File size: 84,186 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 |
{
"paper_id": "R11-1024",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T15:03:46.668927Z"
},
"title": "A Named Entity Recognition Method using Rules Acquired from Unlabeled Data",
"authors": [
{
"first": "Tomoya",
"middle": [],
"last": "Iwakura",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Fujitsu Laboratories Ltd",
"location": {
"addrLine": "1-1, Kamikodanaka 4-chome, Nakahara-ku",
"postCode": "211-8588",
"settlement": "Kawasaki",
"country": "Japan"
}
},
"email": "iwakura.tomoya@jp.fujitsu.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We propose a Named Entity (NE) recognition method using rules acquired from unlabeled data. Rules are acquired from automatically labeled data with an NE recognizer. These rules are used to identify NEs, the beginning of NEs, or the end of NEs. The application results of rules are used as features for machine learning based NE recognizers. In addition, we use word information acquired from unlabeled data as in a previous work. The word information includes the candidate NE classes of each word, the candidate NE classes of co-occurring words of each word, and so on. We evaluate our method with IREX data set for Japanese NE recognition and unlabeled data consisting of more than one billion words. The experimental results show that our method using rules and word information achieves the best accuracy on the GENERAL and ARREST tasks of IREX.",
"pdf_parse": {
"paper_id": "R11-1024",
"_pdf_hash": "",
"abstract": [
{
"text": "We propose a Named Entity (NE) recognition method using rules acquired from unlabeled data. Rules are acquired from automatically labeled data with an NE recognizer. These rules are used to identify NEs, the beginning of NEs, or the end of NEs. The application results of rules are used as features for machine learning based NE recognizers. In addition, we use word information acquired from unlabeled data as in a previous work. The word information includes the candidate NE classes of each word, the candidate NE classes of co-occurring words of each word, and so on. We evaluate our method with IREX data set for Japanese NE recognition and unlabeled data consisting of more than one billion words. The experimental results show that our method using rules and word information achieves the best accuracy on the GENERAL and ARREST tasks of IREX.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Named Entity (NE) recognition aims to recognize proper nouns and numerical expressions in text, such as names of people, locations, organizations, dates, times, and so on. NE recognition is one of the basic technologies used in text processing such as Information Extraction and Question Answering.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To implement NE recognizers, semisupervised-based methods have recently been widely applied. These methods use several different types of information obtained from unlabeled data, such as word clusters (Freitag, 2004; Miller et al., 2004) , the clusters of multi-word nouns (Kazama and Torisawa, 2008) , phrase clusters (Lin and Wu, 2009) , hyponymy relations extracted from WikiPedia (Kazama and Torisawa, 2008) , NE-related word information (Iwakura, 2010) , and the outputs of classifiers or parsers created from unlabeled data (Ando and Zhang, 2005) . These previous works have shown that features acquired from large sets of unlabeled data can contribute to improved accuracy. From the results of these previous works, we see that several types of features augmented with unlabeled data contribute to improved accuracy.",
"cite_spans": [
{
"start": 202,
"end": 217,
"text": "(Freitag, 2004;",
"ref_id": "BIBREF2"
},
{
"start": 218,
"end": 238,
"text": "Miller et al., 2004)",
"ref_id": "BIBREF12"
},
{
"start": 274,
"end": 301,
"text": "(Kazama and Torisawa, 2008)",
"ref_id": "BIBREF10"
},
{
"start": 320,
"end": 338,
"text": "(Lin and Wu, 2009)",
"ref_id": "BIBREF11"
},
{
"start": 375,
"end": 412,
"text": "WikiPedia (Kazama and Torisawa, 2008)",
"ref_id": null
},
{
"start": 443,
"end": 458,
"text": "(Iwakura, 2010)",
"ref_id": "BIBREF9"
},
{
"start": 531,
"end": 553,
"text": "(Ando and Zhang, 2005)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Therefore, if we can incorporate new features augmented with unlabeled data, we expect more improved accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We propose a Named Entity recognition method using rules acquired from unlabeled data. Our method uses rules identifying not only whole NEs, but also the beginning of NEs or the end of NEs. Rules are acquired from automatically labeled data with an NE recognizer. The application results of rules are used as features for machine-learning based NE recognitions. Compared with previous works using rules identifying NEs acquired from manually labeled data (Isozaki, 2001) , or lists of NEs acquired from unlabeled data (Talukdar et al., 2006) , our method uses new features such as identification results of the beginning of NEs and the end of NEs. In addition, we use word information (Iwakura, 2010) . The word information includes the candidate NE classes of each word, the candidate NE classes of co-occurring words of each word, and so on. The word information is also acquired from automatically labeled data with an NE recognizer.",
"cite_spans": [
{
"start": 455,
"end": 470,
"text": "(Isozaki, 2001)",
"ref_id": "BIBREF6"
},
{
"start": 518,
"end": 541,
"text": "(Talukdar et al., 2006)",
"ref_id": "BIBREF17"
},
{
"start": 685,
"end": 700,
"text": "(Iwakura, 2010)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We report experimental results with the IREX Japanese NE recognition data set (IREX, 1999) . The experimental results show that our method using rules and word information achieves the best accuracy on the GENERAL and ARREST tasks. The experimental results also show that our method contributes to fast improvement of accuracy compared with only using manually labeled ",
"cite_spans": [
{
"start": 78,
"end": 90,
"text": "(IREX, 1999)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This section describes our NE recognition method that combines both word-based and characterbased NE recognitions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Japanese Named Entity Recognition",
"sec_num": "2"
},
{
