File size: 64,580 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 |
{
"paper_id": "I08-1006",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:42:04.509980Z"
},
"title": "Story Link Detection based on Dynamic Information Extending",
"authors": [
{
"first": "Xiaoyan",
"middle": [],
"last": "Zhang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Defense Technology No",
"location": {
"addrLine": "137, Yanwachi Street",
"postCode": "410073",
"settlement": "Changsha",
"region": "Hunan",
"country": "P.R.China"
}
},
"email": "zhangxiaoyan@nudt.edu.cn"
},
{
"first": "Ting",
"middle": [],
"last": "Wang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Defense Technology No",
"location": {
"addrLine": "137, Yanwachi Street",
"postCode": "410073",
"settlement": "Changsha",
"region": "Hunan",
"country": "P.R.China"
}
},
"email": "tingwang@nudt.edu.cn"
},
{
"first": "Huowang",
"middle": [],
"last": "Chen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Defense Technology No",
"location": {
"addrLine": "137, Yanwachi Street",
"postCode": "410073",
"settlement": "Changsha",
"region": "Hunan",
"country": "P.R.China"
}
},
"email": "hwchen@nudt.edu.cn"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Topic Detection and Tracking refers to automatic techniques for locating topically related materials in streams of data. As the core technology of it, story link detection is to determine whether two stories are about the same topic. To overcome the limitation of the story length and the topic dynamic evolution problem in data streams, this paper presents a method of applying dynamic information extending to improve the performance of link detection. The proposed method uses previous latest related story to extend current processing story, generates new dynamic models for computing the similarity between the current two stories. The work is evaluated on the TDT4 Chinese corpus, and the experimental results indicate that story link detection using this method can make much better performance on all evaluation metrics.",
"pdf_parse": {
"paper_id": "I08-1006",
"_pdf_hash": "",
"abstract": [
{
"text": "Topic Detection and Tracking refers to automatic techniques for locating topically related materials in streams of data. As the core technology of it, story link detection is to determine whether two stories are about the same topic. To overcome the limitation of the story length and the topic dynamic evolution problem in data streams, this paper presents a method of applying dynamic information extending to improve the performance of link detection. The proposed method uses previous latest related story to extend current processing story, generates new dynamic models for computing the similarity between the current two stories. The work is evaluated on the TDT4 Chinese corpus, and the experimental results indicate that story link detection using this method can make much better performance on all evaluation metrics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Topic Detection and Tracking (TDT) refers to a variety of automatic techniques for discovering and threading together topically related material in streams of data such as newswire or broadcast news. Such automatic discovering and threading could be quite valuable in many applications where people need timely and efficient access to large quantities of information. Supported by such technology, users could be alerted with new events and new information about known events. By examining one or two stories, users define the topic described in them. Then with TDT technologies they could go to a large archive, find all the stories about this topic, and learn how it evolved.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Story link detection, as the core technology defined in TDT, is a task of determining whether two stories are about the same topic, or topically linked. In TDT, a topic is defined as \"something that happens at some specific time and place\" . Link detection is considered as the basis of other event-based TDT tasks, such as topic tracking, topic detection, and first story detection. Since story link detection focuses on the streams of news stories, it has its specific characteristic compared with the traditional Information Retrieval (IR) or Text Classification task: new topics usually come forth frequently during the procedure of the task, but nothing about them is known in advance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The paper is organized as follows: Section 2 describes the procedure of story link detection; Section 3 introduces the related work in story link detection; Section 4 explains a baseline method which will be compared with the proposed dynamic method in Section 5; the experiment results and analysis are given in Section 6; finally, Section 7 concludes the paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(s i1 , s i2 ), s i1 \u2208 S j , s i2 \u2208 S k , 1 \u2264 i \u2264 m, 1 \u2264 j \u2264 k \u2264 n.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The system is required to make decisions on all story pairs to judge if they describe a same topic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We formalize the procedure for processing a pair of stories as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "For a story pair P i = (s i1 , s i2 ):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "1. Get background corpus B i of P i . According to the supposed application situation and the custom that people usually look ahead when they browse something, in TDT research the system is usually allowed to look ahead N (usually 10) source files when deciding whether the current pair is linked. So",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "B i = {S 1 , S 2 , S 3 , . . . , S l }",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": ", where l = k + 10 , s i2 \u2208 S k and (k + 10) \u2264 n n , s i2 \u2208 S k and (k + 10) > n .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(M i1 , M i2 ) for two stories in P i . M = {(f s , w s ) | s \u2265 1},",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Produce the representation models",
"sec_num": "2."
