File size: 71,994 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:12:38.470465Z"
},
"title": "Zero-Shot Cross-Lingual Transfer is a Hard Baseline to Beat in German Fine-Grained Entity Typing",
"authors": [
{
"first": "Sabine",
"middle": [],
"last": "Weber",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Edinburgh",
"location": {
"country": "UK"
}
},
"email": "s.weber@sms.ed.ac.uk"
},
{
"first": "Mark",
"middle": [],
"last": "Steedman",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Edinburgh",
"location": {
"country": "UK"
}
},
"email": "steedman@inf.ed.ac.uk"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The training of NLP models often requires large amounts of labelled training data, which makes it difficult to expand existing models to new languages. While zero-shot cross-lingual transfer relies on multilingual word embeddings to apply a model trained on one language to another, Yarowsky and Ngai (2001) propose the method of annotation projection to generate training data without manual annotation. This method was successfully used for the tasks of named entity recognition and coarse-grained entity typing, but we show that it is outperformed by zero-shot cross-lingual transfer when applied to the similar task of fine-grained entity typing. In our study of fine-grained entity typing with the FIGER type ontology for German, we show that annotation projection amplifies the English model's tendency to underpredict level 2 labels and is beaten by zero-shot cross-lingual transfer on three novel test sets.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "The training of NLP models often requires large amounts of labelled training data, which makes it difficult to expand existing models to new languages. While zero-shot cross-lingual transfer relies on multilingual word embeddings to apply a model trained on one language to another, Yarowsky and Ngai (2001) propose the method of annotation projection to generate training data without manual annotation. This method was successfully used for the tasks of named entity recognition and coarse-grained entity typing, but we show that it is outperformed by zero-shot cross-lingual transfer when applied to the similar task of fine-grained entity typing. In our study of fine-grained entity typing with the FIGER type ontology for German, we show that annotation projection amplifies the English model's tendency to underpredict level 2 labels and is beaten by zero-shot cross-lingual transfer on three novel test sets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The task of fine-grained entity typing (FET) is to assign a semantic label to a span in a text. The task is distinct from coarse-grained entity typing as done by named entity recognition systems because these systems are restricted to a small set of labels like 'person', 'organization' and 'location' which are not helpful for tasks that require more precise information about the entities. For example, FET assigns the label '/location/city' to the named entity 'Berlin' in the sentence 'From 1997 to 2000, it had a permanent exhibition in Berlin.'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Fine-grained entity typing uses a high number of types in a multilevel hierarchy, which can be seen in the level 2 label '/location/city' (see Figure 1) . In this work we use the FIGER type hierarchy which consists of two levels with 112 types in total (37 level 1, 75 level 2). FIGER types are derived from the knowledge graph Freebase (Bollacker et al., 2008) . They are both interpretable by humans and useful in NLP applications such as relation extraction (Kuang et al., 2020) .",
"cite_spans": [
{
"start": 338,
"end": 362,
"text": "(Bollacker et al., 2008)",
"ref_id": "BIBREF2"
},
{
"start": 462,
"end": 482,
"text": "(Kuang et al., 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 143,
"end": 153,
"text": "Figure 1)",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "There are systems for named entity recognition and coarse-grained entity typing in languages other than English (e.g. Stanza (Qi et al., 2020) ), but systems for FET with FIGER types are only available in English, due to the lack of FIGER annotated data in other languages. Because manual annotation is time consuming and expensive, various methods have been proposed to expand NLP models to other languages without additional manual annotation. The method of annotation projection (Yarowsky and Ngai, 2001) uses parallel text to automatically create annotated corpora. Annotations from the resource-rich language are transferred to the resource-poor language using word alignment between translated sentences.",
"cite_spans": [
{
"start": 125,
"end": 142,
"text": "(Qi et al., 2020)",
"ref_id": "BIBREF15"
},
{
"start": 482,
"end": 507,