"text": "Each NE consists of one or more words. To recognize NEs, we have to identify word chunks with their NE classes. We use Start/End (SE) representation (Uchimoto et al., 2000) because an SE representation-based NE recognizer shows the best performance among previous works (Sasano and Kurohashi, 2008) . SE representation uses five tags which are S, B, I, E and O, for representing chunks. S means that the current word is a chunk consisting of only one word. B means the start of a chunk consisting of more than one word. E means the end of a chunk consisting of more than one word. I means the inside of a chunk consisting of more than two words. O means the outside of any chunk. We use the IREX Japanese NE recognition task for our evaluation. The task is to recognize the eight NE classes. The SE based NE label set for IREX task has (8 \u00d7 4) + 1 = 33 labels such as B-PERSON, S-PERSON, and so on.",
"cite_spans": [
{
"start": 149,
"end": 172,
"text": "(Uchimoto et al., 2000)",
"ref_id": "BIBREF18"
},
{
"start": 270,
"end": 298,
"text": "(Sasano and Kurohashi, 2008)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Chunk Representation",
"sec_num": "2.1"
},
{
"text": "We classify each word into one of the NE labels defined by the SE representation for recognizing NEs. Japanese has no word boundary marker. To segment words from Japanese texts, we use MeCab 0.98 with ipadic-2.7.0. 1 Our NE recognizer uses features extracted from the current word, the preceding two words and the two succeeding words (5-word window). The basic features are the word surfaces, the last characters, the base-forms, the readings, the POS tags, and the character types of words within 5-word window size. The base-forms, the readings, and the POS tags are given by MeCab. Base-forms are representative expressions for conjugational words. If the base-form of each word is not equivalent to the word surface, we use the base-form as a feature. If a word consists of only one character, the character type is expressed by using the corresponding character types listed in Table 1 . If a word consists of more than one character, the character type is expressed by a combination of the basic character types listed in Table 1 , such as Kanji-Hiragana. MeCab uses the set of POS tags having at most four levels of subcategories. We use all the levels of POS tags as POS tag features.",
"cite_spans": [],
"ref_spans": [
{
"start": 884,
"end": 891,
"text": "Table 1",
"ref_id": "TABREF0"
},
{
"start": 1029,
"end": 1036,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Word-based NE Recognition",
"sec_num": "2.2"
},
{
"text": "We use outputs of rules to a current word and word information within 5-word window size as features. The rules and the word information are acquired from automatically labeled data with an NE recognizer. We describe rules in section 3. We use the following NE-related labels of words from unlabeled data as word information as in (Iwakura, 2010) .",
"cite_spans": [
{
"start": 331,
"end": 346,
"text": "(Iwakura, 2010)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Word-based NE Recognition",
"sec_num": "2.2"
},
{
"text": "Candidate NE labels: We use NE labels assigned to each word more than or equal to 50 times as candidate NE labels of words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word-based NE Recognition",
"sec_num": "2.2"
},
{
"text": "Candidate co-occurring NE labels: We use NE labels assigned to co-occurring words of each word more than or equal to 50 times as candidate co-occurring NE labels of the word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word-based NE Recognition",
"sec_num": "2.2"
},
{
"text": "Frequency information of candidate NE labels and candidate co-occurring NE labels: These are the frequencies of the NE candidate labels of each word on the automatically labeled data. We categorize the frequencies of these NErelated labels by the frequency of each word n; 50 \u2264 n \u2264 100, 100 < n \u2264 500, 500 < n \u2264 1000, 1000 < n \u2264 5000, 5000 < n \u2264 10000, 10000 < n \u2264 50000, 50000 < n \u2264 100000, and 100000 < n.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word-based NE Recognition",
"sec_num": "2.2"
},
{
"text": "Ranking of candidate NE labels: This information is the ranking of candidate NE class labels for each word. Each ranking is decided according to the label frequencies.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word-based NE Recognition",
"sec_num": "2.2"
},
{
"text": "For example, we obtain the following statistics from automatically labeled data with an NE recognizer for Tanaka: S-PERSON was assigned to Tanaka 10,000 times, B-PERSON was assigned to Tanaka 1,000 times, and I-PERSON was assigned to words appearing next to Tanaka 1,000 times. The following NE-related labels are acquired for Tanaka: Candidate NE labels are S-PERSON and B-ORGANIZATION. Frequency information of candidate NE labels are 5000 < n \u2264 10000 for S-PERSON, and 500 < n \u2264 1000 for B-ORGANIZATION. The ranking of candidate NE labels are the first for S-PERSON, and second for B-ORGANIZATION. Candidate co-occurring NE labels at the next word position is I-PERSON. Frequency information of candidate co-occurring NE labels at the next word position is 500 < n \u2264 1000 for I-PERSON.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word-based NE Recognition",
"sec_num": "2.2"
},
{
"text": "Japanese NEs sometimes include partial words that form the beginning, the end of NE chunks or whole NEs. 2 To recognize Japanese NEs including partial words, we use a character-unitchunking-based NE recognition algorithm (Asahara and Matsumoto, 2003; Nakano and Hirai, 2004) following word-based NE recognition as in (Iwakura, 2010) .",
"cite_spans": [
{
"start": 221,
"end": 250,
"text": "(Asahara and Matsumoto, 2003;",
"ref_id": "BIBREF1"
},
{
"start": 251,
"end": 274,
"text": "Nakano and Hirai, 2004)",
"ref_id": "BIBREF13"
},
{
"start": 317,
"end": 332,
"text": "(Iwakura, 2010)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Character-based NE Recognition",
"sec_num": "2.3"
},
{
"text": "Our character-based NE recognizer uses features extracted from the current character, the preceding two characters and the two succeeding characters (5-character window). The features extracted from each character within the window size are the followings; the character itself, the character type of the character listed in Table 1 , and the NE labels of two preceding recognition results in the direction from the end to the beginning.",
"cite_spans": [],
"ref_spans": [
{
"start": 325,
"end": 332,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Character-based NE Recognition",
"sec_num": "2.3"
},
{
"text": "In addition, we use words including characters within the window size. The features of the words are the character types, the POS tags, and the NE labels assigned by a word-based NE recognizer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Character-based NE Recognition",
