File size: 86,263 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T01:13:11.569127Z"
},
"title": "Restoring the Sister: Reconstructing a Lexicon from Sister Languages using Neural Machine Translation",
"authors": [
{
"first": "Remo",
"middle": [],
"last": "Nitschke",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The University of Arizona",
"location": {}
},
"email": "nitschke@email.arizona.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The historical comparative method has a long history in historical linguists. It describes a process by which historical linguists aim to reverse-engineer the historical developments of language families in order to reconstruct proto-forms and familial relations between languages. In recent years, there have been multiple attempts to replicate this process through machine learning, especially in the realm of cognate detection (List et al., 2016; Ciobanu and Dinu, 2014; Rama et al., 2018). So far, most of these experiments aimed at actual reconstruction have attempted the prediction of a proto-form from the forms of the daughter languages (Ciobanu and Dinu, 2018; Meloni et al., 2019). Here, we propose a reimplementation that uses modern related languages, or sisters, instead, to reconstruct the vocabulary of a target language. In particular, we show that we can reconstruct vocabulary of a target language by using a fairly small data set of parallel cognates from different sister languages, using a neural machine translation (NMT) architecture with a standard encoderdecoder setup. This effort is directly in furtherance of the goal to use machine learning tools to help under-served language communities in their efforts at reclaiming, preserving, or reconstructing their own languages.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "The historical comparative method has a long history in historical linguists. It describes a process by which historical linguists aim to reverse-engineer the historical developments of language families in order to reconstruct proto-forms and familial relations between languages. In recent years, there have been multiple attempts to replicate this process through machine learning, especially in the realm of cognate detection (List et al., 2016; Ciobanu and Dinu, 2014; Rama et al., 2018). So far, most of these experiments aimed at actual reconstruction have attempted the prediction of a proto-form from the forms of the daughter languages (Ciobanu and Dinu, 2018; Meloni et al., 2019). Here, we propose a reimplementation that uses modern related languages, or sisters, instead, to reconstruct the vocabulary of a target language. In particular, we show that we can reconstruct vocabulary of a target language by using a fairly small data set of parallel cognates from different sister languages, using a neural machine translation (NMT) architecture with a standard encoderdecoder setup. This effort is directly in furtherance of the goal to use machine learning tools to help under-served language communities in their efforts at reclaiming, preserving, or reconstructing their own languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Historical linguistics has long employed the historical comparative method to establish familial connections between languages and to reconstruct proto-forms (cf. Klein et al., 2017b; Meillet, 1967) . More recently, the comparative method has been employed by revitalization projects for lexical reconstruction of lost lexical items (cf. Delgado et al., 2019) . In the particular case of Delgado et al. (2019) , lost lexical items of the target language are reconstructed by using equivalent cognates of stillspoken modern sister languages, i.e., languages in the same language family that share some established common ancestor language and a significant amount of cognates with the target language. By reverse-engineering the historical phonological processes that happened between the target language and the sister-languages, one can predict what the lexical item in the target language should be. This is essentially a twist on the comparative method, using the same principles, but to reconstruct a modern sister, as opposed to a proto-antecedent.",
"cite_spans": [
{
"start": 163,
"end": 183,
"text": "Klein et al., 2017b;",
"ref_id": "BIBREF10"
},
{
"start": 184,
"end": 198,
"text": "Meillet, 1967)",
"ref_id": "BIBREF15"
},
{
"start": 338,
"end": 359,
"text": "Delgado et al., 2019)",
"ref_id": "BIBREF4"
},
{
"start": 388,
"end": 409,
"text": "Delgado et al. (2019)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "While neural net systems have been used to emulate the historical comparative method 1 to reconstruct proto-forms (Meloni et al., 2019; Ciobanu and Dinu, 2018) and for cognate detection (List et al., 2016; Ciobanu and Dinu, 2014; Rama et al., 2018) , there have not, to the best of our knowledge, been any attempts to use neural nets to predict/reconstruct lexical items of a sister language for revitalization/reconstruction purposes. Meloni et al. (2019) report success for a similar task (reconstructing Latin proto-forms) by using cognate pattern lists as a training input. Instead of reconstructing Latin proto-forms from only Italian roots, they use Italian, Spanish, Portuguese, Romanian and French cognates of Latin, i.e., mapping from many languages to one. As our intended usecase (see section 1.1) is one that suffers from data sparsity, we explicitly explore the degree to which expanding the list of sister-languages in the manyto-one mapping can compensate for fewer available data-points. Since the long-term goal of this project is to aid language revitalization efforts, the question of available data is of utmost importance. Machine learning often requires vast amounts of data, and languages which are undergoing revitalization usually have very sparse amounts of data available. Hence, the goal for a machine learning approach 1 Due to the nature of neural nets we do not know whether these systems actually emulate the historical comparative method or not. What is meant here is that they were used for the same tasks.",
"cite_spans": [
{
"start": 114,
"end": 135,
"text": "(Meloni et al., 2019;",
"ref_id": "BIBREF16"
},
{
"start": 136,
"end": 159,
"text": "Ciobanu and Dinu, 2018)",
"ref_id": "BIBREF3"
},
{
"start": 186,
"end": 205,
"text": "(List et al., 2016;",
"ref_id": "BIBREF12"
},
{
"start": 206,
"end": 229,
"text": "Ciobanu and Dinu, 2014;",
"ref_id": "BIBREF2"
},
{
"start": 230,
"end": 248,
"text": "Rama et al., 2018)",
"ref_id": "BIBREF19"
},
{
"start": 436,
"end": 456,
"text": "Meloni et al. (2019)",
"ref_id": "BIBREF16"
},
{
"start": 1348,
"end": 1349,
