File size: 91,294 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 |
{
"paper_id": "O08-3003",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:02:24.293896Z"
},
"title": "Cross-Lingual News Group Recommendation Using Cluster-Based Cross-Training",
"authors": [
{
"first": "Cheng-Zen",
"middle": [],
"last": "Yang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Yuan Ze University",
"location": {
"addrLine": "135 Yuan-Tung Rd",
"postCode": "320",
"settlement": "Chungli",
"country": "Taiwan"
}
},
"email": "czyang@syslab.cse.yzu.edu.tw"
},
{
"first": "Ing-Xiang",
"middle": [],
"last": "Chen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Yuan Ze University",
"location": {
"addrLine": "135 Yuan-Tung Rd",
"postCode": "320",
"settlement": "Chungli",
"country": "Taiwan"
}
},
"email": ""
},
{
"first": "Ping-Jung",
"middle": [],
"last": "Wu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Yuan Ze University",
"location": {
"addrLine": "135 Yuan-Tung Rd",
"postCode": "320",
"settlement": "Chungli",
"country": "Taiwan"
}
},
"email": "pjwu@syslab.cse.yzu.edu.tw"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Many Web news portals have provided clustered news categories for readers to browse many related news articles. However, to the best of our knowledge, they only provide monolingual services. For readers who want to find related news articles in different languages, the search process is very cumbersome. In this paper, we propose a cross-lingual news group recommendation framework using the cross-training technique to help readers find related cross-lingual news groups. The framework is studied with different implementations of SVM and Maximum Entropy models. We have conducted several experiments with news articles from Google News as the experimental data sets. From the experimental results, we find that the proposed cross-training framework can achieve accuracy improvement in most cases.",
"pdf_parse": {
"paper_id": "O08-3003",
"_pdf_hash": "",
"abstract": [
{
"text": "Many Web news portals have provided clustered news categories for readers to browse many related news articles. However, to the best of our knowledge, they only provide monolingual services. For readers who want to find related news articles in different languages, the search process is very cumbersome. In this paper, we propose a cross-lingual news group recommendation framework using the cross-training technique to help readers find related cross-lingual news groups. The framework is studied with different implementations of SVM and Maximum Entropy models. We have conducted several experiments with news articles from Google News as the experimental data sets. From the experimental results, we find that the proposed cross-training framework can achieve accuracy improvement in most cases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "As the Web becomes an abundant source of news information, it also becomes an important medium for people to learn recent tidings. To provide readers a convenient way of viewing a news event described by different news agencies, many Web news portals, such as AltaVista News and Google News, cluster news articles according to their relevance with consistent user interfaces. With such news clustering services, readers could easily acquire more details of an interesting news event from numerous reports. Ideally, they can simply click through an entry link to browse many related news reports without need of a cumbersome searching procedure. Nevertheless, if the news event is originally reported by foreign news agencies, the readers usually find that there are only few translated news articles and can only acquire an overview of the news event. If they want to find more related foreign news stories, they may generally get frustrated due to the following two reasons. First, the translated news articles seldom provide as much information as the original news articles. Second, the translation may add more interpretations that can mislead in the searching direction. The following example illustrates these situations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "This news story, reported in BBC News [2006] , is a good example to show these problems. The title of its English version is \"First impressions count for web\" and the article contains 15 paragraphs mainly focused on the impressions in a 20th of a second after first sight [BBC News 2006] . However, the title of its Chinese news story is \"\u597d\u7db2\u9801\u9084\u9700\u8981\u8b93\uf95a \u8005\u4e00\ufa0a\u937e\u60c5\" and may be translated into \"Good web pages need to let readers fall in love at first sight\", which includes additional semantic information related to love. In addition, the Chinese news article has only 7 paragraphs. When readers read the Chinese news article (the source document) and want to find more information from (for example) English news articles (target documents), they will most likely search for the news article entitled with \"fall in love at first sight\" and find nothing related. Apparently, the readers cannot easily find the English news article. Additionally, the amount of information of the source news article may not be equal to that of the corresponding target news article. In this example, the amount of information of the translated Chinese article is much less than that of the original English article. The scant amount of translated information will perplex the readers in other searching operations. These observations suggest the need of a cross-lingual news recommendation framework for readers to get a broader view to a news event.",
"cite_spans": [
{
"start": 29,
"end": 44,
"text": "BBC News [2006]",
"ref_id": "BIBREF18"
},
{
"start": 272,
"end": 287,
"text": "[BBC News 2006]",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "To address the recommendation issue for cross-lingual news groups, the simplest approach is to directly translate the source news article and find the related news group in another language. Unfortunately, the quality of translation and the amount of news information highly influence the recommendation results. Readers may get translated results of poor quality. For instance, using Google Translation (http://www.google.com/translate_t?hl=zh-TW) to translate the Chinese news title of the above example gets \"Readers need to make a good website was love at first sight\". As many Web news portals have provided monolingual cluster-based news browsing interfaces, the quality of cross-lingual news group recommendation can be improved if the cluster information of the source documents is exploited. Such exploration of cluster information has been studied recently in many applications, such as Web catalog integration [Agrawal and Srikan 2001; Tsay et al. 2003; Sarawagi et al. 2003; Zhang and Lee 2004a; Zhang and Lee 2004b; Chen 2005] and title generation [Tseng et al. 2006] .",
"cite_spans": [
{
"start": 921,
"end": 946,
"text": "[Agrawal and Srikan 2001;",
"ref_id": "BIBREF0"
},
{
"start": 947,
"end": 964,
"text": "Tsay et al. 2003;",
"ref_id": "BIBREF10"
},
{
"start": 965,
"end": 986,
"text": "Sarawagi et al. 2003;",
"ref_id": "BIBREF8"
},
{
"start": 987,
"end": 1007,
"text": "Zhang and Lee 2004a;",
"ref_id": "BIBREF15"
},
{
"start": 1008,
"end": 1028,
"text": "Zhang and Lee 2004b;",
"ref_id": "BIBREF16"
},
{
"start": 1029,
"end": 1039,
"text": "Chen 2005]",
"ref_id": "BIBREF3"
},
{
"start": 1061,
"end": 1080,
"text": "[Tseng et al. 2006]",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "In this paper, we propose a cross-lingual news group recommendation framework using the cross-training approach from recent Web taxonomy integration techniques [Sarawagi et al. 2003 ] to find the possible semantic corresponding relationships between news groups of Cluster-Based Cross-Training different languages. With the cross-training approach, the framework explores the implicit clustering information from the source news groups and the target news groups by learning the group features alternately. Then, the framework utilizes the implicit clustering information to improve the mapping accuracy between news groups of different languages.",
"cite_spans": [
{
"start": 160,
"end": 181,
"text": "[Sarawagi et al. 2003",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Such a framework has two major advantages. First, it will save considerable news searching effort resulting from the cumbersome searching procedure in which readers need to query different monolingual news portals in a trial-and-error manner. Second, it mitigates the translation inaccuracy to provide readers a broader panorama of news events from different aspects.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The cross-training framework has been implemented in Support Vector Machines (SVM) and Maximum Entropy (ME) classifiers. We have also conducted experiments to investigate the accuracy improvement of the cross-training approach with a 21-day data set containing English and Chinese news articles collected from Google News. In the experiments, we measured the accuracy performance for different approaches. The experimental results show that the cross-training approach can benefit the mapping accuracy in most cases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The rest of the paper is organized as follows. In Section 2, we present the problem definitions and briefly review previous related research on Web catalog integration. Section 3 elaborates the proposed cross-training framework. Section 4 describes our experiments in which English news and Chinese news articles from Google News were used as the data sets. Section 5 concludes the paper and discusses future directions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "For the recommendation problem of clustered news groups in different languages, we assume that the recommendation process deals with two Web news catalogs in two different languages to find the best semantically correlated relationships between the two news catalogs. We also assume that readers browse one news catalog and want to find related news articles in another news catalog of another language for the sake of simplicity. The catalog browsed by readers is the source S in which the news articles (source documents) are written in language L s and have been classified into m event clusters S 1 , S 2 ,\u2026, S m . The other is the target catalog T in which the news articles (target documents) are written in language L t and have been also classified into n clusters T 1 , T 2 ,\u2026, T n . The terms of the documents of each cluster comprise the feature space of the corresponding news event.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Statement and Related Research",
"sec_num": "2."
