File size: 96,686 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 | {
"paper_id": "P13-1005",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T09:34:58.491153Z"
},
"title": "Smoothed marginal distribution constraints for language modeling",
"authors": [
{
"first": "Brian",
"middle": [],
"last": "Roark",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Oregon Health & Science University",
"location": {
"settlement": "Portland"
}
},
"email": "roarkbr@gmail.com"
},
{
"first": "Cyril",
"middle": [],
"last": "Allauzen",
"suffix": "",
"affiliation": {},
"email": "allauzen@google.com"
},
{
"first": "Michael",
"middle": [],
"last": "Riley",
"suffix": "",
"affiliation": {},
"email": "riley@google.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We present an algorithm for re-estimating parameters of backoff n-gram language models so as to preserve given marginal distributions, along the lines of wellknown Kneser-Ney (1995) smoothing. Unlike Kneser-Ney, our approach is designed to be applied to any given smoothed backoff model, including models that have already been heavily pruned. As a result, the algorithm avoids issues observed when pruning Kneser-Ney models (Siivola et al., 2007; Chelba et al., 2010), while retaining the benefits of such marginal distribution constraints. We present experimental results for heavily pruned backoff ngram models, and demonstrate perplexity and word error rate reductions when used with various baseline smoothing methods. An open-source version of the algorithm has been released as part of the OpenGrm ngram library. 1",
"pdf_parse": {
"paper_id": "P13-1005",
"_pdf_hash": "",
"abstract": [
{
"text": "We present an algorithm for re-estimating parameters of backoff n-gram language models so as to preserve given marginal distributions, along the lines of wellknown Kneser-Ney (1995) smoothing. Unlike Kneser-Ney, our approach is designed to be applied to any given smoothed backoff model, including models that have already been heavily pruned. As a result, the algorithm avoids issues observed when pruning Kneser-Ney models (Siivola et al., 2007; Chelba et al., 2010), while retaining the benefits of such marginal distribution constraints. We present experimental results for heavily pruned backoff ngram models, and demonstrate perplexity and word error rate reductions when used with various baseline smoothing methods. An open-source version of the algorithm has been released as part of the OpenGrm ngram library. 1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Smoothed n-gram language models are the defacto standard statistical models of language for a wide range of natural language applications, including speech recognition and machine translation. Such models are trained on large text corpora, by counting the frequency of n-gram collocations, then normalizing and smoothing (regularizing) the resulting multinomial distributions. Standard techniques store the observed n-grams and derive probabilities of unobserved n-grams via their longest observed suffix and \"backoff\" costs associated with the prefix histories of the unobserved suffixes. Hence the size of the model grows with the number of observed n-grams, which is very large for typical training corpora.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Natural language applications, however, are commonly used in scenarios requiring relatively small footprint models. For example, applications running on mobile devices or in low latency streaming scenarios may be required to limit the complexity of models and algorithms to achieve the desired operating profile. As a result, statistical language models -an important component of many such applications -are often trained on very large corpora, then modified to fit within some pre-specified size bound. One method to achieve significant space reduction is through randomized data structures, such as Bloom (Talbot and Osborne, 2007) or Bloomier (Talbot and Brants, 2008) filters. These data structures permit efficient querying for specific n-grams in a model that has been stored in a fraction of the space required to store the full, exact model, though with some probability of false positives. Another common approach -which we pursue in this paper -is model pruning, whereby some number of the n-grams are removed from explicit storage in the model, so that their probability must be assigned via backoff smoothing. One simple pruning method is count thresholding, i.e., discarding n-grams that occur less than k times in the corpus. Beyond count thresholding, the most widely used pruning methods (Seymore and Rosenfeld, 1996; Stolcke, 1998) employ greedy algorithms to reduce the number of stored n-grams by comparing the stored probabilities to those that would be assigned via the backoff smoothing mechanism, and removing those with the least impact according to some criterion.",
"cite_spans": [
{
"start": 608,
"end": 634,
"text": "(Talbot and Osborne, 2007)",
"ref_id": "BIBREF15"
},
{
"start": 647,
"end": 672,
"text": "(Talbot and Brants, 2008)",
"ref_id": "BIBREF14"
},
{
"start": 1305,
"end": 1334,
"text": "(Seymore and Rosenfeld, 1996;",
"ref_id": "BIBREF9"
},
{
"start": 1335,
"end": 1349,
"text": "Stolcke, 1998)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "While these greedy pruning methods are highly effective for models estimated with most common smoothing approaches, they have been shown to be far less effective with Kneser-Ney trained language models (Siivola et al., 2007; Chelba et al., 2010) , leading to severe degradation in model quality relative to other standard smoothing meth-4-gram models",
"cite_spans": [
{
"start": 202,
"end": 224,
"text": "(Siivola et al., 2007;",
"ref_id": "BIBREF10"
},
{
"start": 225,
"end": 245,
"text": "Chelba et al., 2010)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Interpolated Perplexity n-grams Perplexity n-grams Smoothing method full pruned (\u00d71000) full pruned (\u00d71000) Absolute Discounting (Ney et al., 1994) 120. 5 197.3 383.4 119.8 198.1 386.2 Witten-Bell (Witten and Bell, 1991) 118.8 196.3 380.4 121.6 202.3 396.4 Ristad (1995) 126.4 203.6 395.6 ---N/A --- Katz (1987) 119.8 198.1 386.2 ---N/A ---Kneser-Ney (Kneser and Ney, 1995) 114.5 285.1 388.2 115.8 274.3 398.7 Mod. Kneser-Ney (Chen and Goodman, 1998) Table 3 in Chelba et al. (2010) , demonstrating perplexity degradation of Kneser-Ney smoothed models in contrast to other common smoothing methods. Data: English Broadcast News, 128M words training; 692K words test; 143K word vocabulary. 4-gram language models, pruned using Stolcke (1998) relative entropy pruning to approximately 1.3% of the original size of 31,095,260 n-grams.",
"cite_spans": [
{
"start": 129,
"end": 147,
"text": "(Ney et al., 1994)",
"ref_id": "BIBREF6"
},
{
"start": 153,
"end": 220,
"text": "5 197.3 383.4 119.8 198.1 386.2 Witten-Bell (Witten and Bell, 1991)",
"ref_id": null
},
{
"start": 257,
"end": 270,
"text": "Ristad (1995)",
"ref_id": "BIBREF7"
},
{
"start": 300,
"end": 311,
"text": "Katz (1987)",
"ref_id": "BIBREF4"
},
{
"start": 351,
"end": 373,
"text": "(Kneser and Ney, 1995)",
"ref_id": "BIBREF5"
},
{
"start": 415,
"end": 450,
"text": "Kneser-Ney (Chen and Goodman, 1998)",
"ref_id": null
},
{
"start": 462,
"end": 482,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
},
{
"start": 726,
"end": 740,
"text": "Stolcke (1998)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [
{
"start": 451,