"text": "1",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "here is not necessarily the highest possible accuracy, but rather the ability to operate with as little data as possible, while still retaining a reasonable amount of accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our particular contributions are:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "1. We demonstrate an approach for reframing the historical comparative method to reconstruct a target language from its sisters using a neural machine translation framework. We show that this can be done with easily accessible open source frameworks such as OpenNMT (Klein et al., 2017a) .",
"cite_spans": [
{
"start": 266,
"end": 287,
"text": "(Klein et al., 2017a)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "2. We provide a detailed analysis of the degree to which inputs from additional sister languages can overcome issues of data sparsity. We find that adding more related languages allows for higher accuracy with fewer data points. However, we also find that blindly adding languages to the input stream does not always yield said higher accuracy. The results suggest that there needs to be a significant amount of cognates with the added input language and the target language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This experiment was designed with a specific usecase in mind: Lexical reconstruction for language revitalization projects. Specifically, the situation where this type of model may be most applicable would be a language reclamation project in the definition of Leonhard (2007) or a language revival process in the definition of Mc-Carty and Nicholas (2014) . In essence, a language where there is some need to recover or reconstruct a lexicon. An example of such a case might be the Wampanoag language reclamation project (https://www.wlrp.org/), or comparable projects using the methods outlined in Delgado et al. (2019) . As this is a proof-of-concept, we use the Romance language family, specifically the nonendangered languages of French, Spanish, Italian, Portuguese and Romanian, and operate under assumption that these results can inform how one can use this approach with other languages of interest. However, we are aware that the Romance language morphology may be radically different from some of the languages that may be in the scope of this use case, such as agglutinative and polysynthetic languages, and that we cannot fully predict the performance of this type of system for such languages from the Romance example. Regardless of this, some insights gained here will still be applicable in those cases, such as the question of compensating lack of data by using multiple languages.",
"cite_spans": [
{
"start": 260,
"end": 275,
"text": "Leonhard (2007)",
"ref_id": "BIBREF11"
},
{
"start": 327,
"end": 355,
"text": "Mc-Carty and Nicholas (2014)",
"ref_id": null
},
{
"start": 599,
"end": 620,
"text": "Delgado et al. (2019)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Intended Use-Case and Considerations",
"sec_num": "1.1"
},
{
"text": "Languages that are the focus of language revitalization projects are typically not targets for deep learning projects. One of the reasons for this is the fact that these languages usually do not have large amounts of data available for training state of the art neural approaches. These systems need large amounts of data, and Neural Machine Translation systems, as the one used in this project, are no exception. For example, Cho et al. (2014) use data sets varying between 5.5million and 348million words. However, the task of proto-form reconstruction, which is really a task of cognate prediction, can be achieved with fairly small datasets, if parallel language input is used. This was shown by Meloni et al. (2019) , whose system predicted 84% within an edit distance of 1, meaning that 84% of the predictions were so accurate that only one or 0 edits were necessary to achieve the true target. For example, if the target output is \"grazie\", the machine might predict \"grazia\" (one edit) or \"grazie\" (0 edits). Within a language revitalization context, this level of accuracy would actually be a very good outcome. In this scenario, a linguist or speaker familiar with the language would vet the output regardless, so small edit distances should not pose a big problem. Further, all members of a language revitalization project or language community would ultimately vet the output, as they would make a decision on whether to accept or reject the output as a lexical item of the language.",
"cite_spans": [
{
"start": 427,
"end": 444,
"text": "Cho et al. (2014)",
"ref_id": "BIBREF1"
},
{
"start": 700,
"end": 720,
"text": "Meloni et al. (2019)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Intended Use-Case and Considerations",
"sec_num": "1.1"
},
{
"text": "This begs the question of why a language revitalization project would want to go through the trouble of using such an algorithm in the first place, if they have someone available to vet the output, then that person may as well do the reconstructive work themselves, as proposed in Delgado et al. (2019) . This all depends on two factors: First, how high is the volume of lexical items that need to be reconstructed or predicted? The effort may not be worth it for 10 or even a 100 lexical items, but beyond this an neural machine translation model can potentially outperform the manual labor. Once trained, the model can make thousands of predictions in minutes, as long as input data is available.",
"cite_spans": [
{
"start": 281,
"end": 302,
"text": "Delgado et al. (2019)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Intended Use-Case and Considerations",
"sec_num": "1.1"
},
{
"text": "Second, and potentially more important, it will depend on how well the historical phonological relationships between the languages are understood. For a family like Romance, we have a very good understanding of the historical genesis of the languages and the different phonological processes they underwent, see for example Maiden et al. (2013) . However, there are many language families in the world where these relationships and histories are less than clear. In such situations, a machine learning approach would be beneficial, because the algorithm learns 2 the relationships for us and gives predictions that just need to be vetted. Under this perspective, the best model might not necessarily be the one that produces the most accurate output, but perhaps the one that produces the fewest incorrigible mistakes. An incorrigible mistake here would be the algorithm predicting an item that is completely unrelated to the target root e.g., predicting \"cinque\" for a target of \"grazie\"). Further, ease of usability and accessibility will be another factor for this kind of use-case, as not every project of this type will have a computational linguist to call on. Hence, another aim should be a low-threshold for reproducability and the utilization of easy to use open-source frameworks. In the spirit of the latter, all data and code necessary to reproduce the results are open-source and freely available. This paper is intended for computational linguists and linguists and/or community members who are involved with projects surrounding languages which might benefit from this approach. As such, it is written with both audiences in mind, with Section 6 (\"Warning Labels for Interested Linguists\") specifically aimed at linguists and community members interested in a potential application of this method.",