},
{
"text": "In the recommendation process, therefore, the objective of the framework is to discover all possible cluster-to-cluster mapping relationships between S and T, and report these relationships to the readers for recommendation. For the sake of simplicity in discussion, we only consider the best mapping relationships in this paper, i.e., given a source catalog S i , the best corresponding target catalog T j (S i \u2192T j ) is identified in this study. Ideally, if both news clusters S i and T j focus on the same news event, the news articles in both clusters should have semantic overlap, as shown in Figure 1 . Generally, the mapping relationships are one-to-one and symmetric. However, in our observations, one-to-many situations indeed have occurred because more than one target cluster is overlapped by the same source cluster. Furthermore, source documents will be translated in L t first, and the quality of the feature space of the translated source documents may be hindered due to the poor translation process. These factors may make the symmetric relationships asymmetric. Therefore, the reverse mappings (T j \u2192 S i ) are separately considered.",
"cite_spans": [],
"ref_spans": [
{
"start": 598,
"end": 606,
"text": "Figure 1",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Problem Statement and Related Research",
"sec_num": "2."
},
{
"text": "Generally, the cluster-to-cluster mapping discovery problem can be viewed as a generalization of the Web catalog integration problem on a coarse-grained basis. In the Web catalog integration problem, the objective of the integration process is to classify the documents in the source catalog into the target catalog with the enhancement of the implicit source information. In recent years, there have been many approaches proposed for the general catalog integration problem. For example, the Na\u00efve Bayes approaches [Agrawal and Srikan 2001; Tsay et al. 2003 ], the SVM-based approaches [Sarawagi et al. 2003; Zhang and Lee 2004a; Zhang and Lee 2004b; Chen 2005] , and the Maximum Entropy approach [Wu et al. 2005] have shown that the integration improvement can be effectively achieved. ",
"cite_spans": [
{
"start": 516,
"end": 541,
"text": "[Agrawal and Srikan 2001;",
"ref_id": "BIBREF0"
},
{
"start": 542,
"end": 558,
"text": "Tsay et al. 2003",
"ref_id": "BIBREF10"
},
{
"start": 587,
"end": 609,
"text": "[Sarawagi et al. 2003;",
"ref_id": "BIBREF8"
},
{
"start": 610,
"end": 630,
"text": "Zhang and Lee 2004a;",
"ref_id": "BIBREF15"
},
{
"start": 631,
"end": 651,
"text": "Zhang and Lee 2004b;",
"ref_id": "BIBREF16"
},
{
"start": 652,
"end": 662,
"text": "Chen 2005]",
"ref_id": "BIBREF3"
},
{
"start": 698,
"end": 714,
"text": "[Wu et al. 2005]",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Statement and Related Research",
"sec_num": "2."
},
{
"text": "Some enhancement approaches, however, may not be suitable for the cross-lingual cluster-to-cluster mapping discovery problem. For example, the topic restriction approach proposed in Tsay et al. [2003] requires that the testing target clusters are the clusters containing common documents from the source cluster. Nonetheless, in the cross-lingual cluster-to-cluster mapping discovery problem, there cannot be such a common subset. The enhanced Na\u00efve Bayes (ENB) approach proposed in [Agrawal and Srikan 2001] exploits the implicit source catalog information to enhance the integration accuracy performance. However, due to the diversity of news articles and the translation variety, the iterative algorithm may introduce many false-positive mappings to twist the overlapped space into a larger one. The shrinkage approach adopted in Wu et al. [2005] also needs to be adapted because the news clusters are usually not hierarchically organized.",
"cite_spans": [
{
"start": 182,
"end": 200,
"text": "Tsay et al. [2003]",
"ref_id": "BIBREF10"
},
{
"start": 483,
"end": 508,
"text": "[Agrawal and Srikan 2001]",
"ref_id": "BIBREF0"
},
{
"start": 833,
"end": 849,
"text": "Wu et al. [2005]",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Cluster-Based Cross-Training",
"sec_num": null
},
{
"text": "Our recommendation framework uses the cross-training approach adapted from the cross-training (CT) approach proposed in [Sarawagi et al. 2003 ]. The CT approach is a semi-supervised learning strategy. The idea behind CT is that a better classifier can be built with the assistance of another catalog that has semantic overlap. The overlapped document set is fully-labeled and partitioned into a development set and a test set where the development set is used to tune the system performance and the test set is used to evaluate the system. Through the cross-training process, the implicit information in the source taxonomy is learnt, and more source documents can be accurately integrated into the target taxonomy.",
"cite_spans": [
{
"start": 120,
"end": 141,
"text": "[Sarawagi et al. 2003",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Cluster-Based Cross-Training",
"sec_num": null
},
{
"text": "The proposed framework utilizes the CT approach to first obtain the potential mapping relationships from the reverse mappings (T j \u2192 S i ) through a learning process. The extracted information then is used to augment the feature space in the next learning phase. Finally, the mappings from S i to T j are explored in a classification process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Cluster-Based Cross-Training",
"sec_num": null
},
{
"text": "The main design principle of the cross-training framework is that the implicit mapping relationships are extracted through the first learning phase on reverse mappings. In this phase, the strength of each possible mapping is identified and ranked. For each S i , the framework can find the most possibly corresponding T j . Before the second learning phase, the feature space of each T j is expanded with the discovered mapping information. Then, the augmented classifiers are used to identify the mapping relationships from S i to T j , and give the recommendations. Figure 2 depicts the processing flow in the cross-training framework. Without loss of generality, we use English and Chinese here as two language representatives for L s and L t to explain our bilingual recommendation process in this paper.",
"cite_spans": [],
"ref_spans": [
{
"start": 568,
"end": 576,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Cross-Training for Mapping Discovery",
"sec_num": "3."
