File size: 56,642 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 |
{
"paper_id": "I08-1018",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:42:37.638609Z"
},
"title": "Generic Text Summarization Using Probabilistic Latent Semantic Indexing",
"authors": [
{
"first": "Harendra",
"middle": [],
"last": "Bhandari",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Nara Institute of Science and Technology",
"location": {
"postCode": "630-0192",
"settlement": "Nara",
"country": "Japan"
}
},
"email": "harendra-b@is.naist.jp"
},
{
"first": "Takahiko",
"middle": [],
"last": "Ito",
"suffix": "",
"affiliation": {},
"email": "takahahi-i@is.naist.jp"
},
{
"first": "Masashi",
"middle": [],
"last": "Shimbo",
"suffix": "",
"affiliation": {},
"email": "shimbo@is.naist.jp"
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper presents a strategy to generate generic summary of documents using Probabilistic Latent Semantic Indexing. Generally a document contains several topics rather than a single one. Summaries created by human beings tend to cover several topics to give the readers an overall idea about the original document. Hence we can expect that a summary containing sentences from better part of the topic spectrum should make a better summary. PLSI has proven to be an effective method in topic detection. In this paper we present a method for creating extractive summary of the document by using PLSI to analyze the features of document such as term frequency and graph structure. We also show our results, which was evaluated using ROUGE, and compare the results with other techniques, proposed in the past.",
"pdf_parse": {
"paper_id": "I08-1018",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper presents a strategy to generate generic summary of documents using Probabilistic Latent Semantic Indexing. Generally a document contains several topics rather than a single one. Summaries created by human beings tend to cover several topics to give the readers an overall idea about the original document. Hence we can expect that a summary containing sentences from better part of the topic spectrum should make a better summary. PLSI has proven to be an effective method in topic detection. In this paper we present a method for creating extractive summary of the document by using PLSI to analyze the features of document such as term frequency and graph structure. We also show our results, which was evaluated using ROUGE, and compare the results with other techniques, proposed in the past.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The advent of the Internet has made a wealth of textual data available to everyone. Finding a specific piece of information in this mass of data can be compared with \"finding a small needle in a large heap of straw.\" Search engines do a remarkable job in providing a subset of the original data set which is generally a lot smaller than the original pile of data. However the subset provided by the search engines is still substantial in size. Users need to manually scan through all the information contained in the list of results provided by the search engines until the desired information is found. This makes automatic summarization the task of great importance as the users can then just read the summaries and obtain an overview of the document, hence saving a lot of time during the process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Several methods have been proposed in the field of automatic text summarization. In general two approaches have been taken, extract-based summarization and abstract-based summarization. While extract-based summarization focuses in finding relevant sentences from the original document and using the exact sentences as a summary, abstractbased summaries may contain the words or phrases not present in the original document (Mani, 1999) . The summarization task can also be classified as query-oriented or generic. The query-oriented summary presents text that contains information relevant to the given query, and the generic summarization method presents the summary that gives overall sense of the document (Goldstein et al, 1998) . In this paper, we will focus on extract-based generic single-document summarization.",
"cite_spans": [
{
"start": 423,
"end": 435,
"text": "(Mani, 1999)",
"ref_id": "BIBREF10"
},
{
"start": 709,
"end": 732,