"cite_spans": [
{
"start": 324,
"end": 344,
"text": "Maiden et al. (2013)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Intended Use-Case and Considerations",
"sec_num": "1.1"
},
{
"text": "The data set used for this experiment was provided by Shauli Rafvogel of Meloni et al. (2019) . The initial set consisted of 5420 lines of cognate sextuples of the Romance language family, specifically: Ro- manian, French, Spanish, Portuguese, Italian and Latin. As the aim for this experiment was to reconstruct from sister languages to a sister language, the Latin items were removed from the set and instead Italian was chosen to be the target language for the experiment, since it had the most complete pattern with respect to the other languages in the set. Table 1 illustrates the types of lines present in the initial dataset.",
"cite_spans": [
{
"start": 73,
"end": 93,
"text": "Meloni et al. (2019)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [
{
"start": 563,
"end": 571,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "The Dataset",
"sec_num": "2"
},
{
"text": "Lines with no target and lines with no input were removed. Lines where there was a target but no input (row 3) were also removed, as well as lines where there was input but no target (line 1). After the removal of all lines which lead to empty patterns in the Italian set, and all lines which were empty patterns in the input, 3527 remained. From these, 2466 lines were taken as training data, 345 were taken for validation, and 717 were set aside for testing. Meloni et al. (2019) use both an orthographic and an IPA data set, and show that the orthographic set yielded more accurate results. Here, we use only orthographic representations, which we prefer not for accuracy, but because orthographic datasets are more easily acquired for most languages, particularly those of interest in language reclamation projects. If both an IPA set and an orthographic set are available, one may attempt using both to boost the accuracy of the results. Chen (2018) showed that this is possible with glossing data in the case of sentence level neural machine translation. We will discuss this implementation in Section 5.2.",
"cite_spans": [
{
"start": 461,
"end": 481,
"text": "Meloni et al. (2019)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Dataset",
"sec_num": "2"
},
{
"text": "See Figure 1 for a very simplified phylogenetic tree representation of the familial relations of the Romance languages used in this dataset. This tree was constructed using data from glottolog (Hammarstr\u00f6m et al., 2020) , and is included just for illustrative purposes and not as a statement about the phylogeny of Romance languages. 3",
"cite_spans": [
{
"start": 193,
"end": 219,
"text": "(Hammarstr\u00f6m et al., 2020)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 4,
"end": 12,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "The Dataset",
"sec_num": "2"
},
{
"text": "This experiment was run using the OpenNMTpytorch neural machine translation (Klein et al., 2017a) framework, using the default settings (a 2layer LSTM with 500 hidden units on both the encoder and decoder). The opennmt-py default setup was chosen intentionally; the envisioned use-case requires an easily reproducable approach for interested users or communities who might profit from using this method for their own purposes, but who don't necessarily have deep expertise in machine learning or tuning neural models. A publicly available toolkit, like opennmt, and a no-configuration setup helps lower the bar to entry for these parties.",
"cite_spans": [
{
"start": 76,
"end": 97,
"text": "(Klein et al., 2017a)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "3"
},
{
"text": "Neural machine translation (NMT) frameworks are designed to translate sentences from one language to another, but they can be used for a number of sequential data tasks (Neubig, 2017) . One such task is the prediction of a cognate from a set of input words, as used here. These frameworks are typically an encoder-decoder setup, where both the encoder and decoder are often implemented as LSTM (Long Short-Term Memory) networks (Hochreiter and Schmidhuber, 1997) , which have the advantage of effectively capturing long-distance dependencies (Neubig, 2017) . In an encoder-decoder setup, the encoder reads in the character based input representation and transforms it into a vector representation. The decoder takes this vector representation and transforms it into a character based output representation (Cho et al., 2014) .",
"cite_spans": [
{
"start": 169,
"end": 183,
"text": "(Neubig, 2017)",
"ref_id": "BIBREF17"
},
{
"start": 428,
"end": 462,
"text": "(Hochreiter and Schmidhuber, 1997)",
"ref_id": "BIBREF7"
},
{
"start": 542,
"end": 556,
"text": "(Neubig, 2017)",
"ref_id": "BIBREF17"
},
{
"start": 806,
"end": 824,
"text": "(Cho et al., 2014)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "3"
},
{
"text": "NMT frameworks also employ a \"vocabulary\" set, which contains vocabulary of the language that is being translated from and vocabulary of the language that is being translated to. The size of this vocabulary is often an issue for the effectiveness of NMT models (Hirschmann et al., 2016) . In our case, the source vocabulary simply contains all of the characters that occur in all the input language examples and the target vocabulary contains the characters that occur in the target language example. To illustrate: if this task was about predicting English words, then the target vocabulary would contain all the letters of the English alphabet.",
"cite_spans": [
{
"start": 261,
"end": 286,
"text": "(Hirschmann et al., 2016)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "3"
},
{
"text": "Since the input in our case is a list of cognates from different languages, we need to consider how we feed this input to the machine. There are two obvious options for this task. We can either feed the cognates one by one, or we can merge the cognates first, before feeding them to the machine. In this experiment, we merge the words character by character to construct the input lines. This means that for every line in the input, the first character of each word was concatenated, then the second character of each word was concatenated, and so on. For an illustration:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input Concatenation",