},
{
"text": "In the framework, the classification system first retrieves English and Chinese news articles from news portals, say Google News or Yahoo! News. These news articles have been usually clustered well in the news portals. The framework then performs parsing and preprocessing on each news cluster to get its feature space. The preprocessor parses the Web news, and eliminates stopwords [Fox 1992 ] and HTML tags. After the preprocessing, the source news groups are translated into the target language. For example, if a reader wants to find the possible English news groups for a designated Chinese news group, the English news articles are in the source news groups and will be translated into Chinese. After the translation process, all the source and target news groups are prepared as the data sets for further cross-training operations. ",
"cite_spans": [
{
"start": 383,
"end": 392,
"text": "[Fox 1992",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Processing Flow",
"sec_num": "3.1"
},
{
"text": "A debate may arise about whether the framework should re-cluster the news articles after the translation process. Since the translation process may introduce semantic variety into the news clusters, re-clustering the news articles may produce clusters with better semantic integrity for the following recommendation process. Nonetheless, the observations in show that the re-clustering process can contrarily reduce the quality of the original semantic integrity. Therefore, the proposed framework will not re-cluster the news articles.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Figure 2. The processing flow for bilingual news group recommendation in the cross-training framework.",
"sec_num": null
},
{
"text": "As each Web news article is composed of plain text and HTML tags, it needs to be parsed first to extract useful information. For simplicity sake, the document parsing procedure is currently designed in a conservative manner by ignoring the HTML tags and extracting only the plain text.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing and Preprocessing",
"sec_num": "3.2"
},
{
"text": "Both Chinese and English news articles are then preprocessed. There are four steps for English news articles: (1) tokenization, (2) stopword removal, (3) stemming, and (4) generation of term-frequency vectors. As there is no word boundary in Chinese sentences, the Chinese articles need to be segmented first [Nie and Ren 1999; Nie et al. 2000; Foo and Li 2004] . We use a hybrid approach proposed by Tseng [2002] , which can achieve a high precision rate and a considerably good recall rate by considering unknown words. The hybrid approach combines the longest match dictionary-based segmentation method and a statistical-based approach which is a fast keyword/key-phrase extraction algorithm. With this hybrid approach, each sentence is scanned sequentially and the longest matched words based on the dictionary entries are extracted. This process is repeated until all characters are scanned.",
"cite_spans": [
{
"start": 309,
"end": 327,
"text": "[Nie and Ren 1999;",
"ref_id": "BIBREF6"
},
{
"start": 328,
"end": 344,
"text": "Nie et al. 2000;",
"ref_id": "BIBREF7"
},
{
"start": 345,
"end": 361,
"text": "Foo and Li 2004]",
"ref_id": "BIBREF4"
},
{
"start": 401,
"end": 413,
"text": "Tseng [2002]",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing and Preprocessing",
"sec_num": "3.2"
},
{
"text": "After preprocessing, the Chinese and English news articles in each category are tokenized. Then, the Chinese news documents are translated. The translation can be based on a bilingual dictionary or a well-trained machine translation system. In the translation, we adopt a straightforward word expansion method. Each Chinese word is simply translated to a set of English terms listed in a bilingual dictionary or derived from a machine translation system. The same procedure is also applied to the English news articles. Currently, the translation process does not consider the word choice disambiguation problem when there are several candidates for each word. The translation quality is not further addressed using different translation technologies. Nonetheless, it can be found that the proposed cross-training approach achieves around 90% accuracy performance in top-1 ranking.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Translation and Transformation",
"sec_num": "3.3"
},
{
"text": "Finally, each news article is converted to a feature vector. For each index term in the feature vector, a weight is associated with the term to express the importance. In the current design, the weight of each term is calculated by",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Translation and Transformation",
"sec_num": "3.3"
},
{
"text": "x i TF TF \u2211",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Translation and Transformation",
"sec_num": "3.3"
},
{
"text": ", where i denotes the number of the stemmed terms in each news article.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Translation and Transformation",
"sec_num": "3.3"
},
{
"text": "Previous studies on the general Web catalog integration problem show that, if a source document can be integrated into a target category, there must be a sufficiently large semantic overlap between them [Agrawal and Srikan 2001; Sarawagi et al. 2003; Tsay et al. 2003; Zhang and Lee 2004a; Zhang and Lee 2004b; Wu et al. 2005; Yang 2006] . For the cluster-to-cluster mapping discovery problem, this observation is also an important basis. If an English news category can be associated with a Chinese news category, this mapping must be concluded from a situation in which the semantically overlapped feature space is sufficiently large.",
"cite_spans": [
{
"start": 203,
"end": 228,
"text": "[Agrawal and Srikan 2001;",
"ref_id": "BIBREF0"
},
{
"start": 229,
"end": 250,
"text": "Sarawagi et al. 2003;",
"ref_id": "BIBREF8"
},
{
"start": 251,
"end": 268,
"text": "Tsay et al. 2003;",
"ref_id": "BIBREF10"
},
{
"start": 269,
"end": 289,
"text": "Zhang and Lee 2004a;",
"ref_id": "BIBREF15"
},
{
"start": 290,
"end": 310,
"text": "Zhang and Lee 2004b;",
"ref_id": "BIBREF16"
},
{
"start": 311,
"end": 326,
"text": "Wu et al. 2005;",
"ref_id": "BIBREF13"
},
{
"start": 327,
"end": 337,
"text": "Yang 2006]",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Cross-Training Process",
"sec_num": "3.4"
},
{
"text": "The cross-training process is incorporated mainly for exploring the overlapped feature space. Figure 3 illustrates a cross-training process in which there are two learning phases. In the first phase, the source news clusters are used as the training data sets to train m classifiers, and the target news clusters are used as the testing data sets to extract the implicit mapping information. The m classifiers then calculate the mapping scores (Sc ij ) for n target news clusters to predict the strengths of the semantic overlaps.",
"cite_spans": [],
"ref_spans": [
{
"start": 94,
"end": 102,
"text": "Figure 3",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Learning to Extract the Implicit Information",
"sec_num": "3.4.1"
},
{
"text": "Since SVM and ME are studied in the framework implementations, the mapping score Sc ij of T j \u2192S i can be defined as either the ratio at which the target documents in T j are classified into the source news cluster S i or the average weight derived from the classifier. For example, if the classification scheme used in the framework is SVM, the mapping score Sc ij can be calculated by either Eq. 1where N Tj is the news documents of the target cluster T j or Eq. (2) which is the average of the distance from each document to the hyperplane. This average can be viewed as the discriminative characteristic of all documents to the classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning to Extract the Implicit Information",
"sec_num": "3.4.1"
},
{
