File size: 85,590 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 |
{
"paper_id": "I05-1002",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:25:42.186497Z"
},
"title": "Topic Tracking Based on Linguistic Features",
"authors": [
{
"first": "Fumiyo",
"middle": [],
"last": "Fukumoto",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Univ. of Yamanashi",
"location": {
"addrLine": "4-3-11",
"postCode": "400-8511",
"settlement": "Takeda, Kofu",
"country": "Japan"
}
},
"email": "fukumoto@yamanashi.ac.jp"
},
{
"first": "Yusuke",
"middle": [],
"last": "Yamaji",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Univ. of Yamanashi",
"location": {
"addrLine": "4-3-11",
"postCode": "400-8511",
"settlement": "Takeda, Kofu",
"country": "Japan"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper explores two linguistically motivated restrictions on the set of words used for topic tracking on newspaper articles: named entities and headline words. We assume that named entities is one of the linguistic features for topic tracking, since both topic and event are related to a specific place and time in a story. The basic idea to use headline words for the tracking task is that headline is a compact representation of the original story, which helps people to quickly understand the most important information contained in a story. Headline words are automatically generated using headline generation technique. The method was tested on the Mainichi Shimbun Newspaper in Japanese, and the results of topic tracking show that the system works well even for a small number of positive training data.",
"pdf_parse": {
"paper_id": "I05-1002",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper explores two linguistically motivated restrictions on the set of words used for topic tracking on newspaper articles: named entities and headline words. We assume that named entities is one of the linguistic features for topic tracking, since both topic and event are related to a specific place and time in a story. The basic idea to use headline words for the tracking task is that headline is a compact representation of the original story, which helps people to quickly understand the most important information contained in a story. Headline words are automatically generated using headline generation technique. The method was tested on the Mainichi Shimbun Newspaper in Japanese, and the results of topic tracking show that the system works well even for a small number of positive training data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "With the exponential growth of information on the Internet, it is becoming increasingly difficult to find and organize relevant materials. Tracking task, i.e. starts from a few sample stories and finds all subsequent stories that discuss the target topic, is a new line of research to attack the problem. One of the major problems in the tracking task is how to make a clear distinction between a topic and an event in the story. Here, an event refers to the subject of a story itself, i.e. a writer wants to express, in other words, notions of who, what, where, when, why and how in the story. On the other hand, a topic is some unique thing that occurs at a specific place and time associated with some specific actions [1] . It becomes background among stories. Therefore, an event drifts, but a topic does not. For example, in the stories of 'Kobe Japan quake' from the TDT1 corpus, the event includes early reports of damage, location and nature of quake, rescue efforts, consequences of the quake, and on-site reports, while the topic is Kobe Japan quake.",
"cite_spans": [
{
"start": 722,
"end": 725,
"text": "[1]",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A wide range of statistical and machine learning techniques have been applied to topic tracking, including k-Nearest Neighbor classification, Decision Tree induction [3] , relevance feedback method of IR [12, 13] , hierarchical and non-hierarchical clustering algorithms [20] , and a variety of Language Modeling [15, 5, 10, 17] . The main task of these techniques is to tune the parameters or the threshold for binary decisions to produce optimal results. In the TDT context, however, parameter tuning is a tricky issue for tracking. Because only the small number of labeled positive stories is available for training. Moreover, the well-known past experience from IR that notions of who, what, where, when, why, and how may not make a great contribution to the topic tracking task [1] causes this fact, i.e. a topic and an event are different from each other. This paper explores two linguistically motivated restrictions on the set of words used for topic tracking on newspaper articles: named entities and headline words. A topic is related to a specific place and time, and an event refers to notions of who(person), where(place), when(time) including what, why and how in a story. Therefore, we can assume that named entities is one of the linguistic features for topic tracking. Another linguistic feature is a set of headline words. The basic idea to use headline words for topic tracking is that headline is a compact representation of the original story, which helps people to quickly understand the most important information contained in a story, and therefore, it may include words to understand what the story is about, what is characteristic of this story with respect to other stories, and hopefully include words related to both topic and event in the story. A set of headline words is automatically generated. To do this, we use a technique proposed by Banko [2] . It produces coherent summaries by building statistical models for content selection and surface realization. Another purpose of this work is to create Japanese corpus for topic tracking task. We used Mainichi Shimbun Japanese Newspaper corpus from Oct. to Dec. of 1998 which corresponds to the TDT3 corpus. We annotated these articles against the 60 topics which are defined by the TDT3.",
"cite_spans": [
{
"start": 166,
"end": 169,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 204,
"end": 208,
"text": "[12,",
"ref_id": "BIBREF11"
},
{
"start": 209,
"end": 212,
"text": "13]",
"ref_id": "BIBREF12"
},
{
"start": 271,
"end": 275,
"text": "[20]",
"ref_id": "BIBREF19"
},
{
"start": 313,
