File size: 101,208 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 | {
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:05:10.066095Z"
},
"title": "Translating Knowledge Representations with Monolingual Word Embeddings: the Case of a Thesaurus on Corporate Non-Financial Reporting",
"authors": [],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "A common method of structuring information extracted from textual data is using a knowledge model (e.g. a thesaurus) to organise the information semantically. Creating and managing a knowledge model is already a costly task in terms of human effort, not to mention making it multilingual. Multilingual knowledge modelling is a common problem for both transnational organisations and organisations providing text analytics that want to analyse information in more than one language. Many organisations tend to develop their language resources first in one language (often English). When it comes to analysing data sources in other languages, either a lot of effort has to be invested in recreating the same knowledge base in a different language or the data itself has to be translated into the language of the knowledge model. In this paper, we propose an unsupervised method to automatically induce a given thesaurus into another language using only comparable monolingual corpora. The aim of this proposal is to employ cross-lingual word embeddings to map the set of topics in an already-existing English thesaurus into Spanish. With this in mind, we describe different approaches to generate the Spanish thesaurus terms and offer an extrinsic evaluation by using the obtained thesaurus, which covers non-financial topics in a multi-label document classification task, and we compare the results across these approaches.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "A common method of structuring information extracted from textual data is using a knowledge model (e.g. a thesaurus) to organise the information semantically. Creating and managing a knowledge model is already a costly task in terms of human effort, not to mention making it multilingual. Multilingual knowledge modelling is a common problem for both transnational organisations and organisations providing text analytics that want to analyse information in more than one language. Many organisations tend to develop their language resources first in one language (often English). When it comes to analysing data sources in other languages, either a lot of effort has to be invested in recreating the same knowledge base in a different language or the data itself has to be translated into the language of the knowledge model. In this paper, we propose an unsupervised method to automatically induce a given thesaurus into another language using only comparable monolingual corpora. The aim of this proposal is to employ cross-lingual word embeddings to map the set of topics in an already-existing English thesaurus into Spanish. With this in mind, we describe different approaches to generate the Spanish thesaurus terms and offer an extrinsic evaluation by using the obtained thesaurus, which covers non-financial topics in a multi-label document classification task, and we compare the results across these approaches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Corporate Social Responsibility (CSR) is a concept that aims to understand, categorise, monitor and regulate the actions of corporations regarding environmental, social, governance (ESG) and technological issues. Following Fontaine (Fontaine, 2013) , one of the primary goals of CSR is to encourage corporations to engage in responsible behaviour when it comes to these issues (amongst others). CSR has become extremely relevant in the past decade. Customers, stakeholders and investors have increasingly begun to demand a robust integration of sustainable development practices into the wider business model -making sustainability a financially material issue. A growing number of policies and regulations, both voluntary and otherwise, have pushed companies towards public disclosure of non-financial (i.e. ESG) information in annual reports or stand-alone documents. The latest KPMG Survey of Corporate Responsibility Reporting (KPMG, 2017) indicates that 93% of the 250 largest companies by revenue (based on the Fortune Global 500 2016 index) have adopted non-financial reporting (NFR). These corporate-derived disclosures are not the only factor to consider. The media, which informs about corporations and how they tackle ESG issues, is also a player when it comes to shaping discourse. Individuals, too, share on social media networks their views about organisations and sustainability. All these sources are relatively unstructured (i.e., they are only organised as natural language) textual data. As data scientists, we need to know what information we want to extract and how to organise it in a meaningful way if we want to gain insights and provide evidence for a data-driven decision making process. The sources that we are working with in this paper are courtesy of Datamaran, an ESG focused machine learning platform designed for material analysis of these issues. Datamaran already has an English-language thesaurus built to classify and structure data, which has been manually created and maintained by experts in sustainability matters. It covers over 100 topics and amounts to more than 6000 terms in an ongoing effort that has so far spanned over five years. However, analysing sources in English is only a part of the picture. If we really want to know what is happening in Spain or Latin America, we will need to be able to analyse texts in Spanish. There are basically two options when it comes to annotating Spanish-language data:",
"cite_spans": [
{
"start": 232,
"end": 248,
"text": "(Fontaine, 2013)",
"ref_id": "BIBREF5"
},
{
"start": 931,
"end": 943,
"text": "(KPMG, 2017)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "1. to translate all the texts in Spanish into English and use our English thesaurus and pipeline to annotate the information in English, or 2. to create a thesaurus in Spanish so we can analyse texts in Spanish.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The first option seems at a glance to be the easiest and fastest solution. However, using a third-party translation API at scale is very expensive. Training your own Machine Translation (MT) model is not trivial, especially if you aim to translate from low-resource languages. The crux of the issue is to obtain appropriate machine-training data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Manually creating a thesaurus in Spanish (or in any other language) would allow us to avoid the challenge of accurately translating large amounts of data. However, it would require finding experts in the field with command of the target language, or human translators with extensive ESG knowledge, and going through the process of terminology management and validation. This would be quite costly and slow. However, there is a valid option here if by substituting the thesaurus into our system, we can use the same automatic procedure to analyse text. Our approach is based on using word embedding and crosslingual mapping techniques in order to obtain seeds of terms for the Spanish thesaurus that correspond to the English thesaurus terms. Bearing this in mind, we evaluate the Artetxe et al. (2019) proposal, excluding the exposed unsupervised tuning procedure over the bilingual phrase table extracted from the cross-lingual mapping. Our primary objective is to obtain a mapping between the topics mentioned in English and Spanish. For that, we propose a set of heuristics to generate more terms in Spanish using the initial terms extracted from the cross-lingual mapping. The novelties of this proposal are: (i) an extrinsic evaluation of the Artetxe et al. (2019) approach on a multi-label document classification task and (ii) the creation of metrics to validate the quality of our results. In Section 2. we provide an overview of the different approaches to solve the problem of analysing texts in a target language using a thesaurus in a different language. Next, we present the datasets used, we describe the experiments and the proposed heuristics in Section 3. The evaluation methodology is presented in Section 4.1. Later, we examine the results of the experiments and comment them in Section 5. Finally, we conclude with a summary of what we have learnt and remarks on future work in Section 6.",