"text": "# of classified in #of Tj i ij Tj N S Sc N = (1) # of i i ij Tj w x b Sc N + = \u2211 (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning to Extract the Implicit Information",
"sec_num": "3.4.1"
},
{
"text": "Basically, Equation (1) represents a voting scheme in which the predicted rank of a target cluster T j depends on the number of the positively classified news articles in T j . Equation 2represents a weighting scheme in which the predicted rank of T j depends on the average of the total distance to the hyperplane. For each source cluster, the target cluster with the highest mapping score is qualified as the potential candidate that may have the accurate S i \u2192T j mapping relationship in the second learning phase. The reason the mapping scores are considered in an asymmetric way is that the cross-training approach will adjust the feature vectors back and forth in each learning iteration. Other mapping discovery approaches may provide efficient schemes to consider both mapping scores of S i \u2192T j and T j \u2192 S i as an integrated scoring method. This has been left for our future study.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning to Extract the Implicit Information",
"sec_num": "3.4.1"
},
{
"text": "The implicit information explored in the first learning phase is then used as the prediction information in the second learning phase. The cross training process can be continued until the results converge. The category information of the corresponding source cluster, say S i , for the previously discovered candidate target cluster, say T j , is inserted into the feature space of T j . The category information can be category identifiers or the category title words. For example, we used identifiers starting from 1000001 to 1000040 for categories in the current experiments. Figure 4 depicts the detailed process of concatenating the predicted implicit information to the ordinary feature vectors of the target cluster in the cross-training approach. In the figure, F T is a feature vector for the term features of the target news articles, L T is a feature vector for the label features (category information) of the target cluster, and the test output contains the label features of the predicted source clusters. With the predicted mapping information, the discriminative power of the classifiers of the second phase can be enhanced.",
"cite_spans": [],
"ref_spans": [
{
"start": 580,
"end": 588,
"text": "Figure 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Learning to Find the Corresponding Mappings",
"sec_num": "3.4.2"
},
{
"text": "For controlling the discriminative power of the added semantically-overlapped implicit label information, as in Sarawagi et al. [2003] , the ordinary feature weights in the augmented target vectors are scaled by a factor of f, and the weight of each label attribute by a factor of 1 \u2212 f. The parameter f is used to decide the relative weights of the label and term features and can be tuned for different application environments. In the current experiments, the results show that the best f value ranges from 0.02 to 0.05. The small f values show that the augmented information should not be overemphasized in the cross-training process. This observation for factoring is consistent with previous studies [Sarawagi et al. 2003; Chen et al. 2004] .",
"cite_spans": [
{
"start": 112,
"end": 134,
"text": "Sarawagi et al. [2003]",
"ref_id": "BIBREF8"
},
{
"start": 706,
"end": 728,
"text": "[Sarawagi et al. 2003;",
"ref_id": "BIBREF8"
},
{
"start": 729,
"end": 746,
"text": "Chen et al. 2004]",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Learning to Find the Corresponding Mappings",
"sec_num": "3.4.2"
},
{
"text": "Finally, the second-phase classifiers are trained with the augmented target vectors. The recommended source news groups of the target news groups are calculated using the same mapping scoring method. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning to Find the Corresponding Mappings",
"sec_num": "3.4.2"
},
{
"text": "We have implemented the cross-training framework in SVM and ME classifiers. To rank the predictive corresponding target clusters, we implemented the voting scheme in the cross-training framework of SVM (SVM-VCT) and ME (ME-VCT), and the weighting scheme with SVM (SVM-WCT). As stated in Section 3.4.1, Equation (1) was used to rank the target clusters in SVM-VCT and ME-VCT. Equation 2was used in the weighting scheme SVM-WCT. We also implemented the voting scheme and the weighting scheme in SVM (SVM-V and SVM-W) for comparison. In the experiments, an English news catalog and a Chinese news catalog from Google News were used as the representatives to demonstrate the classification performance of the proposed cross-training framework. We measured the accuracy performance at top-1, top-3, and top-5 ranks. The details of the experiments are presented as follows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4."
},
{
"text": "The framework is currently implemented in Java. The segmentation corpus is based on the Academia Sinica Bilingual Wordnet 1.0 published by the Association for Computational Linguistics and Chinese Language Processing (ACLCLP) [Sinica BOW 2005] . We used SVM light (version 5.00) [Joachims 2002 ] as the SVM tool with a linear kernel, and the maximum-entropy toolkit (version 20041229) [Zhang 2004] as the Maximum Entropy model kernel.",
"cite_spans": [
{
"start": 226,
"end": 243,
"text": "[Sinica BOW 2005]",
"ref_id": "BIBREF9"
},
{
"start": 279,
"end": 293,
"text": "[Joachims 2002",
"ref_id": null
},
{
"start": 385,
"end": 397,
"text": "[Zhang 2004]",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Experimental Environment",
"sec_num": "4.1"
},
{
"text": "The bilingual word lists published by Linguistic Data Consortium (LDC) were used as the bilingual dictionaries. The Chinese-to-English dictionary ver. 2 (ldc2ce) has about 120,000 records, and the English-to-Chinese dictionary (ldc2ec) has about 110,000 records. In ldc2ce and ldc2ec, each entry is composed of a single word and several translated words separated by slashes without any indication of the importance. Therefore, the translated words are treated equally in our experiments. In the translation, each word in the source document was replaced with these translated words. The translation quality issue is not addressed in depth because we want to follow the normal reader behaviors. Furthermore, the implicit semantic information embedded in each category of news articles may mitigate the poorness of the translation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Experimental Environment",
"sec_num": "4.1"
},
{
"text": "In our experiments, two news portals were chosen as the bilingual news sources: Google News U.S. version for English news and Google News Taiwan version for Chinese news. Both the Chinese and English news articles were retrieved from the world news category from May 10 to May 23, 2005 and from October 21 to October 27, 2007. The experiments were performed on the data set of each day. Twenty news categories were collected per day. All the English news articles were translated into Chinese with the bilingual dictionaries. The size of the English-to-Chinese data set is 454.5 Mbytes. All the Chinese news articles were also translated. The size of the Chinese-to-English data set is 341.5 Mbytes. The 21-day data sets contain 36,548 English news articles and 8,224 Chinese news articles.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Sets",
"sec_num": "4.2"
},
{