"sec_num": "2.3"
},
{
"text": "As for words including characters, we extract features as follows. Let W (c i ) be the word including the i-th character c i and P (c i ) be the identifier that indicates the position where c i appears in W (c i ). We combine W (c i ) and P (c i ) to create a feature. P (c i ) is one of the followings: B for a character that is the beginning of a word, I for a character that is in the inside of a word, E for a character that is the end of a word, and S for a character that is a word. 3 We use the POS tags of words including characters within 5-character window. Let P OS(W (c i )) be the POS tag of the word W (c i ) including the ith character c i . We express these features with the position identifier P (c i ) like P (c i )-P OS(W (c i )). In addition, we use the character types of words including characters. To utilize outputs of a wordbased NE recognizer, we use NE labels of words assigned by a word-unit NE recognizer. Each character is classified into one of the 33 NE labels provided by the SE representation.",
"cite_spans": [
{
"start": 489,
"end": 490,
"text": "3",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Character-based NE Recognition",
"sec_num": "2.3"
},
{
"text": "We use a boosting-based learner that learns rules consisting of a feature, or rules represented by combinations of features consisting of more than one feature (Iwakura and Okamoto, 2008) . The boosting algorithm achieves fast training speed by training a weak-learner that learns several rules from a small portion of candidate rules. Candidate rules are generated from a subset of features called bucket. The parameters for the boosting algorithm are as follows. We used the number of rules to be learned as R=100,000, the bucketing size for splitting features into subsets as |B|=1,000, the number of rules learned at each boosting iteration as \u03bd =10, the number of candidate rules used to generate new combinations of features at each rule size as \u03c9=10, and the maximum number of features in rules as \u03b6=2.",
"cite_spans": [
{
"start": 160,
"end": 187,
"text": "(Iwakura and Okamoto, 2008)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning Algorithm",
"sec_num": "2.4"
},
{
"text": "The boosting algorithm operates on binary classification problems. To extend the boosting to multi-class, we used the one-vs-the-rest method. To identify proper tag sequences, we use the Viterbi search. To apply the Viterbi search, we convert the confidence value of each classifier into the range of 0 to 1 with sigmoid function defined as s(X) = 1/(1 + exp(\u2212\u03b2X)), where X is the output of a classifier to an input. We used \u03b2=1 in this experiment. Then we select a tag sequence which maximizes the sum of those log values.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning Algorithm",
"sec_num": "2.4"
},
{
"text": "To obtain a fast processing and training speed, we apply a technique to control the generation of combinations of features (Iwakura, 2009) . This is because fast processing speed is required to obtain word information and rules from large unlabeled data. Using this technique, instead of manually specifying combinations of features to be used, features that are not used in combinations of features are specified as atomic features. The boosting algorithm learns rules consisting of more than one feature from the combinations of features generated from non-atomic features, and rules consisting of only a feature from the atomic and the non-atomic features. We can obtain faster training speed and processing speed because we can reduce the number of combinations of features to be examined by specifying part of features as atomic. We specify features based on word information and rules acquired from unlabeled data as the atomic features.",
"cite_spans": [
{
"start": 123,
"end": 138,
"text": "(Iwakura, 2009)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning Algorithm",
"sec_num": "2.4"
},
{
"text": "This section describes rules and a method to acquire rules.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rules Acquired from Unlabeled Data",
"sec_num": "3"
},
{
"text": "Previous works such as Isozaki (Isozaki, 2001) , Talukdar et al., (Talukdar et al., 2006) , use rules or lists of NEs for only identifying NEs. In addition to rules identifying NEs, we propose to use rules for identifying the beginning of NEs or the end of NEs to capture context information. To acquire rules, an automatically labeled data with an NE recognizer is used. The following types of rules are acquired.",
"cite_spans": [
{
"start": 31,
"end": 46,
"text": "(Isozaki, 2001)",
"ref_id": "BIBREF6"
},
{
"start": 49,
"end": 89,
"text": "Talukdar et al., (Talukdar et al., 2006)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Types",
"sec_num": "3.1"
},
{
"text": "Word N-gram rules for identifying NEs (NE-W-rules, for short): These are word N-grams corresponding to candidate NEs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Types",
"sec_num": "3.1"
},
{
"text": "Word trigram rules for identifying the beginning of NEs (NEB-W-rules): Each rule for identifying the beginning of NEs is represented as a word trigram consisting of the two words preceding the beginning of an NE and the beginning of the NE.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Types",
"sec_num": "3.1"
},
{
"text": "Word trigram rules for identifying the end of NEs (NEE-W-rules): Each rule for identifying the end of NEs is represented as a word trigram consisting of the two words succeeding the end of an NE and the end of the NE.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Types",
"sec_num": "3.1"
},
{
"text": "In addition to word N-gram rules, we acquire Word/POS N-gram rules for achieving higher rule coverage. Word/POS N-gram rules are acquired from N-gram rules by replacing some words in N-gram rules with POS tags. We call NE-W-rules, NEB-W-rules and NEE-W-rules converted to Word/POS N-gram rules NE-WP-rules, NEB-WP-rules and NEE-WP-rules, respectively. Word/POS N-gram rules also identify NEs the beginning of NEs and the end of NEs To acquire Word/POS rules, we replace words having one of the following POS tags with their POS tags as rule constituents: proper noun words, unknown words, and number words. This is because words having these POS tags are usually low frequency words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Types",
"sec_num": "3.1"
},
{