"end": 458,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Backoff",
"sec_num": null
},
{
"text": "ods. Thus, while Kneser-Ney may be the preferred smoothing method for large, unpruned models -where it can achieve real improvements over other smoothing methods -when relatively sparse, pruned models are required, it has severely diminished utility. Table 1 presents a slightly reformatted version of Table 3 from Chelba et al. (2010) . In their experiments (see Table 1 caption for specifics on training/test setup), they trained 4-gram Broadcast News language models using a variety of both backoff and interpolated smoothing methods and measured perplexity before and after Stolcke (1998) relative entropy based pruning. With this size training data, the perplexity of all of the smoothing methods other than Kneser-Ney degrades from around 120 with the full model to around 200 with the heavily pruned model. Kneser-Ney smoothed models have lower perplexity with the full model than the other methods by about 5 points, but degrade with pruning to far higher perplexity, between 270-285.",
"cite_spans": [
{
"start": 315,
"end": 335,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 251,
"end": 258,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 302,
"end": 309,
"text": "Table 3",
"ref_id": null
},
{
"start": 364,
"end": 371,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Backoff",
"sec_num": null
},
{
"text": "The cause of this degradation is Kneser-Ney's unique method for estimating smoothed language models, which will be presented in more detail in Section 3. Briefly, the smoothing method reestimates lower-order n-gram parameters in order to avoid over-estimating the likelihood of n-grams that already have ample probability mass allocated as part of higher-order n-grams. This is done via a marginal distribution constraint which requires the expected frequency of the lower-order n-grams to match their observed frequency in the training data, much as is commonly done for maximum entropy model training. Goodman (2001) proved that, under certain assumptions, such constraints can only improve language models. Lower-order n-gram parameters resulting from Kneser-Ney are not relative frequency estimates, as with other smoothing methods; rather they are parameters estimated specifically for use within the larger smoothed model.",
"cite_spans": [
{
"start": 604,
"end": 618,
"text": "Goodman (2001)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Backoff",
"sec_num": null
},
{
"text": "There are (at least) a couple of reasons why such parameters do not play well with model pruning. First, the pruning methods commonly use lower order n-gram probabilities to derive an estimate of state marginals, and, since these parameters are no longer smoothed relative frequency estimates, they do not serve that purpose well. For this reason, the widely-used SRILM toolkit recently provided switches to modify their pruning algorithm to use another model for state marginal estimates (Stolcke et al., 2011) . Second, and perhaps more importantly, the marginal constraints that were applied prior to smoothing will not in general be consistent with the much smaller pruned model. For example, if a bigram parameter is modified due to the presence of some set of trigrams, and then some or all of those trigrams are pruned from the model, the bigram associated with the modified parameter will be unlikely to have an overall expected frequency equal to its observed frequency anymore. As a result, the resulting model degrades dramatically with pruning.",
"cite_spans": [
{
"start": 489,
"end": 511,
"text": "(Stolcke et al., 2011)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Backoff",
"sec_num": null
},
{
"text": "In this paper, we present an algorithm that imposes marginal distribution constraints of the sort used in Kneser-Ney modeling on arbitrary smoothed backoff n-gram language models. Our approach makes use of the same sort of derivation as the original Kneser-Ney modeling, but, among other differences, relies on smoothed estimates of the empirical relative frequency rather than the unsmoothed observed frequency. The algorithm can be applied after the smoothed model has been pruned, hence avoiding the pitfalls associated with Kneser-Ney modeling. Furthermore, while Kneser-Ney is conventionally defined as a variant of absolute discounting, our method can be applied to models smoothed with any backoff smoothing, including mixtures of models, widely used for domain adaptation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Backoff",
"sec_num": null
},
{
"text": "We next establish formal preliminaries and our smoothed marginal distribution constraints method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Backoff",
"sec_num": null
},
{
"text": "N-gram language models are typically presented mathematically in terms of words w, the strings (histories) h that precede them, and the suffixes of the histories (backoffs) h that are used in the smoothing recursion. Let V be a vocabulary (alphabet), and V * a string of zero or more symbols drawn from V . Let V k denote the set of strings w \u2208 V * of length k, i.e., |w| = k. We will use variables u, v, w, x, y, z \u2208 V to denote single symbols from the vocabulary; h, g \u2208 V * to denote history sequences preceding the specific word; and h , g \u2208 V * the respective backoff histories of h and g as typically defined (see below). For a string w = w 1 . . . w |w| we can calculate the smoothed conditional probability of each word w i in the sequence given the k words that preceded it, depending on the order of the Markov model. Let",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "h k i = w i\u2212k .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": ". . w i\u22121 be the previous k words in the sequence. Then the smoothed model is defined recursively as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "P(wi | h k i ) = P(wi | h k i ) if c(h k i wi) > 0 \u03b1(h k i ) P(wi | h k\u22121 i ) otherwise",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "where c(h k i w i ) is the count of the n-gram sequence w i\u2212k . . . w i in the training corpus; P is a regularized probability estimate that provides some probability mass for unobserved n-grams; and \u03b1(h k i ) is a factor that ensures normalization. Note that for h = h k i , the typically defined backoff history h = h k\u22121 i , i.e., the longest suffix of h that is not h itself. When we use h and g (for notational convenience) in future equations, it is this definition that we are using.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "There are many ways to estimate P, including absolute discounting (Ney et al., 1994) , Katz (1987) and Witten and Bell (1991) . Interpolated models are special cases of this form, where the P is determined using model mixing, and the \u03b1 parameter is exactly the mixing factor value for the lower order model. N-gram language models allow for a sparse representation, so that only a subset of the possible ngrams must be explicitly stored. Probabilities for the rest of the n-grams are calculated through the \"otherwise\" semantics in the equation above. For an n-gram language model G, we will say that an n-gram hw \u2208 G if it is explicitly represented in the model; otherwise hw \u2208 G. In the standard ngram formulation above, the assumption is that if c(h k i w i ) > 0 then the n-gram has a parameter; yet with pruning, we remove many observed n-grams from the model, hence this is no longer the appropriate criterion. We reformulate the standard equation as follows:",
