File size: 88,135 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 | {
"paper_id": "2022",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T01:11:50.821135Z"
},
"title": "Using ASR-Generated Text for Spoken Language Modeling",
"authors": [
{
"first": "Nicolas",
"middle": [],
"last": "Herv\u00e9",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Institut National de l'Audiovisuel (INA)",
"location": {
"country": "France"
}
},
"email": "nherve@ina.fr"
},
{
"first": "Valentin",
"middle": [],
"last": "Pelloin",
"suffix": "",
"affiliation": {
"laboratory": "Laboratoire d'Informatique de l'Universit\u00e9 du Mans (LIUM)",
"institution": "",
"location": {
"country": "France"
}
},
"email": ""
},
{
"first": "Benoit",
"middle": [],
"last": "Favre",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "CNRS",
"location": {
"settlement": "Marseille",
"region": "LIS",
"country": "France"
}
},
"email": ""
},
{
"first": "Franck",
"middle": [],
"last": "Dary",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "CNRS",
"location": {
"settlement": "Marseille",
"region": "LIS",
"country": "France"
}
},
"email": ""
},
{
"first": "Antoine",
"middle": [],
"last": "Laurent",
"suffix": "",
"affiliation": {
"laboratory": "Laboratoire d'Informatique de l'Universit\u00e9 du Mans (LIUM)",
"institution": "",
"location": {
"country": "France"
}
},
"email": ""
},
{
"first": "Sylvain",
"middle": [],
"last": "Meignier",
"suffix": "",
"affiliation": {
"laboratory": "Laboratoire d'Informatique de l'Universit\u00e9 du Mans (LIUM)",
"institution": "",
"location": {
"country": "France"
}
},
"email": ""
},
{
"first": "Laurent",
"middle": [],
"last": "Besacier",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Naver Labs Europe (NLE)",
"location": {
"settlement": "Meylan",
"country": "France"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This papers aims at improving spoken language modeling (LM) using very large amount of automatically transcribed speech. We leverage the INA (French National Audiovisual Institute 1) collection and obtain 19GB of text after applying ASR on 350,000 hours of diverse TV shows. From this, spoken language models are trained either by fine-tuning an existing LM (FlauBERT 2) or through training a LM from scratch. The new models (FlauBERT-Oral) are shared with the community 3 and are evaluated not only in terms of word prediction accuracy but also for two downstream tasks: classification of TV shows and syntactic parsing of speech. Experimental results show that FlauBERT-Oral is better than its initial FlauBERT version demonstrating that, despite its inherent noisy nature, ASR-Generated text can be useful to improve spoken language modeling.",
"pdf_parse": {
"paper_id": "2022",
"_pdf_hash": "",
"abstract": [
{
"text": "This papers aims at improving spoken language modeling (LM) using very large amount of automatically transcribed speech. We leverage the INA (French National Audiovisual Institute 1) collection and obtain 19GB of text after applying ASR on 350,000 hours of diverse TV shows. From this, spoken language models are trained either by fine-tuning an existing LM (FlauBERT 2) or through training a LM from scratch. The new models (FlauBERT-Oral) are shared with the community 3 and are evaluated not only in terms of word prediction accuracy but also for two downstream tasks: classification of TV shows and syntactic parsing of speech. Experimental results show that FlauBERT-Oral is better than its initial FlauBERT version demonstrating that, despite its inherent noisy nature, ASR-Generated text can be useful to improve spoken language modeling.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Large language models are trained with massive texts which do not reflect well the specific aspects of spoken language. Hence, modeling spoken language is challenging as crawling 'oralstyle' transcripts is a difficult task. To overcome this, our pilot study investigates the use of massive automatic speech recognition (ASR) generated text for spoken language modeling. We believe that this methodology could bring diversity (oral/spontaneous style, different topics) to the language modeling data. This might be also useful for languages with fewer text resources but potential high availability of speech recordings. We also see long-term benefits to using ASR generated text as speech recordings convey potentially useful metadata (ex: male/female speech) that could be leveraged for building LMs from more balanced data. Finally, as speech transcripts are naturally grounded with other modalities (if extracted from videos for instance), ASR could help building large scale multimodal language understanding corpora.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The contributions of this paper are the following:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 we build and share FlauBERT-Oral models from a massive amount (350,000 hours) of French TV shows,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 we evaluate them on word prediction (on both written and spoken corpora), automatic classification of TV shows and speech syntactic parsing,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 we demonstrate that ASR-Generated text can be useful for spoken LM.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We mention here related works to better position our approach: learning LMs from spoken transcripts, multimodal models and using LMs to rescore ASR. Learning LMs from spoken transcripts. probes BERT based language models (BERT, RoBERTa) trained on spoken transcripts to investigate their ability to encode properties of spoken language. Their empirical results show that LM is surprisingly good at capturing conversational properties such as pause prediction and overtalk detection from lexical tokens. But their LMs evaluated are mostly trained on clean (non ASR) spoken transcripts except one called ASRoBERTa which is trained on 2000h of transcribed speech only (1k Librispeech + 1k proprietary dataset). As a comparison with this study, we train our models on 175x more ASR data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "2"
},
{
"text": "Multimodal models. While our approach uses ASR to build text-based spoken language models, Chuang et al. (2019) proposed an audio-andtext jointly learned SpeechBERT model for spoken question answering task. They show their model is able to extract information out of audio data that is complementary to (noisy) ASR output text. The architecture proposed by is different in the sense that it learns a joint language model with phoneme sequence and ASR transcript to learn phonetic-aware representations that are robust to ASR errors (not exactly a multimodal model). While speech or multimodal unsupervised representation learning is an interesting direction, this is out of the scope of this paper which focuses on language modeling from text transcripts only.",
