File size: 77,865 Bytes
6fa4bc9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 | {
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T09:42:28.607810Z"
},
"title": "Developing a How-to Tip Machine Comprehension Dataset and its Evaluation in Machine Comprehension by BERT",
"authors": [
{
"first": "Tengyang",
"middle": [],
"last": "Chen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tsukuba",
"location": {
"country": "Japan"
}
},
"email": ""
},
{
"first": "Hongyu",
"middle": [],
"last": "Li",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tsukuba",
"location": {
"country": "Japan"
}
},
"email": ""
},
{
"first": "Miho",
"middle": [],
"last": "Kasamatsu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tsukuba",
"location": {
"country": "Japan"
}
},
"email": ""
},
{
"first": "Takehito",
"middle": [],
"last": "Utsuro",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tsukuba",
"location": {
"country": "Japan"
}
},
"email": ""
},
{
"first": "Yasuhide",
"middle": [],
"last": "Kawada",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Logworks Co., Ltd",
"location": {
"country": "Japan"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In the field of factoid question answering (QA), it is known that the state-of-the-art technology has achieved an accuracy comparable to that of humans in a certain benchmark challenge. On the other hand, in the area of nonfactoid QA, there is still a limited number of datasets for training QA models, i.e., machine comprehension models. Considering such a situation within the field of the non-factoid QA, this paper aims to develop a dataset for training Japanese how-to tip QA models. This paper applies one of the state-of-the-art machine comprehension models to the Japanese how-to tip QA dataset. The trained how-to tip QA model is also compared with a factoid QA model trained with a Japanese factoid QA dataset. Evaluation results revealed that the how-to tip machine comprehension performance was almost comparative with that of the factoid machine comprehension even with the training data size reduced to around 4% of the factoid machine comprehension. Thus, the how-to tip machine comprehension task requires much less training data compared with the factoid machine comprehension task.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "In the field of factoid question answering (QA), it is known that the state-of-the-art technology has achieved an accuracy comparable to that of humans in a certain benchmark challenge. On the other hand, in the area of nonfactoid QA, there is still a limited number of datasets for training QA models, i.e., machine comprehension models. Considering such a situation within the field of the non-factoid QA, this paper aims to develop a dataset for training Japanese how-to tip QA models. This paper applies one of the state-of-the-art machine comprehension models to the Japanese how-to tip QA dataset. The trained how-to tip QA model is also compared with a factoid QA model trained with a Japanese factoid QA dataset. Evaluation results revealed that the how-to tip machine comprehension performance was almost comparative with that of the factoid machine comprehension even with the training data size reduced to around 4% of the factoid machine comprehension. Thus, the how-to tip machine comprehension task requires much less training data compared with the factoid machine comprehension task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Recent advances in the field of QA or machine comprehension are mostly in the domain of factoid QA related to Wikipedia articles and news articles (Yi et al., 2015; Pranav et al., 2016 Pranav et al., , 2018 . One of the most well-known QA datasets and benchmark tests is the Stanford Question Answering Dataset (SQuAD) (Pranav et al., 2016 (Pranav et al., , 2018 , which is a reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a text segment, or span, from the corresponding reading passage, or the question might be unanswerable. It is reported 1 that state-of-theart machine comprehension models trained with SQuAD outperform humans (Devlin et al., 2019; Zhang et al., 2019) .",
"cite_spans": [
{
"start": 147,
"end": 164,
"text": "(Yi et al., 2015;",
"ref_id": "BIBREF12"
},
{
"start": 165,
"end": 184,
"text": "Pranav et al., 2016",
"ref_id": "BIBREF10"
},
{
"start": 185,
"end": 206,
"text": "Pranav et al., , 2018",
"ref_id": "BIBREF11"
},
{
"start": 319,
"end": 339,
"text": "(Pranav et al., 2016",
"ref_id": "BIBREF10"
},
{
"start": 340,
"end": 362,
"text": "(Pranav et al., , 2018",
"ref_id": "BIBREF11"
},
{
"start": 731,
"end": 752,
"text": "(Devlin et al., 2019;",
"ref_id": "BIBREF4"
},
{
"start": 753,
"end": 772,
"text": "Zhang et al., 2019)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "However, apart from the issues related to developing benchmark datasets for factoid QA and improving state-of-the-art general-purpose machine comprehension models, there is a relatively limited number of published literature that handles issues, such as the development of datasets for nonfactoid QA and the application of state-of-the-art general-purpose machine comprehension models to those non-factoid datasets. Typical non-factoid QA tasks include opinion QA, definition QA, reason QA, and how-to tip QA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Among various kinds of non-factoid knowledge which are the key to developing techniques for non-factoid QA tasks, a recent study (Ohkawa et al., 2018) examined the types of Japanese websites which include various howto tips related to job hunting, marriage, and apartment. The study (Ohkawa et al., 2018 ) also aims to automatically identify those how-to tip websites, which will be an important knowledge source for training how-to tip QA models. Considering