"cite_spans": [
{
"start": 66,
"end": 84,
"text": "(Ney et al., 1994)",
"ref_id": "BIBREF6"
},
{
"start": 87,
"end": 98,
"text": "Katz (1987)",
"ref_id": "BIBREF4"
},
{
"start": 103,
"end": 125,
"text": "Witten and Bell (1991)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "P(wi|h k i ) = \u03b2(h k i wi) if h k i wi \u2208 G \u03b1(h k i , h k\u22121 i ) P(wi|h k\u22121 i ) otherwise (1) where \u03b2(h k i w i ) is the parameter associated with the n-gram h k i w i and \u03b1(h k i , h k\u22121 i )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "is the backoff cost associated with going from state h k i to state h k\u22121 i . We assume that, if hw \u2208 G then all prefixes and suffixes of hw are also in G. Figure 1 presents a schema of an automaton representation of an n-gram model, of the sort used in the OpenGrm library (Roark et al., 2012) . States represent histories h, and the words w, whose probabilities are conditioned on h, label the arcs, leading to the history state for the subsequent word. State labels are provided in Figure 1 as a convenience, to show the (implicit) history encoded by the state, e.g., ' xyz' indicates that the state represents a history with the previous three symbols being x, y and z. Failure arcs, labeled with a \u03c6 in Figure 1 , encode an \"otherwise\" semantics and have as destination the origin state's backoff history. Many higher order states will back off to the same lower order state, specifically those that share the same suffix.",
"cite_spans": [
{
"start": 274,
"end": 294,
"text": "(Roark et al., 2012)",
"ref_id": "BIBREF8"
},
{
"start": 571,
"end": 572,
"text": "'",
"ref_id": null
}
],
"ref_spans": [
{
"start": 156,
"end": 164,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 485,
"end": 493,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 708,
"end": 716,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "Note that, in general, the recursive definition of backoff may require the traversal of several back- off arcs before emitting a word, e.g., the highest order states in Figure 1 needing to traverse a couple of \u03c6 arcs to reach state 'z'. We can define the backoff cost between a state h k i and any of its suffix states as follows. Let \u03b1(h, h) = 1 and for m > 1,",
"cite_spans": [],
"ref_spans": [
{
"start": 169,
"end": 177,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "yz z xyz u/\u03b2(xyzu) w/\u03b2(yzw) w/\u03b2(zw) \u03c6/\u03b1(xyz,yz) \u03c6/\u03b1(yz,z) zw yyz \u03c6/\u03b1(yyz,yz) yzw \u03b5 yzu yzv v/\u03b2(yyzv) w/\u03b2(yyzw) \u03c6/\u03b1(z,\u03b5) \u03c6/\u03b1(yzw,zw) z/\u03b2(z)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "\u03b1(h k i , h k\u2212m i ) = m j=1 \u03b1(h k\u2212j+1 i , h k\u2212j i ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "If h k i w \u2208 G then the probability of that n-gram will be defined in terms of backoff to its longest suffix h k\u2212m i w \u2208 G. Let h wG denote the longest suffix of h such that h wG w \u2208 G. Note that this is not necessarily a proper suffix, since h wG could be h itself or it could be . Then",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P(w | h) = \u03b1(h, h wG ) \u03b2(h wG w)",
"eq_num": "(2)"
}
],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "which is equivalent to equation 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "2"
},
{
"text": "Marginal distribution constraints attempt to match the expected frequency of an n-gram with its observed frequency. In other words, if we use the model to randomly generate a very large corpus, the n-grams should occur with the same relative frequency in both the generated and original (training) corpus. Standard smoothing methods overgenerate lower-order n-grams. Using standard n-gram notation (where g is the backoff history for g), this constraint is stated in Kneser and Ney (1995) as",
"cite_spans": [
{
"start": 467,
"end": 488,
"text": "Kneser and Ney (1995)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P(w | h ) = g:g =h P(g, w | h )",
"eq_num": "(3)"
}
],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "where P is the empirical relative frequency estimate. Taking this approach, certain base smoothing methods end up with very nice, easy to calculate solutions based on counts. Absolute discounting (Ney et al., 1994) in particular, using the above approach, leads to the well-known Kneser-Ney smoothing approach (Kneser and Ney, 1995; Chen and Goodman, 1998) . We will follow this same approach, with a couple of changes. First, we will make use of regularized estimates of relative frequency P rather than raw relative frequency P. Second, rather than just looking at observed histories h that back off to h , we will look at all histories (observed or not) of the length of the longest history in the model. For notational simplicity, suppose we have an n+1-gram model, hence the longest history in the model is of length n. Assume the length of the particular backoff history |h | = k. Let V n\u2212k h be the set of strings h \u2208 V n with h as a suffix. Then we can restate the marginal distribution constraint in equation 3 as",
"cite_spans": [
{
"start": 196,
"end": 214,
"text": "(Ney et al., 1994)",
"ref_id": "BIBREF6"
},
{
"start": 310,
"end": 332,
"text": "(Kneser and Ney, 1995;",
"ref_id": "BIBREF5"
},
{
"start": 333,
"end": 356,
"text": "Chen and Goodman, 1998)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P(w | h ) = h\u2208V n\u2212k h P(h, w | h )",
"eq_num": "(4)"
}
],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "Next we solve for \u03b2(h w) parameters used in equation 1. Note that h is a suffix of any h \u2208 V n\u2212k h , so conditioning probabilities on h and h is the same as conditioning on just h. Each of the following derivation steps simply relies on the chain rule or definition of conditional probability, as well as pulling terms out of the summation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "P(w | h ) = h\u2208V n\u2212k h P(h, w | h ) = h\u2208V n\u2212k h P(w | h, h ) P(h | h ) = h\u2208V n\u2212k h P(w | h) P(h) g\u2208V n\u2212k h P(g) = 1 g\u2208V n\u2212k h P(g) h\u2208V n\u2212k h P(w | h) P(h) (5)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "Then, multiplying both sides by the normalizing denominator on the right-hand side and using equation 2 to substitute \u03b1(h, h wG ) \u03b2(h wG w) for P(w | h):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P(w | h ) g\u2208V n\u2212k h P(g) = h\u2208V n\u2212k h P(w | h) P(h) = h\u2208V n\u2212k h \u03b1(h, h wG ) \u03b2(h wG w) P(h)",
"eq_num": "(6)"
}
],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "Note that we are only interested in h w \u2208 G, hence there are two disjoint subsets of histories h \u2208 V n\u2212k h that are being summed over: those such that h wG = h and those such that |h wG | > |h |. We next separate these sums in the next step of the derivation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P(w | h ) g\u2208V n\u2212k h P(g) = h\u2208V n\u2212k h :|h wG |>|h | \u03b1(h, h wG ) \u03b2(h wG w) P(h) + h\u2208V n\u2212k h :h wG =h \u03b1(h, h ) \u03b2(h w) P(h)",
"eq_num": "(7)"
}