"end": 317,
"text": "[15,",
"ref_id": "BIBREF14"
},
{
"start": 318,
"end": 320,
"text": "5,",
"ref_id": "BIBREF4"
},
{
"start": 321,
"end": 324,
"text": "10,",
"ref_id": "BIBREF9"
},
{
"start": 325,
"end": 328,
"text": "17]",
"ref_id": "BIBREF16"
},
{
"start": 783,
"end": 786,
"text": "[1]",
"ref_id": "BIBREF0"
},
{
"start": 1877,
"end": 1880,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The rest of the paper is organized as follows. The next section provides an overview of existing topic tracking techniques. We then describe a brief explanation of a headline generation technique proposed by Banko et al. [2] . Next, we present our method for topic tracking, and finally, we report some experiments using the Japanese newspaper articles with a discussion of evaluation.",
"cite_spans": [
{
"start": 221,
"end": 224,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The approach that relies mainly on corpus statistics is widely studied in the topic tracking task, and an increasing number of machine learning techniques have been applied to the task. CMU proposed two methods: a k-Nearest Neighbor (kNN) classifier and a Decision-Tree Induction (dtree) classifier [1, 20, 3] . Dragon Systems proposed two tracking systems; one is based on standard language modeling technique, i.e. unigram statistics to measure story similarity [18] and another is based on a Beta-Binomial model [10] . UMass viewed the tracking problem as an instance of on-line document classification, i.e. it classifies documents into categories or classes [4, 8, 19, 9, 14] . They proposed a method including query expansion with multi-word features and weight-learning steps for building linear text classifiers for the tracking task [13] . These approaches, described above, seem to be robust and have shown satisfactory performance in stories from different corpora, i.e. TDT1 and TDT2. However, Carbonell claims that something more is needed if the system is intended for recognizing topic drift [3] . Yang et al. addressed the issue of difference between early and later stories related to the target event in the TDT tracking task. They adapted several machine learning techniques, including k-Nearest Neighbor(kNN) algorithm and Rocchio approach [21] . Their method combines the output of a diverse set of classifiers and tuning parameters for the combined system on a retrospective corpus. The idea comes from the well-known practice in information retrieval and speech recognition of combining the output of a large number of systems to yield a better result than the individual system's output. They reported that the new variants of kNN reduced up to 71% in weighted error rates on the TDT3-dryrun corpus. GE R&D proposed a method for topic tracking by using summarization technique, i.e. using content compression rather than on corpus statistics to detect relevance and assess topicality of the source material [16] . Their system operates by first creating a topic tracking query out of the available training stories. Subsequently, it accepts incoming stories, summarizes them topically, scores the summaries(passages) for content, then assesses content relevance to the tracking query. They reported stories whose compressed content summaries clear the empirically established threshold are classified as being 'on topic'. Unlike most previous work on summarization which focused on extractive summarization: selecting text spans -either complete sentences or paragraphs -from the original story, this approach solves a problem for extractive summarization, i.e. in many cases, the most important information in the story is scattered across multiple sentences. However, their approach uses frequency-based term weighting. Therefore, it is not clear if the method can identify the most important information contained in a story.",
"cite_spans": [
{
"start": 299,
"end": 302,
"text": "[1,",
"ref_id": "BIBREF0"
},
{
"start": 303,
"end": 306,
"text": "20,",
"ref_id": "BIBREF19"
},
{
"start": 307,
"end": 309,
"text": "3]",
"ref_id": "BIBREF2"
},
{
"start": 464,
"end": 468,
"text": "[18]",
"ref_id": "BIBREF17"
},
{
"start": 515,
"end": 519,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 663,
"end": 666,
"text": "[4,",
"ref_id": "BIBREF3"
},
{
"start": 667,
"end": 669,
"text": "8,",
"ref_id": "BIBREF7"
},
{
"start": 670,
"end": 673,
"text": "19,",
"ref_id": "BIBREF18"
},
{
"start": 674,
"end": 676,
"text": "9,",
"ref_id": "BIBREF8"
},
{
"start": 677,
"end": 680,
"text": "14]",
"ref_id": "BIBREF13"
},
{
"start": 842,
"end": 846,
"text": "[13]",
"ref_id": "BIBREF12"
},
{
"start": 1107,
"end": 1110,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 1360,
"end": 1364,
"text": "[21]",
"ref_id": "BIBREF20"
},
{
"start": 2031,
"end": 2035,
"text": "[16]",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "These methods, described above, show that it is crucial to develop a method for extracting words related to both topic and event in a story. Like other approaches, our method is based on corpus statistics. However, our method uses two linguistically motivated restrictions on the set of words: named entities and headline words. We assume that named entities is one of the linguistic features for topic tracking, since both topic and event are related to a specific place and time in a story. Another linguistic feature is a set of headline words. The basic idea to use headline words is that headline is a compact representation of the original story, and therefore, it may include words to understand what the story is about, and hopefully include words related to both topic and event in the story.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Banko et al. proposed an approach to summarization capable of generating summaries shorter than a sentence. It produces by building statistical models for content selection and surface realization. We used their method to extract headline words. Content selection requires that the system learns a model of the relationship between the appearance of words in a story and the appearance of corresponding words in the headline. The probability of a candidate headline, H, consisting of words (w 1 ,w 2 ,\u2022 \u2022 \u2022,w n ), can be computed:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Headline",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (w 1 , \u2022 \u2022 \u2022 , w n ) = n i=1 P (w i \u2208 H | w i \u2208 D) \u2022 P (len(H) = n) \u2022 n i=2 P (w i | w 1 , \u2022 \u2022 \u2022 , w i\u22121 )",