"cite_spans": [
{
"start": 780,
"end": 801,
"text": "Artetxe et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Cross-lingual word embedding (CLE) techniques have raised and experienced rapid growth over the past few years, aided by the developments around neural word embeddings such as Word2vec (Mikolov et al., 2013a) . Word embeddings are already a popular tool to induce bilingual lexicon, as continuous word embedding spaces exhibit similar structures across languages, even when considering distant language pairs (Mikolov et al., 2013b) . Cross-lingual word embedding methods exploit these similarities to generate a common representation space that allows transferring information between two languages. Although early CLE techniques relied on partial parallel data, mapping-based CLE approaches only require seed translation dictionaries (Mikolov et al., 2013b; Faruqui and Dyer, 2014; Gardner et al., 2015) or no bilingual information at all (Artetxe et al., 2018a; Lample et al., 2018) . The latter are especially effective in low-resource languages (Ruder et al., 2017) or specific domains. CLE facilitates a number of tasks that can benefit greatly from this unsupervised approach, one of which is bilingual lexicon induction (BLI). Traditional BLI techniques extract word translations from monolingual corpora through a variety of monolingual distributional similarity metrics, such as orthographic, contextual and temporal similarity metrics to discover equivalent words in different languages (Haghighi et al., 2008; Knight, 2002) . The popularity of CLE has encouraged research in applying both techniques together in order to induce a bilingual dictionary capable of obtaining successful results (Zhou et al., 2019; S\u00f8gaard et al., 2018a) . In this proposal we intend to use this latter BLI approach to generate a new thesaurus in a target language from a preexisting thesaurus in a different source language. We already possess an English thesaurus that groups a set of related lexical terms into topics or labels. For example, acid rain, air contamination, air emission, air pollutant, air quality are some of the terms grouped under the topic Air Emissions. Our main objective is to induce the English groups of terms that constitute each topic into Spanish, thus maintaining a topic alignment for both languages, but not necessarily a direct term equivalence.",
"cite_spans": [
{
"start": 185,
"end": 208,
"text": "(Mikolov et al., 2013a)",
"ref_id": "BIBREF16"
},
{
"start": 409,
"end": 432,
"text": "(Mikolov et al., 2013b)",
"ref_id": "BIBREF17"
},
{
"start": 736,
"end": 759,
"text": "(Mikolov et al., 2013b;",
"ref_id": "BIBREF17"
},
{
"start": 760,
"end": 783,
"text": "Faruqui and Dyer, 2014;",
"ref_id": "BIBREF4"
},
{
"start": 784,
"end": 805,
"text": "Gardner et al., 2015)",
"ref_id": "BIBREF6"
},
{
"start": 841,
"end": 864,
"text": "(Artetxe et al., 2018a;",
"ref_id": "BIBREF0"
},
{
"start": 865,
"end": 885,
"text": "Lample et al., 2018)",
"ref_id": "BIBREF15"
},
{
"start": 950,
"end": 970,
"text": "(Ruder et al., 2017)",
"ref_id": "BIBREF18"
},
{
"start": 1398,
"end": 1421,
"text": "(Haghighi et al., 2008;",
"ref_id": "BIBREF9"
},
{
"start": 1422,
"end": 1435,
"text": "Knight, 2002)",
"ref_id": "BIBREF11"
},
{
"start": 1603,
"end": 1622,
"text": "(Zhou et al., 2019;",
"ref_id": "BIBREF21"
},
{
"start": 1623,
"end": 1645,
"text": "S\u00f8gaard et al., 2018a)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2."
},
{
"text": "Previous work on multilingual thesaurus alignment has already taken advantage of the structural correspondence between word embeddings in different languages and the semantic information that they offer to outperform alignment methods based on string similarity metrics (Gromann and Declerck, 2018) . Our proposal further exploits these characteristics through the CLE mapping method VecMap (Artetxe et al., 2018a) . VecMap is currently one of the most effective unsupervised CLE approaches, both in terms of BLI and cross-lingual document classification performance (Glavas et al., 2019) . We chose this cross-lingual word embedding mapping method because of its performance and ease of use, as all of its code is publicly available and it works over the very common Word2vec toolkits. The method allows us to generate a new thesaurus in the Spanish language from a preexisting English thesaurus whilst avoiding any need for bilingual parallel data. To map the different labels or topics of our original thesaurus into another language, we translate each of their terms using a bilingual dictionary induced from a common representation space, according to the procedure described in Artetxe et al. (2018b) . This cross-lingual space was previously generated from two monolingual word embeddings, following Artetxe et al. (2018a) . We employ fastText (Bojanowski et al., 2016) to train the monolingual word embeddings. FastText is a Word2vec implementation that also captures sub-word information (Mikolov et al., 2013a) . Unlike Word2vec, which trains the embedding considering the word as the smallest unit in a corpus, fast-Text learns word vectors at the character level of each word, which has a higher memory and time cost when compared to Word2vec. However, it is generally accepted that fast-Text performs better than Word2vec with out of vocabulary words, as it considers terms that do not appear in the training corpus.",
"cite_spans": [
{
"start": 270,
"end": 298,
"text": "(Gromann and Declerck, 2018)",
"ref_id": "BIBREF8"
},
{
"start": 391,
"end": 414,
"text": "(Artetxe et al., 2018a)",
"ref_id": "BIBREF0"
},
{
"start": 567,
"end": 588,
"text": "(Glavas et al., 2019)",
"ref_id": "BIBREF7"
},
{
"start": 1184,
"end": 1206,
"text": "Artetxe et al. (2018b)",
"ref_id": "BIBREF1"
},
{
"start": 1307,
"end": 1329,
"text": "Artetxe et al. (2018a)",
"ref_id": "BIBREF0"
},
{
"start": 1351,
"end": 1376,
"text": "(Bojanowski et al., 2016)",
"ref_id": "BIBREF3"
},
{
"start": 1497,
"end": 1520,
"text": "(Mikolov et al., 2013a)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2."
},
{
"text": "Although more recent work introduces synthetic Statistical Machine Translation (SMT) and Neural Machine Translation (NMT) models in order to generate sensible multi-word translation from a common representation space (Artetxe et al., 2019) , we instead introduce some heuristics that perform adequately in the described thesaurus translation task and simplify the overall application of the method. Unsupervised CLE tasks often prove to be hard to evaluate, especially when no manual standard exists for the downstream task at hand. In this proposal, we also offer a multilabel document classification evaluation method based on the annotation of a given parallel corpus that can be generalised to different thesauri.",
"cite_spans": [
{
"start": 217,
"end": 239,
"text": "(Artetxe et al., 2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2."