],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "Finally, we solve for \u03b2(h w) in the second sum on the right-hand side of equation 7, yielding the formula in equation 8. Note that this equation is the correlate of equation 6in Kneser and Ney",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03b2(h w) = P(w | h ) g\u2208V n\u2212k h P(g) \u2212 h\u2208V n\u2212k h :|h wG |>|h | \u03b1(h, h wG ) \u03b2(h wG w) P(h) h\u2208V n\u2212k h :h wG =h \u03b1(h, h ) P(h)",
"eq_num": "(8)"
}
],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "(1995), modulo the two differences noted earlier: use of smoothed probability P rather than raw relative frequency; and summing over all history substrings in V n\u2212k h rather than just those with count greater than zero, which is also a change due to smoothing. Keep in mind, P is the target expected frequency from a given smoothed model. Kneser-Ney models are not useful input models, since their P n-gram parameters are not relative frequency estimates. This means that we cannot simply 'repair' pruned Kneser-Ney models, but must use other smoothing methods where the smoothed values are based on relative frequency estimation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "There are, in addition, two other important differences in our approach from that in Kneser and Ney (1995) , which would remain as differences even if our target expected frequency were the unsmoothed relative frequency P instead of the smoothed estimate P. First, the sum in the numerator is over histories of length n, the highest order in the n-gram model, whereas in the Kneser-Ney approach the sum is over histories that immediately back off to h , i.e., from the next highest order in the n-gram model. Thus the unigram distribution is with respect to the bigram model, the bigram model is with respect to the trigram model, and so forth. In our optimization, we sum instead over all possible history sequences of length n. Second, an early assumption made in Kneser and Ney (1995) is that the denominator term in their equation 6(our Eq. 8) is constant across all words for a given history, which is clearly false. We do not make this assumption. Of course, the probabilities must be normalized, hence the final values of \u03b2(h w) will be proportional to the values in Eq. 8.",
"cite_spans": [
{
"start": 85,
"end": 106,
"text": "Kneser and Ney (1995)",
"ref_id": "BIBREF5"
},
{
"start": 766,
"end": 787,
"text": "Kneser and Ney (1995)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "We briefly note that, like Kneser-Ney, if the baseline smoothing method is consistent, then the amount of smoothing in the limit will go to zero and our resulting model will also be consistent.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "The smoothed relative frequency estimate P and higher order \u03b2 values on the right-hand side of Eq. 8 are given values (from the input smoothed model and previous stages in the algorithm, respectively), implying an algorithm that estimates highest orders of the model first. In addition, steady state history probabilities P(h) must be calculated. We turn to the estimation algorithm next.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Marginal distribution constraints",
"sec_num": "3"
},
{
"text": "Our algorithm takes a smoothed backoff n-gram language model in an automaton format (see Figure 1) and returns a smoothed backoff n-gram language model with the same topology. For all ngrams in the model that are suffixes of other ngrams in the model -i.e., that are backed-off to -we calculate the weight provided by equation 8 and assign it (after normalization) to the appropriate n-gram arc in the automaton. There are several important considerations for this algorithm, which we address in this section. First, we must provide a probability for every state in the model. Second, we must memoize summed values that are used repeatedly. Finally, we must iterate the calculation of certain values that depend on the n-gram weights being re-estimated.",
"cite_spans": [],
"ref_spans": [
{
"start": 89,
"end": 98,
"text": "Figure 1)",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Model constraint algorithm",
"sec_num": "4"
},
{
"text": "The steady state probability P(h) is taken to be the probability of observing h after a long word sequence, i.e., the state's relative frequency in a long sequence of randomly-generated sentences from the model:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "P(h) = lim m\u2192\u221e w 1 ...wmP (w 1 . . . w m h) (9)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "whereP is the corpus probability derived as follows: The smoothed n-gram probability model P(w | h) is naturally extended to a sentence s = w 0 . . . w l , where w 0 = <s> and w l = </s> are the sentence initial and final words, by",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "P(s) = l i=1 P(w i | h n i ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "The corpus probability s 1 . . . s r is taken as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P(s 1 . . . s r ) = (1 \u2212 \u03bb)\u03bb r\u22121 r i=1 P(s i )",
"eq_num": "(10)"
}
],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "where \u03bb parameterizes the corpus length distribution. 2 Assuming the n-gram language model automaton G has a single final state </s> into 2P models words in a corpus rather than a single sentence since Equation 9 tends to zero as m \u2192 \u221e otherwise. In Markov chain terms, the corpus distribution is made irreducible to allow a non-trivial stationary distribution. which all </s> arcs enter, adding a \u03bb weighted arc from the </s> state to the initial state and having a final weight 1 \u2212 \u03bb in order to leave the automaton at the </s> state will model this corpus distribution. According to Eq. 9, P (h) is then the stationary distribution of the finite irreducible Markov Chain defined by this altered automaton. There are many methods for computing such a stationary distribution; we use the well-known power method (Stewart, 1999) .",
"cite_spans": [
{
"start": 813,
"end": 828,
"text": "(Stewart, 1999)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "One difficulty remains to be resolved. The backoff arcs have a special interpretation in the automaton: they are traversed only if a word fails to match at the higher order. These failure arcs must be properly handled before applying standard stationary distribution calculations. A simple approach would be for each word w and state h such that hw / \u2208 G, but h w \u2208 G, add a w arc from state h to h w with weight \u03b1(h, h )\u03b2(h w ) and then remove all failure arcs (see Figure 2a ). This however results in an automaton with |V | arcs leaving every state, which is unwieldy with larger vocabularies and n-gram orders. Instead, for each word w and state h such that hw \u2208 G, add a w arc from state h to h w with weight \u2212\u03b1(h, h )\u03b2(h w) and then replace all failure labels with labels (see Figure 2b ). In this case, the added negativelyweighted arcs compensate for the excess probability mass allowed by the epsilon arcs 3 . The number of added arcs is no more than found in the original model.",
"cite_spans": [],
"ref_spans": [
{
"start": 467,
"end": 476,
"text": "Figure 2a",
"ref_id": "FIGREF1"
},
{
"start": 783,
"end": 792,
"text": "Figure 2b",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Steady state probability calculation",
"sec_num": "4.1"
},
{
