": [],
"last": "Kawabata",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Tamoto",
"suffix": ""
}
],
"year": 1996,
"venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "1",
"issue": "",
"pages": "192--195",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kawabata, T. and M. Tamoto, \"Back-off Method for N-gram Smoothing based on Binomial Posteriori Distribution,\" Proc. IEEE International Conference on Acoustics, Speech and Signal Processing, vol. 1, 1996, pp.192-195.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Trigger-based language models: A maximum entropy approach",
"authors": [
{
"first": "R",
"middle": [],
"last": "Lau",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Rosenfeld",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Roukos",
"suffix": ""
}
],
"year": 1993,
"venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "2",
"issue": "",
"pages": "45--48",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lau, R., R. Rosenfeld and S. Roukos, \"Trigger-based language models: A maximum entropy approach,\" Proc. IEEE International Conference on Acoustics, Speech and Signal Processing, vol. 2, 1993, pp. 45-48.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Task adaptation using MAP estimation in n-gram language modeling",
"authors": [
{
"first": "H",
"middle": [],
"last": "Masataki",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Sagisaka",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Hisaki",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Kawahara",
"suffix": ""
}
],
"year": 1997,
"venue": "Proc. IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "2",
"issue": "",
"pages": "783--786",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Masataki, H., Y. Sagisaka, K. Hisaki and T. Kawahara, \"Task adaptation using MAP estimation in n-gram language modeling,\" Proc. IEEE International Conference on Acoustics, Speech and Signal Processing, vol. 2, 1997, pp.783-786.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Use of non-negative matrix factorization for language model adaptation in a lecture transcription task",
"authors": [
{
"first": "M",
"middle": [],
"last": "Novak",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Mammone",
"suffix": ""
}
],
"year": 2001,
"venue": "Proc. of the IEEE International Conference on Acoustics, Speech and Signal Processing",
"volume": "1",
"issue": "",
"pages": "541--544",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Novak, M. and R. Mammone, \"Use of non-negative matrix factorization for language model adaptation in a lecture transcription task,\" Proc. of the IEEE International Conference on Acoustics, Speech and Signal Processing, vol. 1, 2001, pp. 541-544.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Modern information retrieval",
"authors": [
{
"first": "B.-Y",
"middle": [],
"last": "Ricardo",
"suffix": ""
},
{
"first": "R. -N",
"middle": [],
"last": "Berthier",
"suffix": ""
}
],
"year": 2000,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ricardo, B.-Y. and R. -N. Berthier, Modern information retrieval, Addison-Wesley, 2000.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "The zero-frequency problem: Estimating the probabilities of novel events in adaptive text compression",
"authors": [
{
"first": "I",
"middle": [
"H"
],
"last": "Written",
"suffix": ""
},
{
"first": "T",
"middle": [
"C"
],
"last": "Bell",
"suffix": ""
}
],
"year": 1991,
"venue": "IEEE Transaction on Information Theory",
"volume": "37",
"issue": "4",
"pages": "1085--1094",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Written, I. H. and T. C. Bell, \"The zero-frequency problem: Estimating the probabilities of novel events in adaptive text compression,\" IEEE Transaction on Information Theory, 37(4) 1991, pp. 1085-1094.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "A study of smoothing methods for language models applied to ad hoc information retrieval",
"authors": [
{
"first": "C",
"middle": [],
"last": "Zhai",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Lafferty",
"suffix": ""
}
],
"year": 2001,
"venue": "Proc. 24th Annual International ACM SIGIR Conference",
"volume": "",
"issue": "",
"pages": "334--342",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhai, C. and J. Lafferty, \"A study of smoothing methods for language models applied to ad hoc information retrieval,\" Proc. 24th Annual International ACM SIGIR Conference, 2001, pp. 334-342.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Interpolation of n-gram and mutual-information based trigger pair language models for Mandarin speech recognition",
"authors": [
{
"first": "G",
"middle": [
"D"
],
"last": "Zhou",
"suffix": ""
},
{
"first": "K",
"middle": [
"T"
],
"last": "Lua",
"suffix": ""
}
],
"year": 1999,
"venue": "Computer Speech and Language",