},
{
"text": "Our main objective is to use a cross-lingual word embedding and a thesaurus in a source language to generate a thesaurus in a target language without relying on parallel data. The source embeddings can be trained exclusively with data extracted from our ongoing analysis tasks, which is already available, easy to obtain and closely matches the characteristics of the information that will be analysed in the actual downstream use of the translated thesaurus. In this section we will describe the particular features of the knowledge base used in the application of this method, as well as characterise the bilingual lexicon induction techniques ap-plied and our evaluation strategies. Finally, we propose an optimisation strategy for manual validation applied to the translated thesaurus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3."
},
{
"text": "In order to structure the non-financial information, a thesaurus in English has been manually created by experts in sustainability matters. The thesaurus is a formal specification of concepts related to 100 NFR disclosure topics. For example, air pollutant, air pollution, dust emission are some of the concepts covering the topic Air Emissions. Our terms are both words or multi-word expressions and there are a significant quantity of noun phrases. The thesaurus groups into topics more than 6000 terms in an ongoing effort that spans over five years. The terms of the thesaurus are expressed as lexical patterns to build a knowledge base of a matching algorithm responsible to automatically detect the mention of topics in different textual resources. The patterns were created using the spaCy NLP library (Honnibal and Johnson, 2015) . spaCy provides a rule-based matching feature that scans the text at token level according to some predefined patterns. These patterns are built considering some word-level features such as lemmatization, parts-of-speech, dependency parser and others. The matching algorithm compares the token attributes, specified by the pattern, with the attribute of the token in the input text to match or not the sequence. See below examples of the patterns we used.",
"cite_spans": [
{
"start": 809,
"end": 837,
"text": "(Honnibal and Johnson, 2015)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Thesaurus",
"sec_num": "3.1."
},
{
"text": "[ {\"LOWER\" : \"dust\"}, {\"LOWER\" : \"emission\"}, {\"LOWER\" : \"diesel\"}, { \"LOWER\" : \"emissions\"}, {\"LOWER\" : \"air\"}, {\"LOWER\" : \"pollutant\"} ] In the above patterns, each element inside the square brackets represents one or more words that should appear consecutively. Each element inside the curly brackets represents a token. The LOWER: diesel means that we want to match a word whose lower form is diesel. For example, any of the following sequences will be annotated with the second pattern: diesel emissions or DIESEL Emissions or Diesel emissions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Thesaurus",
"sec_num": "3.1."
},
{
"text": "Due to a lack of a Spanish thesaurus, we initially considered different alternatives to extract topics from Spanish texts: (1) maintaining a parallel thesaurus between source and target languages, which is a non-scalable process and required experts in target language;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Thesaurus",
"sec_num": "3.1."
},
{
"text": "(2) using a Commercial Machine Translation System to translate the Spanish text into English. Although using a translation service seems a technically sound solution with adequate quality results, it is not financially feasible; or (3) training our own MT model, which requires too much effort and is also very costly. As a result, we moved on to BLI techniques to derive a Spanish thesaurus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Thesaurus",
"sec_num": "3.1."
},
{
"text": "To generate embeddings that can be used in the CLE method that we have selected for our translation purpose, we need two monolingual datasets: one in the source language in which our original thesaurus was built and another in the tar- get language to which we want to migrate the said thesaurus. We apply lowercase and tokenization to both datasets, with which we then train two fastText embeddings with default hyperparameters and limiting the vocabulary to the 200,000 most frequent tokens as per Artetxe et al. 2019, although any Word2vec-based toolkit should suffice. The English and Spanish spaCy models were used to apply lowercase and to tokenize the datasets in both languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Building monolingual embeddings",
"sec_num": "3.2."
},
{
"text": "To obtain an inducted bilingual dictionary from monolingual data, we recreated the VecMap projection-based CLE method (Artetxe et al., 2018a) using the word embeddings mentioned in the previous section and mapped them into a shared space. We then extracted a source-to-target and target-to-source-phrase table using the technique described in Artetxe et al. (2018b) . A bilingual dictionary is directly induced from the sourceto-target phrase-table by ranking the entries of a given term according to their corresponding likelihood in the phrasetable, thus transforming the quantitative ranking of the phrase-table into a bilingual dictionary with positional ranking. Figure 1 shows a fragment of the phrase-table obtained for the English term market and its Spanish candidate translations. Terms with higher likelihood will appear first in the entry for market in the induced bilingual dictionary dictionary. This dictionary is used to translate the terms that make up our thesaurus. This approach maintains equivalence between source and target at the token level. However, many of the thesaurus terms are multi-word expressions. To cover this limitation and in order to build sensible combinations using the translated words, some heuristics are considered. As a result, token-level equivalence is often ignored.",
"cite_spans": [
{
"start": 118,
"end": 141,
"text": "(Artetxe et al., 2018a)",
"ref_id": "BIBREF0"
},
{
"start": 343,
"end": 365,
"text": "Artetxe et al. (2018b)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 668,
"end": 676,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "CLE method",
"sec_num": "3.3."