"text": "(Yarowsky and Ngai, 2001)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Annotation projection has been used successfully for the task of coarse-grained named entity typing in conjunction with named entity recognition (Agerri et al., 2018; Li et al., 2021; Ni et al., 2017) . We follow these examples by using a parallel English-German corpus, automatic named entity recognition and a state of the art English FET model (Chen et al., 2020) to assign FIGER type labels on the English side for transfer. We then project the labels onto the German half of the corpus. The output of this process is a German corpus annotated with FIGER types, which we use to train a German FET model. Another approach to the same problem is zeroshot cross-lingual transfer, in which a model built on multilingual word embeddings and trained on high-resource language data is applied to test data in a different language. Because the English FET model used in this work (Chen et al., 2020) relies on contextualised multilingual word embeddings (XLM-RoBERTa) (Conneau et al., 2019) it is possible to train it on English data and to test it on German.",
"cite_spans": [
{
"start": 145,
"end": 166,
"text": "(Agerri et al., 2018;",
"ref_id": "BIBREF0"
},
{
"start": 167,
"end": 183,
"text": "Li et al., 2021;",
"ref_id": "BIBREF10"
},
{
"start": 184,
"end": 200,
"text": "Ni et al., 2017)",
"ref_id": "BIBREF12"
},
{
"start": 347,
"end": 366,
"text": "(Chen et al., 2020)",
"ref_id": "BIBREF3"
},
{
"start": 876,
"end": 895,
"text": "(Chen et al., 2020)",
"ref_id": "BIBREF3"
},
{
"start": 964,
"end": 986,
"text": "(Conneau et al., 2019)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We compare the two approaches and show that the annotation projection approach amplifies the model's tendency to underpredict level 2 types, which lowers model performance. We also introduce three new test sets for German FET 1 on which zero-shot cross-lingual transfer performs better than models trained with German or a mix of German and English data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To the best of our knowledge there is no work that compares annotation projection directly against zero-shot cross-lingual transfer. While annotation projection has been used in a variety of tasks, there has not been a study of a case where this approach fails. Authors admit that the quality of the annotating system plays role (e.g. Ehrmann et al. (2011) ; Ni et al. (2017) ), but they don't specify model properties that are necessary for the approach to work, instead focusing on ways to mitigate noise. Pires et al. (2019) ; Hsu et al. (2019) and Artetxe and Schwenk (2019) show the strengths of zeroshot cross-lingual transfer on a variety of different NLP tasks, but they do not address fine-grained entity typing. Zhao et al. (2020) conclude that zeroshot performance can be improved by choosing a small amount of high quality training data from the target language. We test their approach for the FET scenario, but arrive at unclear results.",
"cite_spans": [
{
"start": 335,
"end": 356,
"text": "Ehrmann et al. (2011)",
"ref_id": "BIBREF6"
},
{
"start": 359,
"end": 375,
"text": "Ni et al. (2017)",
"ref_id": "BIBREF12"
},
{
"start": 508,
"end": 527,
"text": "Pires et al. (2019)",
"ref_id": "BIBREF14"
},
{
"start": 530,
"end": 547,
"text": "Hsu et al. (2019)",
"ref_id": "BIBREF7"
},
{
"start": 552,
"end": 578,
"text": "Artetxe and Schwenk (2019)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "In this work we use the hierarchical typing model of Chen et al. (2020) trained on English gold data for the zero-shot approach and also to annotate the English side of the parallel text for annotation projection. We train the model with English silver data to show the amount of noise added by automatic annotation and finally we train it with German data which was produced by annotation projection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method",
"sec_num": "3"
},
{
"text": "In the hierarchical typing model the entity and its context are encoded using multilingual XLM-RoBERTa (Conneau et al., 2019) . For each type in the FIGER ontology the model learns a type embedding. It passes the concatenated entity and context vector through a 2-layer feed-forward network that maps into the same space as the type embedding. The score is the inner product between the transformed entity and context vector and the type embedding. For further model details refer to Chen et al. (2020) .",
"cite_spans": [
{
"start": 103,
"end": 125,
"text": "(Conneau et al., 2019)",
"ref_id": "BIBREF4"
},
{
"start": 484,
"end": 502,
"text": "Chen et al. (2020)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Method",
"sec_num": "3"
},
{
"text": "Training Data To contrast the zero-shot crosslingual transfer approach with models trained on automatically annotated and projected data we use three sources of training data. We use the 2M sentences English FIGER corpus as described by (Ling and Weld, 2012) as a source of English human annotated data, which we will refer to as EN gold. The data set consists of English Wikipedia articles and we use it to train the zero-shot gold model. Second, we use English machine annotated data (EN automatic). Annotating English data using a model is the first step of the annotation projection. We use this data to train the zero-shot automatic model to examine the amount of noise added by automatic annotation. We generate EN automatic from English sentences from the Wiki-Matrix corpus Schwenk et al. (2019) 2 , using the hierarchical typing model trained on 2 M sentences EN gold. Lastly, we use German annotation projected data (DE projected) that was generated by projecting the labels from the EN automatic onto German. For the details of our annotation projection pipeline please refer to appendix A. We use this data to train the annotation projected model.",