},
{
"text": "where f s is a feature extracted from a story and w s is the weight of the feature in the story. They are computed with some parameters counted from current story and the background.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Produce the representation models",
"sec_num": "2."
},
{
"text": "3. Choose a similarity function F and computing the similarity between two models. If t is a predefined threshold and F (M i1 , M i2 ) \u2265 t, then stories in P i are topically linked.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Produce the representation models",
"sec_num": "2."
},
{
"text": "A number of works has been developed on story link detection. It can be classified into two categories: vector-based methods and probabilistic-based methods. The vector space model is widely used in IR and Text Classification research. Cosine similarity between document vectors with tf * idf term weighting (Connell et al., 2004) (Chen et al., 2004) (Allan et al., 2003) is also one of the best technologies for link detection. We have examined a number of similarity measures in story link detection, including cosine, Hellinger and Tanimoto, and found that cosine similarity produced outstanding results. Furthermore, (Allan et al., 2000) also confirms this conclusion among cosine, weighted sum, language modeling and Kullback-Leibler divergence in its story link detection research.",
"cite_spans": [
{
"start": 308,
"end": 330,
"text": "(Connell et al., 2004)",
"ref_id": "BIBREF4"
},
{
"start": 331,
"end": 350,
"text": "(Chen et al., 2004)",
"ref_id": "BIBREF3"
},
{
"start": 351,
"end": 371,
"text": "(Allan et al., 2003)",
"ref_id": "BIBREF1"
},
{
"start": 621,
"end": 641,
"text": "(Allan et al., 2000)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "3"
},
{
"text": "Probabilistic-based method has been proven to be very effective in several IR applications. One of its attractive features is that it is firmly rooted in the theory of probability, thereby allowing the researcher to explore more sophisticated models guided by the theoretical framework. (Nallapati and Allan, 2002) (Lavrenko et al., 2002) (Nallapati, 2003) all apply probability models (language model or relevance model) for story link detection. And the experiment results indicate that the performances are comparable with those using traditional vector space models, if not better.",
"cite_spans": [
{
"start": 287,
"end": 314,
"text": "(Nallapati and Allan, 2002)",
"ref_id": "BIBREF8"
},
{
"start": 315,
"end": 338,
"text": "(Lavrenko et al., 2002)",
"ref_id": "BIBREF6"
},
{
"start": 339,
"end": 356,
"text": "(Nallapati, 2003)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "3"
},
{
"text": "On the basis of vector-based methods, this paper represents a method of dynamic information extending to improve the performance of story link detection. It makes use of the previous latest topically related story to extend the vector model of current being processed story. New dynamic models are generated for computing the similarity between two stories in current pair. This method resolves the problems of information shortage in stories and topic dynamic evolution in streams of data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "3"
},
{
"text": "Before introducing the proposed method, we first describe a method which is implemented with vector model and cosine similarity function. This straight and classic method is used as a baseline to be compared with the proposed method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "3"
},
{
"text": "The related work in story link detection shows that vector representation model with cosine function can be used to build the state-of-the-art story link detection systems. Many research organizations take this as their baseline system (Connell et al., 2004) (Yang et al., 2002) . In this paper, we make a similar choice.",
"cite_spans": [
{
"start": 236,
"end": 258,
"text": "(Connell et al., 2004)",
"ref_id": "BIBREF4"
},
{
"start": 259,
"end": 278,
"text": "(Yang et al., 2002)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline Story Link Detection",
"sec_num": "4"
},
{
"text": "The baseline method represents each story as a vector in term space, where the coordinates represent the weights of the term features in the story. Each vector terms (or feature) is a single word plus its tag which is produced by a segmenter and part of speech tagger for Chinese. So if two tokens with same spelling are tagged with different tags, they will be taken as different terms (or features). It is notable that in it is independent between processing any two comparisons the baseline method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline Story Link Detection",
"sec_num": "4"
},
{
"text": "A preprocessing has been performed for TDT Chinese corpus. For each story we tokenize the text, tag the generated tokens, remove stop words, and then get a candidate set of terms for its vector model. After that, the term-frequency for each token in the story and the length of the story will also be acquired. In the baseline and dynamic methods, both training and test data are preprocessed in this way.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "4.1"
},
{