},
{
"text": "Using the cross-mapping embeddings we obtain a bilingual dictionary containing exclusively unigrams, which means that some techniques have to be applied in order to translate multi-word terms. In this section, we will outline several heuristic techniques that are applied to increase the coverage of the first bilingual dictionary. These heuristics use the phrase-table to generate new terms.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "Literal translation Multi-word expressions are translated term by term, maintaining their original structure. The chosen translation for each word is the first-ranked one in the bilingual dictionary, or a special symbol if there is no possible translation for that term. This is the crudest possible form of translation using a bilingual unigram dictionary, and Table 1 : Mean Reciprocal Rank that evaluates a bilingual dictionary against the full English to Spanish bilingual dictionary found in MUSE (Lample et al., 2018) .",
"cite_spans": [
{
"start": 502,
"end": 523,
"text": "(Lample et al., 2018)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 362,
"end": 369,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "it serves as the baseline for all other heuristic approaches to building expressions. For example, for the English term diesel emissions, the literal translation that is obtained is diesel emisiones, which can be represented as the following pattern: [{\"LOWER\" : \"emisiones\"},{ \"LOWER\" : \"di\u00e9sel\"}]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "Permutations Expressions are first translated term by term, after which all of their possible permutations are added into the thesaurus. In languages that do not share a very similar grammatical structure, translating the expressions maintaining their original order may produce incorrect sentences. Moreover, this technique may help capture all possible variations in languages that present a flexible word order, such as Romance languages, Hungarian, etc. See below an example of the pattern obtained for the English term diesel emissions after obtaining its literal translation in Spanish and applying the permutation heuristic explained in this paragraph.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "[{ \"LOWER\" : \"di\u00e9sel\"},{\"LOWER\" : \"emisiones\"}]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "Lemmatized terms with permutations All terms are translated in their original order, then lemmatized. Finally, like in the previous case, every possible permutation is considered. We lemmatize all terms in an attempt to reduce the variability that morphologically rich languages (that commonly also have a rather flexible word order) might bring, which is often a source of problems for unsupervised bilingual dictionary induction methods, as per S\u00f8gaard et al. (2018b) . The following example shows the patterns generated using the current heuristic.",
"cite_spans": [
{
"start": 447,
"end": 469,
"text": "S\u00f8gaard et al. (2018b)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "[ {\"LEMMA\" : \"emisi\u00f3n\"},{ \"LEMMA\" : \"di\u00e9sel\"}, { \"LEMMA\" : \"di\u00e9sel\"},{\"LEMMA\" : \"emisi\u00f3n\"} ]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "Lemmatized terms with permutations and wildcard inclusion We use the same setup as in the aforementioned approach, but adding a wildcard match before and after every word with the intent of boosting the coverage of the annotation. The longest possible match for each wildcard is selected, where the match can contain multiple tokens, and its sequence within the analysed text is no longer eligible for new matches. That is, we avoid overlap between different term matches. This logic might reduce the overall precision of the system, since overlap between the terms belonging to different labels is possible. We chose to operate in this manner to preserve the structure of our original thesaurus, as it does not present any overlaps between the terms of different labels. See below an example of one of the patterns generated adding the wildcard heuristic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "[ {\"LEMMA\" : \"emisi\u00f3n\"}, { \"OP\" : \"*\", \"IS_ALPHA\" : true}, {\"LEMMA\" : \"di\u00e9sel\"} ] The corpora necessary to build the initial monolingual word embeddings were generated using a preexisting collection of news articles from different online sources that are used in the Datamaran platform 1 . We chose to build these embeddings from news corpora because the English thesaurus that we intend to translate is used within Datamaran to analyse the content of online news, which would also be the purpose of this new translated Spanish thesaurus. Therefore, the domain of the corpora from which the monolingual word embeddings are built matches that of the text analysed in the downstream application of our system. The contents of the employed corpora are detailed below:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "\u2022 Source language corpus, which contains 220,000 English news published during 2019, and more than 137,000,000 tokens.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "\u2022 Target language corpus, composed by 260,000 Spanish news that appeared in online press during 2018-2019, containing around 118,000,000 tokens.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "To validate the quality of the generated Spanish thesaurus we proposed a multi-label document classification task, that will be explained in Section 4.2.2. For that purpose, Version 7 of the English-Spanish Europarl corpus (Koehn, 2005) was used, as it contains a sufficient amount of terminology included in our particular thesaurus (datasets with very sparse annotation would not be very informative). Figure 2 shows an English sentence extracted form the Europarl corpus that mentioned the topic Workforce changes 2 (WFChges). The Europarl corpus contains documents published on the European Parliament's official website, therefore it does not belong to the same domain as the corpus used to build the embeddings, which is a corpus of the news domain. This ensures that the performance obtained in the evaluation task",
"cite_spans": [
{
"start": 223,
"end": 236,
"text": "(Koehn, 2005)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [
{
"start": 404,
"end": 412,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Heuristics to generate terms in the target language",
"sec_num": "3.4."
},
{
"text": "Phrase composition heuristic Precision Recall KLD VecMac (Artetxe et al., 2018a) Literal translation 0.3871 0.2505 5.1149 VecMac (Artetxe et al., 2018a) Permutations 0.5295 0.4590 1.6293 VecMac (Artetxe et al., 2018a) Permutations and lemmatization 0.4236 0.5045 1.2235 VecMac (Artetxe et al., 2018a) Permutations, lemmatization and wildcards 0.4580 0.6976 0.8027 Commercial Machine Translation System None (the whole document is translated) 0.8209 0.8005 0.0233 never surpasses what would be achieved when operating over a dataset that closely matched the information used to generate the embeddings, thus providing a pessimistic estimation for the effectiveness of the evaluated translated thesaurus. We find this property desirable, as it allows us to estimate the quality of the translation in the worst cases with a higher confidence level. Additionally, it can reveal faulty translations that could go undetected in a corpus of the same domain because of context similarities. For instance, the term \"typhoons\" is translated as \"aviones\" (\"airplanes\") in the bilingual dictionary generated with the techniques detailed in 3.3. using the aforementioned datasets. This could be because in news about typhoons it is usually mentioned that there will be delays or cancellations in commercial flights that operate in the affected region. However, airplanes are not necessarily mentioned next to typhoons in the Europarl corpus nearly as often, which means that when performing a multi-label document classification task it will be possible to appreciate that articles that only discuss the effects of reducing commercial flights in pandemics or passenger rights issues are getting labelled as if they were related to natural disasters.",
"cite_spans": [
{
"start": 57,
"end": 80,
"text": "(Artetxe et al., 2018a)",
"ref_id": "BIBREF0"
},
{
"start": 129,
"end": 152,
"text": "(Artetxe et al., 2018a)",
"ref_id": "BIBREF0"
},
{
"start": 194,
"end": 217,
"text": "(Artetxe et al., 2018a)",
"ref_id": "BIBREF0"
},
{
"start": 277,
"end": 300,
"text": "(Artetxe et al., 2018a)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Translation Method",
"sec_num": null
},
{
"text": "Even though CLE models are commonly evaluated considering only BLI tasks, their performance is heavily dependent on their actual application (Glavas et al., 2019) , which highlights the need of using downstream performance evaluation tasks alongside BLI metrics.",
"cite_spans": [
{
"start": 141,
"end": 162,
"text": "(Glavas et al., 2019)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation tasks",
"sec_num": "4.2."
},
{
"text": "A bilingual lexicon induction task is used to assess the quality of the bilingual dictionary generated as detailed in 3.3. This dictionary is compared against a ground-truth translation dictionary over the same language pair. The score of each term is obtained as the position of the first suggested translation for a term in the ground-truth bilingual dictionary within the list of possible translations for the same term of our induced bilingual dictionary, or zero if said translation is not included as an option in the generated dictionary. This scoring method is known as mean reciprocal ranking (MRR). MRR is equivalent to mean average precision (MAP) if only one valid translation per query is considered, in this case the top result. We chose this metric rather than the more common precision at rank k (P @k), which instead scores a term translation with one point if its position in the induced bilingual dictionary is equal or above k. This decision was made because MRR provides a more detailed evaluation, as it does not treat all models that rank the cor-rect translation below k equally Glavas et al. (2019) . In the evaluation, only terms from the ground-truth dictionary that had one or more of their possible translations appear in the induced translation dictionary were considered.",
"cite_spans": [
{
"start": 1103,
"end": 1123,
"text": "Glavas et al. (2019)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "BLI evaluation over a bilingual dictionary",
"sec_num": "4.2.1."