"cite_spans": [
{
"start": 91,
"end": 111,
"text": "Chuang et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "2"
},
{
"text": "BERT for ASR re-ranking. We also mention here LMs to rescore ASR as this could be an interesting application of our proposed spoken language models. Chiu and Chen (2021) used BERT models for reranking of N-best hypotheses produced by automatic speech recognition (ASR). Their experiments on the AMI benchmark demonstrate the effectiveness of the approach in comparison to RNNbased re-ranking. A similar idea is introduced by Fohr and Illina (2021) where BERT features are added to the neural re-ranker used to rescore ASR hypotheses. Even more recently, Xu et al. (2022) showed how to train a BERT-based rescoring model to incorporate a discriminative loss into the finetuning step of deep bidirectional pretrained models for ASR.",
"cite_spans": [
{
"start": 425,
"end": 447,
"text": "Fohr and Illina (2021)",
"ref_id": "BIBREF5"
},
{
"start": 554,
"end": 570,
"text": "Xu et al. (2022)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "2"
},
{
"text": "3 From FlauBERT to FlauBERT-Oral",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "2"
},
{
"text": "The speech recognition system used to produce the text transcripts for this study was built using Kaldi (Povey et al., 2011 ). The acoustic model is based on the lattice-free MMI, so-called \"chain\" model (Povey et al., 2016) . We used a time-delay neural network (Peddinti et al., 2015 ) and a discriminative training on the top of it using the state-level minimum Bayes risk (sMBR) criterion (Vesel\u1ef3 et al., 2013) . For the acoustic model training, we used several TV and RADIO corpora (ESTER 1&2 (Galliano et al., 2009) , REPERE (Giraudel et al., 2012) and VERA (Goryainova et al., 2014) ). A regular backoff n-gram model was estimated using the speech transcripts augmented with several French newspapers (see section 4.2.3 in Del\u00e9glise et al. (2009) ) using SRILM.",
"cite_spans": [
{
"start": 104,
"end": 123,
"text": "(Povey et al., 2011",
"ref_id": "BIBREF18"
},
{
"start": 204,
"end": 224,
"text": "(Povey et al., 2016)",
"ref_id": "BIBREF19"
},
{
"start": 263,
"end": 285,
"text": "(Peddinti et al., 2015",
"ref_id": "BIBREF17"
},
{
"start": 393,
"end": 414,
"text": "(Vesel\u1ef3 et al., 2013)",
"ref_id": "BIBREF21"
},
{
"start": 498,
"end": 521,
"text": "(Galliano et al., 2009)",
"ref_id": "BIBREF6"
},
{
"start": 531,
"end": 554,
"text": "(Giraudel et al., 2012)",
"ref_id": "BIBREF7"
},
{
"start": 564,
"end": 589,
"text": "(Goryainova et al., 2014)",
"ref_id": "BIBREF9"
},
{
"start": 730,
"end": 753,
"text": "Del\u00e9glise et al. (2009)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "ASR system",
"sec_num": "3.1"
},
{
"text": "A 2-gram decoding is performed, followed by a 3-gram and a 4-gram rescoring step. The LM interpolation weights between the different data sources were optimized on the REPERE (Giraudel et al., 2012) development corpus. The vocabulary contains the 160k most frequents words in the manually transcribed train corpus. Automatic speech diarization of the INA collection was performed using the open source toolkit LIUMSpkDiarization (Meignier and Merlin, 2010 The transcripts used in these experiments were taken from time slots corresponding to news programmes on French television and radio between 2013 and 2020. We transcribed the continuous news media between 6am and midnight each day (BFMTV, LCI, CNews, France 24, France Info and franceinfo). For radio, the morning news were used (Europe1, RMC, RTL, France Inter) and for generalist television channels we transcribed the evening news (TF1, France 2, France 3, M6). A total of 350,000 hours were automatically transcribed. The system we use provides us with raw text, without punctuation or capitalization. In order to have a pseudo sentence tokenization, we leverage the speaker diarization output to segment our transcriptions into \"sentences\". We end up with a total of 51M unique speech segments for a total of 3.5G words (19GB of data). The ASR generated text is strongly biased towards news content.",
"cite_spans": [
{
"start": 175,
"end": 198,
"text": "(Giraudel et al., 2012)",
"ref_id": "BIBREF7"
},
{
"start": 429,
"end": 455,
"text": "(Meignier and Merlin, 2010",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "ASR system",
"sec_num": "3.1"
},
{
"text": "The initial French language model (FBU), trained in 2020 on natural text, is FlauBERT (Le et al., 2020) . Models of different sizes were trained using masked language modeling (MLM) following a RoBERTa architecture and using the CNRS Jean Zay supercomputer. They were shared on HuggingFace. 4 For comparison, these models were trained on 71GB of natural text. Following the architecture of Le et al. (2020) , we propose several learning configurations in order to observe the impact of different parameters on the performance of the models obtained. Since we only have lowercase transcripts, we consider the flaubert-base-uncased model as our reference. 5 The first configuration, FlauBERT-O-base_uncased (FT), consists in fine-tuning the public flaubert-base-uncased model for some epochs using our ASR transcripts.",
"cite_spans": [
{
"start": 86,
"end": 103,
"text": "(Le et al., 2020)",
"ref_id": "BIBREF12"
},
{
"start": 291,
"end": 292,
"text": "4",
"ref_id": null
},
{
"start": 390,
"end": 406,
"text": "Le et al. (2020)",
"ref_id": "BIBREF12"
},
{
"start": 654,
"end": 655,
"text": "5",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fine-tuning or re-training FlauBERT-Oral",
"sec_num": "3.3"
},
{
"text": "The second configuration FlauBERT-O-mixed (MIX) is a full model re-trained using a mix of ASR text and written text, as training data. Written text comes from two main sources: the French wikipedia dump and press articles captured by the OTMedia research platform (Herv\u00e9, 2019 ) (online press and AFP agency for the same time period). Overall, this learning dataset is also strongly newsoriented. For the written text, we use the same sentence segmentation tool as the one used for FlauBERT. Our dataset is balanced between ASR and written text: we use 94M randomly selected written text sentences representing 13G of data to which we removed the punctuation and capitalization to make it consistent with our ASR data. For this mixed model, we also retrain the BPE tokenizer (50K sub-word units).",
"cite_spans": [
{
"start": 264,
"end": 276,
"text": "(Herv\u00e9, 2019",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fine-tuning or re-training FlauBERT-Oral",
"sec_num": "3.3"
},
{