"text": "The segmenter and tagger used here is ICTCLAS 1 . The stop word list is composed of 507 terms. Although the term feature in the vector representation is the word plus its corresponding tag, we will ignore the tag information when filtering stop words, because almost all the words in the list should be filtered out whichever part of speech is used to tag them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "4.1"
},
{
"text": "One important issue in the vector model is weighting the individual terms (features) that occur in the vector. Most IR systems employed the traditional tf * idf weighting, which also provide the base for the baseline and dynamic methods in this paper. Furthermore, this paper adopts a dynamic way to compute the tf * idf weighting:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Weighting",
"sec_num": "4.2"
},
{
"text": "w i (f i , d) = tf (f i , d) * idf (f i ) tf = t/(t + 0.5 + 1.5dl/dl avg ) idf = log((N + 0.5)/df )/log(N + 1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Weighting",
"sec_num": "4.2"
},
{
"text": "where t is the term frequency in a story, dl is the length of a story, dl avg is the average length of stories in the background corpus, N is the number of stories in the corpus, df is the number of the stories containing the term in the corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Weighting",
"sec_num": "4.2"
},
{
"text": "The tf shows how much a term represents the story, while the idf reflects the distinctive ability of distinguishing current story from others. The dynamic attribute of the tf * idf weighting lies in the dynamic computation of dl avg , N and df . The background corpus used for statistics is incremental. As more story pairs are processed, more source files could be seen, and the background is expanding as well. Whenever the size of the background has changed, the values of dl avg , N and df will update accordingly. We call this as incremental tf * idf weighting. A story might have different term vectors in different story pairs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Weighting",
"sec_num": "4.2"
},
{
"text": "Another important issue in the vector model is determining the right function to measure the similarity between two vectors. We have firstly tried three functions: cosine, Hellinger and Tanimoto, among which cosine function performs best for its substantial advantages and the most stable performance. So we consider the cosine function in baseline method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity Function",
"sec_num": "4.3"
},
{
"text": "Cosine similarity, as a classic measure and consistent with the vector representation, is simply an inner product of two vectors where each vector is normalized to the unit length. It represents cosine of the angle between two vector models",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity Function",
"sec_num": "4.3"
},
{
"text": "M 1 = {(f 1i , w 1i ), i \u2265 1} and M 2 = {(f 2i , w 2i ), i \u2265 1}. cos(M 1 , M 2 ) = (\u03a3(w 1i \u00d7 w 2i ))/ (\u03a3w 2 1i )(\u03a3w 2 2i )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity Function",
"sec_num": "4.3"
},
{
"text": "Cosine similarity tends to perform best at full dimensionality, as in the case of comparing two stories. Performance degrades as one of the vectors becomes shorter. Because of the built-in length normalization, cosine similarity is less dependent on specific term weighting.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Similarity Function",
"sec_num": "4.3"
},
{
"text": "Investigation on the TDT corpus shows that news stories are usually short, which makes that their representation models are too sparse to reflect topics described in them. A possible method of solving this problem is to extend stories with other related information. The information can be synonym in a dictionary, related documents in external corpora, etc. However, extending with synonym is mainly adding repetitious information, which can not define the topics more clearly. On the other hand, topicbased research should be real-sensitive. The corpora in the same period as the test corpora are not easy to gather, and the number of related documents in previous period is very few. So it is also not feasible to extend the stories with related documents in other corpora. We believe that it is more reasonable that the best extending information may be the story corpus itself. Following the TDT evaluation requirement, we will not use entire corpus at a time. Instead, when we process current pair of stories, we utilize all the stories before the current pair in the story corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Motivation",
"sec_num": "5.1"
},
{
"text": "In addition, topics described by stories usually evolve along with time. A topic usually begins with a seminal event. After that, it will focus mainly on the consequence of the event or other directly related events as the time goes. When the focus in later stories has changed, the words used in them may change remarkably. Keeping topic descriptions unchanged from the beginning to the end is obviously improper. So topic representation models should also be updated as the topic emphases in stories has changed. Formerly we have planed to use related information to extend a story to make up the information shortage in stories. Considering more about topic evolution, we extend a story with its latest related story. In addition, up to now almost all research in story link detection takes the hypothesis that whether two stories in one pair are topically linked is independent of that in another