"text": "In the experiments, the mapping relations between the Chinese and English news reports were first identified by three graduate students manually and independently. The mapping between an English news category and a Chinese news category is recognized if at least two students have the same mapping identification. These manually-identified mapping relations were used to evaluate the accuracy performance of the bilingual classification systems. We found that there were 122 identified mappings in the Chinese-to-English recommendation task and 123 identified mappings in the English-to-Chinese recommendation task. The difference existed because an English category was identified that was to be mapped to two Chinese categories. The data sets collected currently cannot significantly reveal the influences of one-to-many situations. In our future work plan, more news categories need to be collected to verify our scheme for one-to-many cases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Sets",
"sec_num": "4.2"
},
{
"text": "The experiments were conducted in two ways: finding the related Chinese news groups from the English news groups (Chinese-to-English) and finding the related English news groups from the Chinese news groups (English-to-Chinese). Here, we take the Chinese-to-English recommendation process as the example to present the experimental details. The English-to-Chinese recommendation process was conducted in a similar manner.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Sets",
"sec_num": "4.2"
},
{
"text": "In the Chinese-to-English experiments, each Chinese news catalog was first used as the training set in the first learning phase. To find a corresponding Chinese category (S i ) of an English target category (T j ), the news articles in S i were all used as the positive training examples, and the news articles in the other Chinese news categories (S k , k\u2260i) were randomly selected as the negative training examples. Then, all mapping scores between English categories and Chinese categories were measured based on the first-phase classification results. The English category with the highest mapping score was considered as the possibly mapped category.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Sets",
"sec_num": "4.2"
},
{
"text": "In the second learning phase of the Chinese-to-English experiments, the category information of the previously identified English cluster was concatenated to the corresponding Chinese cluster. Then, the English categories were used as the training set to train the second-phase classifiers. The augmented source Chinese categories were classified to calculate the mapping scores for each English news category. Finally, we measured the accuracy performance for each day using the correct mappings at the top-1, top-3, and top-5 recommendation ranks by the following equation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Sets",
"sec_num": "4.2"
},
{
"text": "Total number of the correct mapping in",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Number of the correctly discovered mapping in Accuracy",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "S T S T \u2192 = \u2192 ,",
"eq_num": "(3)"
}
],
"section": "Number of the correctly discovered mapping in Accuracy",
"sec_num": null
},
{
"text": "which is similar to Agrawal and Srikan [2001] . Accuracy, rather than precision or recall, is used because the recommendation process is performed on a cluster-to-cluster basis. The error rate is the complement of the accuracy. In the English-to-Chinese experiments, the roles of two catalogs were switched. 6 6 2 3 6 6 3 3 7 3 1 3 1 0 1 3 8 6 2 5 6 3 2 6 9 5 1 5 3 4 1 5 10 5 1 5 4 4 2 3 11 4 0 4 2 2 2 3 12 7 3 7 4 5 2 7 13 4 3 4 4 3 4 Table 1 lists the experimental results of the correctly discovered Chinese-to-English mappings at the top-1 recommendation lists identified by different approaches. Table 2 lists the correctly discovered Chinese-to-English mappings at the top-3 recommendation lists. From these tables, we can notice that the cross-training approach significantly improves the voting approaches in SVM-V and ME to find correct mappings in the top-1 recommendation results. In addition, it improves SVM-V, SVM-W, and ME entirely to find correct mappings in the top-3 recommendation results. Here, the scaling factor f is 0.05. When f ranged from 0.02 to 0.05, we attained similar results. 6 6 3 3 6 6 4 4 7 3 1 3 3 3 1 3 8 6 2 5 6 6 3 6 9 5 2 5 4 5 3 5 10 5 1 5 5 5 3 3 11 4 1 4 3 3 3 3 12 7 3 7 7 7 5 7 13 4 3 4 4 4 4 Table 1 , it is noticeable that SVM-W outperformed SVM-WCT. The reason the cross-training approach cannot benefit the accuracy performance is because adding more features changes the characteristics of the hyperplanes learned by SVM, thereby affecting the distance summation results in Eq. (2). Therefore, some correct mappings were ranked at the second rank in the recommendation lists for SVM-WCT but at the top rank for SVM-W. For the top-3 recommendation lists as shown in Table 2 , SVM-WCT outperformed SVM-W and got the best accuracy performance. 7 3 0 2 2 2 3 2 8 6 1 5 6 3 5 5 9 5 1 3 4 3 3 4 10 5 1 4 5 4 2 5 11 4 1 4 2 1 1 2 12 7 2 7 6 5 5 6 13 4 0 3 3 3 3 Table 3 and Table 4 list the experimental results of the correct English-to-Chinese mappings in the top-1 and top-3 recommendation lists, respectively. From these two tables, we can see that the cross-training approach significantly improved SVM-V and ME in finding the correct mappings in the top-1 recommendation results. From the top-3 recommendation results, we can observe that SVM-V is highly improved by the cross-training approach. SVM-W and SVM-WCT has the same results and both achieve the best performance. Although the cross-training approach cannot benefit ME more in the top-3 results as in the Chinese-to-English experiments, the performance of ME-VCT is comparable to ME. Table 5 lists the experimental results of the correct Chinese-to-English and English-to-Chinese mappings in the top-5 recommendation lists. Here, we omit the details of the correct mappings of each day and only show the total results. The top-5 results are very similar to the top-3 results.",
"cite_spans": [
{
"start": 20,
"end": 45,
"text": "Agrawal and Srikan [2001]",
"ref_id": "BIBREF0"
}
],
"ref_spans": [
{
"start": 308,
"end": 506,
"text": "6 6 2 3 6 6 3 3 7 3 1 3 1 0 1 3 8 6 2 5 6 3 2 6 9 5 1 5 3 4 1 5 10 5 1 5 4 4 2 3 11 4 0 4 2 2 2 3 12 7 3 7 4 5 2 7 13 4 3 4 4 3 4",
"ref_id": "TABREF3"
},
{
"start": 507,
"end": 514,
"text": "Table 1",
"ref_id": "TABREF3"
},
{
"start": 672,
"end": 679,
"text": "Table 2",
"ref_id": "TABREF5"
},
{
"start": 1178,
"end": 1376,
"text": "6 6 3 3 6 6 4 4 7 3 1 3 3 3 1 3 8 6 2 5 6 6 3 6 9 5 2 5 4 5 3 5 10 5 1 5 5 5 3 3 11 4 1 4 3 3 3 3 12 7 3 7 7 7 5 7 13 4 3 4 4 4 4",
"ref_id": "TABREF3"
},
{
"start": 1377,
"end": 1384,
"text": "Table 1",
"ref_id": "TABREF3"
},
{
"start": 1854,
"end": 1861,
"text": "Table 2",
"ref_id": "TABREF5"
},
{
"start": 1930,
"end": 2103,
"text": "7 3 0 2 2 2 3 2 8 6 1 5 6 3 5 5 9 5 1 3 4 3 3 4 10 5 1 4 5 4 2 5 11 4 1 4 2 1 1 2 12 7 2 7 6 5 5 6 13 4 0 3 3 3 3",
"ref_id": "TABREF3"
},
{
"start": 2104,
"end": 2111,
"text": "Table 3",
"ref_id": "TABREF7"
},
{
"start": 2116,
"end": 2123,
"text": "Table 4",
"ref_id": "TABREF9"
},
{
"start": 2792,
"end": 2799,
"text": "Table 5",
"ref_id": "TABREF10"
}
],
"eq_spans": [],
"section": "Number of the correctly discovered mapping in Accuracy",
"sec_num": null
},
{
"text": "Other improvements can still be introduced in the recommendation framework. For example, unknown name entity recognition (NER) and transliteration processing are two important issues for cross-lingual processing. Improvements to the quality of machine translation in the framework should further enhance the accuracy performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Discussion",
"sec_num": "4.3"
},
{
"text": "As the amount of news information explosively grows over the Internet, on-line Web news services have played an important role in delivering news information to people. Although these Web news portals have provided readers with clustered monolingual news services, cross-lingual news clustering services are still in great demand.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5."