"text": "The third configuration, FlauBERT-O-asr, consists in re-training LMs from scratch using ASR data only. For the first model (ORAL), we use the tokenizer provided with the flaubert-base-uncased model and for the second one (ORAL_NB) we retrain a BPE tokenizer (50K sub-word units). Both tokenizers share 35088 (overlap) out of 67536 (FlauBERT initial) tokens, only 52% overlap.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fine-tuning or re-training FlauBERT-Oral",
"sec_num": "3.3"
},
{
"text": "These different configurations therefore provide us with 4 language models to evaluate. Training was done on a single server with 2 Xeon CPUs of 12 cores each, 256 GB of RAM and 8 Nvidia GeForce RTX 2080 Ti graphics cards with 11 GB of memory. With this hardware, it took us 15 days to train 50 epochs of each model in the flaubert-base configuration (137M parameters) using FlauBERT code.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fine-tuning or re-training FlauBERT-Oral",
"sec_num": "3.3"
},
{
"text": "The first step in evaluating our models is to look at their behaviour for the word prediction task. In addition to the performance on the trained models, we also want to have an idea of the performance on texts of different nature (written style or oral style). We therefore assembled several datasets to measure the word prediction performance of the models we trained.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word Prediction Experiments",
"sec_num": "4"
},
{
"text": "We make sure that these datasets are not included in the training data of the default FlauBERT model nor in our own. We have a first corpus (afp2021) of AFP dispatches from the year 2021, i.e. after the period of our training data collected from the online press. This will allow us to have a measure of performance on written text. Secondly, we want to evaluate our models on oral texts. We use the transcripts of the French National Assembly sessions. 6 We are using the 13th (under Sarkozy parl_13) and 15th (currently under Macron parl_15) mandates. These texts are a manual transcription of what is said in the hemicycle, which are prepared speeches with some degree of spontaneous style as well. A second corpus is constituted with, once again, the manual transcriptions made for educational videos 7 and interviews 8 that INA makes available via its web studio (studio_manual). These transcriptions are of very good quality. We also transcribed these videos from the studio with our ASR system (studio_asr) in order to be able to compare the performance on both types of data.",
"cite_spans": [
{
"start": 454,
"end": 455,
"text": "6",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Word Prediction Experiments",
"sec_num": "4"
},
{
"text": "We report in the graphs the accuracy obtained on the different datasets for a word prediction task after a word has been masked. The masking parameters are the same as those used during training with MLM loss. flaubert_base_uncased model (FBU). For the finetuned FlauBERT-O-base_uncased model, we notice a slight improvement in performance for afp and studio datasets, obtained from the first epoch, which means that adding ASR generated text improves word prediction task on these datasets. We observe that globally, whatever the model, the datasets of the parliamentary sessions are those for which the best performances are obtained on the word prediction task, even exceeding that of the training dataset for the FlauBERT-O-base_uncased and FlauBERT-O-mixed models. These models are trained on written and spoken texts and it is not surprising that the performance is good since the very nature of the parliamentary data is a mix-ture of prepared and spontaneous speech. There is no significant difference between parl_13 and parl_15. On these parlementary speeches, there is no significant performance difference between the 3 models that have seen written text during their training (FBU, FT and MIX). As we observed also that our FlauBERT-O models improve also on written text (afp2021), we explain this by the fact that those texts are strongly related to news events, so they are in a similar context to our ASR data which is focused on news slot transcripts. For the last corpus, from the INA web studio, we have educational videos or interviews of personalities which are more distant from news data. There is a great disparity in performance depending on whether we consider manual (studio_manual) or automatic (stu-dio_asr) transcription. We believe that the different sentence segmentation algorithms have a very clear impact on this corpus. Finally, we notice that the ORAL_NB model performs slightly worse than the ORAL model. The BPE tokenizer obviously has an impact on the overall performance of the LMs and it seems, from this result, that using BPE units extracted from clean data (and not noisy ASR data) is beneficial even if the training material is itself ASR generated text. We evaluate our different models on a news classification task. For the main generalist channels, INA's documentalists finely segment the newscasts and annotate them in order to describe their content. This very rich metadata is used in particular to establish quantitative studies on the news in France. The InaStat barometer 9 has set up a stable method-ology over time to classify these news items into 14 categories (such as society, French politics, sport or environment). We use the news items of 4 channels (TF1, France 2, France 3 and M6) for the years 2017, 2018 and 2019, which gives us a total of 47 867 short TV shows. The average length of these shows is 92 seconds.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word Prediction Experiments",
"sec_num": "4"
},
{
"text": "The objective is to assess to what extent it is possible to classify these topics into the 14 categories solely on the basis of what is pronounced, i.e. from the ASR transcripts. We establish a baseline using a simple SVM classifier (with a non-parametric triangular kernel) on TF-IDF vectors with two vocabulary sizes of 5K and 20K words. To test the FlauBERT models, we use the HuggingFace Transformers library and the FlaubertForSequenceClassification class, which adds a simple dense classification layer on top of our models. To obtain a vector representation of our texts before this classification layer, we use the 'mean' summary type. We do not make any model selection and report the results for all learning epochs. Since the 14 categories are not well-balanced, we use the weighted F1 measure to evaluate the performance. The experiments are systematically performed on 10 different random splits of the dataset, taking into account the cardinality of the 14 categories, so as to have 38K examples for the training set and 5K for the test set. We show the average results and the standard deviation in figure 5. We can see in this configuration the contribution of the LMs compared to the SVMs along the training epochs of the classifier. If we look at the performance at the first epoch, we can see that the flaubert_base_uncased model has almost equivalent performance to the SVM (0.78). It is only after a few iterations of learning that the model fits the ASR data and reaches 0.81. On the other hand, the models that have already seen ASR data during ina-stat-sommaire.html their training have a better performance from the first epoch. The model trained only on ASR data is the best performing (ORAL). After 10 epochs, the 3 FlauBERT-Oral models converge and are equivalent for this task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Standard Learning Setting",