such a situation, within the field of nonfactoid QA, this paper studies how to develop a dataset for training Japanese how-to tip (hereafter throughout the paper, we use the simplified term \"tip\") QA models. As examples in this paper, we developed tip QA datasets for 'job hunting,\" \"marriage,\" \"apartment,\" \"hay fever,\" \"dentist,\" and \"food poisoning,\" where \"job hunting\" and \"marriage\" tip QAs are for both training and testing, while other tip QAs are only for testing. For \"job hunting\", Figure 1 presents a typical example of a tuple of a context, a tip question, and an answer. Furthermore, in order to understand rough idea of Figure 1 : An example of the machine comprehension model of tip QA for \"job hunting\" together with an example of a tuple of a context C, a question Q, and answer A (extracted from a column web page entitled \"Formatting Tips for Your Curriculum Vitae (CV)\" (https://www.thebalancecareers.com/ curriculum-vitae-format-2060351) from a tip website titled \"The Balance Careers\" (https://www.thebalancecareers.com/) ) the (how-to) tip questions we study in this paper in the broader sense, we manually classify them into five types as shown in Table 1 and also shown several examples for each of the five types and their statistics within the dataset we developed in this paper.",
"cite_spans": [
{
"start": 129,
"end": 150,
"text": "(Ohkawa et al., 2018)",
"ref_id": "BIBREF9"
},
{
"start": 283,
"end": 303,
"text": "(Ohkawa et al., 2018",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 953,
"end": 961,
"text": "Figure 1",
"ref_id": null
},
{
"start": 1095,
"end": 1103,
"text": "Figure 1",
"ref_id": null
},
{
"start": 1633,
"end": 1640,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This paper then applies BERT (Devlin et al., 2019) , one of the state-of-the-art machine comprehension models, to a Japanese tip QA dataset. The trained tip QA model is also compared with a factoid QA model which is also trained with a Japanese factoid QA dataset. Evaluation results revealed that the tip machine comprehension performance was almost comparative with that of the factoid machine comprehension even with the training data size reduced to around 4% of the factoid machine comprehension. Thus, the tip machine comprehension task requires much less training data compared with the factoid machine comprehension task.",
"cite_spans": [
{
"start": 29,
"end": 50,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This section briefly describes the workflow of collecting web pages. First, the notion of query focus is a keyword used for every search request related to a specific subject. For example, whenever the aim was to collect web pages about anything related to job hunting, the word \"job hunting\" was always put at the beginning of the query, and all available suggested keywords provided by the search engine were collected, such as \"job hunting self-promotion\" and \"job hunting portfolio.\" Using all such suggested keywords as queries (called search engine suggests or suggests), the search engine is crawled, and top 10 results for each suggest are collected.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Query Focuses and Collecting Web Pages",
"sec_num": "2"
},
{
"text": "Web search engine suggests are the query keywords automatically offered by a search engine when a user types part of a search query. Such suggested keywords can be seen as frequent user activities logged by the search engine, and they mostly lead to pages on trending topics. For a given query focus keyword, about 100 specified types of Japanese hiragana characters were entered into Google R \u20dd search engine from which up to 1,000 suggests were collected. These 100 types of Japanese hiragana characters include the Japanese alphabet consisting of 50 characters, voiced and semi-voiced variants of voiceless characters and Youon. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Search Engine Suggests",
"sec_num": "2.1"
},
{
"text": "Google Custom Search API 2 was used to scrape web pages from the search engine. Using the web search engine suggests collected in the previous section combined with the query focus keyword as queries (in the form of AND search), the first 10 pages returned per search query are collected. The set of web pages queried by suggest s can be represented as D (s, N ) , where N is 10 as a constant standing for top N pages. Additionally, the search engine suggests were saved for every web page. Since different search engine suggests could lead to the same web page, one web page could have multiple suggests. Let S be the set of all suggests about one query focus. Then, the set of web pages scraped using all possible suggests is represented as D.",
"cite_spans": [
{
"start": 355,
"end": 362,
"text": "(s, N )",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages",
"sec_num": "2.2"
},
{
"text": "D = \u222a s\u2208S D(s, N )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages",
"sec_num": "2.2"
},
{
"text": "3 Selecting Candidates of Tip Websites",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages",
"sec_num": "2.2"
},
{
"text": "This paper employs LDA (latent Dirichlet allocation) (Blei et al., 2003) to model topic distributions among documents. Let D be a document set containing all collected web pages and K (= 50 in this paper) be the number of topics. When the topic model is applied, topic distribution P (z",
"cite_spans": [
{
"start": 53,
"end": 72,
"text": "(Blei et al., 2003)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages",
"sec_num": "2.2"
},
{
"text": "n | d) is available for every d (d \u2208 D). Every docu- ment d",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages",
"sec_num": "2.2"
},
{