pair. But we realize that if two stories in a pair describe a same topic, one story can be taken as related information to extend another story in later pairs. Compared with extending with more than one story, extending only with its latest related story can keep representation of the topic as fresh as possible, and avoid extending too much similar information at the same time, which makes the length of the extended vector too long. Since the vector will be renormalized, a too big length means evidently decreasing the weight of an individual feature which will instead cause a lower cosine similarity. This idea has also been confirmed by the experiment showing that the performance extending with one latest related story is superior to that extending with more than one related story, as described in section 6.3. The experiment results also show that this method of dynamic information extending apparently improves the performance of story link detection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Motivation",
"sec_num": "5.1"
},
{
"text": "The proposed dynamic method is actually the baseline method plus dynamic information extending. The preprocessing, feature weighting and similarity computation in dynamic method are similar as those in baseline method. However, the vector representation for a story here is dynamic. This method needs a training corpus to get the extending threshold deciding whether a story should be used to extend another story in a pair. We split the sequence of time-ordered story pairs into two parts: the former is for training and the later is for testing. The following is the processing steps:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "1. Preprocess to create a set of terms for representing each story as a term vector, which is same as baseline method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "2. Run baseline system on the training corpora and find an optimum topically link threshold.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "We take this threshold as extending threshold.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "The topically link threshold used for making link decision in dynamic method is another predefined one.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "3. Along with the ordered story pairs in the test corpora, repeat a) and b):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "(a) When processing a pair of stories P i = (s i1 , s i2 ), if s i1 or s i2 has an extending story, then update the corresponding vector model with its related story to a new dynamic one. The generation procedure of dynamic vector will be described in next subsection. (b) Computing the cosine similarity between the two dynamic term vectors. If it exceeds the extending threshold, then s i1 and s i2 are the latest related stories for each other. If one story already has an extending story, replace the old one with the new one. So a story always has no more than one extending story at any time. If the similarity exceeds topically link threshold, s i1 and s i2 are topically linked.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "From the above description, it is obvious that dynamic method needs two thresholds, one for making extending decision and the other for making link decision. Since in this paper we will focus on the optimum performance of systems, the first threshold is more important. But topically link threshold is also necessary to be properly defined to approach a better performance. In the baseline method, term vectors are dynamic because of the incremental tf * idf weighting. However, dynamic information extending is another more important reason in the dynamic method. Whenever a story has an extending story, its vector representation will update to include the extending information. Having the extending method, the representation model can have more information to describe the topic in a story and make the topic evolve along with time. The dynamic method can define topic description clearer and get a more accurate similarity between stories.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method Description",
"sec_num": "5.2"
},
{
"text": "In the dynamic method, we have tried two ways for the generation of dynamic vector models: increment model and average model. Supposing we use vector model",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dynamic Vector Model",
"sec_num": "5.3"
},
{
"text": "M 1 = {(f 1i , w 1i ), i \u2265 1} of story s 1 to ex- tend vector model M 2 = {(f 2i , w 2i ), i \u2265 1} of story s 2",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dynamic Vector Model",
"sec_num": "5.3"
},
{
"text": ", M 2 will change to representing the latest evolving topic described in current story after extending.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dynamic Vector Model",
"sec_num": "5.3"
},
{
"text": "f 1i in M 1 , if",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Increment Model: For each term",
"sec_num": "1."
},
{
"text": "it also occurs as f 2j in M 2 , then w 2j will not change, otherwise (f 1i , w 1i ) will be added into M 2 . This dynamic vector model only takes interest in the new information that occurs only in M 1 . For features both occurred in M 1 and M 2 , the dynamic model will respect to their original weights.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Increment Model: For each term",
"sec_num": "1."