"sec_num": "3.1"
},
{
"text": "(1) patterns in the input: aille, alha, al, aie",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input Concatenation",
"sec_num": "3.1"
},
{
"text": "(2) target patterns: aglia",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input Concatenation",
"sec_num": "3.1"
},
{
"text": "(3) input: aaaaillilhelae (4) target: aglia This merging delivered marginally better results than simple concatenation in early testing, which is why it was selected. It is unclear as to why this is the case. We suspect that the merged input makes it easier for the model to recognize if the same characters appear in the same position of the input, as is the case with \"a\" in the initial position in the above example. However, we are cautious to recommend this input representation in general, because different morphologies may be better represented in a concatenation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Input Concatenation",
"sec_num": "3.1"
},
{
"text": "To determine the performance gains from simply having more data versus having data from more languages, we create several training scenarios. In each, we use the same aforementioned 2-layer LSTM. To understand the benefit of additional language, we first train with the entire training set with all four languages, then successively remove languages from the input set until only one remains. Next, to compare this to the impact of simply having fewer data points, but from all languages, we generate several impoverished versions of the data set. For these impoverished versions, lines were removed randomly 4 from the set reducing the data by 70%, 50%, 30% and 10% respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Different Training Setups",
"sec_num": "3.2"
},
{
"text": "Machine translation is usually evaluated using the BLEU (Papineni et al., 2002) score, but BLEU is designed with sentence level translations in mind. We instead evaluate the output according to edit distance in the style of Meloni et al. (2019) by calculating the percentage of the output which is within a given edit distance. In addition to this metric, we also use a custom evaluation metric designed to emphasize the usability of the output for the intended use-case, i.e., as predictions to be vetted by an expert to save time over doing the entire analysis manually. In order to calculate this score, we calculate the Damerau-Levenshtein edit distance to the target for each word and assign weights to them by their edit distance. That is:",
"cite_spans": [
{
"start": 56,
"end": 79,
"text": "(Papineni et al., 2002)",
"ref_id": "BIBREF18"
},
{
"start": 224,
"end": 244,
"text": "Meloni et al. (2019)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "4"
},
{
"text": "score = (a + b * .9 + c * .8 + d * .7 + e * .6)/t",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "4"
},
{
"text": "where a is number of predictions with distance 0, b is the number with distance 1, c is the number with distance 2, d is the number with distance 3, e the number with distance 4, and t is the total number of predictions. As an example, consider a scenario where there are three predicted cognates. If system 1 produces 3 output patterns within an edit distance of 2, it would receive a score of 0.8. If system 2 produces two output patterns with edit distance 0 and one within a distance of 5, this would result in a score of 0.67.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "4"
},
{
"text": "The logic behind this metric is that any prediction with an edit distance larger than 4 edits is essentially useless for the proposed task. Since such a large edit distance essentially constitutes an incorrigible mistake as mentioned in (Section 1.1). The edit distance of 4 constitutes an arbitrary cut-off to a degree, but it allows us a simple and informative evaluation metric for our use case. This metric will rank a model that has a large number of items in a and a large number of items beyond 4 edits lower than a model with items mostly in the b-d range. Presumably, the latter is more useful to the task, as small errors can be adjusted by linguists or language users.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "4"
},
{
"text": "Using this metric, we can rank different input combinations according to their assumed useful-ness to the task of lexical reconstruction for revitalization purposes. Table 2 shows the edit distance percentages and scores of different runs at 10,000 steps of training. 5 We can compare the difference in outcome between using fewer languages in the input versus using less input lines overall. This addresses the question of whether adding multiple languages to the input helps compensate for fewer data points (cognate pairs). The runs with successively reduced numbers of languages (top half of the table), are all trained with all available input lines (2466) but excluding specific columns/languages. The \"reduced input\" runs (bottom half of the table), on the other hand, are done with all four languages but with fewer cognates, by excluding rows. These runs had the following amount of training input lines: 10%: 2220 lines of input, 30%: 1793 lines of input, 50%: 1345 lines of input, 70%: 896 lines of input (recall that the total number of input lines available for training was 2466). All runs were tested on the same testing data target.",
"cite_spans": [
{
"start": 268,
"end": 269,
"text": "5",
"ref_id": null
}
],
"ref_spans": [
{
"start": 166,
"end": 173,
"text": "Table 2",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "4"
},
{
"text": "In Table 2 (see following page), we can observe that, unsurprisingly, the training sample with the most languages and data (Span-Fre-Port-Ro) performs best. 44.6% within edit distance 0 means that almost half the predictions the machine makes are correct. In terms of accuracy, this is not incredible, Meloni et al. (2019) report 64.1% within edit distance 0. However, considering that we are using a data set approximately a third the size of theirs for training (2466 cognates compared with 7038), the performance is surprisingly good. The more important measure for the intended use-case is the fact that over 80% of items are within an edit distance of 3, meaning that of the output produced, 80% need only three edits or fewer to meet the target.",
"cite_spans": [
{
"start": 302,
"end": 322,
"text": "Meloni et al. (2019)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [
{
"start": 3,
"end": 10,