"sec_num": "5.1"
},
{
"text": "In order to test the LMs under more challenging conditions, we progressively reduce the number of training examples to get closer to few-shot learning conditions. We thus restart the classification with 5K training examples, then 500 and finally 200. Again, we take into account the cardinality of the 14 categories. For the last experiment with only 200 training examples, the vocabulary is too small and we can only test the SVM baseline with a vocabulary of 5K words, but not the version with 20K words. Moreover, we push to 30 epochs in this latter case. As the number of training examples decreases, the performance gain over SVMs becomes more obvious. This is an expected result. In all cases, the models trained on ASR only text (ORAL) are the best of the FlauBERT-O models. Compared to the ORAL_NB model, only the tokenizer is different. This result may appear counter-intuitive in a first place, as one would expect a model entirely learned on ASR data to perform better on a classification task using only ASR data as input. However, this is probably counterbalanced by the fact that using BPE units extracted from clean data is important (as we have seen in the word prediction experiments). This invites us to further investigate the role of the tokenizer in spoken language modeling. As in the previous case, the Flaubert models converge almost with a 2 F1 point difference in favour of the FlauBERT-O models over the initial FlauBERT model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Few Shot Learning Setting",
"sec_num": "5.2"
},
{
"text": "q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Few Shot Learning Setting",
"sec_num": "5.2"
},
{
"text": "This section is about the downstream task of jointly predicting part of speech tags (POS) and building a labelled dependency tree. The models performing these tasks typically rely on word representations, that are often pretrained, especially when the data is scarce. We will use our different spoken language models to obtain contextual word representations of a syntactically annotated and manually transcribed oral French corpus. For each of these representations, a model will be trained to perform the joint prediction of POS tags and labelled dependencies. We also use as baseline a model trained using non-contextual representations obtained with FastText, 10 and a model learning its own representations without any pretraining.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Downstream Task 2: Syntactic Analysis of Spoken Conversations",
"sec_num": "6"
},
{
"text": "We used the annotated subset of the speech corpus of the Orfeo project (Benzitoun et al., 2016; Nasr et al., 2020) , gathered with the goal of reflecting the contemporary usage of the French language.",
"cite_spans": [
{
"start": 71,
"end": 95,
"text": "(Benzitoun et al., 2016;",
"ref_id": "BIBREF0"
},
{
"start": 96,
"end": 114,
"text": "Nasr et al., 2020)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "6.1"
},
{
"text": "The audio extracts on which this corpus is based come from various origins and modalities: from one to multiple speakers, work meetings, family dinner conversations, narration, political meeting, interview, goal-oriented telephone conversations. Their duration varies from four minutes to an hour.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "6.1"
},
{
"text": "The reference audio transcripts have been obtained after correcting the output of an ASR system. The corpus is annotated in part of speech (POS) tags, lemmas, labeled dependency trees and sentences boundaries. There are 20 possible POS tags and 12 syntactic functions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "6.1"
},
{
"text": "We randomly split the corpus into train/dev/test sets of respective sizes 134,716/27,937/29,529 words; we sampled from each source so that the various origins of the audios are equally represented in each split.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "6.1"
},
{
"text": "The model is a transition based parser using the arc-eager transition system (Nivre, 2008) , which has been extended for the joint prediction of POS tags and parsing transitions (Dary and Nasr, 2021) .",
"cite_spans": [
{
"start": 77,
"end": 90,
"text": "(Nivre, 2008)",
"ref_id": "BIBREF16"
},
{
"start": 178,
"end": 199,
"text": "(Dary and Nasr, 2021)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "It consists of a single classifier, taking as input a numeric representation of the current state of the analysis, called a configuration. The classifier predicts a probability distribution over the set of POS tagging actions or parsing actions, depending of the current state of the configuration. The analysis assume that the text is already tokenized and segmented into sentences; the words of each sentence are considered one by one, in the reading order; a POS action is predicted for the current word, then a sequence of arc-eager actions is predicted until the current word is either attached to a word on its left or shifted to a stack for future attachment to a word on its right. The predictions are greedy: it is always the top scoring action among the allowed ones. We do not use beam search for decoding.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "The numeric representation of the current configuration is comprised of:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "\u2022 The concatenation of the word embeddings, reduced from dimension 768 to dimension 64 by a linear layer, of the following context: the current word, the three preceding ones, the two following ones, the three topmost stack elements and the rightmost and leftmost dependents of the three topmost stack elements,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "\u2022 The output of three different BiLSTM processing sequences of tags of the same nature. The first one is taking as input the sequence of POS tags and syntactic function of the current word, the three previous ones and the three topmost stack elements. The second one is taking the sequence of the last 10 actions that have been applied to this configuration. The last one is taking the sequence of distances (in number of words) between the current word and the three topmost stack elements. In each case, the sequence elements are encoded by learnable and randomly initialized embeddings of size 128, and the output of the BiLSTM is a vector of size 128,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "\u2022 A learnable and randomly initialized embedding encoding the current state of the configuration (POS tagging or dependency parsing).