"text": "This section describes the method to acquire the rules used in this paper. The rule acquisition consists of three main steps: First, we create automatically labeled data. Second, seed rules are acquired. Finally the outputs of rules are decided.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "The first step prepares an automatically labeled data with an NE recognizer. The NE recognizer recognizes NEs from unlabeled data and generates the automatically labeled data by annotating characters recognized as NEs with the NE labels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "The second step acquires seed rules from the automatically labeled data. The following is an automatically labeled sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "[ Tanaka/$PN mission/$N party/$N ] ORG went/ $V to/$P [U.K / $PN] LOC ...\" ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "where $PN (Proper Noun), $N, $V, and $P following / are POS tags, and words between \"[ and ]\" were identified as NEs. ORG and LOC after \"]\" indicate NE types.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "The following seed rules are acquired from the above sentence by following the procedures described in previous sections: NE-W-rules: {Tanaka mission party \u2192 ORG}, NEB-W-rules: {went to U.K \u2192 LW=B-LOC}, NEE-W-rules: {party went to \u2192 FW=E-ORG}, NE-WP-rules: {$PN mission party \u2192 ORG}, NEB-WP-rules: {went to $PN \u2192 LW=B-LOC}, NEE-WP-rules: {$PN mission party \u2192 LW=B-ORG}, where FW, LW, B-LOC, and E-ORG indicate the first words of word sequences that a rule is applied to, the last words of word sequences that a rule is applied to, the beginning word of a LOCATION NE, and the end word of an ORGANIZATION NE, respectively. The left of each \u2192 is the rule condition to apply a rule, and the right of each \u2192 is the seed output of a rule. If the output of a rule is only an NE type, this means the rule identifies an NE. Rules with outputs including = indicate rules for identifying the beginning of NEs or the end of NEs. The left of = indicates the positions of words where the beginning of NEs or the end of NEs exist in the identified word sequences by rules. For example, LW=B-LOC means that LW is B-LOC.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "The final step decides the outputs of each rule. We count the outputs of the rule condition of each seed rule, and the final outputs of each rule are decided by using the frequency of each output. We use outputs assigned to each seed rule more than or equal to 50 times. 4 For example, if LW=B-LOC are obtained 10,000 times, and LW=B-ORG are obtained 1,000 times, as the outputs for {went to $PN}, the followings are acquired as final outputs: LW=B-LOC RANK1, LW=B-ORG RANK2, LW=B-LOC FREQ-5000 < n \u2264 10000, and LW=B-ORG FREQ-500 < n \u2264 1000.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "The LW=B-LOC RANK1 and the LW=B-ORG RANK2 are the ranking of the outputs of rules. LW=B-LOC is 1st ranked output, and LW=B-ORG is 2nd ranked output. Each ranking is decided by the frequency of each output of each rule condition. The most frequent output of each rule is ranked as first.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "LW=B-LOC FREQ-5000 < n \u2264 10000 and LW=B-ORG FREQ-500 < n \u2264 1000 are frequency information. To express the frequency of each rule output as binary features, we categorize the frequency of each rule output by the frequency of each rule output n; 50 \u2264 n \u2264 100, 100 < n \u2264 500, 500 < n \u2264 1000, 1000 < n \u2264 5000, 5000 < n \u2264 10000, 10000 < n \u2264 50000, 50000 < n \u2264 100000, and 100000 < n.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acquiring Rules",
"sec_num": "3.2"
},
{
"text": "We define the rule application by following the method for using phrase clusters in NER (Lin and Wu, 2009) . The application of rules is allowed to overlap with or be nested in one another. If a rule is applied at positions b to e, we add the features combined with the outputs of the rule and matching positions to each word; outputs with B-(beginning) to b-th word, outputs with E-(end) to b-th word, outputs with I-(inside) within b + 1-th to e \u2212 1-th words, outputs with P-(previous) to b \u2212 1-th word, and outputs with F-(following) to e + 1-th word.",
"cite_spans": [
{
"start": 88,
"end": 106,
"text": "(Lin and Wu, 2009)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Application",
"sec_num": "3.3"
},
{
"text": "If a rule having the condition {went to $PN} is applied to {.... Ken/$PN went/$V to/$P Japan/ $PN for/$P ...}, the followings are captured as rule application results: b-th word is went, the word between b-th and e-th is to, e-th word is Japan, b \u2212 1-th is Ken, and e + 1-th is for.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Application",
"sec_num": "3.3"
},
{
"text": "If the output of the rule is LW=B-LOC, the following features are added: B-LW=B-LOC for went, I-LW=B-LOC for to, E-LW=B-LOC for Japan, P-LW=B-LOC for Ken, and F-LW=B-LOC for for.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rule Application",
"sec_num": "3.3"
},
{
"text": "We also apply a method to acquire word information (Iwakura, 2010) to the rule acquisition repeatedly. This is because the previous work reported that better accuracy was obtained by repeating the acquisition of NE-related labels of words. The collection method is as follows.",
"cite_spans": [
{
"start": 51,
"end": 66,
"text": "(Iwakura, 2010)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Repeatedly Acquisition",
"sec_num": "3.4"
},
{
"text": "(1) Create an NE recognizer from training data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Repeatedly Acquisition",
"sec_num": "3.4"
},
{
"text": "(2) Acquire word information and rules from unlabeled data with the current NE recognizer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Repeatedly Acquisition",
"sec_num": "3.4"
},
{
"text": "(3) Create a new NE recognizer with the training data, word information and rules acquired at step (2). This NE recognizer is used for acquiring new word information and rules at the next iteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Repeatedly Acquisition",
"sec_num": "3.4"
},
{
"text": "(4) Go back to step (2) if the termination criterion is not satisfied. The process (2) to (4) is repeated 4 times in this experiment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Repeatedly Acquisition",
"sec_num": "3.4"
},
{