},
{
"text": "In this paper, we propose a cross-lingual news group recommendation framework with the cross-training approach to get high accuracy performance in finding the mapping relationships between two news catalogs in different languages. From the experimental results, we can find that the proposed cross-training recommendation framework comprehensively has the superior accuracy performance. Among all approaches, SVM-WCT can achieve the best accuracy in the top-3 and top-5 recommendation lists for both Chinese-to-English and English-to-Chinese. There are still many research issues left for our future study. For example, feature weighting plays an important role in system performance. Meaningful features should be explored and employed for integration. In addition, we only consider the accuracy rate of correct mappings in current experiments. The correct rejection rate needs to be further studied for independent source/target categories. Furthermore, the scoring method can be discussed to find whether there are other better approaches to discover the correct mapping. In addition, a filtering scheme needs be discussed to screen out incorrect mapping recommendations (negative mappings) for practical use. One of the most challenging issues is how to translate new words which are created daily due to the rapidly changing Web. A better automatic bilingual translation system is needed to fulfill the requirements of effective term translation for the NER problem and the transliteration problem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5."
}
],
"back_matter": [
{
"text": "The authors would like to thank the National Science Council of the Republic of China, Taiwan, for partially supporting this research under Contract No. NSC 95-2745-E-155-008. The authors would also like to express many thanks to the anonymous reviewers for their precious suggestions for this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "On Integrating Catalogs",
"authors": [
{
"first": "R",
"middle": [],
"last": "Agrawal",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Srikan",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of the 10 th International Conference on World Wide Web",
"volume": "",
"issue": "",
"pages": "603--612",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Agrawal, R. and R. Srikan, \"On Integrating Catalogs,\" in Proceedings of the 10 th International Conference on World Wide Web, 2001, pp. 603-612.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Clustering and Visualization in a Multi-lingual Multidocument Summarization System",
"authors": [
{
"first": "H.-H",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "J.-J",
"middle": [],
"last": "Kuo",
"suffix": ""
},
{
"first": "T.-C",
"middle": [],
"last": "Su",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of 25 th European Conference on Information Retrieval Research",
"volume": "",
"issue": "",
"pages": "266--280",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chen, H.-H., J.-J. Kuo, and T.-C. Su, \"Clustering and Visualization in a Multi-lingual Multidocument Summarization System,\" in Proceedings of 25 th European Conference on Information Retrieval Research, 2003, pp. 266-280.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Web Catalog Integration using Support Vector Machines",
"authors": [
{
"first": "I.-X",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "C.-H",
"middle": [],
"last": "Shih",
"suffix": ""
},
{
"first": "C.-Z",
"middle": [],
"last": "Yang",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 1 st Workshop on Intelligent Web Technology (IWT 2004)",
"volume": "",
"issue": "",
"pages": "7--13",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chen, I.-X., C.-H. Shih, and C.-Z. Yang, \"Web Catalog Integration using Support Vector Machines,\" in Proceedings of the 1 st Workshop on Intelligent Web Technology (IWT 2004), Taipei, Taiwan, 2004, pp. 7-13.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "An Iterative Approach for Web Catalog Integration with Support Vector Machines",
"authors": [
{
"first": "I.-X",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "J.-C",
"middle": [],
"last": "Ho",
"suffix": ""
},
{
"first": "C.-Z",
"middle": [],
"last": "Yang",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of 2 nd Asia Information Retrieval Symposium (AIRS 2005)",
"volume": "",
"issue": "",
"pages": "703--708",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chen, I.-X., J.-C. Ho, and C.-Z. Yang, \"An Iterative Approach for Web Catalog Integration with Support Vector Machines,\" in Proceedings of 2 nd Asia Information Retrieval Symposium (AIRS 2005), 2005, pp. 703-708.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Chinese Word Segmentation and Its Effect on Information Retrieval",
"authors": [
{
"first": "S",
"middle": [],
"last": "Foo",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2004,
"venue": "Information Processing and Management",
"volume": "40",
"issue": "1",
"pages": "161--190",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Foo, S. and H. Li, \"Chinese Word Segmentation and Its Effect on Information Retrieval,\" Information Processing and Management, 40(1), 2004, pp. 161-190.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Lexical Analysis and Stop Lists",
"authors": [
{
"first": "C",
"middle": [],
"last": "Fox",
"suffix": ""
}
],
"year": 1992,
"venue": "Information Retrieval: Data Structures and Algorithms",
"volume": "",
"issue": "",
"pages": "102--130",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fox, C., \"Lexical Analysis and Stop Lists\", Information Retrieval: Data Structures and Algorithms, Chapter 7, Frakes, W. and Baeza-Yates, R., (eds.), Prentice-Hall, 1992, pp. 102-130.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Chinese Information Retrieval: Using Characters or Words",
"authors": [
{
"first": "J",
"middle": [
"Y"
],
"last": "Nie",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Ren",
"suffix": ""
}
],
"year": 1999,
"venue": "Information Processing and Management",
"volume": "35",
"issue": "4",
"pages": "443--162",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nie, J.Y. and F. Ren, \"Chinese Information Retrieval: Using Characters or Words,\" Information Processing and Management, 35(4), 1999, pp. 443-162. Cluster-Based Cross-Training",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "On the Use of Words and N-grams for Chinese Information Retrieval",
"authors": [
{
"first": "J",
"middle": [
"Y"
],
"last": "Nie",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Zhou",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 5 th International Workshop on Information Retrieval with Asian Languages",
"volume": "",
"issue": "",