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "A dropout of 50% is applied to the resulting vector; then it passes through two hidden layers of respective sizes 3200 and 1600, both with a dropout of 40% and a ReLU activation. Finally, the network is ended by one of the two decision layers, depending on the current state, which is simply a linear layer of dimension the number of possible actions followed by a softmax.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "Each model was trained for 40 epochs; after every epoch the model was evaluated on the dev set and was saved if it was an improvement. After the fourth epoch, the entire train set was decoded using the model that was being trained, in order to generate and integrate novel configurations in the dataset for the epochs to come. This technique allows the model to be more robust, exploring nonoptimal configurations during its training. It is based on the dynamical oracle model of Goldberg and Nivre (2012) .",
"cite_spans": [
{
"start": 480,
"end": 505,
"text": "Goldberg and Nivre (2012)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing Model",
"sec_num": "6.2"
},
{
"text": "The first set of experiments compares input representations from the FlauBERT variants (FBU, MIX, ORAL) to uncontextual word embeddings (Fasttext) and randomly initialized embeddings. Except for random embeddings, token representations are frozen when the parsing system is trained.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "6.3"
},
{
"text": "As pre-processing, we deanonymize the transcripts by replacing masked proper name tokens with non-ambiguous names randomly chosen for each recording. In the fasttext setting, representations are computed for unknown words from their character n-gram factors. Contextual representations are computed at the whole recording level in chunks of 512 tokens without overlap. The parser is applied on the reference transcript and reference segmentation. We use mean pooling for words that are split in multiple tokens by BPE.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "6.3"
},
{
"text": "Parsing performance is evaluated with Labeled Attachment Score (LAS), the accuracy of predicting the governor of each word and its dependency label, Unlabeled Attachment Score (UAS), which ignores the dependency label, and Part-of-speech tagging accuracy (UPOS). The scoring script is from CoNLL campaigns. Results presented in Table 3 show that pretraining is valuable for syntactic parsing in that setting and that pretraining on ASR (MIX and ORAL) leads to a substancial improvement in LAS over the text-only FlauBERT model (FBU) even though there is no domain overlap between the TV shows on which the earlier is trained and the data of the Orfeo corpus. There is no benefit from retraining BPE (ORAL_NB). As noted earlier, speech recordings do not have punctuation and it is debated whether punctuation is suitable for spontaenous conversations. As punctuation is rather regular in text, it would make sense for LMs trained on text to over-rely on the cues it brings, and representations to be affected by a lack of punctuation. Table 4 shows syntactic parsing results on representations where a simple heuristic is applied to add a period at the end of each sentence prior to extracting representations. This punctuation is stripped before passing the tokens to the syntactic parser and only used at the encoding stage. Results show that most of the difference in performance between the FBU and ORAL models can be compensated by this use of virtual punctuation. Using accurately predicted punctuation with diverse symbols and intra-sentence marks is Repr. 95.55 79.00 -16.55 Table 5 : Syntactic parsing performance on OOV words according to automatic transcription system. The \u2206 column contains the difference between the global accuracy and the accuracy on OOVs only.",
"cite_spans": [],
"ref_spans": [
{
"start": 328,
"end": 335,
"text": "Table 3",
"ref_id": "TABREF5"
},
{
"start": 1034,
"end": 1041,
"text": "Table 4",
"ref_id": "TABREF7"
},
{
"start": 1582,
"end": 1589,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "6.3"
},
{
"text": "LAS UAS UPOS Global OOV \u2206 Global OOV \u2206 Global OOV \u2206 FBU",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Repr",
"sec_num": null
},
{
"text": "left as future work, but we conjecture that it will marginally improve over this crude heuristic. Gauging the impact of speech-to-text errors on representations from LMs trained on such data is difficult since there are no manual references available for large quantities of speech transcripts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Repr",
"sec_num": null
},
{
"text": "Since the system used to transcribe the recordings is closed vocabulary, one way to look at this problem is to compute the accuracy of the syntactic parser on words that are out-of-vocabulary (OOV) for the LM training data. Due to BPE, those words are necesseraly tokenized in smaller units which are pooled prior to passing them to the parser, and might hamper the quality of the associated representations. Table 5 details the performance of the syntactic parser on OOVs. Due to their infrequent nature, OOVs are mainly swear words, proper names, and tokenization artifacts. They are difficult to handle for all models, and suffer from a large performance reduction compared to the global figure, even for the FBU model which has seen a much larger variety of texts. The system fed with representations of the model trained on ASR data only (ORAL) is the most affected despite its better global performance. Finally, Figure 9 shows the learning curve when reducing the training data available to the syntactic parser. For this, we randomly sampled 10 subsets of the training data at the recording level in order to fit a target ratio from 2.5% to 100%. The figure shows that LAS is always better for ORAL representations and that MIX is closer to FBU when less data is available.",
"cite_spans": [],
"ref_spans": [
{
"start": 409,
"end": 416,
"text": "Table 5",
"ref_id": null
},
{
"start": 919,
"end": 927,
"text": "Figure 9",
"ref_id": "FIGREF7"
}
],
"eq_spans": [],
"section": "Repr",
"sec_num": null
},
{
"text": "It seems that exploiting ASR transcripts for learning LMs is beneficial for syntactic parsing of speech transcripts. Analyses presented show that punctuation plays an important role in representations. Our analysis of parsing performance on OOV words (according to the speech-to-text system) reveals that our FlauBERT-O-asr (ORAL) model is more affected than its initial FlauBERT baseline (FBU), despite overall better performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Takeaways",