"text": "is assigned a topic with the highest probability among all its P (z n | d). The net effect is that for every topic z n , there is a group D(z n ) (n = 1, . . . , K) of corresponding documents that are assigned to z n .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages",
"sec_num": "2.2"
},
{
"text": "Then, domain names are extracted from all collected web pages based on their URLs. The domain names that have corresponding web pages reside in 10 or more sets D(z n ) (n = 1, . . . , K), i.e., they have their web pages under more than or equal to 10 topics which are considered as candidates for tip websites 3 . Out of those candidates whose numbers are 31 for job-hunting in this experiment, 14 domain names were randomly selected, for all of which tip QAs were successfully collected. Henceforth, the set of those 14 tips websites will be denoted as T . Similarly, for marriage, 13 domain names have their web pages under more than or equal to 10 topics and are considered as candidates for tip websites. For all of those 13 domain names, tip QAs were successfully collected. Thus, for marriage, the set of those 13 tips websites will be denoted as T .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages",
"sec_num": "2.2"
},
{
"text": "From each website out of the set T of tip websites, web pages are collected as the source for collecting tip QAs. First, from each website of T , all of its web pages are collected into set D inf (T ). Then, the LDA topic model (Blei et al., 2003) 2: # Pairs of QA collected from a web page (for \"job hunting\" and \"marriage\") tion 3 with the set D of web pages scraped using all possible suggests is applied to each web page d within set D inf (T ). According to the probability distribution P (z n |d) of topics z n (n = 1, . . . , K) for each web page d, the topic z n with the highest probability is assigned to d. Then, the set of web pages to which the topic z n is assigned is denoted as D inf (z n , T ):",
"cite_spans": [
{
"start": 228,
"end": 247,
"text": "(Blei et al., 2003)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages of Tip Websites",
"sec_num": "4.1"
},
{
"text": "P (z n | d) (available for every d (d \u2208 D)) trained in Sec-",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages of Tip Websites",
"sec_num": "4.1"
},
{
"text": "D inf (z n , T ) = { d \u2208 D inf (T ) z n = argmax zu (u=1,...,K) P (z u |d) }",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages of Tip Websites",
"sec_num": "4.1"
},
{
"text": "For the query focus \"job hunting,\" out of the total K = 50 topics, |D inf (z n , T )| > 0 holds, i.e., at least one web page is assigned to 42 topics for job hunting and 29 for marriage.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Collecting Web Pages of Tip Websites",
"sec_num": "4.1"
},
{
"text": "This study analyzes the types of web pages which tend to include more and more tips compared with other types of web pages. This paper examines tip websites which include column pages containing various tips and also include other types of web pages, such as pages for commercial sale of products or pages with reviews and experiences. However, most tips are found only in column pages but not in other types of pages. The type of web pages which tend to include tips are mostly column pages. Out of the set D inf (z n , T ) of web pages defined in the previous section, all the column pages are extracted into a subset:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\"Column Pages\" as the Source for Collecting Tip QAs",
"sec_num": "4.2"
},
{
"text": "D inf c (z n , T )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\"Column Pages\" as the Source for Collecting Tip QAs",
"sec_num": "4.2"
},
{
"text": "In the case of the query focus \"job hunting,\" out of 42 topics satisfying |D inf (z n , T )| > 0, 36 topics satisfy |D inf c (z n , T )| > 0, i.e., include column pages. For \"marriage\", all the 29 topics satisfy |D inf c (z n , T )| > 0. For each topic z n , all the web pages in this set are used as a source for collecting tip QAs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\"Column Pages\" as the Source for Collecting Tip QAs",
"sec_num": "4.2"
},
{
"text": "This section describes the procedure for collecting tip questions and examples, such as those presented in Figure 1 . From each web page within the set D inf c (z n , T ) constructed in the previous section, tuples of context C, question Q, and answer A are manually collected. Specifically, within each column web page, every paragraph is examined, and it is decided whether a pair of a question and an answer can be collected from the paragraph. From each column web page, at most 5 pairs of a question and an answer are collected. Figure 1 presents an example of collecting a tuple of a context, a question, and an answer from a column web page of a \"job hunting\" tip website. In this example, context C, the following paragraph about font choice and font size is selected:",
"cite_spans": [],
"ref_spans": [
{
"start": 107,
"end": 115,
"text": "Figure 1",
"ref_id": null
},
{
"start": 534,
"end": 542,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Procedure for Collecting Tip QAs",
"sec_num": "4.3"
},
{
"text": "There's no need to use ornate fonts that are difficult to read; . . . Your font size should be between 10 and 12 points, although . . .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Procedure for Collecting Tip QAs",
"sec_num": "4.3"
},
{
"text": "From this paragraph, a pair of question Q \"What is the font size for CV?\" and answer A \"between 10 and 12 points\" is extracted. Table 2 lists the distribution of the number of the pairs of a question and an answer collected from a web page for \"job hunting\" and \"marriage\".",
"cite_spans": [],
"ref_spans": [
{
"start": 128,