"cite_spans": [
{
"start": 237,
"end": 258,
"text": "(Ling and Weld, 2012)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "4"
},
{
"text": "We portion each training corpus into slices of 100, 200, 300 and 400 K sentences to compare the influence of data size. For DE projected only 300 K sentences are available, because only part of the parallel sentences in the WikiMatrix corpus are of high enough quality for annotation projection. For details of the selection process refer to appendix A.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "4"
},
{
"text": "An important point for our experiments is the label distribution in the training corpora (see to underpredict the finer-grained level 2 labels (e.g. /person/actor, as opposed to level 1 label /person), which leads to a different distribution of labels in EN gold and the other corpora. Compared to approximately 100 K level 2 labels per 100 K sentences in the gold data, we only see about 50 K level 2 labels in the silver data. This tendency does not depend on the different input data: If we use a model trained on 100 K EN gold to predict labels on an unseen portion of EN gold, only 25% of the resulting annotations are level 2 labels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "4"
},
{
"text": "Metrics Following previous FET literature we evaluate the results of our model using strict accuracy (Acc). The strict accuracy is the ratio of instances where the predicted type set is exactly the same as the gold type set. We also evaluate per hierarchy level.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "4"
},
{
"text": "Test sets We compare performance using the following test corpora: 1) a German machine translation of the test split of the English FIGER corpus (Ling and Weld, 2012) using DeepL, which was manually corrected to eliminate translation and labelling errors (DE-FIGER); 2) 500 manually annotated German sentences from the WikiMatrix corpus (DE-Wiki), which we consider to be more challenging than DE-FIGER, because it contains a wider range of type labels; 3) a small challenge set of 135 sentences taken from DE-Wiki, in which we replaced entities with close string matches to English (e.g. 'Pr\u00e4sident Nixon') with specifically German entities of the same type (e.g. 'Bundeskanzler Kohl'), which we call DE-GermEnt; and 4) for experiments where we mix German and English data, we also compare against test split of the English FIGER corpus (Ling and Weld, 2012) (EN-FIGER). Data set statistics can be seen in table 2.",
"cite_spans": [
{
"start": 145,
"end": 166,
"text": "(Ling and Weld, 2012)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "4"
},
{
"text": "Monolingual training Figure 2 automatic and annotation projected at different training data sizes on the DE-FIGER and the DE-Wiki test sets. Zero-shot gold outperforms zero-shot automatic and annotation projected on both test sets and in all training data sizes. Zero-shot gold trained on the full EN gold data set of 2 M sentences performs only 1% better on level 1 labels and 3% better on level 2 labels than a model trained with 400 K sentences, which shows that smaller data slices are sufficient to reach most of the possible performance with this data set. While for level 1 type labels annotation projected gets close to the performance of zero-shot gold on both test sets, on level 2 type labels the system falls behind zero-shot gold, with a wider gap on DE-Wiki. The comparison between zero-shot automatic and annotation projected is less clear. On the DE-Wiki test set annotation projected consistently outperforms zero-shot automatic, while on DE-FIGER both systems perform very similarly.",
"cite_spans": [],
"ref_spans": [
{
"start": 21,
"end": 29,
"text": "Figure 2",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The high performance of zero-shot gold and the noisier zero-shot automatic might be due to the quality of English and German embeddings in XLM-RoBERTa, as both are high resource languages from the same language family. This confirms Lauscher et al. (2020) who show that this method works especially well for close high resource language pairs and low level semantic tasks. The noise introduced by annotation projection affects level 2 label performance the most (see appendix B and table 1). But the amount of level 2 labels in the training data can not be the only reason for this. The total number of labels in the silver corpora (see table 1) shows that 200 K of silver training data contain approximately the same amount of level 2 labels as 100 K of gold data. Nevertheless, the level 2 performance of systems trained on 200 K of silver data lies behind the model trained on 100 K of EN gold. This points towards the possibility, that not only the amount of level 2 labels in the training data, but also their quality and their Multilingual training The underlying XLM-RoBERTa embeddings allow to train a model with both German and English data. For this we combine slices from DE projected with EN automatic, because these data sets have the same distribution of labels. Table 3 shows the performance of a model trained with evenly mixed data (EN+DE) in comparison with monolingually trained models of the same size tested on DE-FIGER and EN-FIGER. German performance benefits from using both German and English training data, while performance in English is best with only English data. The mixed model does not outperform zero-shot gold on these test sets.",