"text": "We are summing over all possible histories of length n in equation 8, and the steady state probability calculation outlined in the previous section includes the probability mass for histories h \u2208 G. The probability mass of states not in G ends up being allocated to the state representing their longest suffix that is explicitly in G. That is the state that would be active when these histories are encountered. Hence, once we have calculated the steady state probabilities for each state in the smoothed model, we only need to sum over states explicitly in the model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Accumulation of higher order values",
"sec_num": "4.2"
},
{
"text": "As stated earlier, the use of \u03b2(h wG w) in the numerator of equation 8 for h wG that are larger than h implies that the longer n-grams must be re-estimated first. Thus we process each history length in descending order, finishing with the unigram state. Since we assume that, for every ngram hw \u2208 G, every prefix and suffix is also in G, we know that if h w \u2208 G then there is no history h such that h is a suffix of h and hw \u2208 G. This allows us to recursively accumulate the \u03b1(h, h ) P(h) in the denominator of Eq. 8.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Accumulation of higher order values",
"sec_num": "4.2"
},
{
"text": "(a) (b) h h' w/\u03b2(hw) w'/\u03b2(h'w') \u03c6/\u03b1(h,h') hw h'w' w'/\u03b1(h,h') \u03b2(h'w') h h' w/\u03b2(hw) w/\u03b2(h'w) \u03b5/\u03b1(h,h') hw h'w w/-\u03b1(h,h') \u03b2(h'w)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Accumulation of higher order values",
"sec_num": "4.2"
},
{
"text": "For every n-gram, we can accumulate values required to calculate the three terms in equation 8, and pass them along to calculate lower order ngram values. Note, however, that a naive implementation of an algorithm to assign these values is O(|V | n ). This is due to the fact that the denominator factor must be accumulated for all higher order states that do not have the given n-gram. Hence, for every state h directly backing off to h (order |V |), and for every n-gram arc leaving state h (order |V |), some value must be accumulated. This can be particularly clearly seen at the unigram state, which has an arc for every unigram (the size of the vocabulary): for every bigram state (also order of the vocabulary), in the naive algorithm we must look for every possible arc. Since there are O(|V | n\u22122 ) lower order histories in the model in the worst case, we have overall complexity O(|V | n ). However, we know that the number of stored n-grams is very sparse relative to the possible number of n-grams, so the typical case complexity is far lower. Importantly, the denominator is calculated by first assuming that all higher order states back off to the current n-gram, then subtract out the mass associated with those that are already observed at the higher order. In such a way, we need only perform work for higher order n-grams hw that are explicitly in the model. This optimization achieves orders-of-magnitude speedups, so that models take seconds to process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Accumulation of higher order values",
"sec_num": "4.2"
},
{
"text": "Because smoothing is not necessarily con-strained across n-gram orders, it is possible that higher-order n-grams could be smoothed less than lower order n-grams, so that the numerator of equation 8 can be less than zero, which is not valid. A value less than zero means that the higher order n-grams will already produce the n-gram more frequently than its smoothed expected frequency. We set a minimum value for the numerator, and any n-gram numerator value less than is replaced with (for the current study, = 0.001). We find this to be relatively infrequent, about 1% of n-grams for most models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Accumulation of higher order values",
"sec_num": "4.2"
},
{
"text": "Recall that P and \u03b2 terms on the right-hand side of equation 8 are given and do not change. But there are two other terms in the equation that change as we update the n-gram parameters. The \u03b1(h, h ) backoff weights in the denominator ensure normalization at the higher order states, and change as the n-gram parameters at the current state are modified. Further, the steady state probabilities will change as the model changes. Hence, at each state, we must iterate the calculation of the denominator term: first adjust n-gram weights and normalize; then recalculate backoff weights at higher order states and iterate. Since this only involves the denominator term, each n-gram weight can be updated by multiplying by the ratio of the old term and the new term. After the entire model has been re-estimated, the steady state probability calculation presented in Section 4.1 is run again and model estimation happens again. As we shall see in the experimental results, this typically converges after just a few iterations. At this time, we have no convergence proofs for either of these iterative components to the algorithm, but expect that something can be said about this, which will be a priority in future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iteration",
"sec_num": "4.3"
},
{
"text": "All results presented here are for English Broadcast News. We received scripts for replicating the Chelba et al. (2010) results from the authors, and we report statistics on our replication of their paper's results in Table 2 . The scripts are distributed in such a way that the user supplies the data from LDC98T31 (1996 CSR HUB4 Language Model corpus) and the script breaks the collection into training and testing sets, normalizes the text, and trains and prunes the language models using the SRILM toolkit (Stolcke et al., 2011) . Presumably due to minor differences in text normalization, resulting in very slightly fewer n-grams in all conditions, we achieve negligibly lower perplexities (one or two tenths of a point) in all conditions, as can be seen when comparing with Table 1 . All of the same trends result, thus that paper's result is successfully replicated here. Note that we ran our Kneser-Ney pruning (noted with a \u2020 in the table), using the new -prune-history-lm switch in SRILM -created in response to the Chelba et al. (2010) paper -which allows the use of another model to calculate the state marginals for pruning. This fixes part of the problem -perplexity does not degrade as much as the Kneser-Ney pruned model in Table 1 -but, as argued earlier in this paper, this is not the sole reason for the degradation and the perplexity remains extremely inflated.",
"cite_spans": [
{
"start": 99,
"end": 119,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
},
{
"start": 510,
"end": 532,
"text": "(Stolcke et al., 2011)",
"ref_id": "BIBREF12"
},
{
"start": 1026,
"end": 1046,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 218,
"end": 225,
"text": "Table 2",
"ref_id": "TABREF3"
},
{
"start": 780,
"end": 787,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 1240,
"end": 1247,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "We follow Chelba et al. (2010) in training and test set definition, vocabulary size, and parameters for reporting perplexity. Note that unigrams in the models are never pruned, hence all models assign probabilities over an identical vocabulary and perplexity is comparable across models. For all results reported here, we use the SRILM toolkit for baseline model training and pruning, then convert from the resulting ARPA format model to an OpenFst format (Allauzen et al., 2007) , as used in the OpenGrm n-gram library (Roark et al., 2012) . We then apply the marginal distribution constraints, and convert the result back to ARPA format for perplexity evaluation with the SRILM toolkit. All models are subjected to full normalization sanity checks, as with typical model functions in the OpenGrm library.",