"text": "Table 2",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "We can also observe that the performance successively drops as we remove languages, with the Spanish only 6 performing worst. However, the way in which this performance drops is not entirely transparent. It appears that in terms of scoring, the Spanish-French (Spa-Fre) sample actu- ally performs better than the Spanish-Portuguese-Romanian sample. Further, while Span-Port-Ro has significantly better values in the 0-2 edit range, it is outperformed by Span-Fre in terms of score because Span-Fre has more items in the \u2264 4 edit range.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The noticeable difference between Span-Fre-Port and Span-Port-Ro is surprising and warrants some examination. The likely explanation is twofold. First, The Romanian set is the one with the most empty patterns. The Romanian training data only includes 930 filled patterns, in comparison, Portuguese includes 1905 patterns, French includes 1790, and Spanish has 2125. It may be the case that the Romanian data is too small in comparison with the others to have a significant impact on the outcome. The other factor may be that Romanian is phylogenetically the most distant from the target language (Italian) (Figure 1 ).",
"cite_spans": [],
"ref_spans": [
{
"start": 606,
"end": 615,
"text": "(Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "This becomes even more apparent in Figure 2 , which. shows the performance of different models over time. 7 Here we can observe that there is hardly any difference between the performance of Span-Fre-Port-Ro and Span-Fre-Port over time, and it is only at 10,000 steps that they start to diverge. This divergence at the 10,000 step mark is likely random, the graph suggest that their overall performance is almost identical in regards to scoring. Another point in this direction are the seemingly convergent graphs of Span-Fre and Span-Port-Ro, suggesting that there is no difference between using 2 or 3 languages as input if the third language is Romanian.",
"cite_spans": [
{
"start": 106,
"end": 107,
"text": "7",
"ref_id": null
}
],
"ref_spans": [
{
"start": 35,
"end": 43,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Discounting the performance of the exclusion/inclusion of Romanian, we can observe that performance overall tends to increase with each parallel language added. This is especially evident with the obvious drop-off in performance of the Spanish only input. If we assume that Romanian has no impact, then we can see that 3 languages (blue and orange) perform similarly and two languages (red and green) perform similarly, and there is an obvious drop-off between those two patterns. This suggests that using parallel language input can compensate smaller datasets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Due to the small dataset, the scores plateau fairly early, around the 3000 epoch mark for most. This suggests that it would be sufficient to run these models at 3000 epochs, which would save some time on low-end hardware. However, with these small datasets, training time should rarely exceed 5 hours on consumer grade PCs. 8",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Let us now consider the second question of this paper: Can parallel language input compensate for small dataset size? We know that performance reduces if we reduce the number of languages in the input mix. Now we compare this drop-off to the reduction in performance caused by reducing the overall amount of input data. This can be seen in Figure 3 , which shows the performance at different training steps for models trained on decreasing amounts of data. Included for comparison are models trained on all data using all four (Span-Fra-Port-Ro), three (Span-Port-Ro), and one (Span) input language. 9 First, we observe that a 10% reduction in training data (grey) does not seem to have a strong impact, as this performs mostly equal to Span-Fre-Port-Ro. Further, we can see is that the 30% reduced case performs marginally better than Span-Port-Ro. This is a good result, as it suggests that we can compensate for a fair amount of data by using additional languages. Essentially, in this case we can observe that removing a language from the input can be equivalent to removing 30% of the input or more. Even the 50% reduced case (brown) still performs better than using just one language (Spanish only).",
"cite_spans": [
{
"start": 600,
"end": 601,
"text": "9",
"ref_id": null
}
],
"ref_spans": [
{
"start": 340,
"end": 348,
"text": "Figure 3",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Parallel Languages vs Input Reduction",
"sec_num": "5.1"
},
{
"text": "The extreme fall-off between the 50% reduction and the 70% reduction suggests that there is some point beyond which even multiple languages cannot compensate for lack of data points. Where this fall-off point is exactly, will likely fluctuate depending on the data set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parallel Languages vs Input Reduction",
"sec_num": "5.1"
},
{
"text": "Chen 2018shows that neural machine translation tasks can be greatly improved by adding glossing data to the input mix (We will gloss over the technical details of the implementation here). While there is no direct equivalent to the gloss-sentence relationship, there might be a close analog for words: phonetic transcriptions. Orthography may be conservative and often misleading, but phonetic representations are not. Meloni et al. (2019) use a phonetic dataset in their experiment, but they map from phonetic representations to phonetic representations, so their input and their target items are represented in IPA. This performs worse than the orthographic task. An interesting further experiment would be to blend orthographic representations and phonetic representations in the input, in the style of Chen (2018), mapping that to an orthographic output. This would be a close analog to the sentence-gloss to sentence mapping that Chen (2018) reports success with.",
"cite_spans": [
{
"start": 419,
"end": 439,
"text": "Meloni et al. (2019)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Potential Improvements",
"sec_num": "5.2"
},
{
"text": "One thing to consider, is that this may be not ideal for the use-case. Phonetic datasets are not easy to produce and the orthography is often more readily available. While this might improve performance, needing a phonetic as well as an orthographic dataset would likely increase the threshold of reproducability for interested parties.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Potential Improvements",
"sec_num": "5.2"
},
{
"text": "There are some important aspects of this kind of approach that linguists, or community members who are interested in utilizing it for their purposes, should be aware of. There are certain things that this type of approach can and cannot do for a community or project. The model does not so much reconstruct a word for the community, but rather proposes what the word could be, according to the data it has been fed. The model will propose these recommendations on the basis of an abstract notion of what the historic phonological and morphological differences are between languages ABC and language D. This does not necessarily mean that the model learns or understands the historical phonological and morphological processes that separate the input sister languages from the target languages. It has simply learned a way to generalize from the input to the output with some degree of accuracy. What is learned need not necessarily overlap with what linguists believe to have happened.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Warning Labels for Interested Linguists",