"cite_spans": [
{
"start": 233,
"end": 255,
"text": "Lauscher et al. (2020)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 1277,
"end": 1284,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The low performance in the data mixing scenario compared to zero-shot gold can be explained with the distribution of labels in the silver corpora. Due to the noise added when labels are projected from English to German, the mixed model tested in German profits from the addition of higher quality English data, but not vice versa.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Few-shot training Zhao et al. 2020 mance. To test this we take a model trained on 100 K sentences EN gold and fine-tune it by training on the 135 sentence manually annotated DE-GermEnt data set. We evaluate the resulting model's performance on DE-FIGER. In comparison with the model trained on 100 K EN gold only, the performance of the resulting model is 10% lower in accuracy of level 1 labels and 12% lower on level 2 labels. We did not specifically select which sentences to use like Zhao et al. 2020, which is an avenue for future work. The low performance of the few-shot model could be due to the high number of different labels, only a few of which can be observed during few-shot training, but further work is needed to confirm this. German entities To challenge zero-shot gold, we test a model trained 2 M sentences EN gold on the test set DE-GermEnt. Surprisingly, we find that the model performs better on DE-GermEnt than on its English entity counter part, with 1% higher performance on level 1 labels and 3% higher performance on level 2 labels. It is unclear why zero-shot gold behaves this way, and examining this with larger challenge data sets it an avenue for future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Our results show that zero-shot cross-lingual transfer building upon XLM-RoBERTa is a strong baseline for the task of FET and the language pair of English and German. It outperforms annotation projection on three new test sets. We also show that in our specific scenario annotation projection using the hierarchical typing model amplifies the models tendency to underpredict level 2 types.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and Conclusions",
"sec_num": "6"
},
{
"text": "One way to mitigate these shortcomings would be to sample level 1 and level 2 labels in a training corpus so that they have the same distribution as in the gold data, although this would not control for data quality. Another way could be to machine translate the manually annotated English corpus into German and then use annotation projection, as suggested by Ehrmann et al. (2011) . This way the label distribution of the human annotated data could be preserved as well. Lastly, improving the few-shot approach and designing more challenging test sets are other avenues to explore. ",
"cite_spans": [
{
"start": 361,
"end": 382,
"text": "Ehrmann et al. (2011)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and Conclusions",
"sec_num": "6"
},
{
"text": "A.1 Preprocessing A diagram of our pipeline can be seen in figure 3. To annotate the English halves of our parallel corpora with FIGER types preprocessing is necessary. Due to its automatic creation the WikiMatrix corpus contains a small amount of German sentences in its English half and English sentences in its German half, the translations of which are assigned very high confidence. We remove these by discarding the 5000 highest-confidence sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A German Data Creation",
"sec_num": null
},
{
"text": "To enable annotation by the English FET system, we run a named entity recognition system over the English input sentences (see the second box of Figure 3 ). We used the NER component of Stanza (Qi et al., 2020) for this task. We then use the English FET model to assign FIGER types to the named entities (see the third box of Figure 3 ). The FET model only annotates one entity per sentence. Sentences that contain more than one named entity occur multiple times in the English input, so that each entity receives an annotation.",
"cite_spans": [
{
"start": 193,
"end": 210,
"text": "(Qi et al., 2020)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 145,
"end": 153,
"text": "Figure 3",
"ref_id": "FIGREF3"
},
{
"start": 326,
"end": 334,
"text": "Figure 3",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "A German Data Creation",
"sec_num": null
},
{