"eq_num": "(1)"
}
],
"section": "Generating Headline",
"sec_num": "3"
},
{
"text": "In formula (1), the first term denotes the words selected for the headline, and can be computed:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Headline",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (w i \u2208 H | w i \u2208 D) = P (w i \u2208 D | w i \u2208 H) \u2022 P (w i \u2208 H) P (w i \u2208 D)",
"eq_num": "(2)"
}
],
"section": "Generating Headline",
"sec_num": "3"
},
{
"text": "where H and D represent the bags of words that the headline and the story contain. Formula (2) shows the conditional probability of a word occurring in the headline given that the word appeared in the story. It has been estimated from a suitable story/headline corpus. The second term in formula (1) shows the length of the resulting headline, and can also be learned from the source story. The third term shows the most likely sequencing of the words in the content set. Banko et al. assumed that the likelihood of a word in the story is independent of other words in the headline. Surface realization is to estimate the probability of any particular surface ordering as a headline candidate. It can be computed by modeling the probability of word sequences. Banko et al. used a bigram language model. When they estimate probabilities for sequences that have not been seen in the training data, they used back-off weights [6] . Headline generation can be obtained as a weighted combination of the content and structure model log probabilities which is shown in formula (3) .",
"cite_spans": [
{
"start": 91,
"end": 94,
"text": "(2)",
"ref_id": "BIBREF1"
},
{
"start": 296,
"end": 299,
"text": "(1)",
"ref_id": "BIBREF0"
},
{
"start": 923,
"end": 926,
"text": "[6]",
"ref_id": "BIBREF5"
},
{
"start": 1070,
"end": 1073,
"text": "(3)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Headline",
"sec_num": "3"
},
{
"text": "arg max H (\u03b1 \u2022 n i=1 log(P (w i \u2208 H | w i \u2208 D)) + \u03b2 \u2022 log(P (len(H) = n)) + \u03b3 \u2022 n i=2 log(P (w i | w i\u22121 ))) (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Headline",
"sec_num": "3"
},
{
"text": "To generate a headline, it is necessary to find a sequence of words that maximizes the probability, under the content selection and surface realization models, that it was generated from the story to be summarized. In formula (3), cross-validation is used to learn weights, \u03b1, \u03b2 and \u03b3 for a particular story genre.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Headline",
"sec_num": "3"
},
{
"text": "We explore two linguistically motivated restrictions on the set of words used for tracking: named entities and headline words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extracting Linguistic Features and Tracking",
"sec_num": "4"
},
{
"text": "For identifying named entities, we use CaboCha [7] for Japanese Mainichi Shimbun corpus, and extracted Person Name, Organization, Place, and Proper Name.",
"cite_spans": [
{
"start": 47,
"end": 50,
"text": "[7]",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Extracting Named Entities and Generating Headline Words",
"sec_num": "4.1"
},
{
"text": "Headline generation can be obtained as a weighted combination of the content and structure model log probabilities shown in formula (3). The system was trained on the 3 months Mainichi Shimbun articles((27,133 articles from Jan. to Mar. 1999) for Japanese corpus. We estimate \u03b1, \u03b2 and \u03b3 in formula (3) using 5 cross-validation 1 . Fig. 1 illustrates sample output using Mainichi Shimbun corpus. Numbers to the right are log probabilities of the word sequence.",
"cite_spans": [
{
"start": 327,
"end": 328,
"text": "1",
"ref_id": "BIBREF0"
}
],
"ref_spans": [
{
"start": 331,
"end": 337,
"text": "Fig. 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Extracting Named Entities and Generating Headline Words",
"sec_num": "4.1"
},
{
"text": "In the TDT tracking task, the number of labeled positive training stories is small (at most 16 stories) compared to the negative training stories. Therefore, the choice of good negative stories from a large number of training data is an important issue to detect subject shifts for a binary classifier such as a machine learning technique, Support Vector Machines(SVMs) [22] . We apply hierarchical classification technique to the training data.",
"cite_spans": [
{
"start": 370,
"end": 374,
"text": "[22]",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "h Headlinei j l m n o (Pakistan) p q r t u v w y { | (Kashimir issue) } 3 j t (third party mediation) (meeting) h /Headlinei m t , j l m n o 2 , o \u00a1 , \u00a3 \u00a4 \u00a5 \u00a6 j l m n o } 3 \u00a7 \u00a9 o j l m n o \u00ac \u00ae\u00b0\u00b1 \u00b2 \u00a7 3 \u00b5 \u00a3 \u00b6 p q r t u v w \u00a9 \u2022 \u00a3 \u00b6 \u00b9 \u00ba \u00bb \u00bc \u00be \u00c0\u00b0",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": ". .. (ISLAMABAD, Pakistan, Less than two weeks ahead of fresh talks with its uneasy neighbor India, Pakistan pressed on Saturday for international mediation in the thorny Kashmir issue, the flashpoint of two previous wars between the two countries...) \u00c2 Generated title words\u00c3 2: A hierarchical decomposition of a classification problem can be used to set the negative set for discriminative training. We use partitioning clustering algorithm, k-means (k = 2) which partitions a training data into clusters where similar stories are found in the same cluster and separated from dissimilar stories. Fig. 2 illustrates hierarchical classification of training data with k-means. Each level in Fig. 2 denotes the result obtained by a simple k-means (k=2) algorithm, and consists of two clusters: one is a cluster which includes positive and negative stories. Another is a cluster with only negative stories, each of these are dissimilar with the positive stories. The algorithm involves iterating through the data that the system is permitted to classify during each iteration. More specifically:",
"cite_spans": [],
"ref_spans": [
{
"start": 598,
"end": 604,
"text": "Fig. 2",
"ref_id": null
},
{