},
{
"text": "2. Average Model: For each term f 1i in M 1 , if it also occurs as f 2j in M 2 , then w 2j = 0.5 * (w 1i + w 2j ), otherwise (f 1i , w 1i ) will be added into M 2 . This dynamic model will take account of all information in M 1 . So the difference between those two dynamic models is the weight recalculation method of the feature occurred in both M 1 and M 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Increment Model: For each term",
"sec_num": "1."
},
{
"text": "Both the above two dynamic models can take account of information extending and topic evolution. Increment Model is closer to topic description since it is more dependent on latest term weights, while Average Model makes more reference to the centroid concept. The experiment results show that dynamic method with Average Model is a little superior to that with Increment Model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Increment Model: For each term",
"sec_num": "1."
},
{
"text": "To evaluate the proposed method, we use the Chinese subset of TDT4 corpus (LDC, 2003) developed by the Linguistic Data Consortium (LDC) for TDT research. This subset contains 27145 stories all in Chinese from October 2000 through January 2001, which are gathered from news, broadcast or TV shows.",
"cite_spans": [
{
"start": 74,
"end": 85,
"text": "(LDC, 2003)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment Data",
"sec_num": "6.1"
},
{
"text": "LDC totally labeled 40 topics on TDT4 for 2003 evaluation. There are totally 12334 stories pairs from 1151 source files in the experiment data. The answers for these pairs are based on 28 topics of these topics, generated from the LDC 2003 annotation documents. The first 2334 pairs are used for training and finding extending threshold of dynamic method. The rest 10000 pairs are testing data used for comparing performances of baseline and the dynamic methods.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment Data",
"sec_num": "6.1"
},
{
"text": "The work is measured by the TDT evaluation software, which could be referred to (Hoogma, 2005) for detail. Here is a brief description. The goal of link detection is to minimize the cost due to errors caused by the system. The TDT tasks are evaluated by computing a \"detection cost\":",
"cite_spans": [
{
"start": 80,
"end": 94,
"text": "(Hoogma, 2005)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "6.2"
},
{
"text": "C det = C miss \u2022P miss \u2022P target +C f a \u2022P f a \u2022P non\u2212target",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "6.2"
},
{
"text": "where C miss is the cost of a miss, P miss is the estimated probability of a miss, P target is the prior probability under which a pair of stories are linked, C f a is the cost of a false alarm, P f a is the estimated probability of a false alarm, and P non\u2212target is the prior probability under which a pair of stories are not linked. A miss occurs when a linked story pair is not identified as being linked by the system. A false alarm occurs when the pair of stories that are not linked are identified as being linked by the system. A target is a pair of linked stories; conversely a nontarget is a pair of stories that are not linked. For the link detection task these parameters are set as follows: C miss is 1, P target is 0.02, and C f a is 0.1. The cost for each topic is equally weighted (usually the cost of topic-weighted is the mainly evaluation parameter) and normalized so that for a given system, the normalized value (C det ) norm can be no less than one without extracting information from the source data:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "6.2"
},
{
"text": "(C det ) norm = C det min(C miss P target , C f a P non\u2212target ) (C det ) overall = \u03a3 i (C i det )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "6.2"
},
{
"text": "norm /#topics where the sum is over topics i. A detection curve (DET curve) is computed by sweeping a threshold over the range of scores, and the minimum cost over the DET curve is identified as the minimum detection cost or min DET. The topic-weighted DET cost is dependent on both a good minimum cost and a good method for selecting an operating point, which is usually implemented by selecting a threshold. A system with a very low min DET cost can have a much larger topic-weighted DET score. Therefore, we focus on the minimum DET cost for the experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Measures",
"sec_num": "6.2"
},
{