"text": "We use ZAP to obtain a word alignment between the English and German halves of our parallel corpora (see the fourth box of Figure 3 ). While the similar tools fast_align (Dyer et al., 2013) and Giza++ (Och and Ney, 2003) are language agnostic, ZAP's model for English-German word alignment uses probabilities computed from large-scale parallel corpora. We then use our own code to project the fine grained entity type labels from the annotated English text to its German translation. We use static rules to filter out misalignments, e.g. discarding all cases where not all words of an entity were aligned. We then use the resulting German FET annotated corpus to train our FET model. Because of the ordering by alignment quality in the machine-aligned WikiMatrix corpus, we introduce a preprocessing epoch to the training to mitigate noisy input. During training the model receives the sentences in exactly the order that they occur in the corpus. In the WikiMatrix corpus the sentences are sorted by the confidence of the alignment algorithm. This means that the sentences towards the bottom of the corpus are more likely to be incorrectly aligned. Incorrectly aligned sentences are more likely to have incorrectly projected labels. Therefor the quality of FIGER type annotations in the resulting German Figure 5 : Loss rises and level 1 label accuracy deteriorates as the quality of samples gets worse towards the end of the automatically aligned and sorted corpus. The graphs show a cut-off point at approximately 300 thousand sentences. We use this information to select a high-quality slice of the corpus to train our system. data is higher towards the beginning of the corpus and lower towards its end.",
"cite_spans": [
{
"start": 170,
"end": 189,
"text": "(Dyer et al., 2013)",
"ref_id": "BIBREF5"
},
{
"start": 194,
"end": 220,
"text": "Giza++ (Och and Ney, 2003)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 123,
"end": 131,
"text": "Figure 3",
"ref_id": "FIGREF3"
},
{
"start": 1305,
"end": 1313,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "A.2 Annotation projection and training with noise mitigation",
"sec_num": null
},
{
"text": "During the first epoch of training this drop in quality can be observed in the change of learning rate and the accuracy of predictions after approximately 300,000 sentences (see Figure 5 ). These curves give us important information about what portion of the data is clean enough to be used in the following epochs of training. It gives us a possible cut off point for our data set at 300,000 sentences, so that in the next epochs we only train on a slice of the corpus before this point.",
"cite_spans": [],
"ref_spans": [
{
"start": 178,
"end": 186,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "A.2 Annotation projection and training with noise mitigation",
"sec_num": null
},
{
"text": "To show the effect of increasing training data size we select for our experiments 3 slices of data that were processed before the cut off point: the first 100K, 200K and 300K sentences of the corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A.2 Annotation projection and training with noise mitigation",
"sec_num": null
},
{
"text": "The hierarchical typing model of Chen et al. (2020) 's model tends to under-predict level 2 labels. This property of the model is exacerbated by our annotation projection approach, because we use data generated by this model in English to train the same model in German. Figure 4 shows the confusion matrices for level 1 and level 2 labels for EN gold 2 Mil and DE silver 200K. While for level 1 labels in EN gold there is no dominant class that labels are misclassified to, the most common misprediction for level 2 labels is to assign no label at all, which can be seen as the dotted vertical line in the second upper confusion matrix. When comparing the upper confusion matrices to the lower ones, it becomes clear that this trend to under-predict level 2 labels is even stronger in the German model. This makes sense because the German model is trained on output from the English model, which contains fewer level 2 labels than the human annotated data used to train the English model in he first place. The German model sees less level 2 labels in its training data and therefore doesn't learn to predict them.",
"cite_spans": [
{
"start": 33,
"end": 51,
"text": "Chen et al. (2020)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [
{
"start": 271,
"end": 279,
"text": "Figure 4",
"ref_id": "FIGREF4"
}
],
"eq_spans": [],
"section": "B Confusion matrices",
"sec_num": null
},
{
"text": "In keeping with the EMNLP reproducibility guidelines we report the specifications of the systems that our models where trained on. We trained all models using a single GeForce RTX 2080 Ti GPU. Running the largest model (EN zero-shot trained on 2 million sentences) took approximately 8 hours. Training the other models took under an hour per model. The number of model parameters is 50484362. All hyperparameters of the model were taken from the implementation of Chen et al. (2020) . For the few-shot experiment we increased the training epoch number to stop if there was no more improvement on the development set.",
"cite_spans": [
{
"start": 464,
"end": 482,
"text": "Chen et al. (2020)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "C System and model specifications",
"sec_num": null
},
{
"text": "all test data sets and relevant code are available under https://github.com/webersab/german_FET",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "While higher quality parallel data sets are available, this is the only one in the domain of Wikipedia articles. Preliminary experiments have shown that domain is an important factor for the quality of automatic FET, which is why we chose domain consistency over data quality for our experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This work was funded by the ERC H2020 Advanced Fellowship GA 742137 SEMANTAX.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": "7"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Building named entity recognition taggers via parallel corpora",