"end": 135,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Procedure for Collecting Tip QAs",
"sec_num": "4.3"
},
{
"text": "For the query focus \"job hunting,\" out of the overall 1,268 column web pages collected following the procedure of this paper, 352 pages were actually examined, out of which 907 pairs of tip QAs are collected. For the query focus \"marriage,\" out of the overall 3,075 column web pages collected following the procedure of this paper, 114 pages were actually examined, out of which 432 pairs of tip QAs are collected. For \"apartment\" query focuses, 50 pairs of tip QAs are collected. For other query focuses \"hay fever,\" \"dentist,\" and \"food poisoning,\" a total of 50 pairs of tip QAs are collected. Table 3 presents an example of Japanese tip QAs for each of \"job hunting,\" \"marriage,\" and \"hay fever.\" These numbers and examples are all for SQuAD1.1 type answerable questions only. : Examples of Japanese tip QAs selected from training and test datasets used in evaluation (tuples of Context C, Question Q, and Answer A for query focuses \"job hunting,\" \"marriage,\" and \"hay fever\", for SQuAD1.1: answerable questions)",
"cite_spans": [],
"ref_spans": [
{
"start": 597,
"end": 604,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Procedure for Collecting Tip QAs",
"sec_num": "4.3"
},
{
"text": "From these tip QAs of SQuAD1.1 type with answerable questions, tip QAs of SQuAD2.0 type with unanswerable questions are manually created. From a tuple of a context C, a question Q, and an answer A of SQuAD1.1 type, which is answerable in that the context C includes the answer A to the question Q, the annotator manually created another tuple, which is an unanswerable QA, of a context C \u2032 (\u0338 = C), a question Q \u2032 (= Q), and the answer A \u2032 = \u27e8null\u27e9. Here, within exactly the same column web page of the tip website, from which the context C is extracted, the annotator searched for another paragraph other than C, which does not include any answer to the original question Q. The selected paragraph C \u2032 constitutes the context of a tip QA of SQuAD2.0 type with an unanswerable question. Note that it is quite important to search for C \u2032 within exactly the same column web page of the tip website, from which the context C is extracted. For example, in the case of the tip QA on \"job hunting\" in Figure 1 , for the question Q \"What is the font size for CV?\", within the same column web page about \"job hunting\", another paragraph C \u2032 other than C is selected. The selected paragraph C \u2032 still presents a certain tip about job hunting and CV, while it does not include any tip about the font size for CV. We follow this strategy simply because it avoids tip QAs with unanswerable questions becoming much easier to answer compared with tip QAs with answerable questions. With this strategy, for each of almost all the tip QAs of SQuAD1.1 type answerable questions, we successfully created at least one tip QA of SQuAD2.0 type with an unanswerable question.",
"cite_spans": [],
"ref_spans": [
{
"start": 995,
"end": 1003,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Procedure for Collecting Tip QAs",
"sec_num": "4.3"
},
{
"text": "5 Applying BERT to Tip Machine Comprehension",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Procedure for Collecting Tip QAs",
"sec_num": "4.3"
},
{
"text": "In this paper, we developed two types of datasets for evaluation: one for SQuAD1.1 type answerable questions only and another for SQuAD2.0 type answerable and unanswerable questions. This paper presents evaluation results with the SQuAD2.0 type dataset. For the SQuAD2.0 type dataset, Table 4 presents the statistics of training and test datasets for evaluation in this paper. Table 4 (a) presents those of the training and test datasets for Japanese factoid QAs 4 . Those Japanese factoid QAs, which are of SQuAD2.0 type, are manually collected from Japanese quiz data by automatically identifying context texts from Japanese version of Wikipedia and then manually judging whether each identified context includes the answer to the question. Table 4 (b) and Table 4 (c) present the statistics of training and test datasets for Japanese tip QAs about \"job hunting\" and \"marriage\". Similarly, Table 4 (d) presents those for test datasets for Japanese tip QAs about \"apartment,\" \"hay fever,\" \"dentist,\" and \"food poisoning\" 5 . 4 http://www.cl.ecei.tohoku.ac.jp/rcqa/ (in Japanese)",
"cite_spans": [
{
"start": 1027,
"end": 1028,
"text": "4",
"ref_id": null
}
],
"ref_spans": [
{
"start": 285,
"end": 292,
"text": "Table 4",
"ref_id": "TABREF7"
},
{
"start": 377,
"end": 384,
"text": "Table 4",
"ref_id": "TABREF7"
},
{
"start": 743,
"end": 767,
"text": "Table 4 (b) and Table 4",
"ref_id": "TABREF7"
},
{
"start": 893,
"end": 900,
"text": "Table 4",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Dataset for Evaluation",
"sec_num": "5.1"
},
{
"text": "5 Four annotators participated in the procedure of collecting tip QAs, where we measured inter-annotator agreement rate according to AC1 (Gwet, 2008) , but not to kappa (Cohen, 1960) , mainly because two or more annotators tend to have high overall agreement rate, causing imbalanced class label distribution and instability of kappa. AC1 inter-annotator agreement is measured through the two sub-procedures: i.e., i) manually judging whether the questions selected by two out of three annotators are semantically equivalent when exactly the same context paragraph is given to the three anno- ",
"cite_spans": [
{
"start": 137,
"end": 149,
"text": "(Gwet, 2008)",
"ref_id": "BIBREF5"
},
{
"start": 169,
"end": 182,
"text": "(Cohen, 1960)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset for Evaluation",
"sec_num": "5.1"