"sec_num": "6.4"
},
{
"text": "We investigated spoken language modeling using ASR generated text (350,000 hours of diverse TV shows). The new models for French (FlauBERT-O) are shared with the community. Experimental results show that FlauBERT-O is generally better than its initial FlauBERT version for the downstream speech tasks we experimented with. However we should also check its performance on text downstream tasks (such as (Le et al., 2020) ) and on more downstream speech tasks (SLU or ASR re-scoring).",
"cite_spans": [
{
"start": 402,
"end": 419,
"text": "(Le et al., 2020)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and future work",
"sec_num": "7"
},
{
"text": "In this work, all our texts were uncased as our ASR only generates lowercased transcripts. We believe that applying massively re-capitalisation (and restoring punctuation as well) might be beneficial to train stronger spoken LMs. We also plan to analyze more the specificities of our ASR-generated texts (do they contain more oral features such as word repetitions, more interjections?). Finally, some of the results obtained lead us to believe that it is important to further evaluate the impact of BPE units for spoken language modeling.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and future work",
"sec_num": "7"
},
{
"text": "https://www.ina.fr 2 https://github.com/getalp/Flaubert 3 https://huggingface.co/nherve",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://huggingface.co/flaubert",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://huggingface.co/flaubert/ flaubert_base_uncased",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://data.assemblee-nationale.fr/ 7 https://www.ina.fr/ offres-et-services/fresques-numeriques 8 https://entretiens.ina.fr/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://www.inatheque.fr/ publications-evenements/ina-stat/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://fasttext.cc",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Le projet orf\u00e9o: un corpus d'\u00e9tude pour le fran\u00e7ais contemporain",
"authors": [
{
"first": "Christophe",
"middle": [],
"last": "Benzitoun",
"suffix": ""
},
{
"first": "Jeanne-Marie",
"middle": [],
"last": "Debaisieux",
"suffix": ""
},
{
"first": "Henri-Jos\u00e9",
"middle": [],
"last": "Deulofeu",
"suffix": ""
}
],
"year": 2016,
"venue": "Corpus",
"volume": "",
"issue": "15",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christophe Benzitoun, Jeanne-Marie Debaisieux, and Henri-Jos\u00e9 Deulofeu. 2016. Le projet orf\u00e9o: un cor- pus d'\u00e9tude pour le fran\u00e7ais contemporain. Corpus, (15).",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Innovative bert-based reranking language models for speech recognition",
"authors": [
{
"first": "",
"middle": [],
"last": "Shih-Hsuan",
"suffix": ""
},
{
"first": "Berlin",
"middle": [],
"last": "Chiu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2021,
"venue": "IEEE Spoken Language Technology Workshop (SLT)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1109/slt48900.2021.9383557"
]
},
"num": null,
"urls": [],
"raw_text": "Shih-Hsuan Chiu and Berlin Chen. 2021. Innovative bert-based reranking language models for speech recognition. 2021 IEEE Spoken Language Technol- ogy Workshop (SLT).",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Speechbert: Cross-modal pre-trained language model for end-to-end spoken question answering",
"authors": [
{
"first": "Yung-Sung",
"middle": [],
"last": "Chuang",
"suffix": ""
},
{
"first": "Chi-Liang",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Hung-Yi",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yung-Sung Chuang, Chi-Liang Liu, and Hung-yi Lee. 2019. Speechbert: Cross-modal pre-trained language model for end-to-end spoken question answering. CoRR, abs/1910.11559.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "The reading machine: A versatile framework for studying incremental parsing strategies",
"authors": [
{
"first": "Franck",
"middle": [],
"last": "Dary",
"suffix": ""
},
{
"first": "Alexis",
"middle": [],
"last": "Nasr",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the 17th International Conference on Parsing Technologies and the IWPT 2021 Shared Task on Parsing into Enhanced Universal Dependencies (IWPT 2021)",
"volume": "",
"issue": "",
"pages": "26--37",
"other_ids": {
"DOI": [
"10.18653/v1/2021.iwpt-1.3"
]
},
"num": null,
"urls": [],
"raw_text": "Franck Dary and Alexis Nasr. 2021. The reading ma- chine: A versatile framework for studying incremen- tal parsing strategies. In Proceedings of the 17th International Conference on Parsing Technologies and the IWPT 2021 Shared Task on Parsing into En- hanced Universal Dependencies (IWPT 2021), pages 26-37, Online. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Improvements to the lium french asr system based on cmu sphinx: what helps to significantly reduce the word error rate?",
"authors": [
{
"first": "Paul",
"middle": [],
"last": "Del\u00e9glise",
"suffix": ""
},
{
"first": "Yannick",
"middle": [],
"last": "Esteve",
"suffix": ""
},
{
"first": "Sylvain",
"middle": [],
"last": "Meignier",
"suffix": ""
},
{
"first": "Teva",
"middle": [],
"last": "Merlin",
"suffix": ""
}
],
"year": 2009,
"venue": "Tenth Annual Conference of the International Speech Communication Association",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Paul Del\u00e9glise, Yannick Esteve, Sylvain Meignier, and Teva Merlin. 2009. Improvements to the lium french asr system based on cmu sphinx: what helps to signif- icantly reduce the word error rate? In Tenth Annual Conference of the International Speech Communica- tion Association.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "BERT-based Semantic Model for Rescoring N-best Speech Recognition List",
"authors": [
{
"first": "Dominique",
"middle": [],
"last": "Fohr",
"suffix": ""
},
{
"first": "Irina",
"middle": [],
"last": "Illina",
"suffix": ""
}
],
"year": 2021,
"venue": "INTERSPEECH 2021, Proceedings of INTERSPEECH 2021",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dominique Fohr and Irina Illina. 2021. BERT-based Semantic Model for Rescoring N-best Speech Recog- nition List. In INTERSPEECH 2021, Proceedings of INTERSPEECH 2021, Brno, Czech Republic.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "The ester 2 evaluation campaign for the rich transcription of french radio broadcasts",
"authors": [
{
"first": "Sylvain",
"middle": [],
"last": "Galliano",
"suffix": ""
},
{
"first": "Guillaume",
"middle": [],
"last": "Gravier",
"suffix": ""
},
{
"first": "Laura",
"middle": [],