"text": "In this paper, we have tried three methods for story link detection: the baseline method described in Section 4 and two dynamic methods with different dynamic vectors introduced in Section 5. The following In the table, Clink min is the minimum (C det ) overall , DET Graph Minimum Detection Cost (topic-weighted), Clink norm is the normalized minimum (C det ) overall , the dynamic 1 is the dynamic method which uses Increment Model and the dynamic 2 is the dynamic method which uses Average Model. We can see that the proposed two dynamic methods are both much better than baseline method on all four metrics. The Clink N orm of dynamic 1 and 2 are improved individually by 27.2% and 27.8% as compared to that of baseline method. The difference between two dynamic methods is due to different in the P miss . However, it is too little to compare the two dynamic systems. We also make additional experiments in which a story is extended with all of its previous related stories. The minimum (Cdet)overall is 0.0614 for the system using Increment Model, and 0.0608 for the system using Average Model. Although the performances are also much superior to baseline, it is still a little poorer than that with only one latest related story, which confirm the ideal described in section 5.1. Figure 1, 2 and 3 show the detail evaluation information for individual topic on Minimum Norm Detection Cost, P miss and P f a . From Figure 1 we know these two dynamic methods have improved the performance on almost all the topic, except topic 12, 26 and 32. Note that detection cost is a function of P miss and P f a . Figure 2 shows that both two dynamic methods reduce the false alarm rates on all evaluation topics. In Figure 3 there are 20 topics on which the miss rates remain zero or unchange. The dynamic methods reduce the miss rates on 5 topics. However, dynamic methods get relatively poorer results on topic 12, 26 and 32 . Altogether dynamic methods can notably improve system performance on evaluation metrics of both individual and weighted topic, especially the false alarm rate, but on some topics, it gets poorer results.",
"cite_spans": [],
"ref_spans": [
{
"start": 1287,
"end": 1304,
"text": "Figure 1, 2 and 3",
"ref_id": "FIGREF0"
},
{
"start": 1421,
"end": 1429,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 1608,
"end": 1616,
"text": "Figure 2",
"ref_id": null
},
{
"start": 1711,
"end": 1719,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiment Results",
"sec_num": "6.3"
},
{
"text": "Further investigation shows that topic 12, 26 and 32 are about Presidential election in Ivory Coast on October 25, 2000, Airplane Crash in Chiang Kai Shek International Airport in Taiwan on October 31, 2000, and APEC Conference on November 12-15, 2000 at Brunei. After analyzing those story pairs with error link decision, we can split them into two sets. One is that two stories in a pair are general linked but not TDT specific topically linked. Here general linked means that there are many common words in two stories, but the events described in them happened in different times or different places. For example, Airplane Crash is a general topic, while Airplane Crash in certain location at specification time is a TDT topic. The other is that two stories in a pair are TDT topically linked while they describe the topic from different perspectives. In this condition they will have few common words. These may be due to that the information extracted from stories is still not accurate enough to represent them. It also may be because of the 1 3 5 7 9 1 1 1 3 1 5 1 7 1 9 2 1 2 3 2 5 2 7 2 9 3 1 3 3 3 5 3 7 3 9 T o p i c I 1 3 5 7 9 1 1 1 3 1 5 1 7 1 9 2 1 2 3 2 5 2 7 2 9 3 1 3 3 3 5 3 7 3 Figure 3: P miss for individual topic deficiency of vector model itself. Furthermore, we know that the extending story is chosen by cosine similarity, which results that the extending story and the extended story are usually topically linked from the same perspectives, seldom from different perspectives. Therefore the method of information extending may sometimes turn the above first problem worse and have no impact on the second problem. So mining more useful information or making more use of other useful resources to solve these problems will be the next work. In addition, how to represent this information with a proper model and seeking better or more proper representation models for TDT stories are also important issues. In a word, the method of information extending has been verified efficient in story link detection and can provide a reference to improve the performance of some other similar systems whose data must be processed serially, and it is also hopeful to combined with other improvement technologies.",
"cite_spans": [],
"ref_spans": [
{
"start": 1049,
"end": 1170,
"text": "1 3 5 7 9 1 1 1 3 1 5 1 7 1 9 2 1 2 3 2 5 2 7 2 9 3 1 3 3 3 5 3 7 3 9 T o p i c I",
"ref_id": null
},
{
"start": 1171,
"end": 1271,
"text": "1 3 5 7 9 1 1 1 3 1 5 1 7 1 9 2 1 2 3 2 5 2 7 2 9 3 1 3 3 3 5 3 7 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiment Results",
"sec_num": "6.3"
},
{