},
{
"text": "As the version of BERT (Devlin et al., 2019) implementation which can handle a text in Japanese, the TensorFlow version 6 and the Multilingual Cased model 7 were used as the pre-trained model. tators, and ii) manually judging whether the answers selected by two out of three annotators are semantically equivalent when exactly the same pair of a question and a context is given to the three annotators (their detailed procedures are omitted for space restriction). Average of AC1 is 0.61 for the sub-procedure i) and 0.92 for the sub-procedure ii), which are sufficiently high. 6 https://github.com/google-research/bert 7 Trained with 104 languages, available from https://github.com/google-research/bert/ blob/master/multilingual.md.",
"cite_spans": [
{
"start": 23,
"end": 44,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF4"
},
{
"start": 578,
"end": 579,
"text": "6",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "BERT Implementation",
"sec_num": "5.2"
},
{
"text": "Before applying BERT modules, MeCab 8 was applied with IPAdic dictionary, and the Japanese text was segmented into a morpheme sequence. Then, within the BERT fine-tuning module, the Word-Piece module with 110k shared WordPiece vocabulary was applied, and the Japanese text was further segmented into a subword unit sequence. Finally, the BERT fine-tuning module for machine comprehension 9 was applied as well as the finetuned model. The BERT pre-trained model was fine-tuned with the following three types of train- Table 4 (a).",
"cite_spans": [],
"ref_spans": [
{
"start": 517,
"end": 524,
"text": "Table 4",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "BERT Implementation",
"sec_num": "5.2"
},
{
"text": "(ii) The training datasets of the tip QA about \"job hunting\" in Table 4 (b) and \"marriage\" in Table 4 (c).",
"cite_spans": [],
"ref_spans": [
{
"start": 64,
"end": 71,
"text": "Table 4",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "BERT Implementation",
"sec_num": "5.2"
},
{
"text": "(iii) Mix of (i) and (ii).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "BERT Implementation",
"sec_num": "5.2"
},
{
"text": "Here, note that we train a single model with each of these three training datasets (i)\u223c(iii), i.e., a single factoid machine comprehension model with (i), a single tip machine comprehension model with (ii), and a single machine comprehension model for the mixture of factoid and tip with (iii). It is especially important to note that we train a single tip machine comprehension model with the tip QA datasets about \"job hunting\" and \"marriage\", then evaluate it against the tip QA test datasets about all the query focuses, i.e., 'job hunting,\" \"marriage,\" \"apartment,\" \"hay fever,\" \"dentist,\" and \"food poisoning.\"",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "BERT Implementation",
"sec_num": "5.2"
},
{
"text": "In the evaluation, it is manually judged whether the answer predicted by the fine-tuned model and the reference answer partially match or not. We prefer manual evaluation rather than automatic evaluation, mainly because we prefer the quality of evaluation than avoiding the cost of evaluation. Figure 2 presents the evaluation results for the tip QA test datasets about \"job hunting,\" \"marriage,\" \"apartment,\" and a mix of \"hay fever,\" \"dentist,\" and \"food poisoning,\" as well as for the factoid QA test dataset. As clearly seen from these results, for all the tips test datasets, (ii) training only with tip QAs and (iii) training with a mix of tip QA and factoid QA training datasets outperforms and (i) training only with factoid QAs. For the factoid QA test datasets, on the other hand, (i) training only with factoid QAs and (iii) training with a mix of tip QA and factoid QA training datasets outperforms (ii) training only with tip QAs. This result supports the conclusion that the tip machine comprehension task is essentially different from the factoid machine comprehension task. But, still, for tips on \"job hunting,\" \"marriage,\" and the mix of \"hay fever,\" \"dentist,\" and \"food poisoning,\" training with a mix of tip QA and factoid QA training datasets slightly outperforms training only with tip QAs. This result indicates that the tip machine comprehension task still to some extent benefits from a large-scale factoid QA training dataset when only small-scale tip QAs are available.",
"cite_spans": [],
"ref_spans": [
{
"start": 294,
"end": 302,
"text": "Figure 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Evaluation Result",
"sec_num": "5.3"
},
{
"text": "Another interesting finding is that, in tip machine comprehension, the single model fine-tuned with tip QA training datasets on \"job hunting\" and \"marriage\" performed well in tip machine comprehension of other query focuses, such as \"apartment,\" \"hay fever,\" \"dentist,\" and \"food poisoning.\" Thus, in tip comprehension, it is sufficient to collect tip QA only for one or two query fo- cuses, such as \"job hunting\" and \"marriage,\" and then to fine-tune the tip machine comprehension model which is applicable to tip machine comprehension on any query focus. Figure 3 also presents a comparison of learning curves for the following three cases:",
"cite_spans": [],
"ref_spans": [
{
"start": 557,
"end": 565,
"text": "Figure 3",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Evaluation Result",
"sec_num": "5.3"
},
{
"text": "(a) Training with 5%, 10%, . . ., 95%, and 100%",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Result",
"sec_num": "5.3"
},
{
"text": "of factoid QA training dataset of (i) in the previous section and testing with the factoid QA test dataset from Table 4 (a) (plotted in blue).",
"cite_spans": [],
"ref_spans": [
{
"start": 112,
"end": 119,