"sec_num": "6"
},
{
"text": "Therefore, this type of model will only ever generate cognates of the input. It cannot generate novel items. This is an important factor to consider for any community or linguist planning on using this approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Warning Labels for Interested Linguists",
"sec_num": "6"
},
{
"text": "Consider the following case: Imagine we are trying to use this approach to reconstruct English from other Germanic languages. A large part of the English lexicon is not of Germanic ancestry. However, any lexicon we would try to reconstruct using this trained algorithm would give us approximations of a Germanic derived lexeme for the word we are trying to reconstruct. This is a potentially undesirable effect of the way the model was trained. Linguists and interested community members need to be aware of this and implement their own quality control.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Warning Labels for Interested Linguists",
"sec_num": "6"
},
{
"text": "However, this approach can potentially be useful for any language project where a community and or linguists are working with an incomplete lexicon for a language. The prerequisite for this being a useful tool in such a scenario is the assumption that the sister languages to the target language are somewhat well documented and have at least dictionaries available from which data can be extracted. A final prerequisite is the presence of minimally a small dictionary of the target language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Warning Labels for Interested Linguists",
"sec_num": "6"
},
{
"text": "The model would then be trained using the sister languages as input, and the target language list as a target output. After training confirms a reasonable accuracy, the model can then be fed with other known words in the sister language to get a prediction of those words in the target language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Warning Labels for Interested Linguists",
"sec_num": "6"
},
{
"text": "After producing said output, the linguist, or language community, needs to subject the output to a quality control and decide on a series of questions: Do the output patterns match what we know of the target language? Can we assume that these words are cognates in the target language, or is there some evidence that other forms were present? Finally, if this is used by a community to fill in empty patterns in their language, the community needs to decide whether the output is something that the community wants in their language. The algorithm is not infallible, and only proposes. Ultimately, a language community using this tool must make a decision whether to accept or reject the algorithm's recommendations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Warning Labels for Interested Linguists",
"sec_num": "6"
},
{
"text": "In this paper, we have shown that NMT frameworks can be used to predict cognates of a target language from cognates of its sister languages. We have further shown that adding or removing input languages has interesting effects on the accuracy of the model. This indicates that we can use additional sister languages to compensate the lack of data in a given situation, though, as demonstrated in the case of Romanian, we cannot blindly add sister languages, nor assume that all additions are equally useful. This might be a promising method for situations where not a lot of data is present, but there are multiple well-documented related languages of the target language. The next step for this line of research is to move from a proof of concept to an implementation in an actual language revitalization scenario. This is something we are currently working on. A further question that need to be addressed as well, is how well this approach performs with languages that exhibit a different morphology from the Romance languages, such as agglutinative and polysynthetic languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "All code and data used for this project are opensource and can be found here, in order to reproduce these results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "Something we would like to address in this final paragraphs is that machine learning is a potential tool. Like every tool, it has its uses and cases where it is not useful. The decision of using such a tool to expand the lexicon of a language is a decision of that language community, and not of a linguist.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "We also acknowledge that tree representations are not necessarily the most accurate way to represent these relationships(Kaylan and Fran\u00e7ois, 2019).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "This was done by simply removing every n th line depending on how much reduction was needed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "One step of training means that the algorithm has gone through one batch of input lines. The default batch-size for opennmt is 64.6 Spanish only was only trained for 5000 steps, as the model plateaus around 1000 steps. The performance of the Spanish only model was measured every 500 steps forFigure 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "This can give a better representation of the performance, because a neural net constantly adjusts its weights, so looking at just one point in time can be deceiving.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "These were trained on an i5-5200 CPU with 2.2GHz, and training took anywhere between 4-7 hours for 10,000 steps.9 Since inFigure 2we observe that Span-Port-Ro and Span-Fre perform quite similarly, and Span-Fre-Port performs similarly to Span-Fre-Port-Fro, to make the graph easier to read, we remove Span-Fre and Span-Fre-Port from this graph.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The author would like to thank the anonymous reviewers for their comments and give special thanks to Becky Sharp for helping with last minute edits.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Improving Neural Net Machine Translation Systems with Linguistic Information",
"authors": [
{
"first": "Yuan",
"middle": [],
"last": "Lu",
"suffix": ""
},
{