"text": "Story link detection is a key technique in TDT research. Though many approaches have been tried, there are still some characters ignored. After analyzing the characters and deficiency in TDT stories and story link detection, this paper presents a method of dynamic information extending to improve the system performance by focus on two problems: information deficiency and topic evolution. The experiment results indicate that this method can effectively improve the performance on both miss and false alarm rates, especially the later one. However, we should realize that there are still some problems to solve in story link detection. How to compare general topically linked stories and how to compare stories describing a TDT topic from different angles will be very vital to improve system performance. The next work will focus on mining more and deeper useful information in TDT stories and exploiting more proper models to represent them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "Problem DefinitionIn the task definition of story link detection(NIST, 2003), a link detection system is given a sequence of time-ordered news source files S = S 1 , S 2 , S 3 , . . . , S n where each S i includes a set of stories, and a sequence of time-ordered story pairs P = P 1 , P 2 , P 3 , . . . , P m where P i =",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://sewm.pku.edu.cn/QA/reference/ICTCLAS/FreeICT-CLAS/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This research is supported by the National Natural Science Foundation of China (60403050), Program for New Century Excellent Talents in University (NCET-06-0926) and the National Grand ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Detections, bounds, and timelines: Umass and tdt-3",
"authors": [
{
"first": "James",
"middle": [],
"last": "Allan",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Lavrenko",
"suffix": ""
},
{
"first": "Daniella",
"middle": [],
"last": "Malin",
"suffix": ""
},
{
"first": "Russell",
"middle": [],
"last": "Swan",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of Topic Detection and Tracking (TDT-3)",
"volume": "",
"issue": "",
"pages": "167--174",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "James Allan, Victor Lavrenko, Daniella Malin, and Rus- sell Swan. 2000. Detections, bounds, and timelines: Umass and tdt-3. In Proceedings of Topic Detection and Tracking (TDT-3), pages 167-174.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Umass tdt 2003 research summary. In proceedings of TDT workshop",
"authors": [
{
"first": "J",
"middle": [],
"last": "Allan",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bolivar",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Connell",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Cronen-Townsend",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Feng",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Feng",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Kumaran",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Larkey",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Lavrenko",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Raghavan",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Allan, A. Bolivar, M. Connell, S. Cronen-Townsend, A Feng, F. Feng, G. Kumaran, L. Larkey, V. Lavrenko, and H. Raghavan. 2003. Umass tdt 2003 research summary. In proceedings of TDT workshop.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Topic Detection and Tracking: Event-based Information Organization",
"authors": [],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "James Allan, editor. 2002. Topic Detection and Track- ing: Event-based Information Organization. Kluwer Academic Publishers, Norvell, Massachusetts.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Multiple similarity measures and source-pair information in story link detection",
"authors": [
{
"first": "Francine",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Ayman",
"middle": [],
"last": "Farahat",
"suffix": ""
},
{
"first": "Thorsten",
"middle": [],
"last": "Brants",
"suffix": ""
}
],
"year": 2004,
"venue": "HLT-NAACL",
"volume": "",
"issue": "",
"pages": "313--320",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Francine Chen, Ayman Farahat, and Thorsten Brants. 2004. Multiple similarity measures and source-pair information in story link detection. In HLT-NAACL, pages 313-320.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Umass at tdt",
"authors": [
{
"first": "Margaret",
"middle": [],
"last": "Connell",
"suffix": ""
},
{
"first": "Ao",
"middle": [],
"last": "Feng",
"suffix": ""
},
{
"first": "Giridhar",
"middle": [],
"last": "Kumaran",
"suffix": ""
},
{
"first": "Hema",
"middle": [],
"last": "Raghavan",
"suffix": ""
},
{
"first": "Chirag",
"middle": [],
"last": "Shah",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Allan",
"suffix": ""
}
],
"year": 2004,
"venue": "TDT2004 Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Margaret Connell, Ao Feng, Giridhar Kumaran, Hema Raghavan, Chirag Shah, and James Allan. 2004. Umass at tdt 2004. In TDT2004 Workshop.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "The modules and methods of topic detection and tracking",