"pages": "141--148",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nie, J.Y., J. Gao, J. Zhang, and M. Zhou, \"On the Use of Words and N-grams for Chinese Information Retrieval,\" in Proceedings of the 5 th International Workshop on Information Retrieval with Asian Languages, 2000, pp. 141-148.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Cross-training: Learning Probabilistic Mappings between Topics",
"authors": [
{
"first": "S",
"middle": [],
"last": "Sarawagi",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Chakrabarti",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Godbole",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 9 th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining",
"volume": "",
"issue": "",
"pages": "177--186",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sarawagi, S., S. Chakrabarti, and S. Godbole, \"Cross-training: Learning Probabilistic Mappings between Topics,\" in Proceedings of the 9 th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2003, pp. 177-186.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The Academia Sinica Bilingual Wordnet. Ver. 1.0, The Association for Computational Linguistics and Chinese Language Processing",
"authors": [
{
"first": "Bow",
"middle": [],
"last": "Sinica",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sinica BOW, The Academia Sinica Bilingual Wordnet. Ver. 1.0, The Association for Computational Linguistics and Chinese Language Processing, 2005.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Enhancing Techniques for Efficient Topic Hierarchy Integration",
"authors": [
{
"first": "J.-J",
"middle": [],
"last": "Tsay",
"suffix": ""
},
{
"first": "H.-Y",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "C.-F",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "C.-H",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 3 rd International Conference on Data Mining (ICDM'03)",
"volume": "",
"issue": "",
"pages": "657--660",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tsay, J.-J., H.-Y. Chen, C.-F. Chang, and C.-H. Lin, \"Enhancing Techniques for Efficient Topic Hierarchy Integration,\" in Proceedings of the 3 rd International Conference on Data Mining (ICDM'03), 2003, pp. 657-660.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Automatic Thesaurus Generation for Chinese Documents",
"authors": [
{
"first": "Y.-H",
"middle": [],
"last": "Tseng",
"suffix": ""
}
],
"year": 2002,
"venue": "Journal of the American Society for Information Science and Technology",
"volume": "53",
"issue": "13",
"pages": "1130--1138",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tseng, Y.-H., \"Automatic Thesaurus Generation for Chinese Documents,\" Journal of the American Society for Information Science and Technology, 53(13), 2002, pp. 1130-1138.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Toward Generic Title Generation for Clustered Documents",
"authors": [
{
"first": "Y.-H",
"middle": [],
"last": "Tseng",
"suffix": ""
},
{
"first": "C.-J",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "H.-H",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Y.-I",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 3 rd Asia Information Retrieval Symposium (AIRS 2006)",
"volume": "",
"issue": "",
"pages": "145--157",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tseng, Y.-H., C.-J. Lin, H.-H. Chen, and Y.-I Lin, \"Toward Generic Title Generation for Clustered Documents,\" in Proceedings of the 3 rd Asia Information Retrieval Symposium (AIRS 2006), 2006, Singapore, pp. 145-157.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Learning to Integrate Web Taxonomies with Fine-Grained Relations: A Case Study Using Maximum Entropy Model",
"authors": [
{
"first": "C",
"middle": [
"W"
],
"last": "Wu",
"suffix": ""
},
{
"first": "T",
"middle": [
"H"
],
"last": "Tsai",
"suffix": ""
},
{
"first": "W",
"middle": [
"L"
],
"last": "Hsu",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of 2 nd Asia Information Retrieval Symposium (AIRS 2005)",
"volume": "",
"issue": "",
"pages": "190--205",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wu, C. W., T. H. Tsai, and W. L. Hsu, \"Learning to Integrate Web Taxonomies with Fine-Grained Relations: A Case Study Using Maximum Entropy Model,\" in Proceedings of 2 nd Asia Information Retrieval Symposium (AIRS 2005), 2005, pp. 190-205.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A Cross-Lingual Framework for Web News Taxonomy Integration",
"authors": [
{
"first": "C.-Z",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "C.-M",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "I.-X",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 3 rd Asia Information Retrieval Symposium (AIRS 2006)",
"volume": "",
"issue": "",
"pages": "270--283",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yang, C.-Z., C.-M. Chen, and I.-X. Chen, \"A Cross-Lingual Framework for Web News Taxonomy Integration,\" in Proceedings of the 3 rd Asia Information Retrieval Symposium (AIRS 2006), 2006, Singapore, pp. 270-283.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Web Taxonomy Integration using Support Vector Machines",
"authors": [
{
"first": "D",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "W",
"middle": [
"S"
],
"last": "Lee",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 13 th International Conference on World Wide Web",
"volume": "",
"issue": "",
"pages": "472--481",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhang, D. and W. S. Lee, \"Web Taxonomy Integration using Support Vector Machines,\" in Proceedings of the 13 th International Conference on World Wide Web, 2004a, pp. 472-481.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Web Taxonomy Integration through Co-Bootstrapping",
"authors": [
{
"first": "D",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "W",
"middle": [
"S"
],
"last": "Lee",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 27 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval",
"volume": "",
"issue": "",
"pages": "410--417",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhang, D. and W. S. Lee, \"Web Taxonomy Integration through Co-Bootstrapping,\" in Proceedings of the 27 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 2004b, pp. 410-417.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "First impressions count for web",
"authors": [
{
"first": "",
"middle": [],
"last": "Bbc News",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "BBC News, \"First impressions count for web.\" English version is available at http://news.bbc.co.uk/2/hi/technology/4616700.stm; Chinese version is available at http://news.bbc.co.uk/chinese/trad/hi/newsid_4610000/newsid_4618500/4618552.stm, 2006.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Google Translation",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Google Translation, http://www.google.com/translate_t?hl=zh-TW.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Linguistic Data Consortium",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Linguistic Data Consortium, http://projects.ldc.upenn.edu/Chinese/LDCch.htm.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "The Maximum Entropy model toolkit",