"cite_spans": [
{
"start": 10,
"end": 30,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
},
{
"start": 456,
"end": 479,
"text": "(Allauzen et al., 2007)",
"ref_id": "BIBREF0"
},
{
"start": 520,
"end": 540,
"text": "(Roark et al., 2012)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "Recall that our algorithm assumes that, for every n-gram in the model, all prefix and suffix ngrams are also in the model. For pruned models, the SRILM toolkit does not impose such a requirement, hence explicit arcs are added to the Table 3 : Perplexity reductions achieved with marginal distribution constraints (MDC) on the heavily pruned models from Chelba et al. (2010) , and a mixture model. WFST ngram counts are slightly higher than ARPA format in Table 2 due to adding prefix and suffix n-grams.",
"cite_spans": [
{
"start": 353,
"end": 373,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 233,
"end": 240,
"text": "Table 3",
"ref_id": null
},
{
"start": 455,
"end": 462,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "model during conversion, with probability equal to what they would receive in the the original model. The resulting model is equivalent, but with a small number of additional arcs in the explicit representation (around 1% for the most heavily pruned models). Table 3 presents perplexity results for models that result from applying our marginal distribution constraints to the four heavily pruned models from Table 2 . In all four cases, we get perplexity reductions of around 10 points. We present the number of n-grams represented explicitly in the WFST, which is a slight increase from those presented in Table 2 due to the reintroduction of prefix and suffix n-grams.",
"cite_spans": [],
"ref_spans": [
{
"start": 259,
"end": 266,
"text": "Table 3",
"ref_id": null
},
{
"start": 409,
"end": 416,
"text": "Table 2",
"ref_id": "TABREF3"
},
{
"start": 608,
"end": 615,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "In addition to the four models reported in Chelba et al. (2010) , we produced a mixture model by interpolating (with equal weight) smoothed ngram probabilities from the full (unpruned) absolute discounting, Witten-Bell and Katz models, which share the same set of n-grams. After renormalizing and pruning to approximately the same size as the other models, we get commensurate gains using this model as with the other models. Figure 3 demonstrates the importance of iterating the steady state history calculation. All of the methods achieve perplexity reductions with subsequent iterations. Katz and absolute discounting achieve very little reduction in the first iteration, but catch back up in the second iteration.",
"cite_spans": [
{
"start": 43,
"end": 63,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 426,
"end": 434,
"text": "Figure 3",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "The other iterative part of the algorithm, discussed in Section 4.3, is the denominator of equation 8, which changes due to adjustments in the backoff weights required by the revised n-gram probabilities. If we do not iteratively update the backoff weights when reestimating the weights, the 'Pruned+MDC' perplexities in Table 3 increase by between 0.2-0.4 points. Hence, iterating the steady state probability calculation is quite important, as illustrated by Figure 3 ; iterating the denominator calculation much less so, at least for these models. We noted in Section 3 that a key difference between our approach and Kneser and Ney (1995) is that their approach treated the denominator as a constant. If we do this, the 'Pruned+MDC' perplexities increase by between 4.5-5.6 points, i.e., about half of the perplexity reduction is lost for each method. Thus, while iteration of denominator calculation may not be critical, it should not be treated as a constant.",
"cite_spans": [
{
"start": 620,
"end": 641,
"text": "Kneser and Ney (1995)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [
{
"start": 321,
"end": 328,
"text": "Table 3",
"ref_id": null
},
{
"start": 461,
"end": 469,
"text": "Figure 3",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "We now look at the impacts on system performance we can achieve with these new models 4 , and whether the perplexity differences that we observe translate to real error rate reductions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "For automatic speech recognition experiments, we used as test set the 1997 Hub4 evaluation set consisting of 32,689 words. The acoustic model is a tied-state triphone GMM-based HMM whose input features are 9-frame stacked 13-dimensional PLP-cepstral coefficients projected down to 39 dimensions using LDA. The model was trained on the 1996 and 1997 Hub4 acoustic model training sets (about 150 hours of data) using semi-tied covariance modeling and CMLLR-based speaker adaptive training and 4 iterations of boosted MMI.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "We used a multi-pass decoding strategy: two quick passes for adaptation supervision, CMLLR and MLLR estimation; then a slower full decoding pass running about 3 times slower than real time. Table 4 presents recognition results for the heavily pruned models that we have been considering, both for first pass decoding and rescoring of the resulting lattices using failure transitions rather than epsilon backoff approximations. Chelba et al. (2010) , and a mixture model. Kneser-Ney results are shown for: a) original pruning; and b) with -prune-history-lm switch.",
"cite_spans": [
{
"start": 427,
"end": 447,
"text": "Chelba et al. (2010)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 190,
"end": 197,
"text": "Table 4",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "The perplexity reductions that were achieved for these models do translate to real word error rate reductions at both stages of between 0.5 and 0.9 percent absolute. All of these gains are statistically significant at p < 0.0001 using the stratified shuffling test (Yeh, 2000) . For pruned Kneser-Ney models, fixing the state marginals with the -prune-history-lm switch reduces the WER versus the original pruned model, but no reductions were achieved vs. baseline methods. Table 5 presents perplexity and WER results for less heavily pruned models, where the pruning thresholds were set to yield approximately 1.5 million n-grams (4 times more than the previous models); and another set at around 5 million n-grams, as well as the full, unpruned models. While the robust gains we've observed up to now persist with the 1.5M n-gram models (WER reductions significant, Witten-Bell at p < 0.02, others at p < 0.0001), the larger models yield diminishing gains, with no real WER improvements. Performance of Witten-Bell models with the marginal distribution constraints degrade badly for the larger models, indicating that this method of regularization, unmodified by aggressive pruning, does not provide a well suited distribution for this sort of optimization. We speculate that this is due to underregularization, having noted some floating point precision issues when allowing the backoff recalculation to run indefinitely.",