"authors": [
{
"first": "Niek",
"middle": [],
"last": "Hoogma",
"suffix": ""
}
],
"year": 2005,
"venue": "2nd Twente Student Conference on IT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Niek Hoogma. 2005. The modules and methods of topic detection and tracking. In 2nd Twente Student Confer- ence on IT.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Relevance models for topic detection and tracking",
"authors": [
{
"first": "Victor",
"middle": [],
"last": "Lavrenko",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Allan",
"suffix": ""
},
{
"first": "Edward",
"middle": [],
"last": "Deguzman",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Laflamme",
"suffix": ""
},
{
"first": "Veera",
"middle": [],
"last": "Pollard",
"suffix": ""
},
{
"first": "Stephen",
"middle": [],
"last": "Thomas",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of Human Language Technology Conference (HLT)",
"volume": "",
"issue": "",
"pages": "104--110",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Victor Lavrenko, James Allan, Edward DeGuzman, Daniel LaFlamme, Veera Pollard, and Stephen Thomas. 2002. Relevance models for topic detec- tion and tracking. In Proceedings of Human Language Technology Conference (HLT), pages 104-110.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Topic detection and tracking -phase 4",
"authors": [
{
"first": "",
"middle": [],
"last": "Ldc",
"suffix": ""
}
],
"year": 2003,
"venue": "Linguistic Data Consortium",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "LDC. 2003. Topic detection and tracking -phase 4. Technical report, Linguistic Data Consortium.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Capturing term dependencies using a language model based on sentence trees",
"authors": [
{
"first": "Ramesh",
"middle": [],
"last": "Nallapati",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Allan",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the eleventh international conference on Information and knowledge management",
"volume": "",
"issue": "",
"pages": "383--390",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ramesh Nallapati and James Allan. 2002. Capturing term dependencies using a language model based on sentence trees. In Proceedings of the eleventh interna- tional conference on Information and knowledge man- agement, pages 383-390. ACM Press.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Semantic language models for topic detection and tracking",
"authors": [
{
"first": "Ramesh",
"middle": [],
"last": "Nallapati",
"suffix": ""
}
],
"year": 2003,
"venue": "HLT-NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ramesh Nallapati. 2003. Semantic language models for topic detection and tracking. In HLT-NAACL.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "The 2003 topic detection and tracking task definition and evaluation plan",
"authors": [
{
"first": "",
"middle": [],
"last": "Nist",
"suffix": ""
}
],
"year": 2003,
"venue": "National Institute of Standards and Technology(NIST)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "NIST. 2003. The 2003 topic detection and tracking task definition and evaluation plan. Technical report, Na- tional Institute of Standards and Technology(NIST).",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Topic-conditioned novelty detection",
"authors": [
{
"first": "Yiming",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jaime",
"middle": [],
"last": "Carbonell",
"suffix": ""
},
{
"first": "Chun",
"middle": [],
"last": "Jin",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining",
"volume": "",
"issue": "",
"pages": "688--693",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yiming Yang, Jian Zhang, Jaime Carbonell, and Chun Jin. 2002. Topic-conditioned novelty detection. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 688-693. ACM Press.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"uris": null,
"type_str": "figure",
"text": "Normalized Minimum Detection Cost for individual topic"
},
"FIGREF1": {
"num": null,
"uris": null,
"type_str": "figure",
"text": "Figure 2: P f a for individual topic"
},
"TABREF0": {
"type_str": "table",
"content": "<table><tr><td/><td colspan=\"3\">their evaluation results.</td></tr><tr><td>metrics</td><td colspan=\"3\">baseline dynamic 1 dynamic 2</td></tr><tr><td>P miss</td><td>0.0514</td><td>0.0348</td><td>0.0345</td></tr><tr><td>P f a</td><td>0.0067</td><td>0.0050</td><td>0.0050</td></tr><tr><td>Clink min</td><td>0.0017</td><td>0.0012</td><td>0.0012</td></tr><tr><td colspan=\"2\">Clink norm 0.0840</td><td>0.0591</td><td>0.0588</td></tr><tr><td colspan=\"4\">Table 1: Experiment Results of Baseline System and</td></tr><tr><td colspan=\"2\">Dynamic Systems</td><td/><td/></tr></table>",
"html": null,
"text": "",
"num": null
}
}
}
} |