"text": "(Goldstein et al, 1998)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In the recent years graph based techinques have become very popular in automatic text summariza-tion (Erkan and Radev, 2004) , (Mihalcea, 2005) . These techniques view each sentence as a node of a graph and the similarities between each sentences as the links between those sentences. Generally the links are retained only if the similarity values between the sentences exceed a pre-determined threshold value; the links are discarded otherwise. The sentences are then ranked using some graph ranking algorithms such as HITS (Kleinberg, 1998) or PageRank (Brin and Page, 1998) etc. However the graph ranking algorithms tend to give the highest ranking to the sentences related to one central topic in the document. So if a document contains several topics, these algorithms will only choose one central topic and rank the sentences related to those topic higher than any other topics, ignoring the importance of other topics present. This will create summaries that may not cover the overall topics of the document and hence cannot be considered generic enough. We will focus on that problem and present a way to create better generic summary of the document using PLSI (Hofmann 1999) which covers several topics in the document and is closer to the summaries created by human beings. The benchmarking done using DUC 2 2002 data set showed that our technique improves over other proposed methods in terms of ROUGE 1 evaluation score.",
"cite_spans": [
{
"start": 101,
"end": 124,
"text": "(Erkan and Radev, 2004)",
"ref_id": "BIBREF5"
},
{
"start": 127,
"end": 143,
"text": "(Mihalcea, 2005)",
"ref_id": "BIBREF11"
},
{
"start": 525,
"end": 542,
"text": "(Kleinberg, 1998)",
"ref_id": "BIBREF9"
},
{
"start": 555,
"end": 576,
"text": "(Brin and Page, 1998)",
"ref_id": "BIBREF1"
},
{
"start": 1170,
"end": 1184,
"text": "(Hofmann 1999)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Maximal Marginal Relevance(MMR) MMR is a summarization procedure based on vector-space model and is suited to generic summarization (Goldstein et al, 1999) . In MMR the sentence are chosen according to the weighed combination of their general relevance in the document and their redundancy with the sentences already chosen. Both the relevance and redundancy are measured using cosine similarity. Relevance is the cosine similarity of a sentence with rest of the sentence in the document whereas redundancy is measured using cosine similarity between the sentence and the sentences already chosen for the summary.",
"cite_spans": [
{
"start": 132,
"end": 155,
"text": "(Goldstein et al, 1999)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "2.1",
"sec_num": null
},
{
"text": "Graph Based Summarization The graph-based summarization procedure are be _________________________________________ coming increasingly popular in recent years. Lex PageRank (Erkan and Radev, 2004) is one of such methods. LexPageRank constructs a graph where each sentence is a node and links are the similarities between the sentences. Similarity is measured using cosine similarity of the word vectors, and if the similarity value is more than certain threshold value the link is kept otherwise the links are removed. PageRank is an algorithm which has been successfully applied by Google search engine to rank the search results. Similarly PageRank is applied in LexPageRank to rank the nodes (or, sentences) of the resultant graph. A similar summarization method has been proposed by Mihalcea (2005) .",
"cite_spans": [
{
"start": 173,
"end": 196,
"text": "(Erkan and Radev, 2004)",
"ref_id": "BIBREF5"
},
{
"start": 787,
"end": 802,
"text": "Mihalcea (2005)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "2.2",
"sec_num": null
},
{
"text": "Algorithms like HITS and PageRank calculate the principal eigenvector (hence find the principal community) of the matrix representing the graph. But as illustrated in Figure 1 , another eigenvector which is slightly smaller than the principal eigenvector may exist. In documents, each community represented by the eigenvectors can be considered as a topic present in the document. As these algorithms tend to ignore the influence of eigenvectors other than largest one, the sentences related to topics other than a central one can be ignored, and creating the possibility for the inclusion of redundant sentences as well. This kind of summary cannot be considered as a generic one. Figure 1 . In algorithms like HITS and PageRank only the principal eigenvectors are considered. In the figure the vector EV1 is slightly larger than vector EV2, but the score commanded by members of EV2 communities are ignored.",
"cite_spans": [],
"ref_spans": [
{
"start": 167,
"end": 175,
"text": "Figure 1",
"ref_id": null
},
{
"start": 682,
"end": 690,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "2.2",
"sec_num": null
},
{
"text": "As we mentioned in section 1, we take into consideration the sentences from all the topics generated by PLSI in the summary, hence getting a more generic summary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.2",
"sec_num": null
},
{
"text": "Latent Semantic Analysis Latent Semantic Analysis (LSA) (Deerwester et al.,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3",
"sec_num": null
},
{