"first": "Chen",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuan Lu Chen. 2018. Improving Neural Net Ma- chine Translation Systems with Linguistic Informa- tion. Phd thesis, University of Arizona.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "On the properties of neural machine translation: Encoder-decoder approaches",
"authors": [
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Bart",
"middle": [],
"last": "Van Merrienboer",
"suffix": ""
},
{
"first": "Dzmitry",
"middle": [],
"last": "Bahdanau",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "KyungHyun Cho, Bart van Merrienboer, Dzmitry Bah- danau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder-decoder ap- proaches. CoRR, abs/1409.1259.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Automatic detection of cognates using orthographic alignment",
"authors": [
{
"first": "Alina",
"middle": [
"Maria"
],
"last": "Ciobanu",
"suffix": ""
},
{
"first": "Liviu",
"middle": [
"P"
],
"last": "Dinu",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "99--105",
"other_ids": {
"DOI": [
"10.3115/v1/P14-2017"
]
},
"num": null,
"urls": [],
"raw_text": "Alina Maria Ciobanu and Liviu P. Dinu. 2014. Auto- matic detection of cognates using orthographic align- ment. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Vol- ume 2: Short Papers), pages 99-105, Baltimore, Maryland. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Ab initio: Automatic Latin proto-word reconstruction",
"authors": [
{
"first": "Alina",
"middle": [
"Maria"
],
"last": "Ciobanu",
"suffix": ""
},
{
"first": "Liviu",
"middle": [
"P"
],
"last": "Dinu",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 27th International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1604--1614",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alina Maria Ciobanu and Liviu P. Dinu. 2018. Ab ini- tio: Automatic Latin proto-word reconstruction. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1604-1614, Santa Fe, New Mexico, USA. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Digital documentation training for long island algonquian community language researchers: a new paradigm for community linguistics",
"authors": [
{
"first": "Leighton",
"middle": [],
"last": "Delgado",
"suffix": ""
},
{
"first": "Irene",
"middle": [],
"last": "Navas",
"suffix": ""
},
{
"first": "Conor",
"middle": [],
"last": "Quinn",
"suffix": ""
},
{
"first": "Tina",
"middle": [],
"last": "Tarrant",
"suffix": ""
},
{
"first": "Wunetu",
"middle": [],
"last": "Tarrant",
"suffix": ""
},
{
"first": "Harry",
"middle": [],
"last": "Wallace",
"suffix": ""
}
],
"year": 2019,
"venue": "Presented at: 51st Algonquian Conference",
"volume": "",
"issue": "",
"pages": "24--27",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Leighton Delgado, Irene Navas, Conor Quinn, Tina Tarrant, Wunetu Tarrant, and Harry Wallace. 2019. Digital documentation training for long island al- gonquian community language researchers: a new paradigm for community linguistics. Presented at: 51st Algonquian Conference, McGill University, Montr\u00e9al, QC, 24-27 October.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Martin Haspelmath",
"authors": [
{
"first": "Harald",
"middle": [],
"last": "Hammarstr\u00f6m",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Forkel",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.5281/zenodo.4061162"
]
},
"num": null,
"urls": [],
"raw_text": "Harald Hammarstr\u00f6m, Robert Forkel, Martin Haspel- math, and Sebastian Bank. 2020. Glottolog 4.3. Jena.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "What makes word-level neural machine translation hard: A case study on English-German translation",
"authors": [
{
"first": "Fabian",
"middle": [],
"last": "Hirschmann",
"suffix": ""
},
{
"first": "Jinseok",
"middle": [],
"last": "Nam",
"suffix": ""
},
{
"first": "Johannes",
"middle": [],
"last": "F\u00fcrnkranz",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers",
"volume": "",
"issue": "",
"pages": "3199--3208",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fabian Hirschmann, Jinseok Nam, and Johannes F\u00fcrnkranz. 2016. What makes word-level neural machine translation hard: A case study on English- German translation. In Proceedings of COLING 2016, the 26th International Conference on Compu- tational Linguistics: Technical Papers, pages 3199- 3208, Osaka, Japan. The COLING 2016 Organizing Committee.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Long short-term memory",
"authors": [
{
"first": "Sepp",
"middle": [],
"last": "Hochreiter",
"suffix": ""
},
{
"first": "J\u00fcrgen",
"middle": [],
"last": "Schmidhuber",
"suffix": ""
}
],
"year": 1997,
"venue": "Neural Computation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Freeing the comparative method from the tree model: A framework for historical glottometry. In Let's talk about trees: Genetic relationships of languages and their phylogenetic representation",
"authors": [
{
"first": "Siva",
"middle": [],
"last": "Kaylan",
"suffix": ""
},
{
"first": "Alexandre",
"middle": [],
"last": "Fran\u00e7ois",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Siva Kaylan and Alexandre Fran\u00e7ois. 2019. Freeing the comparative method from the tree model: A framework for historical glottometry. In Let's talk about trees: Genetic relationships of languages and their phylogenetic representation. Cambridge Uni- versity Press, online.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Opennmt: Open-source toolkit for neural machine translation",
"authors": [
{
"first": "Guillaume",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Yoon",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Yuntian",
"middle": [],
"last": "Deng",
"suffix": ""
},
{
"first": "Jean",
"middle": [],
"last": "Senellart",
"suffix": ""
},
{
"first": "Alexander",
"middle": [
"M"
],
"last": "Rush",
"suffix": ""
}
],
"year": 2017,
"venue": "Proc. ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.18653/v1/P17-4012"
]
},
"num": null,
"urls": [],
"raw_text": "Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senel- lart, and Alexander M. Rush. 2017a. Opennmt: Open-source toolkit for neural machine translation. In Proc. ACL.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Handbook of Comparative and Historical Indo-European Linguistics : An International Handbook",
"authors": [
{
"first": "Jared",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Brian",
"middle": [],
"last": "Joseph",
"suffix": ""
},
{
"first": "Matthias",
"middle": [],
"last": "Fritz",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jared Klein, Brian Joseph, and Matthias Fritz. 2017b. Handbook of Comparative and Historical Indo- European Linguistics : An International Handbook. De Gruyter, Berlin.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Miami Language Reclamation in the Home: A Case Study",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Wesley",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Leonhard",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wesley Y. Leonhard. 2007. Miami Language Reclama- tion in the Home: A Case Study. Phd thesis, Univer- sity of California, Berkeley.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Using sequence similarity networks to identify partial cognates in multilingual wordlists",