"text": "The following data prepared for IREX (IREX, 1999) were used in our experiment. We used the CRL data for the training. CRL data has 18,677 NEs on 1,174 stories from Mainichi Newspaper. In addition, to investigate the effectiveness of unlabeled data and labeled data, we prepared another labeled 7,000 news stories including 143,598 NEs from Mainichi Shinbun between 2007 and 2008 according to IREX definition. We have, in total, 8,174 news stories including 162,859 NEs that are about 8 times of CRL data. To create the additional labeled 7,000 news stories, about 509 hours were required. The average time for creating a labeled news story is 260 seconds, which means only 14 labeled news stories are created in an hour.",
"cite_spans": [
{
"start": 37,
"end": 49,
"text": "(IREX, 1999)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental settings",
"sec_num": "4.1"
},
{
"text": "For evaluation, we used formal-run data of IREX: GENERAL task including 1,581 NEs, and ARREST task including 389 NEs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental settings",
"sec_num": "4.1"
},
{
"text": "We compared performance of NE recognizers by using the F-measure (FM) defined as follows with Recall (RE) and Precision (PR); FM = 2 \u00d7 RE \u00d7 PR / ( RE + PR ), where, RE = NUM / (the number of correct NEs), PR = NUM / (the number of NEs extracted by an NE recognizer), and NUM is the number of NEs correctly identified by an NE recognizer. The news stories from the Mainichi Shinbun between 1991 and 2008 and Japanese WikiPedia entries of July 13, 2009, were used as unlabeled data for acquiring word information and rules. The total number of words segmented by MeCab from these unlabeled data was 1,161,758,003, more than one billion words. 5",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental settings",
"sec_num": "4.1"
},
{
"text": "We evaluated the effectiveness of the combination of word information and rules. Table 2 shows experimental results obtained with an NE recognizer without any word information and rules (NER-BASE, for short), an NE recognizer using word information (NER-W for short), an NE recognizer using rules (NER-R, for short), and an NE recognizer using word information and rules (NER-WR, for short), which is based on our proposed method",
"cite_spans": [],
"ref_spans": [
{
"start": 81,
"end": 88,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Evaluation of Our Proposed Method",
"sec_num": "4.2"
},
{
"text": "We used word information and rules obtained with the NER-BASE, which was created from CRL data without word information and rules. We see that we obtain better accuracy by using word information and rules acquired from unlabeled data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation of Our Proposed Method",
"sec_num": "4.2"
},
{
"text": "The NER-WR shows the best average Fmeasure (FM). The average FM of the NER-WR is 3.6 points higher than that of the NER-BASE. The average FM of the NER-WR is 0.44 points higher than that of NER-W, and 2.78 points higher than that of the NER-R. These results show that combination of word information and rules contributes to improved accuracy. We also evaluated the effec- Table 3 : Experimental Results obtained with NE recognizers using word information and rules: G., A., and AV. indicate GENERAL, ARREST, and a micro average obtained with each NE recognizer at each iteration, respectively. 1 2 3 4 5 G. 85.35 88.43 88.22 88.20 88.31 A. 85.64 91.33 91.52 91.49 92.19 AV. 85.40 89.00 88.88 88.85 89.08 tiveness of the combination of rules for identifying NEs, and rules for identifying beginning of NEs or end of NEs. The micro average FM values for an NE recognizer using rules for identifying NEs, an NE recognizer using rules for identifying beginning of NEs or end of NEs, and the NE recognizer using the both types of rules are 85. 77, 84.19 and 86.22 . This result shows using the two types of rules are effective.",
"cite_spans": [
{
"start": 608,
"end": 704,
"text": "85.35 88.43 88.22 88.20 88.31 A. 85.64 91.33 91.52 91.49 92.19 AV. 85.40 89.00 88.88 88.85 89.08",
"ref_id": null
},
{
"start": 1040,
"end": 1059,
"text": "77, 84.19 and 86.22",
"ref_id": null
}
],
"ref_spans": [
{
"start": 373,
"end": 380,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation of Our Proposed Method",
"sec_num": "4.2"
},
{
"text": "Then we evaluate the effectiveness of the acquisition method described in section 3.4. Table 3 shows the accuracy obtained with each NE recognizer at each iteration. The results at iteration 1 is the results obtained with the base line NE recognizer not using word information and rules. We obtained the best average accuracy at iteration 5. The results obtained with the NE recognizer at iteration 5 shows 4.76 points higher average Fmeasure than that of the NE recognizer at iteration 1, and 0.37 points higher average F-measure than that of the NE recognizer at iteration 2. Table 4 shows the results of the previous works using IREX Japanese NE recognition tasks. All the results were obtained with CRL data as manually labeled training data. Our results are Fmeasure values obtained with the NE recognizer at iteration 5 on Table 3 .",
"cite_spans": [],
"ref_spans": [
{
"start": 87,
"end": 94,
"text": "Table 3",
"ref_id": null
},
{
"start": 578,
"end": 585,
"text": "Table 4",
"ref_id": null
},
{
"start": 829,
"end": 836,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation of Our Proposed Method",
"sec_num": "4.2"
},
{
"text": "We see that our NE recognizer shows the best F-measure values for GENERAL and ARREST. Compared with our method only using unlabeled data, most previous works use handcrafted resources, such as a set of NEs are used in (Uchimoto et al., 2000) , and NTT GOI Taikei (Ikehara et al., 1999) , which is a handcrafted thesaurus, is used in (Isozaki and Kazawa, 2002; Sasano and Kurohashi, 2008) . These results indicate that word information and rules acquired from large unlabeled data are also useful as well as handcrafted resources. In addition, we see that our method with large labeled data show much better perfor- Table 4 : Comparison with previous works. GE and AR indicate GENERAL and ARREST. GE AR (Uchimoto et al., 2000) 80.17 85.75 (Takemoto et al., 2001) 83.86 - (Utsuro et al., 2002) 84.07 - (Isozaki and Kazawa, 2002) 85.77 - (Sasano and Kurohashi, 2008) 87.72 - (Iwakura, 2010) 87 ",
"cite_spans": [
{
"start": 218,
"end": 241,
"text": "(Uchimoto et al., 2000)",
"ref_id": "BIBREF18"
},
{
"start": 263,
"end": 285,
"text": "(Ikehara et al., 1999)",
"ref_id": "BIBREF3"
},
{
"start": 333,
"end": 359,
"text": "(Isozaki and Kazawa, 2002;",
"ref_id": "BIBREF5"
},
{
"start": 360,
"end": 387,
"text": "Sasano and Kurohashi, 2008)",
"ref_id": "BIBREF14"
},
{
"start": 702,
"end": 725,
"text": "(Uchimoto et al., 2000)",
"ref_id": "BIBREF18"
},
{
"start": 738,
"end": 761,
"text": "(Takemoto et al., 2001)",
"ref_id": "BIBREF16"
},
{
"start": 770,
"end": 791,
"text": "(Utsuro et al., 2002)",
"ref_id": "BIBREF19"
},
{
"start": 800,