"authors": [
{
"first": "Rodrigo",
"middle": [],
"last": "Agerri",
"suffix": ""
},
{
"first": "Yiling",
"middle": [],
"last": "Chung",
"suffix": ""
},
{
"first": "Itziar",
"middle": [],
"last": "Aldabe",
"suffix": ""
},
{
"first": "Nora",
"middle": [],
"last": "Aranberri",
"suffix": ""
},
{
"first": "Gorka",
"middle": [],
"last": "Labaka",
"suffix": ""
},
{
"first": "German",
"middle": [],
"last": "Rigau",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rodrigo Agerri, Yiling Chung, Itziar Aldabe, Nora Aranberri, Gorka Labaka, and German Rigau. 2018. Building named entity recognition taggers via paral- lel corpora. In Proceedings of the Eleventh Interna- tional Conference on Language Resources and Eval- uation (LREC 2018).",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond",
"authors": [
{
"first": "Mikel",
"middle": [],
"last": "Artetxe",
"suffix": ""
},
{
"first": "Holger",
"middle": [],
"last": "Schwenk",
"suffix": ""
}
],
"year": 2019,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "7",
"issue": "",
"pages": "597--610",
"other_ids": {
"DOI": [
"10.1162/tacl_a_00288"
]
},
"num": null,
"urls": [],
"raw_text": "Mikel Artetxe and Holger Schwenk. 2019. Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond. Transactions of the Association for Computational Linguistics, 7:597-610.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Freebase: a collaboratively created graph database for structuring human knowledge",
"authors": [
{
"first": "Kurt",
"middle": [],
"last": "Bollacker",
"suffix": ""
},
{
"first": "Colin",
"middle": [],
"last": "Evans",
"suffix": ""
},
{
"first": "Praveen",
"middle": [],
"last": "Paritosh",
"suffix": ""
},
{
"first": "Tim",
"middle": [],
"last": "Sturge",
"suffix": ""
},
{
"first": "Jamie",
"middle": [],
"last": "Taylor",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 2008 ACM SIGMOD international conference on Management of data",
"volume": "",
"issue": "",
"pages": "1247--1250",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collab- oratively created graph database for structuring hu- man knowledge. In Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 1247-1250.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Hierarchical entity typing via multi-level learning to rank",
"authors": [
{
"first": "Tongfei",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Yunmo",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Van Durme",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "2020",
"issue": "",
"pages": "8465--8475",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tongfei Chen, Yunmo Chen, and Benjamin Van Durme. 2020. Hierarchical entity typing via multi-level learning to rank. In Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics, ACL 2020, Online, July 5-10, 2020, pages 8465-8475.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Unsupervised cross-lingual representation learning at scale",
"authors": [
{
"first": "Alexis",
"middle": [],
"last": "Conneau",
"suffix": ""
},
{
"first": "Kartikay",
"middle": [],
"last": "Khandelwal",
"suffix": ""
},
{
"first": "Naman",
"middle": [],
"last": "Goyal",
"suffix": ""
},
{
"first": "Vishrav",
"middle": [],
"last": "Chaudhary",
"suffix": ""
},
{
"first": "Guillaume",
"middle": [],
"last": "Wenzek",
"suffix": ""
},
{
"first": "Francisco",
"middle": [],
"last": "Guzm\u00e1n",
"suffix": ""
},
{
"first": "Edouard",
"middle": [],
"last": "Grave",
"suffix": ""
},
{
"first": "Myle",
"middle": [],
"last": "Ott",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Veselin",
"middle": [],
"last": "Stoyanov",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1911.02116"
]
},
"num": null,
"urls": [],
"raw_text": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm\u00e1n, Edouard Grave, Myle Ott, Luke Zettle- moyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A simple, fast, and effective reparameterization of ibm model 2",
"authors": [
{
"first": "Chris",
"middle": [],
"last": "Dyer",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Chahuneau",
"suffix": ""
},
{
"first": "Noah A",
"middle": [],
"last": "Smith",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "644--648",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chris Dyer, Victor Chahuneau, and Noah A Smith. 2013. A simple, fast, and effective reparameteriza- tion of ibm model 2. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 644-648.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Building a multilingual named entityannotated corpus using annotation projection",
"authors": [
{
"first": "Maud",
"middle": [],
"last": "Ehrmann",
"suffix": ""
},
{
"first": "Marco",
"middle": [],
"last": "Turchi",
"suffix": ""
},
{
"first": "Ralf",
"middle": [],
"last": "Steinberger",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the International Conference Recent Advances in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "118--124",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maud Ehrmann, Marco Turchi, and Ralf Steinberger. 2011. Building a multilingual named entity- annotated corpus using annotation projection. In Proceedings of the International Conference Recent Advances in Natural Language Processing 2011, pages 118-124.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Zero-shot reading comprehension by crosslingual transfer learning with multi-lingual language representation model",