"start": 690,
"end": 696,
"text": "Fig. 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "p q r t u (Kashimir) v",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "p q r t u (Kashimir) o (India) AE \u00c7 (resume) m t (Islamabad) p q r t u (Kashimir) m \u00c9 \u00ca (Muslim) -38.32",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "1. In the training data which includes all the initial positive training stories, select two initial seeds g ands i , where g is a vector of the center of gravity on positive training stories, ands i is a vector of the negative training story which has the smallest value(as measured by cosine similarity) betweens i and g. The center of gravity g is defined as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "g = (g 1 , \u2022 \u2022 \u2022 , g n ) = ( 1 p p i=1 s i1 , \u2022 \u2022 \u2022 , 1 p p i=1 s in )",
"eq_num": "(4)"
}
],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "where s ij (1 \u2264 j \u2264 n) is the TF * IDF value of word j in the positive story s i . 2. Apply k-means (k=2) to the training data. 3. For the cluster which includes positive stories, iterate step 1 and 2 until positive training stories are divided into two clusters 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "Tracking involves a training phase and a testing phase. During the training phase, we employ the hierarchy which is shown in Fig. 2 by learning separate classifiers trained by SVMs. '\u00b11' in Fig. 2 denotes binary classification for stories at each level of the hierarchy. Each test story is judged to be negative or positive by using these classifiers to greedily select sub-branches until a leaf is reached. Once, the test story is judged to be negative, tracking is terminated. When the test story is judged to be positive by using a classifier of the bottom cluster, a cluster is divided into two: positive and negative stories. For each training data in the bottom cluster and test stories, we extract named entities and headline words. The result of training data is used to train SVMs and a classifier is induced. Each test story which also consists of a set of words produced by named entities and generating headline word procedures is judged to be negative or positive by using the classifier. This procedure, tracking, is repeated until the last test story is judged.",
"cite_spans": [],
"ref_spans": [
{
"start": 125,
"end": 131,
"text": "Fig. 2",
"ref_id": null
},
{
"start": 190,
"end": 196,
"text": "Fig. 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Tracking by Hierarchical Classification",
"sec_num": "4.2"
},
{
"text": "We chose the TDT3 corpus covering October 1, 1998 to December 31, 1998 as our gold standard corpus for creating Japanese corpus. The TDT3 corpus, developed at LDC, is a larger and richer collection, consisting of 34,600 stories with 60 manually identified topics. The stories were collected from 2 newswire, 3 radio programs and 4 television programs. We then create a Japanese corpus, i.e. we annotate Mainichi Shimbun Japanese Newspaper stories from October 1, 1998 to December 31, 1998 against the 60 topics. Not all the topics could have seen over the 3 months Japanese Newspaper stories. Table 1 shows 20 topics which are included in the Japanese Newspaper corpus.",
"cite_spans": [],
"ref_spans": [
{
"start": 593,
"end": 601,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments Set Up",
"sec_num": "5.1"
},
{
"text": "'Topic ID' in Table 1 denotes ID number defined by the TDT3. The evaluation for annotation is made by three humans. The classification is determined to be correct if the majority of three human judges agrees. The Japanese corpus consists of 27,133 stories. We used it in the experiment. We obtained a vocabulary of 52,065 unique words after tagging by a morphological analysis, Chasen [11] . Table 2 summarizes the results using all words for each sequence that maximizes the probability, i.e. 14 sequences in all. The results were obtained using the standard TDT . The test set is always the collection minus the N t = 16 stories. 'Miss' denotes Miss rate, which is the ratio of the stories that were judged as YES but were not evaluated as YES for the run in question. 'F/A' shows false alarm rate, which is the ratio of the stories judged as NO but were evaluated as YES. 'Prec.' is the ratio of correct assignments by the system divided by the total number of system's assignments. 'F'(pooled avg) is a measure that balances recall(Rec.) and precision, where recall denotes the ratio of correct assignments by the system divided by the total number of correct assignments. We recall that a generated headline is a sequence of words that maximizes the probability. We set the maximum number of word sequence by calculating the average number of the original titles, and obtained the number of 15 words. The minimum number of words in a sequence is two. Fig. 3 illustrates the extracted headline for each sequence. Box in Fig. 3 shows a word, and 'arg max P(x)' denotes the maximum probability of a candidate headline. For example, the extracted sequence ",
"cite_spans": [
{
"start": 385,
"end": 389,
"text": "[11]",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 14,
"end": 21,
"text": "Table 1",
"ref_id": null
},
{
"start": 392,
"end": 399,
"text": "Table 2",
"ref_id": null
},
{
"start": 1456,
"end": 1462,
"text": "Fig. 3",
"ref_id": null
},
{
"start": 1524,
"end": 1530,
"text": "Fig. 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments Set Up",
"sec_num": "5.1"
},
{
"text": "The extracted sequences Fig. 3 . The extracted headline for each sequence of two words is the sequence whose maximum probability is 1 2 . Table 2 shows that our method is more likely to be effective for higher values of N t , while F-score was 0 when N t = 1.",
"cite_spans": [],
"ref_spans": [
{
"start": 24,
"end": 30,
"text": "Fig. 3",
"ref_id": null
},
{
"start": 138,
"end": 145,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "{",
"sec_num": null
},
{