"end": 826,
"text": "(Isozaki and Kazawa, 2002)",
"ref_id": "BIBREF5"
},
{
"start": 835,
"end": 863,
"text": "(Sasano and Kurohashi, 2008)",
"ref_id": "BIBREF14"
},
{
"start": 872,
"end": 887,
"text": "(Iwakura, 2010)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 615,
"end": 622,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation of Our Proposed Method",
"sec_num": "4.2"
},
{
"text": "This section describes the performances of NE recognizers trained with larger training data than CRL-data. Figure 1 shows the performance of each NE recognizer trained with different size of labeled training data. The leftmost points are the performance of the NE recognizers trained with CRL data (1,174 news stories). The other points are the performances of NE recognizers trained with training data larger than CRL data. The size of the additional training data is increased by 500 news stories.",
"cite_spans": [],
"ref_spans": [
{
"start": 107,
"end": 115,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Evaluating Effectiveness of Our Method",
"sec_num": "4.3"
},
{
"text": "We examined NE recognizers using our proposed method (semi), and NE recognizers not using our method (non-semi). In the following, semi-NER indicates NE recognizers using unlabeled data based on our method, and non-semi-NER indicates NE recognizers not using unlabeled data. Figure 1 shows that the semi-NER trained with CRL data shows competitive perfor-mance of the non-semi-NER trained with about 1.5 time larger training data consisting of CRL data and additional labeled 500 news stories. To create manually labeled 500 news stories, about 36 hours are required. 6 To achieve the competitive performance of the non-semi-NER trained with CRL data and the labeled 7,000 news stories, semi-NER requires only 2,000 news stories in addition to CRL data. This result shows that our proposed method significantly reduces the number of labeled data to achieve a competitive performance obtained with only using labeled data. Figure 1 also shows that our method contributes to improved accuracy when using the large labeled training data consisting of CRL data and 7,000 news stories. The accuracy is 90.47 for GEN-ERAL, and 94.30 for ARREST. In contrast, when without word information and rules acquired from unlabeled data, the accuracy is 89.43 for GEN-ERAL, and 93.44 for ARREST.",
"cite_spans": [],
"ref_spans": [
{
"start": 275,
"end": 283,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 922,
"end": 930,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Evaluating Effectiveness of Our Method",
"sec_num": "4.3"
},
{
"text": "To augment features, methods for using information obtained with clustering algorithms were proposed. These methods used word clusters (Freitag, 2004; Miller et al., 2004) , the clusters of multi-word nouns (Kazama and Torisawa, 2008) , or phrase clusters (Lin and Wu, 2009) . In contrast, to collect rules, we use an automatically tagged data with an NE recognizer. Therefore, we expect to obtain more target-task-oriented information with our method than that of previous works. Although there are differences between our method and the previous works, our method and previous works are complementary .",
"cite_spans": [
{
"start": 135,
"end": 150,
"text": "(Freitag, 2004;",
"ref_id": "BIBREF2"
},
{
"start": 151,
"end": 171,
"text": "Miller et al., 2004)",
"ref_id": "BIBREF12"
},
{
"start": 207,
"end": 234,
"text": "(Kazama and Torisawa, 2008)",
"ref_id": "BIBREF10"
},
{
"start": 256,
"end": 274,
"text": "(Lin and Wu, 2009)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "To use rules in machine-learning-based NE recognitions, Isozaki proposed a Japanese NE recognition method based on a simple rule generator and decision tree learning. The method generates rules from supervised training data (Isozaki, 2001) . Talukdar et al., proposed a method to use lists of NEs acquired from unlabeled data for NE recognition (Talukdar et al., 2006) . Starting with a few NE seed examples, the method extends lists of NEs. These methods use rules or lists of NEs for identifying only NEs. Compared with these methods, our method uses rules for identifying the beginning of NEs and the end of NEs in addition 6 We estimate the hours by using the average labeling time of a news story. The average time is 260 seconds per news story.",
"cite_spans": [
{
"start": 224,
"end": 239,
"text": "(Isozaki, 2001)",
"ref_id": "BIBREF6"
},
{
"start": 345,
"end": 368,
"text": "(Talukdar et al., 2006)",
"ref_id": "BIBREF17"
},
{
"start": 627,
"end": 628,
"text": "6",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "to rules identifying whole NEs. Therefore, our methods can use new features not used in previous works.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "This paper proposed an NE recognition method using rules acquired from unlabeled data. Our method acquires rules for identifying NEs, the beginning of NEs, and the end of NEs from an automatically labeled data with an NE recognizer. In addition, we use word information including the candidate NE classes, and so on. We evaluated our method with IREX data set for Japanese NE recognition and unlabeled data consisting of more than one billion words. The experimental results showed that our method using rules and word information achieved the best accuracy on the GEN-ERAL and ARREST tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "http://mecab.sourceforge.net/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "For example, Japanese word \"houbei\" (visit U.S.) does not match with LOCATION \"bei (U.S)\".3 If \"Gaimusyouha\", is segmented as \"Gaimusyou (the Ministry of Foreign Affairs) / ha (particle)\", then words including characters are follows; W (Gai) = Gaimusyou, W (mu) = Gaimusyou, W (syou) = Gaimusyou, and W (ha)=ha. The identifiers that indicate positions where characters appear are follows; P (Gai) =B, P (mu) = I, P (syou) = E, and P (ha)=S.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We conducted experiments using word information and rules obtained from training data with different frequency threshold parameters. The parameters are 1, 3, 5, 10, 20, 30, 40, and 50. We select 50 as the threshold because the parameter shows the best result among the results obtained with these parameters on a pilot study.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We used WikiPedia in addition to news stories becauseSuzuki and Isozaki (Suzuki and Isozaki, 2008) reported that the use of more unlabeled data in their learning algorithm can really lead to further improvements. We treated a successive numbers and alphabets as a word in this experiment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A high-performance semi-supervised learning method for text chunking",