"text": "1990) takes the high dimensional vector space representation of the document based on term frequency and projects it to lesser dimension space. It is thought that the similarities between the documents can be more reliably estimated in the reduced latent space representation than original representation. LSA has been applied in areas of text retrieval (Deerwester et al., 1990) and automatic text summarization (Gong and Liu, 2001) . LSA is based on Singular Value Decomposition (SVD) of m\u00d7n term-document matrix A. Each entry in A, A ij , represents the frequency of term i in document j. Using SVD, the matrix A is decomposed into U,S,V as, A=USV T U=Matrix of n left singular vectors S=diag(\u03c3 i )=Diagonal matrix of singular values where with \u03c3 i \u2265 \u03c3 i+1 for all i. V T =Matrix of right singular vectors. Each row represents a topic and the values in each row represent the score of documents, represented by each columns, for the topic represented by the row. Gong and Liu (2001) have proposed a scheme for automatic text summarization using LSA. Their algorithm can be stated below.",
"cite_spans": [
{
"start": 354,
"end": 379,
"text": "(Deerwester et al., 1990)",
"ref_id": "BIBREF4"
},
{
"start": 413,
"end": 433,
"text": "(Gong and Liu, 2001)",
"ref_id": "BIBREF6"
},
{
"start": 966,
"end": 985,
"text": "Gong and Liu (2001)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "EV1 EV2",
"sec_num": null
},
{
"text": "a. Choose the highest ranked sentence from k th right singular vector in matrix V T and use the sentence in summary. b. If k reaches the predefined number, terminate the process; otherwise, go to step a again. LSA categorizes sentences on the basis of the topics they belong to. Gong and Liu's method picks sentences from various topics hence producing the summaries that are generic in nature.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "EV1 EV2",
"sec_num": null
},
{
"text": "In section 3 we explain how PLSI is more advanced form of LSA. In section 5, we compare our summarization results with that of LSA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "EV1 EV2",
"sec_num": null
},
{
"text": "Probabilistic Latent Semantic Indexing (PLSI) (Hofmann, 1999 ) is a new approach to automated document indexing, and is based on a statistical latent class model for factor analysis of count data. PLSI is considered to be a probabilistic analogue of Latent Semantic Indexing (LSI), which is a document indexing technique based on LSA. Despite the success of LSI, it is not devoid of deficits. The main argument against LSI is pointed to its unsatisfactory statistical foundations. In contrast, PLSI has solid statistical foundations, as it is based on the maximum likelihood principle and defines a proper generative model of data. Hofmann (1999) has shown that PLSI indeed performs better than LSI in several text retrieval experiments. The factor representation obtained in PLSI allows us to classify sentences according to the topics they belong to. We will use this ability of PLSI to generate summary of document that are more generic in nature by picking sentences from different topics.",
"cite_spans": [
{
"start": 46,
"end": 60,
"text": "(Hofmann, 1999",
"ref_id": "BIBREF7"
},
{
"start": 632,
"end": 646,
"text": "Hofmann (1999)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Probabilistic Latent Semantic Indexing",
"sec_num": "3"
},
{
"text": "Our document model is similar to Aspect Model (Hofmann et al, 1999 , Saul and Pereira, 1997 ) used by Hoffman (1999 . The model attempts to associate an unobserved class variable z\u2208Z={z 1 , ..., z k } (in our case the topics contained in the document), with two sets of observables, documents (d \u2208 D={d 1 ,\u2026..d m }, sentences in our case) and words (w \u2208W={w 1 ,\u2026,w n }) contained in documents. In terms of generative model it can be defined as follows:",
"cite_spans": [
{
"start": 46,
"end": 66,
"text": "(Hofmann et al, 1999",
"ref_id": "BIBREF7"
},
{
"start": 67,
"end": 91,
"text": ", Saul and Pereira, 1997",
"ref_id": null
},
{
"start": 92,
"end": 115,
"text": ") used by Hoffman (1999",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization with PLSI 4.1 The Latent Variable Model for Document",
"sec_num": "4"
},
{
"text": "-A document d is selected with probability P(d) -A latent class z is selected with probability P(z|d)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization with PLSI 4.1 The Latent Variable Model for Document",
"sec_num": "4"
},
{
"text": "-A word w is selected with probability P(w|z) For each document-word pair (d,w), the likelihood for each pair can be represented as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization with PLSI 4.1 The Latent Variable Model for Document",
"sec_num": "4"
},
{
"text": "P(d,w)=P(d)P(w|d)=P(d) \u2211 z P(w|z)P(z|d).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization with PLSI 4.1 The Latent Variable Model for Document",
"sec_num": "4"
},
{
"text": "Following the maximum likelihood principle P(d), P(z|d), P(w|z) are determined by the maximization of of log-likelihood function,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization with PLSI 4.1 The Latent Variable Model for Document",