"authors": [
{
"first": "L",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhang, L., The Maximum Entropy model toolkit, version 20041229, http://homepag es.inf.ed.ac.uk/s0450736/maxent_toolkit.html, 2004.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"text": "The relation of the news event and the correspondent news clusters in L s and L t .",
"num": null,
"uris": null,
"type_str": "figure"
},
"FIGREF2": {
"text": "The basic concept of the cross-training process.",
"num": null,
"uris": null,
"type_str": "figure"
},
"FIGREF3": {
"text": "Adding the predicted implicit information in the cross-training process.",
"num": null,
"uris": null,
"type_str": "figure"
},
"TABREF3": {
"html": null,
"num": null,
"text": "",
"content": "<table><tr><td/><td colspan=\"6\">. Experimental results of the correctly discovered Chinese-to-English</td><td/></tr><tr><td/><td colspan=\"4\">mappings in the top-1 recommendation lists.</td><td/><td/><td/></tr><tr><td>Day</td><td>Tagged Mappings</td><td colspan=\"4\">SVM-V SVM-VCT SVM-W SVM-WCT</td><td>ME</td><td>ME-VCT</td></tr><tr><td>1</td><td>7</td><td>0</td><td>6</td><td>5</td><td>5</td><td>4</td><td>7</td></tr><tr><td>2</td><td>6</td><td>1</td><td>6</td><td>6</td><td>4</td><td>3</td><td>6</td></tr><tr><td>3</td><td>6</td><td>0</td><td>5</td><td>4</td><td>5</td><td>3</td><td>6</td></tr><tr><td>4</td><td>6</td><td>1</td><td>6</td><td>6</td><td>5</td><td>5</td><td>6</td></tr><tr><td>5</td><td>6</td><td>0</td><td>6</td><td>6</td><td>5</td><td>3</td><td>5</td></tr></table>",
"type_str": "table"
},
"TABREF5": {
"html": null,
"num": null,
"text": "",
"content": "<table><tr><td/><td colspan=\"6\">. Experimental results of the correctly discovered Chinese-to-English</td><td/></tr><tr><td/><td colspan=\"4\">mappings in the top-3 recommendation lists.</td><td/><td/><td/></tr><tr><td>Day</td><td>Tagged Mappings</td><td colspan=\"4\">SVM-V SVM-VCT SVM-W SVM-WCT</td><td>ME</td><td>ME-VCT</td></tr><tr><td>1</td><td>7</td><td>0</td><td>6</td><td>7</td><td>7</td><td>5</td><td>7</td></tr><tr><td>2</td><td>6</td><td>1</td><td>6</td><td>6</td><td>6</td><td>6</td><td>6</td></tr><tr><td>3</td><td>6</td><td>0</td><td>5</td><td>6</td><td>6</td><td>5</td><td>6</td></tr><tr><td>4</td><td>6</td><td>1</td><td>6</td><td>6</td><td>6</td><td>5</td><td>6</td></tr><tr><td>5</td><td>6</td><td>0</td><td>6</td><td>6</td><td>6</td><td>4</td><td>6</td></tr></table>",
"type_str": "table"
},
"TABREF7": {
"html": null,
"num": null,
"text": "",
"content": "<table><tr><td>Day</td><td>Tagged Mappings</td><td colspan=\"4\">SVM-V SVM-VCT SVM-W SVM-WCT</td><td>ME</td><td>ME-VCT</td></tr><tr><td>1</td><td>7</td><td>0</td><td>4</td><td>7</td><td>5</td><td>6</td><td>5</td></tr><tr><td>2</td><td>6</td><td>3</td><td>6</td><td>6</td><td>4</td><td>6</td><td>6</td></tr><tr><td>3</td><td>6</td><td>0</td><td>5</td><td>6</td><td>6</td><td>6</td><td>5</td></tr><tr><td>4</td><td>6</td><td>0</td><td>4</td><td>6</td><td>5</td><td>5</td><td>6</td></tr><tr><td>5</td><td>6</td><td>0</td><td>2</td><td>5</td><td>5</td><td>6</td><td>6</td></tr><tr><td>6</td><td>6</td><td>2</td><td>5</td><td>5</td><td>5</td><td>4</td><td>6</td></tr></table>",
"type_str": "table"
},
"TABREF9": {
"html": null,
"num": null,
"text": "",
"content": "<table><tr><td>Day</td><td>Tagged Mappings</td><td colspan=\"4\">SVM-V SVM-VCT SVM-W SVM-WCT</td><td>ME</td><td>ME-VCT</td></tr><tr><td>1</td><td>7</td><td>0</td><td>4</td><td>7</td><td>7</td><td>7</td><td>6</td></tr><tr><td>2</td><td>6</td><td>3</td><td>6</td><td>6</td><td>6</td><td>6</td><td>6</td></tr><tr><td>3</td><td>6</td><td>0</td><td>5</td><td>6</td><td>6</td><td>6</td><td>6</td></tr><tr><td>4</td><td>6</td><td>0</td><td>4</td><td>6</td><td>5</td><td>6</td><td>6</td></tr><tr><td>5</td><td>6</td><td>0</td><td>2</td><td>5</td><td>6</td><td>6</td><td>6</td></tr><tr><td>6</td><td>6</td><td>2</td><td>5</td><td>6</td><td>6</td><td>5</td><td>6</td></tr><tr><td>7</td><td>3</td><td>0</td><td>2</td><td>3</td><td>3</td><td>3</td><td>2</td></tr><tr><td>8</td><td>6</td><td>2</td><td>6</td><td>6</td><td>6</td><td>6</td><td>5</td></tr><tr><td>9</td><td>5</td><td>1</td><td>3</td><td>4</td><td>4</td><td>3</td><td>4</td></tr><tr><td>10</td><td>5</td><td>1</td><td>4</td><td>5</td><td>5</td><td>5</td><td>5</td></tr><tr><td>11</td><td>4</td><td>1</td><td>4</td><td>3</td><td>3</td><td>2</td><td>3</td></tr><tr><td>12</td><td>7</td><td>2</td><td>7</td><td>6</td><td>6</td><td>6</td><td>6</td></tr><tr><td>13</td><td>4</td><td>1</td><td>3</td><td>4</td><td>4</td><td>4</td><td>4</td></tr><tr><td>14</td><td>8</td><td>1</td><td>8</td><td>8</td><td>8</td><td>8</td><td>8</td></tr><tr><td>15</td><td>10</td><td>5</td><td>10</td><td>10</td><td>10</td><td>10</td><td>10</td></tr><tr><td>16</td><td>8</td><td>0</td><td>6</td><td>7</td><td>7</td><td>7</td><td>6</td></tr><tr><td>17</td><td>5</td><td>2</td><td>5</td><td>5</td><td>5</td><td>5</td><td>3</td></tr><tr><td>18</td><td>7</td><td>6</td><td>7</td><td>7</td><td>7</td><td>6</td><td>6</td></tr><tr><td>19</td><td>2</td><td>1</td><td>2</td><td>2</td><td>2</td><td>2</td><td>2</td></tr><tr><td>20</td><td>5</td><td>2</td><td>5</td><td>5</td><td>5</td><td>5</td><td>5</td></tr><tr><td>21</td><td>7</td><td>6</td><td>7</td><td>7</td><td>7</td><td>7</td><td>7</td></tr><tr><td>Total</td><td>123</td><td>36</td><td>105</td><td>118</td><td>118</td><td>115</td><td>112</td></tr><tr><td>Avg. Acc.</td><td/><td>29.27%</td><td>85.37%</td><td>95.93%</td><td>95.93%</td><td>93.50%</td><td>91.06%</td></tr></table>",
"type_str": "table"
},
"TABREF10": {
"html": null,
"num": null,
"text": "",
"content": "<table><tr><td/><td/><td colspan=\"4\">(a) Results of Chinese-to-English mappings</td><td/><td/></tr><tr><td>Day</td><td>Tagged Mappings</td><td colspan=\"4\">SVM-V SVM-VCT SVM-W SVM-WCT</td><td>ME</td><td>ME-VCT</td></tr><tr><td>Total</td><td>122</td><td>37</td><td>115</td><td>119</td><td>120</td><td>103</td><td>117</td></tr><tr><td>Avg. Acc.</td><td/><td>30.33%</td><td>94.26%</td><td>97.54%</td><td>98.36%</td><td>84.43%</td><td>95.90%</td></tr><tr><td/><td/><td colspan=\"4\">(b) Results of English-to-Chinese mappings</td><td/><td/></tr><tr><td>Day</td><td>Tagged Mappings</td><td colspan=\"4\">SVM-V SVM-VCT SVM-W SVM-WCT</td><td>ME</td><td>ME-VCT</td></tr><tr><td>Total</td><td>123</td><td>36</td><td>105</td><td>120</td><td>120</td><td>117</td><td>113</td></tr><tr><td>Avg. Acc.</td><td/><td>29.27%</td><td>85.37%</td><td>97.56%</td><td>97.56%</td><td>95.12%</td><td>91.87%</td></tr></table>",
"type_str": "table"
}
}
}
} |