},
{
"text": "Although BLI evaluation is a decent indicator our cross-lingual embedding quality along with the bilingual dictionary induce from it and can help the developer fine-tune this particular piece of the translation system, it does not directly correlate with the downstream performance of the system at hand, which in this particular use case corresponds to document classification. We propose a multi-label document classification task that directly matches the intended use of a translated thesaurus (where the classes of the task directly correspond to the topics of said thesaurus) and can be easily applied to other similar setups because of the simplicity of its logic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-label document classification",
"sec_num": "4.2.2."
},
{
"text": "The parallel bilingual corpus (i.e. the Europarl corpus referenced in section 4.1.) is considered, divided in different documents using an arbitrary window and one of the monolingual sections is annotated with the preexisting source language thesaurus (for our application, the source language in which the original thesaurus was written was English, so we would score the English section of the parallel corpus). This process will yield a score per document, which may have a different representation depending on the specific analysis criteria, be it mentions of a certain topic or the frequency of the terms with which it is related (i.e. combined incidences for all the terms that belong to a certain topic in a text). The source language thesaurus is then translated using a bilingual dictionary induced from two monolingual embeddings mapped into a common space, which are represented by two source-to-target and target-to-source cross-lingual embeddings, as seen in section 3.3.. This new thesaurus is used to annotate the section of the parallel corpus written in the target language (the Spanish section of the corpus in our case), thus obtaining a new list of document-score tuples. To get a better idea of how this parallel scoring would look in our case, we can see Figure 3 , which shows two fragments of English and Spanish news extracted from https://elpais.com in which the topic Renewables alternatives 3 (Renew) was mentioned. Next to each highlighted term in Figure 3 we can see a label that indicates the topic to which it belongs (for instance \"photovoltaic\" is a term included in the topic Renew in our English thesaurus, and it appears in the English version of the article). We now have two different scores per document, one obtained using the original thesaurus over the source language version of the article, and the other extracted with the induced thesaurus to rank the target language version of this same article. Based on the difference in label scoring per document we can obtain recall and precision at label (only scoring related to a specific topic is considered, i.e. hits from terms that belong to a specific topic) and thesaurus level (all topics are taken into consideration). We use micro averaging for both metrics, as the labels or topics of our thesaurus can present differences in the number of terms that they contain and how common those are. Furthermore, extracting the relative frequency of each label allows us to calculate the binary log of the ratio of relative frequencies (log-ratio) at label level and Kullback-Leibler divergence (KLD) (Kullback and Leibler, 1951) at thesaurus level. Equation 1 is used to estimate the KLD, which quantifies how much a probability distribution diverges from another one in terms of the amount of additional information needed, where P (x) and Q(x) are the relative frequency of a category x in each corpora P or Q with respect to the categories in the same corpora. The higher the KLD value, the greater the divergence. If the value is 0, it means that both distributions are identical.",
"cite_spans": [
{
"start": 2592,
"end": 2620,
"text": "(Kullback and Leibler, 1951)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [
{
"start": 1278,
"end": 1286,
"text": "Figure 3",
"ref_id": null
},
{
"start": 1478,
"end": 1486,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Multi-label document classification",
"sec_num": "4.2.2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "KLD(P ||Q) = P (x)log( P (x) Q(x) )",
"eq_num": "(1)"
}
],
"section": "Multi-label document classification",
"sec_num": "4.2.2."
},
{
"text": "Log-ratio (LR) is a metric commonly used for the task of keyword extraction, as it tells us how relevant the difference between two probability distributions is. It is estimated using Equation 2. We used the binary logarithm to better represent the dissimilitude pointed by the log-ratio measure.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-label document classification",
"sec_num": "4.2.2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "LR(P ||Q) = log( P (x) Q(x) )",
"eq_num": "(2)"
}
],
"section": "Multi-label document classification",
"sec_num": "4.2.2."
},
{
"text": "In our case, KLD compares the source and target language thesauri as a distribution of probabilities, where the relative frequency of each topic acts as the dependent variable, and the labels themselves are a qualitative factor (that is, the frequency is a topic or label-level metric, so its value will be different depending on the chosen topic). This comparison yields the expectation of the log difference between the probability of a topic in the original thesaurus distribution with the generated thesaurus distribution, which is the amount of information that is lost when we approximate the source language thesaurus with the target language one. The log-ratio is also given per topic and its estimation is based on the ratio of the relative frequency of a topic in the source and target corpora, providing a measure of how many times a topic is more frequent in one corpus compared with the other.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multi-label document classification",
"sec_num": "4.2.2."
},
{
"text": "No matter the performance of the technique in charge of translating a knowledge base, subsequent human validation will need to be applied in order to ensure the quality of the final product. This usually means that the thesaurus goes through a number of iterations before reaching its final state. However, knowledge bases can contain a tremendous volume of information, which complicates obtaining a complete human validation. With the objective of achieving an optimal partial validation, we establish a priority for each label or topic in our thesaurus and work over them in the resulting descending order. This priority metric guides the manual validation of the topic in the sense that topics with higher values should be the first to be reviewed manually, as they have a more significant impact over the quality of the translated thesaurus when compared with topics with lower priority. We achieve this by multiplying the absolute value of the log-likelihood ratio detailed in 4.2.2. with the source or target language frequency of the topic, depending on the sign of said log-ratio. For example, if the log-ratio is positive there are instances of the topic in the source language that are not being registered when analysing the text with the translated thesaurus in the target language. We multiply frequency of the topic in the original language with the absolute value of this positive log-ratio in order to get an idea of the negative impact of the translation of the aforementioned topic over the quality of our translated thesaurus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Optimising manual validation",
"sec_num": "4.3."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P rio i = |lr i | * (of i * (lr i >= 0) + tf i * (lr i < 0))",
"eq_num": "(3)"
}
],
"section": "Optimising manual validation",
"sec_num": "4.3."