"volume": "13",
"issue": "2",
"pages": "125--141",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhou, G. D. and K. T. Lua, \"Interpolation of n-gram and mutual-information based trigger pair language models for Mandarin speech recognition,\" Computer Speech and Language, 13(2) 1999, pp.125-141.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"uris": null,
"num": null,
"text": "The N-gram model )Pr(W is written as ),...,,Pr(),...,Pr()Pr( .",
"type_str": "figure"
},
"FIGREF2": {
"uris": null,
"num": null,
"text": "document history significantly affects the probability of occurring B w .The trigger pairs provide long distance information because the triggering and triggered words might be separated by several words. However, trigger pair selection neglects the possibility of word triggers, which might contain useful semantic information. The LSA method was developed to resolve this problem. A schematic diagram of a speech recognition system.",
"type_str": "figure"
},
"FIGREF3": {
"uris": null,
"num": null,
"text": "A diagram of the truncated SVD. 3. New language modeling and smoothing techniques 3.1 LSA Parameter Modeling N-gram language models are useful for modeling the local dependencies of word occurrences but not for capturing global word dependencies. The modeling process leads to the estimation of the",
"type_str": "figure"
},
"FIGREF9": {
"uris": null,
"num": null,
"text": "the current word q w , the higher is the weighting coefficient that q j \u03b2 produces. Also, it is reasonable to adopt the interpolation coefficient q \u03b1 in(21), which is proportional to the closeness between q w and1 \u2212 q win the semantic space. The smoothing method proposed in (19) should be performed when the current word q w is trained using LSA. Different from the Jelinek-Mercer and Witten-Bell smoothing methods that adopt the maximum likelihood language model, the proposed smoothing technique is combined with the LSA framework, and the probabilities )",
"type_str": "figure"
},
"FIGREF10": {
"uris": null,
"num": null,
"text": "The perplexity of the standard bigram with LSA smoothing was calculated as 102. We then fixed 5 = k in the subsequent comparison experiment.",
"type_str": "figure"
},
"FIGREF11": {
"uris": null,
"num": null,
"text": "Figure 3. Comparison of perplexity results for different SVD dimensions.",
"type_str": "figure"
},
"FIGREF12": {
"uris": null,
"num": null,
"text": "Comparison of computation times for different SVD dimensions.",
"type_str": "figure"
},
"FIGREF13": {
"uris": null,
"num": null,
"text": "Perplexity versus the k nearest neighbor words when LSA smoothing was applied to the standard bigram and proposed LSA bigram.Furthermore, different language modeling and smoothing methods were compared, and the results are shown inTable 2. Besides the standard bigram, we implemented Bellegarda 's LSA bigram[Bellegarda 1998] and the proposed LSA bigram to evaluate the effect of language modeling. The main difference is that proposed LSA bigram aims to retrieve the most likely relevance document vector in order to represent the historical words. In addition, the language models with and without parameter smoothing were examined. The algorithms of Witten-Bell smoothing and the proposed LSA smoothing were also used for the purpose of comparison. The Witten-Bell smoothed bigram is estimated by interpolating with the corresponding unigram.",
"type_str": "figure"
},
"TABREF7": {
"num": null,
"type_str": "table",
"content": "<table><tr><td/><td colspan=\"8\">Technology Social International Leisure Politics Financial Entertain Sports</td></tr><tr><td>Training data</td><td>289</td><td>2,658</td><td>330</td><td>1,106</td><td>1,299</td><td>2,605</td><td>430</td><td>431</td></tr><tr><td>Testing data</td><td>30</td><td>24</td><td>23</td><td>24</td><td>24</td><td>49</td><td>23</td><td>27</td></tr></table>",
"text": "",
"html": null
},
"TABREF8": {
"num": null,
"type_str": "table",
"content": "<table><tr><td colspan=\"2\">Language Model Modeling Method Smoothing Method</td><td>Perplexity</td><td>Reduction Rate (%)</td><td>Computation Time (minutes)</td></tr><tr><td>Bigram</td><td>N/A</td><td>158.3</td><td>N/A</td><td>48.3</td></tr><tr><td>Bigram</td><td>Witten-Bell Smoothing</td><td>122.6</td><td>22.6</td><td>51.3</td></tr><tr><td>Bellegarda's LSA Bigram</td><td>N/A</td><td>128.7</td><td>18.7</td><td>176.7</td></tr><tr><td>Proposed LSA Bigram</td><td>N/A</td><td>124.4</td><td>21.4</td><td>161.2</td></tr><tr><td>Proposed LSA Bigram</td><td>Witten-Bell Smoothing</td><td>108.7</td><td>31.3</td><td>163.3</td></tr><tr><td>Bigram</td><td>LSA Smoothing</td><td>102</td><td>35.6</td><td>52.2</td></tr><tr><td>Proposed LSA Bigram</td><td>LSA Smoothing</td><td>81</td><td>48.8</td><td>163.4</td></tr></table>",
"text": "",
"html": null
}
}
}
} |