"authors": [
{
"first": "Rie",
"middle": [],
"last": "Ando",
"suffix": ""
},
{
"first": "Tong",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. of ACL 2005",
"volume": "",
"issue": "",
"pages": "1--9",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rie Ando and Tong Zhang. 2005. A high-performance semi-supervised learning method for text chunking. In Proc. of ACL 2005, pages 1-9.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Japanese named entity extraction with redundant morphological analysis",
"authors": [
{
"first": "Masayuki",
"middle": [],
"last": "Asahara",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. of HLT-NAACL 2003",
"volume": "",
"issue": "",
"pages": "8--15",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Masayuki Asahara and Yuji Matsumoto. 2003. Japanese named entity extraction with redundant morphological analysis. In Proc. of HLT-NAACL 2003, pages 8-15.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Trained named entity recognition using distributional clusters",
"authors": [
{
"first": "Dayne",
"middle": [],
"last": "Freitag",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of EMNLP",
"volume": "",
"issue": "",
"pages": "262--269",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dayne Freitag. 2004. Trained named entity recog- nition using distributional clusters. In Proc. of EMNLP 2004, pages 262-269.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Goi-Taikei -A Japanese Lexicon CDROM. Iwanami Shoten",
"authors": [
{
"first": "Satoru",
"middle": [],
"last": "Ikehara",
"suffix": ""
},
{
"first": "Masahiro",
"middle": [],
"last": "Miyazaki",
"suffix": ""
},
{
"first": "Satoshi",
"middle": [],
"last": "Shirai",
"suffix": ""
},
{
"first": "Akio",
"middle": [],
"last": "Yokoo",
"suffix": ""
},
{
"first": "Hiromi",
"middle": [],
"last": "Nakaiwa",
"suffix": ""
},
{
"first": "Kentaro",
"middle": [],
"last": "Ogura",
"suffix": ""
},
{
"first": "Yoshifumi",
"middle": [],
"last": "Ooyama",
"suffix": ""
},
{
"first": "Yoshihiki",
"middle": [],
"last": "Hayashi",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai, Akio Yokoo, Hiromi Nakaiwa, Kentaro Ogura, Yoshifumi Ooyama, and Yoshihiki Hayashi. 1999. Goi-Taikei -A Japanese Lexicon CDROM. Iwanami Shoten.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Proc. of the IREX workshop",
"authors": [
{
"first": "Irex",
"middle": [],
"last": "Committee",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Committee IREX. 1999. Proc. of the IREX workshop.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Speeding up named entity recognition based on Support Vector Machines",
"authors": [
{
"first": "Hideki",
"middle": [],
"last": "Isozaki",
"suffix": ""
},
{
"first": "Hideto",
"middle": [],
"last": "Kazawa",
"suffix": ""
}
],
"year": 2002,
"venue": "IPSJ SIG notes NL-149-1",
"volume": "",
"issue": "",
"pages": "1--8",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hideki Isozaki and Hideto Kazawa. 2002. Speeding up named entity recognition based on Support Vector Machines (in Japanese). In IPSJ SIG notes NL-149- 1, pages 1-8.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Japanese named entity recognition based on a simple rule generator and decision tree learning",
"authors": [
{
"first": "Hideki",
"middle": [],
"last": "Isozaki",
"suffix": ""
}
],
"year": 2001,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "314--321",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hideki Isozaki. 2001. Japanese named entity recogni- tion based on a simple rule generator and decision tree learning. In Proc. of ACL 2001, pages 314-321.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "A fast boosting-based learner for feature-rich tagging and chunking",
"authors": [
{
"first": "Tomoya",
"middle": [],
"last": "Iwakura",
"suffix": ""
},
{
"first": "Seishi",
"middle": [],
"last": "Okamoto",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. of CoNLL",
"volume": "",
"issue": "",
"pages": "17--24",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tomoya Iwakura and Seishi Okamoto. 2008. A fast boosting-based learner for feature-rich tagging and chunking. In Proc. of CoNLL 2008, pages 17-24.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Fast boosting-based part-ofspeech tagging and text chunking with efficient rule representation for sequential labeling",
"authors": [
{
"first": "Tomoya",
"middle": [],
"last": "Iwakura",
"suffix": ""
}
],
"year": 2009,
"venue": "Proc. of RANLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tomoya Iwakura. 2009. Fast boosting-based part-of- speech tagging and text chunking with efficient rule representation for sequential labeling. In Proc. of RANLP 2009.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A named entity extraction using word information repeatedly collected from unlabeled data",
"authors": [
{
"first": "Tomoya",
"middle": [],
"last": "Iwakura",
"suffix": ""
}
],
"year": 2010,
"venue": "Proc. of CICLing",
"volume": "",
"issue": "",
"pages": "212--223",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tomoya Iwakura. 2010. A named entity extraction us- ing word information repeatedly collected from un- labeled data. In Proc. of CICLing 2010, pages 212- 223.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Inducing gazetteers for named entity recognition by largescale clustering of dependency relations",
"authors": [
{
"first": "Kentaro",
"middle": [],
"last": "Jun'ichi Kazama",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Torisawa",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. of ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "407--415",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jun'ichi Kazama and Kentaro Torisawa. 2008. Induc- ing gazetteers for named entity recognition by large- scale clustering of dependency relations. In Proc. of ACL-08: HLT, pages 407-415.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Phrase clustering for discriminative learning",
"authors": [
{
"first": "Dekang",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Xiaoyun",
"middle": [],
"last": "Wu",
"suffix": ""
}
],
"year": 2009,
"venue": "Proc. of ACL-IJCNLP 2009",
"volume": "",
"issue": "",
"pages": "1030--1038",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dekang Lin and Xiaoyun Wu. 2009. Phrase cluster- ing for discriminative learning. In Proc. of ACL- IJCNLP 2009, pages 1030-1038.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Name tagging with word clusters and discriminative training",
"authors": [
{
"first": "Scott",
"middle": [],