"sec_num": "4"
},
{
"text": "L= \u2211 d \u2211 w n(d,w)logP(d,w)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization with PLSI 4.1 The Latent Variable Model for Document",
"sec_num": "4"
},
{
"text": "where n(d,w) denotes the term frequency, i.e., the number of time w occurred in d.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization with PLSI 4.1 The Latent Variable Model for Document",
"sec_num": "4"
},
{
"text": "Expectation Maximization (EM) is the standard procedure for maximizing likelihood estimation in the presence of latent variables. EM is an iterative procedure and each of the iteration contains two steps. (a) An Expectation (E) step, where the posterior probabilities for latent variable z are computed and (b) Maximization (M) step, where parameters for given posterior probabilities are computed. The aspect model can be re-parameterized using the Bayes' rule as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "P(d,w)= \u2211 z P(z) P(d|z) P(w|z) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "Then using the re-parameterized equation the Estep calculates the posterior for z by",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "' ( ) ( | ) ( | ) ( ') ( | ') ( | ') ( | , ) z P z P d z P w z P z P d z P w z P z d w = \u2211",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "This step calculates the probability that word w present in document d can be described by the factor corresponding to z. Subsequently, the M-step re-evaluates the parameters using following equations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": ", '",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "( , ) ( | , ) ( , ') ( | , ') ( | ) , d d w n d w P z d w n d w P z d w P w z = \u2211 \u2211 (1) ', ( , ) ( | , ) ( ', ) ( | ', ) ( | ) , w d w n d w P z d w n d w P z d w P d z = \u2211 \u2211 (2) , ,",
"eq_num": "( , )"
}
],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "( | , )",
"eq_num": "( , )"
}
],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "( ) d w d w n d w P z d w n d w P z = \u2211 \u2211",
"eq_num": "(3)"
}
],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "Alternating the E-and M-steps one approaches a converging point which describes local maximum of the log-likelihood. We used the tempered EM (TEM) as described by Hofmann (1999) . TEM basically introduces a control parameter B, upon which the E-step is modified as, The TEM reduces to original EM if B=1.",
"cite_spans": [
{
"start": 163,
"end": 177,
"text": "Hofmann (1999)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Maximizing Model Likelihood",
"sec_num": "4.2"
},
{
"text": "We applied PLSI in 4 different ways during the summarization process. We will denote each of the 4 ways as PROC1, PROC2, PROC3, PROC4. Each of the four summarization procedure is discussed below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summarization procedure",
"sec_num": "4.3"
},
{
"text": "PROC1 consists of the following steps: a. Each document is represented as termfrequency matrix.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PROC1 (Dominant topic only):",
"sec_num": null
},
{
"text": "b. P(w|z), P(d|z), and P(z) (as in (1), (2), (3)) are calculated until the convergence criteria for EM-algorithm is met. P(d|z) represents the importance of document d in given topic represented by z and P(z) represents the importance of the topic z itself in the document d. c. z with highest probability P(z) is picked as the central topic of the document and then the sentences with highest P(d|z) score contained in selected topic are picked. d. The top scoring sentences are used in the summary. PROC2 (Dominant topic only): PROC2 is the graph based method. PROC2 is similar to PROC1 except for the fact that instead of using term-frequency matrix we use sentence-similarity matrix. Sentence-similarity matrix A is n\u00d7n matrix where n is the number of sentences present in the document. Cosine similarity of each sentence present in the document with respect to all the sentences is calculated. The cosine-similarity values calculated are used instead of term-frequency values as in PROC1. Each entry A ij in matrix A is 0 if the cosine similarity value between sentence i and sentence j is less than threshold value and 1 if greater. We used 0.2 as the threshold value in our experiments after normalizing cosine similarity value. Steps b, c, d from PROC1 are followed after the initial procedure is complete.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PROC1 (Dominant topic only):",
"sec_num": null
},
{
"text": "This method is analogous to PHITS (Cohn and Chang (2001) ) method where the authors utilized PLSI to find communities in hyperlinked environment.",
"cite_spans": [
{
"start": 34,
"end": 56,
"text": "(Cohn and Chang (2001)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "PROC1 (Dominant topic only):",
"sec_num": null
},
{