"last": "Chaubard",
"suffix": ""
}
],
"year": 2009,
"venue": "Tenth Annual Conference of the International Speech Communication Association",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sylvain Galliano, Guillaume Gravier, and Laura Chaubard. 2009. The ester 2 evaluation campaign for the rich transcription of french radio broadcasts. In Tenth Annual Conference of the International Speech Communication Association.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The repere corpus: a multimodal corpus for person recognition",
"authors": [
{
"first": "Aude",
"middle": [],
"last": "Giraudel",
"suffix": ""
},
{
"first": "Matthieu",
"middle": [],
"last": "Carr\u00e9",
"suffix": ""
},
{
"first": "Val\u00e9rie",
"middle": [],
"last": "Mapelli",
"suffix": ""
},
{
"first": "Juliette",
"middle": [],
"last": "Kahn",
"suffix": ""
},
{
"first": "Olivier",
"middle": [],
"last": "Galibert",
"suffix": ""
},
{
"first": "Ludovic",
"middle": [],
"last": "Quintard",
"suffix": ""
}
],
"year": 2012,
"venue": "LREC",
"volume": "",
"issue": "",
"pages": "1102--1107",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aude Giraudel, Matthieu Carr\u00e9, Val\u00e9rie Mapelli, Juliette Kahn, Olivier Galibert, and Ludovic Quintard. 2012. The repere corpus: a multimodal corpus for person recognition. In LREC, pages 1102-1107.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "A dynamic oracle for arc-eager dependency parsing",
"authors": [
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
},
{
"first": "Joakim",
"middle": [],
"last": "Nivre",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of COLING 2012",
"volume": "",
"issue": "",
"pages": "959--976",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoav Goldberg and Joakim Nivre. 2012. A dynamic ora- cle for arc-eager dependency parsing. In Proceedings of COLING 2012, pages 959-976.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Morpho-syntactic study of errors from speech recognition system",
"authors": [
{
"first": "Maria",
"middle": [],
"last": "Goryainova",
"suffix": ""
},
{
"first": "Cyril",
"middle": [],
"last": "Grouin",
"suffix": ""
},
{
"first": "Sophie",
"middle": [],
"last": "Rosset",
"suffix": ""
},
{
"first": "Ioana",
"middle": [],
"last": "Vasilescu",
"suffix": ""
}
],
"year": 2014,
"venue": "LREC",
"volume": "14",
"issue": "",
"pages": "3050--3056",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maria Goryainova, Cyril Grouin, Sophie Rosset, and Ioana Vasilescu. 2014. Morpho-syntactic study of errors from speech recognition system. In LREC, volume 14, pages 3050-3056.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "OTMedia, the TransMedia news observatory",
"authors": [
{
"first": "Nicolas",
"middle": [],
"last": "Herv\u00e9",
"suffix": ""
}
],
"year": 2019,
"venue": "FIAT/IFTA Media Management Seminar",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nicolas Herv\u00e9. 2019. OTMedia, the TransMedia news observatory. In FIAT/IFTA Media Management Sem- inar 2019.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "What BERT based language model learns in spoken transcripts: An empirical study",
"authors": [
{
"first": "Ayush",
"middle": [],
"last": "Kumar",
"suffix": ""
},
{
"first": "Jithendra",
"middle": [],
"last": "Mukuntha Narayanan Sundararaman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Vepa",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP",
"volume": "",
"issue": "",
"pages": "322--336",
"other_ids": {
"DOI": [
"10.18653/v1/2021.blackboxnlp-1.25"
]
},
"num": null,
"urls": [],
"raw_text": "Ayush Kumar, Mukuntha Narayanan Sundararaman, and Jithendra Vepa. 2021. What BERT based lan- guage model learns in spoken transcripts: An empiri- cal study. In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Net- works for NLP, pages 322-336, Punta Cana, Do- minican Republic. Association for Computational Linguistics.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "FlauBERT: Unsupervised language model pre-training for French",
"authors": [
{
"first": "Hang",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Lo\u00efc",
"middle": [],
"last": "Vial",
"suffix": ""
},
{
"first": "Jibril",
"middle": [],
"last": "Frej",
"suffix": ""
},
{
"first": "Vincent",
"middle": [],
"last": "Segonne",
"suffix": ""
},
{
"first": "Maximin",
"middle": [],
"last": "Coavoux",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Lecouteux",
"suffix": ""
},
{
"first": "Alexandre",
"middle": [],
"last": "Allauzen",
"suffix": ""
},
{
"first": "Benoit",
"middle": [],
"last": "Crabb\u00e9",
"suffix": ""
},
{
"first": "Laurent",
"middle": [],
"last": "Besacier",
"suffix": ""
},
{
"first": "Didier",
"middle": [],
"last": "Schwab",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 12th Language Resources and Evaluation Conference",
"volume": "",
"issue": "",
"pages": "2479--2490",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hang Le, Lo\u00efc Vial, Jibril Frej, Vincent Segonne, Max- imin Coavoux, Benjamin Lecouteux, Alexandre Al- lauzen, Benoit Crabb\u00e9, Laurent Besacier, and Didier Schwab. 2020. FlauBERT: Unsupervised language model pre-training for French. In Proceedings of the 12th Language Resources and Evaluation Confer- ence, pages 2479-2490, Marseille, France. European Language Resources Association.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Lium spkdiarization: an open source toolkit for diarization",
"authors": [
{
"first": "Sylvain",
"middle": [],
"last": "Meignier",
"suffix": ""
},
{
"first": "Teva",
"middle": [],
"last": "Merlin",
"suffix": ""
}
],
"year": 2010,
"venue": "CMU SPUD Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sylvain Meignier and Teva Merlin. 2010. Lium spkdi- arization: an open source toolkit for diarization. In CMU SPUD Workshop.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Annotation syntaxique automatique de la partie orale du orf\u00e9o",
"authors": [
{
"first": "Alexis",
"middle": [],
"last": "Nasr",
"suffix": ""
},
{
"first": "Franck",
"middle": [],
"last": "Dary",
"suffix": ""
}
],
"year": 2020,
"venue": "Langages",
"volume": "",
"issue": "3",
"pages": "87--102",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexis Nasr, Franck Dary, Fr\u00e9d\u00e9ric Bechet, and Beno\u00eet Fabre. 2020. Annotation syntaxique automatique de la partie orale du orf\u00e9o. Langages, (3):87-102.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Algorithms for deterministic incremental dependency parsing",
"authors": [
{
"first": "Joakim",
"middle": [],
"last": "Nivre",
"suffix": ""
}
],
"year": 2008,
"venue": "Computational Linguistics",
"volume": "34",
"issue": "4",
"pages": "513--553",
"other_ids": {
"DOI": [
"10.1162/coli.07-056-R1-07-027"
]
},
"num": null,
"urls": [],