"text": "Our approach using the headline generation is to find a sequence of words that maximizes the probability. It can be produced for an arbitrary number of words. We recall that Table 2 shows the result using each sequence that maximizes the probability. However, when N t = 1, the result was not good, as the F-score was zero. We thus conducted the following two experiments to examine the effect of the number of words in a sequence: (1) the tracking task using all words, each of which is the element of only one sequence that maximizes the probability (Fig. 4) and (2) the tracking using various number of word sequences (Fig. 5) . In (2), we tested different number of words in a sequence, and we chose six words that optimized the global F-score. The results are shown in Tables 3 and 4 . ",
"cite_spans": [],
"ref_spans": [
{
"start": 174,
"end": 181,
"text": "Table 2",
"ref_id": null
},
{
"start": 552,
"end": 560,
"text": "(Fig. 4)",
"ref_id": "FIGREF3"
},
{
"start": 621,
"end": 629,
"text": "(Fig. 5)",
"ref_id": "FIGREF4"
},
{
"start": 774,
"end": 789,
"text": "Tables 3 and 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Title Words",
"sec_num": "5.3"
},
{
"text": "The extracted sequence Table 3 shows the tracking result using only one sequence of words that maximizes the probability, and Table 4 shows the result of six words. In Table 3 , the average number of words which maximizes the probability for all the training data is 4.4, and the result is similar to that of Table 4 . We can see from both Tables 3 and 4 that when the number of words in a sequence is small, the result has no effect with the number of positive training data, since the range of F-score in Table 3 is 0.415 \u223c 0.478, and that in Table 4 is 0.453 \u223c 0.516. On the other hand, as we can see from Table 2 , when the number of title words is large, the smaller the number of positive training data is, the worse the result is. To summarize the evaluation, the best result is when we use a sequence which consists of a small number of words, six words.",
"cite_spans": [],
"ref_spans": [
{
"start": 23,
"end": 30,
"text": "Table 3",
"ref_id": null
},
{
"start": 126,
"end": 133,
"text": "Table 4",
"ref_id": null
},
{
"start": 168,
"end": 175,
"text": "Table 3",
"ref_id": null
},
{
"start": 309,
"end": 316,
"text": "Table 4",
"ref_id": null
},
{
"start": 340,
"end": 354,
"text": "Tables 3 and 4",
"ref_id": null
},
{
"start": 507,
"end": 514,
"text": "Table 3",
"ref_id": null
},
{
"start": 545,
"end": 552,
"text": "Table 4",
"ref_id": null
},
{
"start": 609,
"end": 616,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "{",
"sec_num": null
},
{
"text": "We assume that named entities is effective for topic tracking, since both topic and event are related to a specific place and time in a story. We conducted an experiment using various types of named entities. The results are shown in Table 5 . Table 5 shows the tracking result using six words which is the output of the headline generation with some named entities. In Table 5 , 'Org', 'Per', 'Loc', 'Proper' denotes organization, person, location, and proper name, respectively. 'None' denotes the baseline, i.e. we use only the output of the headline generation, six words. Table 5 shows that the best result was when we use 'Org', 'Person', and 'Proper' with N t = 16, and the F-score is 0.717. When N t is larger than 8 positive training stories, the method which uses six title words with named entities consistently outperforms the baseline. When N t is smaller than 4 positive training stories, the result was improved when we add 'Per' and 'Proper' to the baseline. This indicates that these two named entities are especially effective for topic tracking.",
"cite_spans": [],
"ref_spans": [
{
"start": 234,
"end": 241,
"text": "Table 5",
"ref_id": null
},
{
"start": 244,
"end": 251,
"text": "Table 5",
"ref_id": null
},
{
"start": 370,
"end": 377,
"text": "Table 5",
"ref_id": null
},
{
"start": 577,
"end": 584,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Named Entities",
"sec_num": "5.4"
},
{
"text": "We recall that we used partitioning clustering algorithm, k-means (k = 2) to balance the amount of positive and negative training stories used per estimate. To examine the effect of hierarchical classification using k-means, we compare the result with and without a hierarchy. Table 6 shows the results using the same data, i.e. we use the output of headline generation, six words, and named entities, Person name, and Proper name.",
"cite_spans": [],
"ref_spans": [
{
"start": 277,
"end": 284,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Hierarchical Classification",
"sec_num": "5.5"
},
{
"text": "Overall, the result of 'with hierarchy' was better than that of 'without hierarchy' in all N t values. On the other hand, there are four topics/N t patterns whose results with hierarchical classification were worse than those of without a hierarchy. Table 7 shows the result. The F/A for all results with a hierarchy were lower than those without a hierarchy. One reason behind this lies iteration of a hierarchical classification, i.e. our algorithm involves iterating through the data that the system is permitted to classify during each iteration. As a result, there are a few negative training data in the bottom cluster, and the test stories were judged as NO but were evaluated as YES. We need to explore a method for determining the depth of the tree in the hierarchical classification, and this is a rich space for further investigation.",
"cite_spans": [],
"ref_spans": [
{
"start": 250,
"end": 257,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Hierarchical Classification",
"sec_num": "5.5"
},
{
"text": "The contribution of two linguistically motivated restrictions on the set of words is best explained by looking at other features. We thus compared our method with two baselines: -5% -5% -+45% +61% Headlines and NE -2% -2% -+2% +11% Original headlines -26% -16% -+22% +34%",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparative Experiments",
"sec_num": "5.6"
},
{
"text": "(1) all words in the stories as features, and (2) the original headlines in the stories as features 3 . Table 8 shows each result, when N t = 4. 'Stories' shows the result using all words in the stories and 'Original headlines' shows the result using the original headlines in the stories. 'Headlines and NE' denotes the best result obtained by our method, i.e. the output of headline generation, six words, and named entities, Person and Proper name. Table 8 shows that our method outperformed the other two methods, especially attained a better balance between recall and precision. Table 9 illustrates changes in pooled F1 measure as N t varies, with N t = 4 as the baseline. Table 9 shows that our method is the most stable all N t training instances before N t = 16, especially our method is effective even for a small number of positive training instances for per-source training: it learns a good topic representation and gains almost nothing in effectiveness beyond N t = 16.",