"text": "Table 4",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Evaluation Result",
"sec_num": "5.3"
},
{
"text": "(b) Training with 10%, 20%, . . ., 90%, and 100% of the tip QA training datasets on \"job hunting\" and \"marriage\" of (ii) in the previous section and testing with the tip QA test dataset on \"job hunting\" of Table 4 (b) (plotted in orange).",
"cite_spans": [],
"ref_spans": [
{
"start": 206,
"end": 213,
"text": "Table 4",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Evaluation Result",
"sec_num": "5.3"
},
{
"text": "(c) Training with a mix of (a) and (b), where the factoid QA training dataset of (a) is always with its 100% size, whereas the tip QA training dataset on \"job hunting\" of (b) has a size of 10%, 20%, . . ., 90%, and 100% sizes and testing with the tip QA test dataset on \"job hunting\" of Table 4 (b) (plotted in green).",
"cite_spans": [],
"ref_spans": [
{
"start": 287,
"end": 294,
"text": "Table 4",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Evaluation Result",
"sec_num": "5.3"
},
{
"text": "As can be seen from these results, the learning curve (b) of tip QAs and that (c) of the mix of factoid and tip QAs perform comparatively well and outperform that of factoid QAs (a) in the range of around a few thousand training data size. This result indicates that, at least for tip machine com-prehension of \"job hunting\", benefit from a largescale factoid QA training dataset is very little. Far more important finding in this result is that the tip machine comprehension performance is almost comparative with that of the factoid machine comprehension even when trained with as little as around 4% (\u2252 2,364/56,551) of the training data size of the factoid machine comprehension. Thus, it can be concluded that the tip machine comprehension task requires much less training data compared with the factoid machine comprehension task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Result",
"sec_num": "5.3"
},
{
"text": "In the field of developing QA datasets or machine comprehension datasets which may include nonfactoid QAs, quite a limited number of datasets are publicly available in any language. In English, MS MARCO (Nguyen et al., 2016) has been developed using Bing's search logs and passages of retrieved web pages, which may include non-factoid QAs. Question types in MS MARCO are classified into numeric, entity, location, person, and description (phrase) . In Chinese, DuReader (He et al., 2018) has been developed using Baidu Search and Baidu Zhidao, which is a Chinese community-based QA site. DuReader's question types are classified into entity, description, and yes-no questions on fact or opinion. DuReader's QAs definitely include non-factoid ones. Another type of non-factoid QA dataset is NarrativeQA (Ko\u010disk\u00fd et al., 2018) dataset (in En-glish), which contains questions created by editors based on summaries of movie scripts and books. In the case of the Japanese language QA dataset, there is quite a limited number of publicly available factoid QA datasets, and one of them was introduced in Section 5.1. There is no publicly available Japanese non-factoid QA dataset.",
"cite_spans": [
{
"start": 203,
"end": 224,
"text": "(Nguyen et al., 2016)",
"ref_id": "BIBREF8"
},
{
"start": 405,
"end": 447,
"text": "location, person, and description (phrase)",
"ref_id": null
},
{
"start": 471,
"end": 488,
"text": "(He et al., 2018)",
"ref_id": "BIBREF6"
},
{
"start": 803,
"end": 825,
"text": "(Ko\u010disk\u00fd et al., 2018)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "6"
},
{
"text": "This paper explored a way to develop a dataset for training Japanese tip QA models, and it applied BERT (Devlin et al., 2019) to a Japanese tip QA dataset. Evaluation results revealed that the tip machine comprehension performance was almost comparative with that of the factoid machine comprehension even with the training data size reduced to around 4% of the factoid machine comprehension. Thus, the tip machine comprehension task requires much less training data compared with the factoid machine comprehension task. Future direction of this work includes applying the proposed framework of tip machine comprehension to other languages, such as English and Chinese. In both languages, factoid QA datasets are publicly available (e.g., SQuAD (Pranav et al., 2016 (Pranav et al., , 2018 for English and CMRC2018 (Cui et al., 2018) for Chinese), and it is quite attainable to train a factoid machine comprehension model by fine-tuning the BERT pre-trained model and then to directly apply the factoid machine comprehension model to the tip machine comprehension task. Actually, as a preliminary work, a Chinese factoid machine comprehension model is trained by finetuning the pre-trained Multilingual Cased model with CMRC2018 Chinese factoid QA dataset 1011 , and then applying it to 30 Chinese tip questions on \"marriage\" with context texts. As a result, around 50% accuracy for manual evaluation is achieved by exact and partial match, which is almost comparative to the performance achieved in the Japanese tip machine comprehension task reported in this paper. Thus, it is expected that extending the proposed framework of tip machine comprehension to other languages, such as English and Chinese, is quite straightforward.",
"cite_spans": [
{
"start": 104,
"end": 125,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF4"
},
{
"start": 745,
"end": 765,
"text": "(Pranav et al., 2016",
"ref_id": "BIBREF10"
},
{
"start": 766,
"end": 788,
"text": "(Pranav et al., , 2018",
"ref_id": "BIBREF11"
},
{
"start": 814,
"end": 832,