"raw_text": "Joakim Nivre. 2008. Algorithms for deterministic incre- mental dependency parsing. Computational Linguis- tics, 34(4):513-553.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "A time delay neural network architecture for efficient modeling of long temporal contexts",
"authors": [
{
"first": "Vijayaditya",
"middle": [],
"last": "Peddinti",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Povey",
"suffix": ""
},
{
"first": "Sanjeev",
"middle": [],
"last": "Khudanpur",
"suffix": ""
}
],
"year": 2015,
"venue": "Sixteenth Annual Conference of the International Speech Communication Association",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khu- danpur. 2015. A time delay neural network architec- ture for efficient modeling of long temporal contexts. In Sixteenth Annual Conference of the International Speech Communication Association.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "The kaldi speech recognition toolkit",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Povey",
"suffix": ""
},
{
"first": "Arnab",
"middle": [],
"last": "Ghoshal",
"suffix": ""
},
{
"first": "Gilles",
"middle": [],
"last": "Boulianne",
"suffix": ""
},
{
"first": "Lukas",
"middle": [],
"last": "Burget",
"suffix": ""
},
{
"first": "Ondrej",
"middle": [],
"last": "Glembek",
"suffix": ""
},
{
"first": "Nagendra",
"middle": [],
"last": "Goel",
"suffix": ""
},
{
"first": "Mirko",
"middle": [],
"last": "Hannemann",
"suffix": ""
},
{
"first": "Petr",
"middle": [],
"last": "Motlicek",
"suffix": ""
},
{
"first": "Yanmin",
"middle": [],
"last": "Qian",
"suffix": ""
},
{
"first": "Petr",
"middle": [],
"last": "Schwarz",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukas Burget, Ondrej Glembek, Nagendra Goel, Mirko Hannemann, Petr Motlicek, Yanmin Qian, Petr Schwarz, et al. 2011. The kaldi speech recognition toolkit. Technical report, IEEE Signal Processing Society.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Purely sequence-trained neural networks for asr based on lattice-free mmi",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Povey",
"suffix": ""
},
{
"first": "Vijayaditya",
"middle": [],
"last": "Peddinti",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Galvez",
"suffix": ""
},
{
"first": "Pegah",
"middle": [],
"last": "Ghahremani",
"suffix": ""
},
{
"first": "Vimal",
"middle": [],
"last": "Manohar",
"suffix": ""
},
{
"first": "Xingyu",
"middle": [],
"last": "Na",
"suffix": ""
},
{
"first": "Yiming",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Sanjeev",
"middle": [],
"last": "Khudanpur",
"suffix": ""
}
],
"year": 2016,
"venue": "Interspeech",
"volume": "",
"issue": "",
"pages": "2751--2755",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Povey, Vijayaditya Peddinti, Daniel Galvez, Pe- gah Ghahremani, Vimal Manohar, Xingyu Na, Yim- ing Wang, and Sanjeev Khudanpur. 2016. Purely sequence-trained neural networks for asr based on lattice-free mmi. In Interspeech, pages 2751-2755.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Phoneme-bert: Joint language modelling of phoneme sequence and asr transcript",
"authors": [
{
"first": "Ayush",
"middle": [],
"last": "Mukuntha Narayanan Sundararaman",
"suffix": ""
},
{
"first": "Jithendra",
"middle": [],
"last": "Kumar",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Vepa",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mukuntha Narayanan Sundararaman, Ayush Kumar, and Jithendra Vepa. 2021. Phoneme-bert: Joint lan- guage modelling of phoneme sequence and asr tran- script.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Sequence-discriminative training of deep neural networks",
"authors": [
{
"first": "Karel",
"middle": [],
"last": "Vesel\u1ef3",
"suffix": ""
},
{
"first": "Arnab",
"middle": [],
"last": "Ghoshal",
"suffix": ""
},
{
"first": "Luk\u00e1s",
"middle": [],
"last": "Burget",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Povey",
"suffix": ""
}
],
"year": 2013,
"venue": "In Interspeech",
"volume": "2013",
"issue": "",
"pages": "2345--2349",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Karel Vesel\u1ef3, Arnab Ghoshal, Luk\u00e1s Burget, and Daniel Povey. 2013. Sequence-discriminative training of deep neural networks. In Interspeech, volume 2013, pages 2345-2349.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Ariya Rastrow, Andreas Stolcke, and Ivan Bulyko. 2022. Rescorebert: Discriminative speech recognition rescoring with bert",
"authors": [
{
"first": "Liyan",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Yile",
"middle": [],
"last": "Gu",
"suffix": ""
},
{
"first": "Jari",
"middle": [],
"last": "Kolehmainen",
"suffix": ""
},
{
"first": "Haidar",
"middle": [],
"last": "Khan",
"suffix": ""
},
{
"first": "Ankur",
"middle": [],
"last": "Gandhe",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Liyan Xu, Yile Gu, Jari Kolehmainen, Haidar Khan, Ankur Gandhe, Ariya Rastrow, Andreas Stolcke, and Ivan Bulyko. 2022. Rescorebert: Discriminative speech recognition rescoring with bert.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"type_str": "figure",
"num": null,
"text": "FT -Word prediction accuracy of FlauBERT-O-base_uncasedFigures 1 to 4 show the results assessed at each epoch. In table 2, we summarise the results for the last epoch and also for the default ORAL -Word prediction accuracy of FlauBERT-O-asr, using the initial flaubert-baseuncased BPE tokenizer ORAL_NB -Word prediction accuracy of FlauBERT-O-asr_newbpe, using a new BPE tokenizer trained on ASR data",
"uris": null
},
"FIGREF3": {
"type_str": "figure",
"num": null,
"text": "TV news classification -train 38K, test 5K",
"uris": null
},
"FIGREF4": {
"type_str": "figure",
"num": null,
"text": "Figure 6: TV news classification -train 5K, test 38K",
"uris": null
},
"FIGREF5": {
"type_str": "figure",
"num": null,
"text": "Figure 7: TV news classification -train 500, test 47K",
"uris": null
},
"FIGREF6": {
"type_str": "figure",
"num": null,
"text": "TV news classification -train 200, test 47K",
"uris": null
},
"FIGREF7": {
"type_str": "figure",
"num": null,
"text": "LAS learning curve for syntactic parser according to quantity of training data. Similar shape is obtained for UAS and UPOS.",
"uris": null
},
"TABREF2": {
"type_str": "table",
"text": "Word prediction task accuracies",
"html": null,
"content": "<table><tr><td>5 Downstream Task 1: Automatic Classification of TV Shows</td></tr></table>",
"num": null
},
"TABREF5": {
"type_str": "table",
"text": "",
"html": null,
"content": "<table><tr><td>: Main result on syntax prediction. Metrics are Labeled Attachment Score (LAS), Unlabeled Attach-ment Score (UAS) and Part-of-speech tagging accuracy (UPOS). Higher is better, highest figure in bold.</td></tr></table>",
"num": null
},
"TABREF7": {
"type_str": "table",
"text": "Effect of repunctuating speech transcripts on syntactic parsing prior to extracting representations. Results from the ORAL representations are given for reference.",
"html": null,
"content": "<table/>",
"num": null
}
}
}
} |