"cite_spans": [
{
"start": 100,
"end": 101,
"text": "3",
"ref_id": "BIBREF2"
}
],
"ref_spans": [
{
"start": 104,
"end": 111,
"text": "Table 8",
"ref_id": null
},
{
"start": 452,
"end": 459,
"text": "Table 8",
"ref_id": null
},
{
"start": 585,
"end": 592,
"text": "Table 9",
"ref_id": null
},
{
"start": 679,
"end": 686,
"text": "Table 9",
"ref_id": null
}
],
"eq_spans": [],
"section": "Comparative Experiments",
"sec_num": "5.6"
},
{
"text": "We have reported an approach for topic tracking on newspaper articles based on the two linguistic features, named entities and headlines. The result was 0.776 average precision and 0.481 recall, especially our method is effective even for a small number of positive training instances for per-source training in the tracking task. Future work includes (i) optimal decision of seed points for k-means clustering algorithm, (ii) exploring a method to determine the depth of the tree in the hierarchical classification, and (iii) applying the method to the TDT3 corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 10-21, 2005. c Springer-Verlag Berlin Heidelberg 2005",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "In the experiment, we set \u03b1, \u03b2, \u03b3 to 1.0, 1.0, 0.8, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "When the number of positive training stories(Nt) is 1, iterate step 1 and 2 until the depth of the tree in the hierarchy is identical to that of Nt=2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "In both cases, we used hierarchical classification to make our results comparable with these two results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Topic Detection and Tracking Pilot Study Final Report. Proc. of the DARPA Workshop",
"authors": [
{
"first": "J",
"middle": [],
"last": "Allan",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Carbonell",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Doddington",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Yamron",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Yang",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.Allan and J.Carbonell and G.Doddington and J.Yamron and Y.Yang: Topic Detection and Tracking Pilot Study Final Report. Proc. of the DARPA Workshop. (1997)",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Headline Generation Based on Statistical Translation. Proc. of ACL-2000",
"authors": [
{
"first": "M",
"middle": [],
"last": "Banko",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Mittal",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Witbrock",
"suffix": ""
}
],
"year": 2000,
"venue": "",
"volume": "",
"issue": "",
"pages": "318--325",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M.Banko and V.Mittal and M.Witbrock: Headline Generation Based on Statistical Transla- tion. Proc. of ACL-2000. (2000) 318-325",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "CMU Report on TDT-2: Segmentation, Detection and Tracking, Proc. of the DARPA Workshop",
"authors": [
{
"first": "J",
"middle": [],
"last": "Carbonell",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Lafferty",
"suffix": ""
},
{
"first": "R",
"middle": [
"D"
],
"last": "Brown",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Pierce",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.Carbonell and Y.Yang and J.Lafferty and R.D.Brown and T.Pierce and X.Liu: CMU Report on TDT-2: Segmentation, Detection and Tracking, Proc. of the DARPA Workshop, (1999)",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Scatter/Gather: a Cluster-based Approach to Browsing Large Document Collections . Proc. of ACM SIGIR-1992",
"authors": [
{
"first": "D",
"middle": [
"R"
],
"last": "Cutting",
"suffix": ""
},
{
"first": "D",
"middle": [
"R"
],
"last": "Karger",
"suffix": ""
},
{
"first": "L",
"middle": [
"O"
],
"last": "Pedersen",
"suffix": ""
},
{
"first": "J",
"middle": [
"W"
],
"last": "Tukey",
"suffix": ""
}
],
"year": 1992,
"venue": "",
"volume": "",
"issue": "",
"pages": "318--329",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D.R.Cutting, D.R.Karger and L.O.Pedersen and J.W.Tukey: Scatter/Gather: a Cluster-based Approach to Browsing Large Document Collections . Proc. of ACM SIGIR-1992. (1992) 318-329",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Topic Tracking for Radio, TV Broadcast, and Newswire. Proc. of the DARPA Broadcast News Transcription and Understanding Workshop",
"authors": [
{
"first": "H",
"middle": [],
"last": "Jin",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Schwartz",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Sista",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Walls",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H.Jin and R.Schwartz and S.Sista and F.Walls: Topic Tracking for Radio, TV Broadcast, and Newswire. Proc. of the DARPA Broadcast News Transcription and Understanding Workshop. (1999)",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Estimation of Probabilities from Sparse Data for the Language Model Component of a Speech Recognizer",
"authors": [
{
"first": "S",
"middle": [],
"last": "Katz",
"suffix": ""
}
],
"year": 1987,
"venue": "IEEE Transactions on Acoustics, Speech and Signal Processing",
"volume": "24",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S.Katz: Estimation of Probabilities from Sparse Data for the Language Model Component of a Speech Recognizer. IEEE Transactions on Acoustics, Speech and Signal Processing. 24 (1987)",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Fast Methods for Kernel-Based Text Analysis. Proc. of the ACL-2003",
"authors": [
{
"first": "T",
"middle": [],
"last": "Kudo",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "24--31",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T.Kudo and Y.Matsumoto: Fast Methods for Kernel-Based Text Analysis. Proc. of the ACL- 2003. (2003) 24-31",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "An Evaluation of Phrasal and Clustered Representations on a Text Categorization Task",