"text": "(Cui et al., 2018)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "Another future direction is to extending the proposed framework of tip machine comprehension 10 https://hfl-rc.github.io/cmrc2018/ english/ 11 https://github.com/ymcui/cmrc2018",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "to open domain tip machine comprehension. This extension is similar to the extension of existing factoid machine comprehension with Wikipedia texts' paragraphs as contexts to open domain machine comprehension with the whole Wikipedia articles (Chen et al., 2017) .",
"cite_spans": [
{
"start": 243,
"end": 262,
"text": "(Chen et al., 2017)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "In the extended open domain tip machine comprehension framework, the document retriever module is realized based on the tip websites search and column web page collection architectures proposed in this paper. The document reader module can be easily realized by simply applying the tip machine comprehension model of this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "Another definitely important future direction should be to invent a technique of how to automate the procedure of collecting column web pages and generating the tuple of a context C, a question Q, and answer A. This task can be regarded as that of training a tip machine comprehension model from a noisy training dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "https://rajpurkar.github.io/ SQuAD-explorer/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://developers.google.com/ custom-search/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Ohkawa et al. (2018) examined quantitative characteristics of tip websites. Furthermore, it is reported inOhkawa et al. (2018) that domain names of candidate tip websites can be collected from those that have corresponding web pages in sets D(zn) of web pages for multiple topics zn (n = 1, . . . , K). It is observed, however, that typical tip websites actually have their web pages under far more than two topics and typically, more than or equal to 10 topics. Thus, in this paper, it was decided to select domain names which have their web pages under more than or equal to 10 topics as candidates for tip websites.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Latent Dirichlet allocation",
"authors": [
{
"first": "D",
"middle": [
"M"
],
"last": "Blei",
"suffix": ""
},
{
"first": "A",
"middle": [
"Y"
],
"last": "Ng",
"suffix": ""
},
{
"first": "M",
"middle": [
"I"
],
"last": "Jordan",
"suffix": ""
}
],
"year": 2003,
"venue": "Journal of Machine Learning Research",
"volume": "3",
"issue": "",
"pages": "993--1022",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. M. Blei, A. Y. Ng, and M. I. Jordan. 2003. Latent Dirichlet allocation. Journal of Machine Learning Research, 3:993-1022.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Reading Wikipedia to answer open-domain questions",
"authors": [
{
"first": "D",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Fisch",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Weston",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bordes",
"suffix": ""
}
],
"year": 2017,
"venue": "Proc. 55th ACL",
"volume": "",
"issue": "",
"pages": "1870--1879",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Chen, A. Fisch, J. Weston, and A. Bordes. 2017. Reading Wikipedia to answer open-domain ques- tions. In Proc. 55th ACL, pages 1870-1879.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement",
"authors": [
{
"first": "J",
"middle": [],
"last": "Cohen",
"suffix": ""
}
],
"year": 1960,
"venue": "",
"volume": "20",
"issue": "",
"pages": "37--46",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Cohen. 1960. A Coefficient of Agreement for Nom- inal Scales. Educational and Psychological Mea- surement, 20(1):37-46.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A span-extraction dataset for Chinese machine reading comprehension",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Cui",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Xiao",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Che",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Hu",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Cui, T. Liu, L. Xiao, Z. Chen, W. Ma, W. Che, S. Wang, and G. Hu. 2018. A span-extraction dataset for Chinese machine reading comprehen- sion. CoRR, abs/1810.07366.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "J",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "M.-W",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2019,
"venue": "Proc. NAACL-HLT",
"volume": "",
"issue": "",
"pages": "4171--4186",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL-HLT, pages 4171-4186.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Computing inter-rater reliability and its variance in the presence of high agreement",
"authors": [
{
"first": "K",
"middle": [],
"last": "Gwet",
"suffix": ""
}
],
"year": 2008,
"venue": "The British Journal of Mathematical and Statistical Psychology",
"volume": "61",
"issue": "",
"pages": "29--48",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Gwet. 2008. Computing inter-rater reliability and its variance in the presence of high agreement. The British Journal of Mathematical and Statistical Psy- chology, 61:29-48.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "DuReader: a Chinese machine reading comprehension dataset from real-world applications",
"authors": [
{
"first": "W",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Lyu",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Xiao",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Q",
"middle": [],
"last": "She",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2018,
"venue": "Proc. MRQA",
"volume": "",
"issue": "",