},
{
"text": "Where P rio i corresponds to the priority given to topic i, lr i is the log-ratio obtained for topic i and of i and tf i are the original and target frequency for topic i respectively. One of the side effects of using this formula is that topics that have close to no frequency at source and target will be classified as having low priority, even if their recall and Figure 3 : Multi-level classification for English and Spanish news using the thesauri in both languages priority are zero or close to zero. We considered that, although recall and priority can be very low for a topic, if the source and target frequency are too low it becomes hard to assess the quality of the translation for the group of terms grouped under this topic, so human validation is not as useful as in other cases. Additionally, we can consider that uncommon terms will have a lower impact over the quality of the translated thesaurus even if their translations are not very good. Table 1 shows the MRR obtained from evaluating the bilingual dictionary generated from the base corpus described in the previous section (as well as a similar, smaller dataset) against the full English to Spanish bilingual dictionary provided in the MUSE specification (Lample et al., 2018) . The bilingual dictionary evaluated is obtained according to the procedure described in 3.3. We compare a smaller corpus of online news against another dataset with a bigger volume that contains news from the same sources, the latter one being our main experimental corpus. Mean reciprocal rank for the generated bilingual dictionary does not always correlate directly with the actual downstream performance of the system, and some authors use it as a threshold of quality of the BLI procedure, like in Glavas et al. (2019) , where 0.05 was established as a minimal value to consider a language pair translation run as acceptable.",
"cite_spans": [
{
"start": 1229,
"end": 1250,
"text": "(Lample et al., 2018)",
"ref_id": "BIBREF15"
},
{
"start": 1755,
"end": 1775,
"text": "Glavas et al. (2019)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [
{
"start": 367,
"end": 375,
"text": "Figure 3",
"ref_id": null
},
{
"start": 960,
"end": 967,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Optimising manual validation",
"sec_num": "4.3."
},
{
"text": "During our experiments, we have only considered MRR when measuring the impact of the size of the monolingual datasets with which our word embeddings are generated over the induced CLE-phrase table. It is displayed here to show how it can help developers evaluate certain pieces of the translation system individually (in this case our induced bilingual dictionary), and as a reference for future CLE-related tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and discussion",
"sec_num": "5."
},
{
"text": "Conversely, our multi-label document classification evaluation (Table 2 ) yields much more informative results about the performance of both the source-to-target language alignment and the heuristic used to build terms from the induced unigram bilingual dictionary. As expected, literal translation returns low precision and recall scores, paired with a high KLD value, which indicates that most of the information contained in the original thesaurus is being lost. Part of the reason for this outcome can be attributed to the grammatical differences between Spanish and English, which are not properly accounted for when translating token by token.",
"cite_spans": [],
"ref_spans": [
{
"start": 63,
"end": 71,
"text": "(Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Results and discussion",
"sec_num": "5."
},
{
"text": "Providing all possible permutations for each term has a notable impact for all metrics, but especially over the Kullback-Leibler divergence. Because KLD is a measure of information loss between two probability distributions (in this case modelled after the frequency of the topics in each thesaurus), we can infer that, although precision and recall are still relative low, this information loss is distributed more evenly across all the labels of the thesaurus. That is, the probability distributions that are modelled after topic annotation in the source and target language present a more similar shape. Lemmatization seems to increase recall, which is expected, especially when working with a highly-inflectional language such as Spanish. However, it might introduce some noise, because it amplifies the coverage of all terms. This means that terms that were originally meaningful but that have been translated into common expressions will have a noticeable negative impact in the quality of the translated thesaurus. For instance, the term \"unionized\" that belongs to the topic \"Union\" is translated into Spanish as \"trabajadores\" (workers) in the bilingual English-Spanish dictionary obtained using the procedures detailed in 3.3. with the experimental settings mentioned in 4.1. \"Trabajadores\" is a much more common word that does not only appear in news articles concerning unionisation issues. This faulty translation already caused a loss in precision when using literal translation and permutation heuristics, but only in instances where the exact word appeared in the text. Moreover, the translation procedure depends on the actual contents of the used monolingual word embeddings, so it is possible that \"trabajadores\" often appears in a similar context to \"unionized\" and the precision is not affected excessively. However, after applying lemmatization, all possible forms of this term (\"trabajadoras\", \"trabajador\", \"trabajadora\") will produce a hit for the topic \"Union\". Lastly, the addition of wildcards on top of the previous heuristics provides the best overall scores save for precision, which is still improved over using only permutations and lemmatization heuristics. The remarkable improvement of the recall is to be expected when applying this kind of \"loose matching\" (multiple tokens can appear in between the word that make up a multi-word term) over the Spanish language, which presents a flexible phrase structure. Even so, precision and KLD are still relatively far from the results obtained with the commercial machine translation system. In terms of precision, we have observed that our resulting bilingual dictionary has a tendency to place common terms as the most likely translation over more scarce expressions that may match the original term better. This phenomenon is likely related to the noise that results from the word embeddings cross-mapping procedure (Artetxe et al., 2018a) . Further refinements in such processes and integration of the CLE-generated phrase-table into statistical or neural machine translation models may mitigate the issue, among other possibilities that we will briefly explore at the end of this section. The results of our experiments show that the proposal does not perform equally for all the topics. This could be due to some topics being more or less specific, or due to factors that affect the number of occurrences of each topic in the training corpus. In Table 3 it is possible to see the frequency of each topic at source and target languages. For example, in the Europarl corpus we did not find mentions for any of the terms grouped under the topics BrdComp and BuildAct in the original English thesaurus, so the source frequency for both of these topics is 0. As a result, their precision and recall are zero independently of whether there are incidences for the same terms when translated with the method in our proposal (that is, their target frequency). However, this value for precision and recall does not imply that the translation of these two terms is necessarily bad. Instead, in cases where the source frequency for a topic is relatively low when compared to other topics, our confidence about the recall and precision values obtained will be lower. To reflect this, source and target frequency of the terms grouped under a topic contributes to the estimation of the priority of said topic, and the priority metric guides the manual validation of the topic in the sense that topics with higher values should be the first to review manually because they have worse results. For instance, the topic AntiCorr has a higher priority value, although it presents better precision and recall. In this case the priority metric is telling us that, even though this topic has been translated better than others, it appears very frequently in the analysed text, which means that it has a big impact over the quality of the translated thesaurus and should be reviewed before other topics. We can get to this conclusion because the priority is a function of the absolute value of the log-ratio and the frequency, which itself affects this calculation of the log-ratio the most. Consequentially, some topics have similar precision and recall values (i.e. Biod and AltAccnt), but the priority of one of them is lower (AltAccnt) because its terms are not very frequent. For cases where a topic has low values of precision and recall but its priority is still low, recommending additional terms for this topic could be useful. Future improvements could include refining the phrase-table obtained from cross-lingual embeddings so as to obtain a better bilingual dictionary, as it has already been proposed in Artetxe et al. (2019) , which also reduces the need for heuristics that build multi-word expressions. Term matching overlap can be tuned in order to maximise performance, although it would mean that the logic behind some of the terms of the original thesaurus might be compromised, which in some cases might be a better fit for the target language. It could also be of interest to evaluate terms that are found commonly as false positives according to their relevance (i.e. relying on tf-idf), discarding those that are too general by establishing a threshold and speeding up manual validation without losing meaningful terms.",
"cite_spans": [
{
"start": 2899,
"end": 2922,
"text": "(Artetxe et al., 2018a)",
"ref_id": "BIBREF0"
},
{
"start": 5680,
"end": 5701,
"text": "Artetxe et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [
{
"start": 3432,
"end": 3439,
"text": "Table 3",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Results and discussion",
"sec_num": "5."