"authors": [
{
"first": "Tsung-Yuan",
"middle": [],
"last": "Hsu",
"suffix": ""
},
{
"first": "Chi-Liang",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Hung-Yi",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "5933--5940",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1607"
]
},
"num": null,
"urls": [],
"raw_text": "Tsung-Yuan Hsu, Chi-Liang Liu, and Hung-yi Lee. 2019. Zero-shot reading comprehension by cross- lingual transfer learning with multi-lingual lan- guage representation model. In Proceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5933-5940, Hong Kong, China. Association for Computational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Improving neural relation extraction with implicit mutual relations",
"authors": [
{
"first": "Jun",
"middle": [],
"last": "Kuang",
"suffix": ""
},
{
"first": "Yixin",
"middle": [],
"last": "Cao",
"suffix": ""
},
{
"first": "Jianbing",
"middle": [],
"last": "Zheng",
"suffix": ""
},
{
"first": "Xiangnan",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Ming",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Aoying",
"middle": [],
"last": "Zhou",
"suffix": ""
}
],
"year": 2020,
"venue": "2020 IEEE 36th International Conference on Data Engineering (ICDE)",
"volume": "",
"issue": "",
"pages": "1021--1032",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jun Kuang, Yixin Cao, Jianbing Zheng, Xiangnan He, Ming Gao, and Aoying Zhou. 2020. Improving neu- ral relation extraction with implicit mutual relations. In 2020 IEEE 36th International Conference on Data Engineering (ICDE), pages 1021-1032. IEEE.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "From zero to hero: On the limitations of zero-shot language transfer with multilingual transformers",
"authors": [
{
"first": "Anne",
"middle": [],
"last": "Lauscher",
"suffix": ""
},
{
"first": "Vinit",
"middle": [],
"last": "Ravishankar",
"suffix": ""
},
{
"first": "Ivan",
"middle": [],
"last": "Vuli\u0107",
"suffix": ""
},
{
"first": "Goran",
"middle": [],
"last": "Glava\u0161",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "4483--4499",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anne Lauscher, Vinit Ravishankar, Ivan Vuli\u0107, and Goran Glava\u0161. 2020. From zero to hero: On the lim- itations of zero-shot language transfer with multilin- gual transformers. In Proceedings of the 2020 Con- ference on Empirical Methods in Natural Language Processing (EMNLP), pages 4483-4499.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Cross-lingual named entity recognition using parallel corpus: A new approach using xlm-roberta alignment",
"authors": [
{
"first": "Bing",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Yujie",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Wenjin",
"middle": [],
"last": "Xu",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2101.11112"
]
},
"num": null,
"urls": [],
"raw_text": "Bing Li, Yujie He, and Wenjin Xu. 2021. Cross-lingual named entity recognition using parallel corpus: A new approach using xlm-roberta alignment. arXiv preprint arXiv:2101.11112.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Fine-grained entity recognition",
"authors": [
{
"first": "Xiao",
"middle": [],
"last": "Ling",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Daniel S Weld",
"suffix": ""
}
],
"year": 2012,
"venue": "AAAI",
"volume": "12",
"issue": "",
"pages": "94--100",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Xiao Ling and Daniel S Weld. 2012. Fine-grained en- tity recognition. In AAAI, volume 12, pages 94-100.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Weakly supervised cross-lingual named entity recognition via effective annotation and representation projection",
"authors": [
{
"first": "Jian",
"middle": [],
"last": "Ni",
"suffix": ""
},
{
"first": "Georgiana",
"middle": [],
"last": "Dinu",
"suffix": ""
},
{
"first": "Radu",
"middle": [],
"last": "Florian",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "1470--1480",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jian Ni, Georgiana Dinu, and Radu Florian. 2017. Weakly supervised cross-lingual named entity recog- nition via effective annotation and representation projection. In Proceedings of the 55th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1470-1480.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "A systematic comparison of various statistical alignment models",
"authors": [
{
"first": "Josef",
"middle": [],
"last": "Franz",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2003,
"venue": "Computational Linguistics",
"volume": "29",
"issue": "1",
"pages": "19--51",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19-51.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "How multilingual is multilingual BERT?",
"authors": [
{
"first": "Telmo",
"middle": [],
"last": "Pires",
"suffix": ""
},
{
"first": "Eva",
"middle": [],
"last": "Schlinger",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Garrette",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "4996--5001",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1493"
]
},
"num": null,
"urls": [],