"pages": "37--46",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. He, K. Liu, J. Liu, Y. Lyu, S. Zhao, X. Xiao, Y. Liu, Y. Wang, H. Wu, Q. She, X. Liu, T. Wu, and H. Wang. 2018. DuReader: a Chinese machine reading comprehension dataset from real-world ap- plications. In Proc. MRQA, pages 37-46.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The NarrativeQA reading comprehension challenge",
"authors": [
{
"first": "T",
"middle": [],
"last": "Ko\u010disk\u00fd",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Schwarz",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Blunsom",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Dyer",
"suffix": ""
},
{
"first": "K",
"middle": [
"M"
],
"last": "Hermann",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Melis",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Grefenstette",
"suffix": ""
}
],
"year": 2018,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "6",
"issue": "",
"pages": "317--328",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Ko\u010disk\u00fd, J. Schwarz, P. Blunsom, C. Dyer, K. M. Hermann, G. Melis, and E. Grefenstette. 2018. The NarrativeQA reading comprehension challenge. Transactions of the Association for Computational Linguistics, 6:317-328.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "MS MARCO: A human generated machine reading comprehension dataset",
"authors": [
{
"first": "T",
"middle": [],
"last": "Nguyen",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Rosenberg",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Song",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Tiwary",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Majumder",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Deng",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Nguyen, M. Rosenberg, X. Song, J. Gao, S. Tiwary, R. Majumder, and L. Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. CoRR, abs/1611.09268.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Identifying tips Web sites of a specific query based on search engine suggests and the topic distribution",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Ohkawa",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Kawabata",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Niu",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Utsuro",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Kawada",
"suffix": ""
}
],
"year": 2018,
"venue": "Proc. 3rd ABCSS",
"volume": "",
"issue": "",
"pages": "4347--4353",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Ohkawa, S. Kawabata, C. Zhao, W. Niu, Y. Lin, T. Utsuro, and Y. Kawada. 2018. Identifying tips Web sites of a specific query based on search en- gine suggests and the topic distribution. In Proc. 3rd ABCSS, pages 4347-4353.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "SQuAD: 100,000+ questions for machine comprehension of text",
"authors": [
{
"first": "R",
"middle": [],
"last": "Pranav",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Jian",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Konstantin",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Percy",
"suffix": ""
}
],
"year": 2016,
"venue": "Proc. EMNLP",
"volume": "",
"issue": "",
"pages": "2383--2392",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Pranav, Z. Jian, L. Konstantin, and L. Percy. 2016. SQuAD: 100,000+ questions for machine compre- hension of text. In Proc. EMNLP, pages 2383-2392.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Know what you don't know: Unanswerable questions for SQuAD",
"authors": [
{
"first": "R",
"middle": [],
"last": "Pranav",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Robin",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Percy",
"suffix": ""
}
],
"year": 2018,
"venue": "Proc. 56th ACL",
"volume": "",
"issue": "",
"pages": "784--789",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Pranav, J. Robin, and L. Percy. 2018. Know what you don't know: Unanswerable questions for SQuAD. In Proc. 56th ACL, pages 784-789.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Wik-iQA: A challenge dataset for open-domain question answering",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Yi",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Wen-Tau",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Christopher",
"suffix": ""
}
],
"year": 2015,
"venue": "Proc. EMNLP",
"volume": "",
"issue": "",
"pages": "2013--2018",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Yi, Y. Wen-tau, and M. Christopher. 2015. Wik- iQA: A challenge dataset for open-domain question answering. In Proc. EMNLP, pages 2013-2018.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "SG-Net: Syntax-guided machine reading comprehension",
"authors": [
{
"first": "Z",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Duan",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Zhao",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Z. Zhang, Y. Wu, J. Zhou, S. Duan, and H. Zhao. 2019. SG-Net: Syntax-guided machine reading compre- hension. CoRR, abs/1908.05147.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "Evaluation results (exact match + partial match) ing datasets: (i) The training dataset of factoid QAs in",
"type_str": "figure",
"num": null
},
"FIGREF1": {
"uris": null,
"text": "Comparing learning curves of factoid QAs, tip QAs, and training with a mix of factoid and tip QAs (exact match + partial match)",
"type_str": "figure",
"num": null
},
"TABREF1": {
"text": "Statistics of the Classification of Tip Question Types",
"num": null,
"type_str": "table",
"html": null,
"content": "<table/>"
},
"TABREF3": {
"text": "",
"num": null,
"type_str": "table",
"html": null,
"content": "<table/>"
},
"TABREF5": {
"text": "",
"num": null,
"type_str": "table",
"html": null,
"content": "<table/>"
},
"TABREF7": {
"text": "Statistics of training and test datasets",
"num": null,
"type_str": "table",
"html": null,
"content": "<table/>"
}
}
}
} |