},
{
"text": "In this work we offer a practical application of a bilingual lexicon induction (BLI) method based on cross lingual embeddings (CLE) (Artetxe et al., 2018a; Artetxe et al., 2018b) that allows us to induce a domain specific Spanish thesaurus from a preexisting English thesaurus used for multi-label document classification within Non-Financial Reporting. We include some possible heuristics that may help build sensible expressions from a unigram translation dictionary, which is itself induced from the aforementioned CLE procedure, and compare their performance against each other and a commercial machine translation system. To this end, we also offer some evaluation metrics that measure the performance of the proposed multi-label document classification task, along with a term prioritisation strategy for manual annotation. We hope that some of the strategies proposed here pave the way for an easier application of CLE-based BLI techniques, especially for tasks that rely on transferring information across multilingual knowledge representations, and help understand better the behaviour of these methods for similar use cases.",
"cite_spans": [
{
"start": 132,
"end": 155,
"text": "(Artetxe et al., 2018a;",
"ref_id": "BIBREF0"
},
{
"start": 156,
"end": 178,
"text": "Artetxe et al., 2018b)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6."
},
{
"text": "https://www.datamaran.com/ 2 References to variation in number of people employed by an entity. Includes changes due to restructuring. E.g. reorganisations, turnover rates, outsourcing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "References to energy from natural processes and/or nontraditional sources that are replenished on a human timescale. E.g. alternative energy sources, photovoltaic, biomass.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This paper was developed by a research team at Datamaran, Ltd. Datamaran is a company that develops AI-driven software for identifying and monitoring external risks and opportunities tied to ESG issues. We are thankful to our colleagues who provided expertise that greatly assisted the research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": "7."
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings",
"authors": [
{
"first": "M",
"middle": [],
"last": "Artetxe",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Labaka",
"suffix": ""
},
{
"first": "Agirre",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "789--798",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Artetxe, M., Labaka, G., and Agirre, E. (2018a). A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. In Proceedings of the 56th Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers), pages 789- 798, Melbourne, Australia, July. Association for Compu- tational Linguistics.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Unsupervised Statistical Machine Translation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Artetxe",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Labaka",
"suffix": ""
},
{
"first": "Agirre",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Artetxe, M., Labaka, G., and Agirre, E. (2018b). Unsuper- vised Statistical Machine Translation. Technical report.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Bilingual Lexicon Induction through Unsupervised Machine Translation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Artetxe",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Labaka",
"suffix": ""
},
{
"first": "Agirre",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "5002--5007",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Artetxe, M., Labaka, G., and Agirre, E. (2019). Bilingual Lexicon Induction through Unsupervised Machine Trans- lation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5002- 5007, Stroudsburg, PA, USA. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Enriching Word Vectors with Subword Information",
"authors": [
{
"first": "P",
"middle": [],
"last": "Bojanowski",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Grave",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Joulin",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Mikolov",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "7",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bojanowski, P., Grave, E., Joulin, A., and Mikolov, T. (2016). Enriching Word Vectors with Subword Informa- tion. 7.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Improving vector space word representations using multilingual correlation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Faruqui",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Dyer",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "462--471",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Faruqui, M. and Dyer, C. (2014). Improving vector space word representations using multilingual correlation. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguis- tics, pages 462-471, Gothenburg, Sweden, April. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Corporate social responsibility and sustainability: the new bottom line? International Journal of Business and Social Science",
"authors": [
{
"first": "M",
"middle": [],
"last": "Fontaine",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "4",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fontaine, M. (2013). Corporate social responsibility and sustainability: the new bottom line? International Jour- nal of Business and Social Science, 4(4).",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Translation Invariant Word Embeddings",
"authors": [
{
"first": "M",
"middle": [],
"last": "Gardner",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Papalexakis",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Fu",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Talukdar",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Faloutsos",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Sidiropoulos",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Mitchell",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gardner, M., Huang, K., Papalexakis, E., Fu, X., Taluk- dar, P., Faloutsos, C., Sidiropoulos, N., and Mitchell, T. (2015). Translation Invariant Word Embeddings. Tech- nical report.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "How to (Properly) Evaluate Cross-Lingual Word Embeddings: On Strong Baselines, Comparative Analyses, and Some Misconceptions",
"authors": [
{
"first": "G",
"middle": [],
"last": "Glavas",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Litschko",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Ruder",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Vulic",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Glavas, G., Litschko, R., Ruder, S., and Vulic, I. (2019). How to (Properly) Evaluate Cross-Lingual Word Embed- dings: On Strong Baselines, Comparative Analyses, and Some Misconceptions. 2.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Comparing pretrained multilingual word embeddings on an ontology alignment task",
"authors": [
{
"first": "D",
"middle": [],
"last": "Gromann",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Declerck",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gromann, D. and Declerck, T. (2018). Comparing pre- trained multilingual word embeddings on an ontology alignment task. In Proceedings of the Eleventh Inter- national Conference on Language Resources and Eval- uation (LREC 2018), Miyazaki, Japan, May. European Language Resources Association (ELRA).",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Learning bilingual lexicons from monolingual corpora",
"authors": [
{
"first": "A",
"middle": [],