"raw_text": "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Pro- ceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 4996- 5001, Florence, Italy. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Stanza: A Python natural language processing toolkit for many human languages",
"authors": [
{
"first": "Peng",
"middle": [],
"last": "Qi",
"suffix": ""
},
{
"first": "Yuhao",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Yuhui",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Bolton",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python natural language processing toolkit for many human languages. In Proceedings of the 58th An- nual Meeting of the Association for Computational Linguistics: System Demonstrations.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Wikimatrix: Mining 135m parallel sentences in 1620 language pairs from wikipedia",
"authors": [
{
"first": "Holger",
"middle": [],
"last": "Schwenk",
"suffix": ""
},
{
"first": "Vishrav",
"middle": [],
"last": "Chaudhary",
"suffix": ""
},
{
"first": "Shuo",
"middle": [],
"last": "Sun",
"suffix": ""
},
{
"first": "Hongyu",
"middle": [],
"last": "Gong",
"suffix": ""
},
{
"first": "Francisco",
"middle": [],
"last": "Guzm\u00e1n",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1907.05791"
]
},
"num": null,
"urls": [],
"raw_text": "Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzm\u00e1n. 2019. Wiki- matrix: Mining 135m parallel sentences in 1620 language pairs from wikipedia. arXiv preprint arXiv:1907.05791.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Inducing multilingual pos taggers and NP bracketers via robust projection across aligned corpora",
"authors": [
{
"first": "David",
"middle": [],
"last": "Yarowsky",
"suffix": ""
},
{
"first": "Grace",
"middle": [],
"last": "Ngai",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Yarowsky and Grace Ngai. 2001. Inducing mul- tilingual pos taggers and NP bracketers via robust projection across aligned corpora. In Proceedings",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"num": null,
"text": "An example of fine-grained entity typing with the FIGER ontology. Correct types are highlighted.",
"uris": null
},
"FIGREF2": {
"type_str": "figure",
"num": null,
"text": "Zero-shot cross-lingual transfer performs best on both German data sets. EN automatic and DE projected perform similar on both data sets, with a wider gap in level 2 performance on DE-Wiki proportion with level 1 labels play a role here.",
"uris": null
},
"FIGREF3": {
"type_str": "figure",
"num": null,
"text": "Our annotation projection setup uses parallel text and an automatic named entity recognition component to generate an annotated corpus in German.of the 2nd Meeting of the North American Chapter of the Association for Computational Linguistics, pages XXX-XXX. ACL.Mengjie Zhao, Yi Zhu, Ehsan Shareghi, Roi Reichart, Anna Korhonen, and Hinrich Sch\u00fctze. 2020. A closer look at few-shot crosslingual transfer: Variance, benchmarks and baselines. arXiv preprint arXiv:2012.15682.",
"uris": null
},
"FIGREF4": {
"type_str": "figure",
"num": null,
"text": "The upper two matrices show level 1 and level 2 performance of EN gold 2 M, the lower two matrices show the same for DE silver 200K. While the English model has a slight tendency to predict no label for level 2 label, this tendency is stronger in the German model. The yellow vertical line shows this effect.",
"uris": null
},
"TABREF0": {
"text": ". The hierarchical typing model has the tendency EN gold EN aut. DE proj.",
"content": "<table><tr><td>Lvl1 labels 60%</td><td>78%</td><td>77%</td></tr><tr><td>Lvl2 labels 40%</td><td>22%</td><td>23%</td></tr><tr><td colspan=\"3\">Lvl1 labels 155679 148571 150166</td></tr><tr><td colspan=\"2\">Lvl2 labels 104807 42008</td><td>43604</td></tr></table>",
"type_str": "table",
"num": null,
"html": null
},
"TABREF1": {
"text": "Percentage and total numbers of level 1 and level 2 labels in 100 K sentences of the training corpora. Data created by annotation with the hierarchicaltyping model contains fewer level 2 labels than human annotated gold data.",
"content": "<table/>",
"type_str": "table",
"num": null,
"html": null
},
"TABREF2": {
"text": "compares the performance of the models zero-shot gold, zero-shot",
"content": "<table><tr><td/><td>size</td><td colspan=\"2\">unique lab. t. lab.</td></tr><tr><td>EN-FIGER</td><td>563 sent.</td><td>42</td><td>624</td></tr><tr><td>DE-FIGER</td><td>563 sent.</td><td>42</td><td>624</td></tr><tr><td>DE-Wiki</td><td>500 sent.</td><td>57</td><td>771</td></tr><tr><td colspan=\"2\">DE-GermEnt 135 sent.</td><td>34</td><td>213</td></tr></table>",
"type_str": "table",
"num": null,
"html": null
},
"TABREF3": {
"text": "Statistics of the different test sets listing size, number of unique labels and total number of labels. While DE-FIGER is parallel to the commonly used English FIGER test set, DE-Wiki contains more unique labels and more labels in total.",
"content": "<table/>",
"type_str": "table",
"num": null,
"html": null
},
"TABREF5": {
"text": "Performance of a model trained with both English and German data, in comparison with monolingual data tested on parallel test sets. While performance in German is best with mixed data, performance in English is best with only English training data.",
"content": "<table/>",
"type_str": "table",
"num": null,
"html": null
}
}
}
} |