"authors": [
{
"first": "D",
"middle": [
"D"
],
"last": "Lewis",
"suffix": ""
}
],
"year": 1994,
"venue": "Proc. of the ACM SIGIR-1994",
"volume": "",
"issue": "",
"pages": "37--50",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D.D.Lewis: An Evaluation of Phrasal and Clustered Representations on a Text Categorization Task. Proc. of the ACM SIGIR-1994. (1994) 37-50",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Training Algorithms for Linear Text Classifiers",
"authors": [
{
"first": "D",
"middle": [
"D"
],
"last": "Lewis",
"suffix": ""
},
{
"first": "R",
"middle": [
"E"
],
"last": "Schapire",
"suffix": ""
},
{
"first": "J",
"middle": [
"P"
],
"last": "Callan",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Papka",
"suffix": ""
}
],
"year": 1996,
"venue": "Proc. of the ACM SIGIR-1996",
"volume": "",
"issue": "",
"pages": "298--306",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D.D.Lewis and R.E.Schapire and J.P.Callan and R.Papka: Training Algorithms for Linear Text Classifiers. Proc. of the ACM SIGIR-1996. (1996) 298-306",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The Beta-binomial Mixture Model and its Application to TDT Tracking and Detection",
"authors": [
{
"first": "S",
"middle": [
"A"
],
"last": "Lowe",
"suffix": ""
}
],
"year": 1999,
"venue": "Proc. of the DARPA Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S.A.Lowe: The Beta-binomial Mixture Model and its Application to TDT Tracking and De- tection. Proc. of the DARPA Workshop. (1999)",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Japanese Morphological Analysis System Chasen Mannual",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Matsumoto",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Kitauchi",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Yamashita",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Haruno",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Imaichi",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Imamura",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y.Matsumoto and A.Kitauchi and T.Yamashita and Y.Haruno and O.Imaichi and T.Imamura: Japanese Morphological Analysis System Chasen Mannual. NAIST Technical Report NAIST-IS-TR97007. (1997)",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Topic Tracking with the PRISE Information Retrieval System",
"authors": [
{
"first": "D",
"middle": [
"W"
],
"last": "Oard",
"suffix": ""
}
],
"year": 1999,
"venue": "Proc. of the DARPA Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D.W.Oard: Topic Tracking with the PRISE Information Retrieval System. Proc. of the DARPA Workshop. (1999)",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "UMASS Approaches to Detection and Tracking at TDT2. Proc. of the DARPA Workshop",
"authors": [
{
"first": "R",
"middle": [],
"last": "Papka",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Allan",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R.Papka and J.Allan: UMASS Approaches to Detection and Tracking at TDT2. Proc. of the DARPA Workshop. (1999)",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "BoosTexter: A Boosting-based System for Text Categorization",
"authors": [
{
"first": "R",
"middle": [
"E"
],
"last": "Schapire",
"suffix": ""
}
],
"year": 1999,
"venue": "Journal of Machine Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R.E.Schapire: BoosTexter: A Boosting-based System for Text Categorization. Journal of Ma- chine Learning. (1999)",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A Maximum Likelihood Model for Topic Classification of Broadcast News",
"authors": [
{
"first": "R",
"middle": [],
"last": "Schwartz",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Imai",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Nguyen",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Makhoul",
"suffix": ""
}
],
"year": 1996,
"venue": "Proc. of Eurospeech",
"volume": "",
"issue": "",
"pages": "270--278",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R.Schwartz and T.Imai and L.Nguyen and J.Makhoul: A Maximum Likelihood Model for Topic Classification of Broadcast News. Proc. of Eurospeech. (1996) 270-278",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "GE.Tracker: A Robust, Lightweight Topic Tracking System. Proc. of the DARPA Workshop",
"authors": [
{
"first": "T",
"middle": [],
"last": "Strzalkowski",
"suffix": ""
},
{
"first": "G",
"middle": [
"C"
],
"last": "Stein",
"suffix": ""
},
{
"first": "G",
"middle": [
"B"
],
"last": "Wise",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T.Strzalkowski and G.C.Stein and G.B.Wise: GE.Tracker: A Robust, Lightweight Topic Tracking System. Proc. of the DARPA Workshop. (1999)",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Topic Tracking in a News Stream. Proc. of the DARPA Broadcast News Transcription and Understanding Workshop",
"authors": [
{
"first": "Carp",
"middle": [],
"last": "Yamron",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yamron and Carp: Topic Tracking in a News Stream. Proc. of the DARPA Broadcast News Transcription and Understanding Workshop. (1999)",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Topic Tracking in a News Stream. Proc. of the DARPA Workshop",
"authors": [
{
"first": "J",
"middle": [
"P"
],
"last": "Yamron",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Carp",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Gillick",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Lowe",
"suffix": ""
},
{
"first": "P",
"middle": [
"V"
],
"last": "Mulbregt",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.P.Yamron and I.Carp and L.Gillick and S.Lowe and P.V.Mulbregt: Topic Tracking in a News Stream. Proc. of the DARPA Workshop. (1999)",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Expert Network: Effective and Efficient Learning from Human Decisions in Text Categorization and Retrieval",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Yang",
"suffix": ""
}
],
"year": 1994,
"venue": "Proc. of the ACM SIGIR-1994",
"volume": "",
"issue": "",