"cite_spans": [
{
"start": 265,
"end": 276,
"text": "(Yeh, 2000)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [
{
"start": 474,
"end": 481,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "5"
},
{
"text": "The presented method reestimates lower order n-gram model parameters for a given smoothed backoff model, achieving perplexity and WER reductions for many smoothed models. There remain a number of open questions to investigate in the future. Recall that the numerator in Eq. 8 can be less than zero, meaning that no parameterization would lead to a model with the target frequency of the lower order n-gram, presumably due to over-or under-regularization. We anticipate a pre-constraint on the baseline smoothing method, that would recognize this problem and adjust the smoothing to ensure that a solution does exist. Additionally, it is clear that different regularization methods yield different behaviors, notably that large, relatively lightly pruned Witten-Bell models yield poor results. We will look to identify the issues with such models and provide general guidelines for prepping models prior to processing. Finally, we would like to perform extensive controlled experimentation to examine the relative contribution of the various aspects of our approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summary and Future Directions",
"sec_num": "6"
},
{
"text": "Thanks to Ciprian Chelba and colleagues for the scripts to replicate their results. This work was supported in part by a Google Faculty Research Award and NSF grant #IIS-0964102. Any opinions, findings, conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of the NSF. Table 5 : Perplexity (PPL) and both first pass (FP) and rescoring (RS) WER reductions for less heavily pruned models using marginal distribution constraints (MDC).",
"cite_spans": [],
"ref_spans": [
{
"start": 342,
"end": 349,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Summary and Future Directions",
"sec_num": "6"
},
{
"text": "www.opengrm.org",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Since each negatively-weighted arc leaving a state exactly cancels an epsilon arc followed by a matching positively-weighted arc in each iteration of the power method, convergence is assured.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "For space purposes, we exclude the Ristad method from this point forward since it is not competitive with the others.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "OpenFst: A general and efficient weighted finite-state transducer library",
"authors": [
{
"first": "Cyril",
"middle": [],
"last": "Allauzen",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Riley",
"suffix": ""
},
{
"first": "Johan",
"middle": [],
"last": "Schalkwyk",
"suffix": ""
},
{
"first": "Wojciech",
"middle": [],
"last": "Skut",
"suffix": ""
},
{
"first": "Mehryar",
"middle": [],
"last": "Mohri",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the Twelfth International Conference on Implementation and Application of Automata (CIAA 2007)",
"volume": "4793",
"issue": "",
"pages": "11--23",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cyril Allauzen, Michael Riley, Johan Schalkwyk, Wo- jciech Skut, and Mehryar Mohri. 2007. OpenFst: A general and efficient weighted finite-state transducer library. In Proceedings of the Twelfth International Conference on Implementation and Application of Automata (CIAA 2007), Lecture Notes in Computer Science, volume 4793, pages 11-23.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Study on interaction between entropy pruning and Kneser-Ney smoothing",
"authors": [
{
"first": "Ciprian",
"middle": [],
"last": "Chelba",
"suffix": ""
},
{
"first": "Thorsten",
"middle": [],
"last": "Brants",
"suffix": ""
},
{
"first": "Will",
"middle": [],
"last": "Neveitt",
"suffix": ""
},
{
"first": "Peng",
"middle": [],
"last": "Xu",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of Interspeech",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ciprian Chelba, Thorsten Brants, Will Neveitt, and Peng Xu. 2010. Study on interaction between en- tropy pruning and Kneser-Ney smoothing. In Pro- ceedings of Interspeech, page 24222425.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "An empirical study of smoothing techniques for language modeling",
"authors": [
{
"first": "Stanley",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Joshua",
"middle": [],
"last": "Goodman",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Stanley Chen and Joshua Goodman. 1998. An em- pirical study of smoothing techniques for language modeling. Technical Report, TR-10-98, Harvard University.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A bit of progress in language modeling",
"authors": [
{
"first": "Joshua",
"middle": [],
"last": "Goodman",
"suffix": ""
}
],
"year": 2001,
"venue": "Computer Speech and Language",
"volume": "15",
"issue": "4",
"pages": "403--434",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Joshua Goodman. 2001. A bit of progress in lan- guage modeling. Computer Speech and Language, 15(4):403-434.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Estimation of probabilities from sparse data for the language model component of a speech recogniser",
"authors": [
{
"first": "M",
"middle": [],
"last": "Slava",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Katz",
"suffix": ""
}
],
"year": 1987,
"venue": "IEEE Transactions on Acoustic, Speech, and Signal Processing",
"volume": "35",
"issue": "3",
"pages": "400--401",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Slava M. Katz. 1987. Estimation of probabilities from sparse data for the language model component of a speech recogniser. IEEE Transactions on Acoustic, Speech, and Signal Processing, 35(3):400-401.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Improved backing-off for m-gram language modeling",
"authors": [
{
"first": "Reinhard",
"middle": [],
"last": "Kneser",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 1995,
"venue": "Proceedings of the International Conference on Acoustics, Speech, and Signal Processing (ICASSP)",
"volume": "",
"issue": "",
"pages": "181--184",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Reinhard Kneser and Hermann Ney. 1995. Improved backing-off for m-gram language modeling. In Pro- ceedings of the International Conference on Acous- tics, Speech, and Signal Processing (ICASSP), pages 181-184.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "On structuring probabilistic dependences in stochastic language modeling",
"authors": [
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
},
{
"first": "Ute",
"middle": [],
"last": "Essen",
"suffix": ""
},
{
"first": "Reinhard",
"middle": [],
"last": "Kneser",
"suffix": ""
}
],
"year": 1994,
"venue": "Computer Speech and Language",
"volume": "8",
"issue": "",
"pages": "1--38",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hermann Ney, Ute Essen, and Reinhard Kneser. 1994. On structuring probabilistic dependences in stochas- tic language modeling. Computer Speech and Lan- guage, 8:1-38.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "A natural law of succession",
"authors": [
{
"first": "Eric",
"middle": [
"S"
],
"last": "Ristad",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eric S. Ristad. 1995. A natural law of succession. Technical Report, CS-TR-495-95, Princeton Univer- sity.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "The OpenGrm open-source finite-state grammar software libraries",