"last": "Miller",
"suffix": ""
},
{
"first": "Jethran",
"middle": [],
"last": "Guinness",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Zamanian",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of HLT-NAACL 2004",
"volume": "",
"issue": "",
"pages": "337--342",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Scott Miller, Jethran Guinness, and Alex Zamanian. 2004. Name tagging with word clusters and dis- criminative training. In Proc. of HLT-NAACL 2004, pages 337-342.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Japanese named entity extraction with bunsetsu features",
"authors": [
{
"first": "Keigo",
"middle": [],
"last": "Nakano",
"suffix": ""
},
{
"first": "Yuzo",
"middle": [],
"last": "Hirai",
"suffix": ""
}
],
"year": 2004,
"venue": "IPSJ Journal",
"volume": "45",
"issue": "3",
"pages": "934--941",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Keigo Nakano and Yuzo Hirai. 2004. Japanese named entity extraction with bunsetsu features (in Japanese). In IPSJ Journal, 45(3), pages 934-941.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Japanese named entity recognition using structural natural language processing",
"authors": [
{
"first": "Ryohei",
"middle": [],
"last": "Sasano",
"suffix": ""
},
{
"first": "Sadao",
"middle": [],
"last": "Kurohashi",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. of IJCNLP",
"volume": "",
"issue": "",
"pages": "607--612",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ryohei Sasano and Sadao Kurohashi. 2008. Japanese named entity recognition using structural natural language processing. In Proc. of IJCNLP 2008, pages 607-612.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Semi-supervised sequential labeling and segmentation using gigaword scale unlabeled data",
"authors": [
{
"first": "Jun",
"middle": [],
"last": "Suzuki",
"suffix": ""
},
{
"first": "Hideki",
"middle": [],
"last": "Isozaki",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. of ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "665--673",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jun Suzuki and Hideki Isozaki. 2008. Semi-supervised sequential labeling and segmentation using giga- word scale unlabeled data. In Proc. of ACL-08: HLT, pages 665-673.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "A Japanese named entity extraction system based on building a large-scale and high quality dictionary and pattern-matching rules",
"authors": [
{
"first": "Yoshikazu",
"middle": [],
"last": "Takemoto",
"suffix": ""
},
{
"first": "Toshikazu",
"middle": [],
"last": "Fukushima",
"suffix": ""
},
{
"first": "Hiroshi",
"middle": [],
"last": "Yamada",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "42",
"issue": "",
"pages": "1580--1591",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoshikazu Takemoto, Toshikazu Fukushima, and Hi- roshi Yamada. 2001. A Japanese named entity ex- traction system based on building a large-scale and high quality dictionary and pattern-matching rules (in Japanese). 42(6):1580-1591.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "A context pattern induction method for named entity extraction",
"authors": [
{
"first": "Partha",
"middle": [],
"last": "Pratim Talukdar",
"suffix": ""
},
{
"first": "Thorsten",
"middle": [],
"last": "Brants",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Liberman",
"suffix": ""
},
{
"first": "Fernando",
"middle": [],
"last": "Pereira",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. of CoNLL",
"volume": "",
"issue": "",
"pages": "141--148",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Partha Pratim Talukdar, Thorsten Brants, Mark Liber- man, and Fernando Pereira. 2006. A context pat- tern induction method for named entity extraction. In Proc. of CoNLL 2006, pages 141-148.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Named entity extraction based on a maximum entropy model and transformati on rules",
"authors": [
{
"first": "Kiyotaka",
"middle": [],
"last": "Uchimoto",
"suffix": ""
},
{
"first": "Qing",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "Masaki",
"middle": [],
"last": "Murata",
"suffix": ""
},
{
"first": "Hiromi",
"middle": [],
"last": "Ozaku",
"suffix": ""
},
{
"first": "Masao",
"middle": [],
"last": "Utiyama",
"suffix": ""
},
{
"first": "Hitoshi",
"middle": [],
"last": "Isahara",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. of the ACL 2000",
"volume": "",
"issue": "",
"pages": "326--335",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kiyotaka Uchimoto, Qing Ma, Masaki Murata, Hiromi Ozaku, Masao Utiyama, and Hitoshi Isahara. 2000. Named entity extraction based on a maximum en- tropy model and transformati on rules. In Proc. of the ACL 2000, pages 326-335.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Combining outputs of multiple Japanese named entity chunkers by stacking",
"authors": [
{
"first": "Takehito",
"middle": [],
"last": "Utsuro",
"suffix": ""
},
{
"first": "Manabu",
"middle": [],
"last": "Sassano",
"suffix": ""
},
{
"first": "Kiyotaka",
"middle": [],
"last": "Uchimoto",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of EMNLP 2002",
"volume": "",
"issue": "",
"pages": "281--288",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Takehito Utsuro, Manabu Sassano, and Kiyotaka Uchi- moto. 2002. Combining outputs of multiple Japanese named entity chunkers by stacking. In Proc. of EMNLP 2002, pages 281-288.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "Experimental results obtained with different size of training data. Each point indicates the micro average F-measure of an NE recognizer. mance than the other methods.",
"type_str": "figure",
"num": null
},
"TABREF0": {
"content": "<table><tr><td>Hiragana (Japanese syllabary characters), Katakana,</td></tr><tr><td>Kanji (Chinese letter), Capital alphabet,</td></tr><tr><td>Lower alphabet, number and Others</td></tr><tr><td>training data.</td></tr></table>",
"html": null,
"type_str": "table",
"text": "Basic character types",
"num": null
},
"TABREF1": {
"content": "<table><tr><td/><td>B.</td><td>+ W</td><td>+ R</td><td>+WR</td></tr><tr><td colspan=\"5\">GENERAL 85.35 88.04 85.93 88.43</td></tr><tr><td>ARREST</td><td colspan=\"4\">85.64 89.35 87.39 91.33</td></tr><tr><td>AV.</td><td colspan=\"4\">85.40 88.56 86.22 89.00</td></tr></table>",
"html": null,
"type_str": "table",
"text": "Experimental Results: Each AV. indicates a micro average F-measure obtained with each NE recognizer. B., +W, +R, and +WR indicate the base line recognizer, using word information, using rules, and using word information and rules. Base indicates the base line NE recognizer not using word information and rules.",
"num": null
}
}
}
} |