"pages": "13--22",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Yang: Expert Network: Effective and Efficient Learning from Human Decisions in Text Categorization and Retrieval. Proc. of the ACM SIGIR-1994. (1994) 13-22",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A Study on Retrospective and On-Line Event Detection",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Pierce",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Carbonell",
"suffix": ""
}
],
"year": 1998,
"venue": "Proc. of the ACM SIGIR-1998",
"volume": "",
"issue": "",
"pages": "28--36",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y.Yang and T.Pierce and J.Carbonell: A Study on Retrospective and On-Line Event Detection. Proc. of the ACM SIGIR-1998. (1998) 28-36",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Lattimer: Improving Text Categorization Methods for Event Tracking",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Ault",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Pierce",
"suffix": ""
},
{
"first": "C",
"middle": [
"W"
],
"last": "",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. of the ACM SIGIR-2000",
"volume": "",
"issue": "",
"pages": "65--72",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y.Yang and T.Ault and T.Pierce and C.W.Lattimer: Improving Text Categorization Methods for Event Tracking. Proc. of the ACM SIGIR-2000. (2000) 65-72",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "The Nature of Statistical Learning Theory",
"authors": [
{
"first": "V",
"middle": [],
"last": "Vapnik",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "V.Vapnik: The Nature of Statistical Learning Theory. Springer. (1995)",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Simple story with original headline and generated output Graphical representation of hierarchical classification"
},
"FIGREF3": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "The extracted headline for maximizing the probability arg max P(x)"
},
"FIGREF4": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "The extracted headline for various sequences"
},
"TABREF1": {
"type_str": "table",
"num": null,
"content": "<table><tr><td>Topic ID Topic name</td><td>Topic ID Topic name</td></tr><tr><td>30001 Cambodian government coalition</td><td>30003 Pinochet trial</td></tr><tr><td>30006 NBA labor disputes</td><td>30014 Nigerian gas line fire</td></tr><tr><td>30017 North Korean food shortages</td><td>30018 Tony Blair visits China in Oct.</td></tr><tr><td>30022 Chinese dissidents sentenced</td><td>30030 Taipei Mayoral elections</td></tr><tr><td colspan=\"2\">30031 Shuttle Endeavour mission for space station 30033 Euro Introduced</td></tr><tr><td>30034 Indonesia-East Timor conflict</td><td>30038 Olympic bribery scandal</td></tr><tr><td>30042 PanAm lockerbie bombing trial</td><td>30047 Space station module Zarya launched</td></tr><tr><td>30048 IMF bailout of Brazil</td><td>30049 North Korean nuclear facility?</td></tr><tr><td>30050 U.S. Mid-term elections</td><td>30053 Clinton's Gaza trip</td></tr><tr><td>30055 D'Alema's new Italian government</td><td>30057 India train derailment</td></tr></table>",
"text": "Topic Name The results Nt Prec. Rec. F Miss F/A Nt Prec. Rec. F Miss F/A 1 .000 .000 .000 1.000 .0000 8 .858 .432 .575 .568 .0001 2 .846 .040 .077 .960 .0000 16 .788 .520 .626 .480 .0004 4 .905 .142 .245 .858 .0000 Avg. .679 .227 .305 .663 .0001",
"html": null
},
"TABREF2": {
"type_str": "table",
"num": null,
"content": "<table/>",
"text": "The result using title words with high probabilities Nt Prec. Rec. F Miss F/A Nt Prec. Rec. F Miss F/A 1 .466 .375 .415 .626 .0005 8 .702 .372 .487 .628 .0003 2 .591 .402 .478 .599 .0003 16 .604 .393 .476 .607 .0007 4 .674 .340 .452 .660 .0003 Avg. .607 .376 .462 .624 .0004 The result using 3 title words Nt Prec. Rec. F Miss F/A Nt Prec. Rec. F Miss F/A 1 .608 .378 .465 .622 .0003 8 .687 .334 .453 .662 .0003 2 .652 .365 .466 .635 .0002 16 .734 .397 .516 .603 .0004 4 .709 .336 .456 .664 .0002 Avg. .678 .362 .471 .637 .0003",
"html": null
},
"TABREF3": {
"type_str": "table",
"num": null,
"content": "<table><tr><td>Named entities</td><td/><td colspan=\"3\">Nt [F-measure]</td><td colspan=\"2\">Avg. Named entities</td><td/><td>Nt [F-measure]</td><td>Avg.</td></tr><tr><td/><td>1</td><td>2</td><td>4</td><td>8 16</td><td/><td/><td>1</td><td>2</td><td>4</td><td>8 16</td></tr><tr><td colspan=\"6\">Org Per Loc Proper .138 .302 .377 .589 .673 .416</td><td>Per Loc</td><td colspan=\"2\">.237 .379 .453 .565 .647 .456</td></tr><tr><td>Org Per Loc</td><td colspan=\"8\">.138 .307 .498 .497 .517 .543 .629 .537</td></tr><tr><td>Org Per</td><td colspan=\"5\">.112 .178 .288 .579 .704 .372</td><td>Loc</td><td colspan=\"2\">.439 .459 .485 .561 .612 .511</td></tr><tr><td>Org Loc</td><td colspan=\"5\">.165 .350 .342 .594 .657 .422</td><td>Proper</td><td colspan=\"2\">.486 .473 .470 .453 .557 .488</td></tr><tr><td>Org Proper</td><td colspan=\"5\">.143 .229 .235 .548 .638 .359</td><td>None</td><td colspan=\"2\">.465 .466 .456 .453 .516 .471</td></tr></table>",
"text": "Combination of Named Entities .391 .586 .668 .418 Per Proper .437 .474 .542 .580 .671 .541 Org Per Loc .118 .187 .296 .590 .717 .382 Loc Proper .440 .461 .496 .647 .633 .535 Org Loc Proper .159 .342 .350 .607 .667 .471 Org .143 .205 .270 .561 .606 .357 Per Loc Proper .239 .397 .458 .574 .652 .464 Per The result with v.s. without hierarchical classification With hierarchy Without hierarchy Nt Prec. Rec. F Miss F/A Nt Prec. Rec. F Miss F/A 1 .695 .422 .525 .578 .0002 1 .669 .396 .498 .604 .0002 2 .707 .475 .568 .526 .0002 2 .671 .394 .497 .606 .0002 4 .835 .414 .554 .586 .0001 4 .747 .396 .517 .605 .0002 8 .823 .523 .639 .477 .0002 8 .709 .440 .543 .560 .0003 16 .819 .573 .674 .428 .0003 16 .818 .511 .629 .489 .0003 Avg. .776 .481 .592 .519 .0001 Avg. .723 .427 .537 .573 .0002 The result with a hierarchy was worse than that of without a hierarchy",
"html": null
},
"TABREF4": {
"type_str": "table",
"num": null,
"content": "<table><tr><td>Method</td><td>Nt</td><td/><td>Method</td><td>Nt</td><td/><td/></tr><tr><td>1</td><td>2 4</td><td>8</td><td>16</td><td>1 2 4</td><td>8</td><td>16</td></tr><tr><td>Stories</td><td/><td/><td/><td/><td/><td/></tr></table>",
"text": "Comparative experiment Method Prec. Rec. F Miss F/A Method Prec. Rec. F Miss F/A Stories .875 .026 .057 .974 .0000 Headlines and NE .835 .414 .554 .586 .0001 Original headlines .911 .190 .315 .810 .0000 Nt and F-measure",
"html": null
}
}
}
} |