"authors": [
{
"first": "Johann-Mattis",
"middle": [],
"last": "List",
"suffix": ""
},
{
"first": "Philippe",
"middle": [],
"last": "Lopez",
"suffix": ""
},
{
"first": "Eric",
"middle": [],
"last": "Bapteste",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "599--605",
"other_ids": {
"DOI": [
"10.18653/v1/P16-2097"
]
},
"num": null,
"urls": [],
"raw_text": "Johann-Mattis List, Philippe Lopez, and Eric Bapteste. 2016. Using sequence similarity networks to iden- tify partial cognates in multilingual wordlists. In Proceedings of the 54th Annual Meeting of the As- sociation for Computational Linguistics (Volume 2: Short Papers), pages 599-605.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "The Cambridge History of the Romance Languages",
"authors": [
{
"first": "M",
"middle": [],
"last": "Maiden",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Smith",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Ledgeway",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "2",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Maiden, J. Smith, and A. Ledgeway, editors. 2013. The Cambridge History of the Romance Languages, volume 2. Cambridge University Press, Cambridge.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Reclaiming indigenous languages: A reconsideration of the roles and responsibilities of schools",
"authors": [
{
"first": "Teresa",
"middle": [
"L"
],
"last": "Mccarty",
"suffix": ""
},
{
"first": "Sheilah",
"middle": [
"E"
],
"last": "Nicholas",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Teresa L. McCarty and Sheilah E. Nicholas. 2014. Re- claiming indigenous languages: A reconsideration of the roles and responsibilities of schools. Review of Research in Education, 31.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "The comparative method in historical linguistics",
"authors": [
{
"first": "Antoine",
"middle": [],
"last": "Meillet",
"suffix": ""
}
],
"year": 1967,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Antoine Meillet. 1967. The comparative method in historical linguistics. Librairie Honor\u00e9 Champion, Paris.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Ab antiquo: Proto-language reconstruction with rnns",
"authors": [
{
"first": "Carlo",
"middle": [],
"last": "Meloni",
"suffix": ""
},
{
"first": "Shauli",
"middle": [],
"last": "Ravfogel",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Carlo Meloni, Shauli Ravfogel, and Yoav Goldberg. 2019. Ab antiquo: Proto-language reconstruction with rnns.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Neural machine translation and sequence-to-sequence models: A tutorial",
"authors": [
{
"first": "Graham",
"middle": [],
"last": "Neubig",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Graham Neubig. 2017. Neural machine translation and sequence-to-sequence models: A tutorial. CoRR, abs/1703.01619.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Bleu: a method for automatic evaluation of machine translation",
"authors": [
{
"first": "Kishore",
"middle": [],
"last": "Papineni",
"suffix": ""
},
{
"first": "Salim",
"middle": [],
"last": "Roukos",
"suffix": ""
},
{
"first": "Todd",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "Wei-Jing",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "311--318",
"other_ids": {
"DOI": [
"10.3115/1073083.1073135"
]
},
"num": null,
"urls": [],
"raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic eval- uation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Com- putational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Are automatic methods for cognate detection good enough for phylogenetic reconstruction in historical linguistics? CoRR",
"authors": [
{
"first": "Taraka",
"middle": [],
"last": "Rama",
"suffix": ""
},
{
"first": "Johann-Mattis",
"middle": [],
"last": "List",
"suffix": ""
},
{
"first": "Johannes",
"middle": [],
"last": "Wahle",
"suffix": ""
},
{
"first": "Gerhard",
"middle": [],
"last": "J\u00e4ger",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Taraka Rama, Johann-Mattis List, Johannes Wahle, and Gerhard J\u00e4ger. 2018. Are automatic meth- ods for cognate detection good enough for phyloge- netic reconstruction in historical linguistics? CoRR, abs/1804.05416.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "An abridged family tree of the relevant Romance languages. Adapted from glottolog(Hammarstr\u00f6m et al., 2020)."
},
"FIGREF1": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Performance at different training steps for models with different combinations of input languages, plotted by custom score. All scores are calculated from the testing data."
},
"FIGREF2": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Performance of models trained on all four languages, but with varying levels of downsampled data. Included for comparison are models trained with all data on different language combinations. Plotted is the custom score over steps. Scores are calculated every 1000 training steps. All models were run on OpenNMTpy default parameters."
},
"TABREF1": {
"type_str": "table",
"text": "Examples of data patterns, including types of data removed during cleanup (e.g., rows 1 and 3).",
"num": null,
"html": null,
"content": "<table/>"
},
"TABREF3": {
"type_str": "table",
"text": "63% 57.74% 69.6% 80.33% 88.42% 0.82 Span-Fre-Port 42.68% 53.27% 68.34% 77.68% 84.94% 0.78 Span-Port-Ro 42.54% 53.28% 66.39% 74.76% 81.59% 0.75",
"num": null,
"html": null,
"content": "<table><tr><td>Edit Distance</td><td>0</td><td>\u22641</td><td>\u2264 2</td><td>\u22643</td><td>\u22644</td><td>score</td></tr><tr><td colspan=\"2\">Span-Fre-Port-Ro 44.Span-Fre 39.9%</td><td colspan=\"4\">50.9% 63.88% 74.62% 83.4%</td><td>0.76</td></tr><tr><td>Spanish only</td><td colspan=\"6\">35.6% 47.98% 60.25% 69.03% 74.76% 0.68</td></tr><tr><td colspan=\"6\">10% Reduced Input 40.17% 54.25% 69.6% 81.31% 87.59%</td><td>0.8</td></tr><tr><td colspan=\"7\">30% Reduced Input 39.75% 50.91% 66.11% 73.36% 83.12% 0.77</td></tr><tr><td colspan=\"6\">50% Reduced Input 33.19% 45.61% 60.95% 71.27% 82.4%</td><td>0.75</td></tr><tr><td colspan=\"3\">70% Reduced Input 17.02% 26.08%</td><td>41%</td><td colspan=\"3\">50.77% 65.97% 0.59</td></tr></table>"
},
"TABREF4": {
"type_str": "table",
"text": "Edit distance percentiles at 10,000 training steps. Shown are the results from using all data points with different combinations of languages (top), as well as using all languages but with random downsampling of the data from each (bottom). All scores are calculated from the testing data.",
"num": null,
"html": null,
"content": "<table/>"
}
}
}
} |