"authors": [
{
"first": "Brian",
"middle": [],
"last": "Roark",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Sproat",
"suffix": ""
},
{
"first": "Cyril",
"middle": [],
"last": "Allauzen",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Riley",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Sorensen",
"suffix": ""
},
{
"first": "Terry",
"middle": [],
"last": "Tai",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the ACL 2012 System Demonstrations",
"volume": "",
"issue": "",
"pages": "61--66",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Brian Roark, Richard Sproat, Cyril Allauzen, Michael Riley, Jeffrey Sorensen, and Terry Tai. 2012. The OpenGrm open-source finite-state grammar soft- ware libraries. In Proceedings of the ACL 2012 Sys- tem Demonstrations, pages 61-66.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Scalable backoff language models",
"authors": [
{
"first": "Kristie",
"middle": [],
"last": "Seymore",
"suffix": ""
},
{
"first": "Ronald",
"middle": [],
"last": "Rosenfeld",
"suffix": ""
}
],
"year": 1996,
"venue": "Proceedings of the International Conference on Spoken Language Processing (ICSLP)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kristie Seymore and Ronald Rosenfeld. 1996. Scal- able backoff language models. In Proceedings of the International Conference on Spoken Language Processing (ICSLP).",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "On growing and pruning kneserney smoothed n-gram models",
"authors": [
{
"first": "Vesa",
"middle": [],
"last": "Siivola",
"suffix": ""
},
{
"first": "Teemu",
"middle": [],
"last": "Hirsimaki",
"suffix": ""
},
{
"first": "Sami",
"middle": [],
"last": "Virpioja",
"suffix": ""
}
],
"year": 2007,
"venue": "IEEE Transactions on Audio, Speech, and Language Processing",
"volume": "15",
"issue": "5",
"pages": "1617--1624",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vesa Siivola, Teemu Hirsimaki, and Sami Virpioja. 2007. On growing and pruning kneserney smoothed n-gram models. IEEE Transactions on Audio, Speech, and Language Processing, 15(5):1617- 1624.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Numerical methods for computing stationary distributions of finite irreducible markov chains. Computational Probability",
"authors": [
{
"first": "J",
"middle": [],
"last": "William",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Stewart",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "81--111",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "William J Stewart. 1999. Numerical methods for com- puting stationary distributions of finite irreducible markov chains. Computational Probability, pages 81-111.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Srilm at sixteen: Update and outlook",
"authors": [
{
"first": "Andreas",
"middle": [],
"last": "Stolcke",
"suffix": ""
},
{
"first": "Jing",
"middle": [],
"last": "Zheng",
"suffix": ""
},
{
"first": "Wen",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Abrash",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andreas Stolcke, Jing Zheng, Wen Wang, and Victor Abrash. 2011. Srilm at sixteen: Update and out- look. In Proceedings of the IEEE Automatic Speech Recognition and Understanding Workshop (ASRU).",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Entropy-based pruning of backoff language models",
"authors": [
{
"first": "Andreas",
"middle": [],
"last": "Stolcke",
"suffix": ""
}
],
"year": 1998,
"venue": "Proc. DARPA Broadcast News Transcription and Understanding Workshop",
"volume": "",
"issue": "",
"pages": "270--274",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andreas Stolcke. 1998. Entropy-based pruning of backoff language models. In Proc. DARPA Broad- cast News Transcription and Understanding Work- shop, pages 270-274.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Randomized language models via perfect hash functions",
"authors": [
{
"first": "David",
"middle": [],
"last": "Talbot",
"suffix": ""
},
{
"first": "Thorsten",
"middle": [],
"last": "Brants",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "505--513",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Talbot and Thorsten Brants. 2008. Randomized language models via perfect hash functions. In Pro- ceedings of ACL-08: HLT, pages 505-513.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Smoothed Bloom filter language models: Tera-scale LMs on the cheap",
"authors": [
{
"first": "David",
"middle": [],
"last": "Talbot",
"suffix": ""
},
{
"first": "Miles",
"middle": [],
"last": "Osborne",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)",
"volume": "",
"issue": "",
"pages": "468--476",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Talbot and Miles Osborne. 2007. Smoothed Bloom filter language models: Tera-scale LMs on the cheap. In Proceedings of the 2007 Joint Con- ference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 468-476.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "The zerofrequency problem: Estimating the probabilities of novel events in adaptive text compression",
"authors": [
{
"first": "H",
"middle": [],
"last": "Ian",
"suffix": ""
},
{
"first": "Timothy",
"middle": [
"C"
],
"last": "Witten",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Bell",
"suffix": ""
}
],
"year": 1991,
"venue": "IEEE Transactions on Information Theory",
"volume": "37",
"issue": "4",
"pages": "1085--1094",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ian H. Witten and Timothy C. Bell. 1991. The zero- frequency problem: Estimating the probabilities of novel events in adaptive text compression. IEEE Transactions on Information Theory, 37(4):1085- 1094.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "More accurate tests for the statistical significance of result differences",
"authors": [
{
"first": "A",
"middle": [],
"last": "Yeh",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 18th International COLING",
"volume": "",
"issue": "",
"pages": "947--953",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Yeh. 2000. More accurate tests for the statistical significance of result differences. In Proceedings of the 18th International COLING, pages 947-953.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "N-gram weighted automaton schema. State labels are presented for convenience, to specify the history implicitly encoded by the state."
},
"FIGREF1": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Schemata showing failure arc handling: (a) \u03c6 removal: add w arc (red), delete \u03c6 arc; (b) \u03c6 replacement: add w arc (red), replace \u03c6 by (red)"
},
"FIGREF3": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Models resulting from different numbers of parameter re-estimation iterations. Iteration 0 is the baseline pruned model."
},
"TABREF1": {
"html": null,
"text": "",
"content": "<table/>",
"type_str": "table",
"num": null
},
"TABREF3": {
"html": null,
"text": "Replication ofChelba et al. (2010) using provided script. Using the script, the size of the unpruned model is 31,091,219 ngrams, 4,041 fewer thanChelba et al. (2010).",
"content": "<table><tr><td>\u2020 Kneser-Ney model pruned using -prune-history-lm switch in SRILM.</td></tr></table>",
"type_str": "table",
"num": null
},
"TABREF6": {
"html": null,
"text": "WER reductions achieved with marginal distribution constraints (MDC) on the heavily pruned models from",
"content": "<table/>",
"type_str": "table",
"num": null
}
}
}
} |