"last": "Haghighi",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Berg-Kirkpatrick",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Klein",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Haghighi, A., Liang, P., Berg-Kirkpatrick, T., and Klein, D. (2008). Learning bilingual lexicons from monolingual corpora. pages 771-779, 01.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "An improved nonmonotonic transition system for dependency parsing",
"authors": [
{
"first": "M",
"middle": [],
"last": "Honnibal",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Johnson",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1373--1378",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Honnibal, M. and Johnson, M. (2015). An improved non- monotonic transition system for dependency parsing. In Proceedings of the 2015 Conference on Empirical Meth- ods in Natural Language Processing, pages 1373-1378, Lisbon, Portugal, September. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Learning a translation lexicon from monolingual corpora",
"authors": [
{
"first": "K",
"middle": [],
"last": "Knight",
"suffix": ""
}
],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Knight, K. (2002). Learning a translation lexicon from monolingual corpora. 05.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Europarl: A Parallel Corpus for Statistical Machine Translation",
"authors": [
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Koehn, P. (2005). Europarl: A Parallel Corpus for Statisti- cal Machine Translation. Technical report.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "The KPMG Survey of Corporate Responsibility Reporting",
"authors": [],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "KPMG. (2017). The KPMG Survey of Corporate Respon- sibility Reporting 2017. Technical report.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "On information and sufficiency",
"authors": [
{
"first": "S",
"middle": [],
"last": "Kullback",
"suffix": ""
},
{
"first": "R",
"middle": [
"A"
],
"last": "Leibler",
"suffix": ""
}
],
"year": 1951,
"venue": "Ann. Math. Statist",
"volume": "22",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kullback, S. and Leibler, R. A. (1951). On information and sufficiency. Ann. Math. Statist., 22(1):79-86, 03.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Word translation without parallel data",
"authors": [
{
"first": "G",
"middle": [],
"last": "Lample",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Conneau",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Ranzato",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Denoyer",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "J\u00e9gou",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lample, G., Conneau, A., Ranzato, A., Denoyer, L., and J\u00e9gou, H. (2018). Word translation without parallel data. Technical report.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Efficient Estimation of Word Representations in Vector Space",
"authors": [
{
"first": "T",
"middle": [],
"last": "Mikolov",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Corrado",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Dean",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013a). Efficient Estimation of Word Representations in Vector Space. 1.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Exploiting Similarities among Languages for Machine Translation",
"authors": [
{
"first": "T",
"middle": [],
"last": "Mikolov",
"suffix": ""
},
{
"first": "Q",
"middle": [
"V"
],
"last": "Le",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Sutskever",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mikolov, T., Le, Q. V., and Sutskever, I. (2013b). Exploiting Similarities among Languages for Machine Translation.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "A Survey Of Cross-lingual Word Embedding Models",
"authors": [
{
"first": "S",
"middle": [],
"last": "Ruder",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Vuli\u0107",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "S\u00f8gaard",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ruder, S., Vuli\u0107, I., and S\u00f8gaard, A. (2017). A Survey Of Cross-lingual Word Embedding Models. 6.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "On the limitations of unsupervised bilingual dictionary induction",
"authors": [
{
"first": "A",
"middle": [],
"last": "S\u00f8gaard",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Ruder",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Vuli\u0107",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "778--788",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S\u00f8gaard, A., Ruder, S., and Vuli\u0107, I. (2018a). On the lim- itations of unsupervised bilingual dictionary induction. In Proceedings of the 56th Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 1: Long Papers), pages 778-788, Melbourne, Australia, jul. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "On the Limitations of Unsupervised Bilingual Dictionary Induction",
"authors": [
{
"first": "A",
"middle": [],
"last": "S\u00f8gaard",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Ruder",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Vuli\u0107",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S\u00f8gaard, A., Ruder, S., and Vuli\u0107, I. (2018b). On the Lim- itations of Unsupervised Bilingual Dictionary Induction.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Density matching for bilingual word embedding",
"authors": [
{
"first": "C",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Neubig",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhou, C., Ma, X., Wang, D., and Neubig, G. (2019). Den- sity matching for bilingual word embedding. In NAACL- HLT.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"num": null,
"text": "Fragment of market phrase-table and its candidate translations",
"uris": null
},
"FIGREF1": {
"type_str": "figure",
"num": null,
"text": "Figure 2: Topic Workforce changes (WFChges) mentioned in an English Europarl sentence 4. Experimental settings",
"uris": null
},
"TABREF0": {
"type_str": "table",
"content": "<table><tr><td>108,000 English news</td><td>84,000 Spanish news</td><td>0.093</td></tr><tr><td>220,000 English news</td><td>260,000 Spanish news</td><td>0.107</td></tr></table>",
"text": "",
"html": null,
"num": null
},
"TABREF1": {
"type_str": "table",
"content": "<table/>",
"text": "Multi-label document classification comparison over the parallel corpus Europarl for the English-Spanish pair. The embeddings used for the CLE approach (VecMap) were built from the corpora detailed in 4.1. Validation metrics at thesaurus level.",
"html": null,
"num": null
},
"TABREF2": {
"type_str": "table",
"content": "<table><tr><td>AEmiss</td><td>0.0112</td><td>0.0045</td><td>0.3269</td><td>0.2602</td><td>1.3067</td><td>0.1664</td></tr><tr><td>AltAccnt</td><td>0.0015</td><td>0.0012</td><td>0.5476</td><td>0.8214</td><td>0.2971</td><td>0.0006</td></tr><tr><td>AltFuel</td><td>0.0019</td><td>0.0103</td><td>0.0507</td><td>0.5625</td><td>-2.4477</td><td>0.2637</td></tr><tr><td>AntUse</td><td>0.0008</td><td>0.0046</td><td>0.1320</td><td>0.84</td><td>-2.4465</td><td>0.0528</td></tr><tr><td>AntiCorr</td><td>0.0314</td><td>0.0177</td><td>0.8649</td><td>0.8913</td><td>0.8272</td><td>0.8204</td></tr><tr><td>Biod</td><td>0.0075</td><td>0.0073</td><td>0.4701</td><td>0.8939</td><td>0.0398</td><td>0.0022</td></tr><tr><td>BrdComp</td><td>0</td><td>0.0002</td><td>0</td><td>0</td><td>-12.9659</td><td>0.0007</td></tr><tr><td>BuildAct</td><td>0</td><td>0.00005</td><td>0</td><td>0</td><td>-10.9665</td><td>0.00003</td></tr><tr><td>CChg</td><td>0.0234</td><td>0.0304</td><td>0.1633</td><td>0.4899</td><td>-0.3740</td><td>0.3465</td></tr></table>",
"text": "",
"html": null,
"num": null
},
"TABREF3": {
"type_str": "table",
"content": "<table/>",
"text": "Multi-label document classification comparison over the parallel corpus Europarl for the English-Spanish pair. Validation metrics at topic level",
"html": null,
"num": null
}
}
}
} |