"text": "In both PROC1 and PROC2 we did not take the advantage of the fact that PLSI divides a document into several topics. We only used the sentences from highest ranked topic. In PROC3 we attempt to combine the sentences from different topics while forming the summary. PROC3 can be explained in the following steps. a. Steps a and b from PROC1 are taken as normal. b. We mentioned that P(d|z) represents the score of the sentence d in topic z. In this procedure we will create new score R for each sentence using following relation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PROC3 (Multiple topics):",
"sec_num": null
},
{
"text": "R= \u2211 z P(d|z)P(z)=P(d)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PROC3 (Multiple topics):",
"sec_num": null
},
{
"text": "This will essentially score the sentences with generic values or the sentences which have good influence ranging over several topics better. c. We pick the sentences that score highest score R as the summary. PROC3 will pick sentences from several topics resulting in better generic summary of the document.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PROC3 (Multiple topics):",
"sec_num": null
},
{
"text": "PROC4 is essentially PROC3 except for the first few steps. PROC4 does not use the matrix created in PROC1 instead it uses the similarity-matrix produced in PROC2. Once the similarity matrix is created P(z) and P(d|z) are calculated as in step b of PROC1. Then steps b and c of PROC3 are taken to produce the summary of the document.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PROC4 (Multiple Topics):",
"sec_num": null
},
{
"text": "We produced summaries for all the procedures mentioned in section 4.3. We used DUC 2002 data set for summarization. DUC 2002 contains test data for both multiple document and single document summarization. It also contains summaries created by human beings for both single document and multiple document summarization. Our focus in this paper is single document summarization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "5"
},
{
"text": "After creating summaries we evaluated summaries using ROUGE. ROUGE has been the standard benchmarking technique for summarization tasks adopted by Document Understanding Conference (DUC). We also compared our results with other summarization methods such as LexPageRank (Erkan and Radev, 2004) and Gong and Liu's (2001) LSA-based method. We also compared the results with HITS based method which is similar to Lex-PageRank but instead of PageRank, HITS is used as ranking algorithm (Klienberg 1998).The results are listed in Table 1 . We used five measures for evaluation, Rouge-L Rouge1, Rouge2, Rouge-SU4 and F 1 . These methods are standard methods used in DUC evaluation Table 1 : Evaluation of summaries The table shows the score of summaries generated using methods described in section 4.3. On the table n means number of topics into which the document has been divided into. Control parameter B from (4) was fixed to 0.75 in this case. tests and these schemes are known to be very effect tive to calculate the correlation between the summaries. All of the scores can be calculated using Rouge package. Rouge is based on N-gram statistics (Lin and Hovy, 2003) . Rouge has been known to highly correlate with human evaluations. According to (Lin and Hovy, 2003) , among the methods implemented in ROUGE, ROUGE-N (N=1,2), ROUGE-L, ROUGE-S are relatively simple and work very well even when the length of summary is quite short, which is mostly the case in single document summarization. ROUGE-N, ROUGE-L and ROUGE-S are all basically the recall scores. As DUC keeps the length of the summaries constant recall is the main evaluation criterion. Fmeasure is also shown in the table as a reference parameter, but since we kept the length of our summaries constant, too, the ROUGE-L, ROUGE-N and ROUGE-S scores carry the highest weight. As seen on Table 1 , the scores gained by PROC1 and PROC2 are less than others. This is mainly because the sentences chosen by these methods were simply chosen from one topic. As PROC3 and PROC4 use sentences from several topics the score of PROC3 and PROC4 were better than PROC1 and PROC2. For methods PROC3 and PROC4 we took the summaries for topics 2 through 6 and found that the method performed well when the number of topics was kept between 2 to 4. But the difference was very small, and in general the performance was quite stable. We also compared our results to other methods such as LexPageRank and LSA and found that PROC3 performed quite well when compared to those methods. LexPageRank was marginally better in F-measure (F 1 ) but PROC3 got best recall scores. PROC3 also outperformed LSA by 0.16 in recall (ROUGE-L) scores. Comparison to HITS also shows PROC3 more advantageous.",
"cite_spans": [
{
"start": 281,
"end": 293,
"text": "Radev, 2004)",
"ref_id": "BIBREF5"
},
{
"start": 298,
"end": 319,
"text": "Gong and Liu's (2001)",
"ref_id": "BIBREF6"
},
{
"start": 1146,
"end": 1166,
"text": "(Lin and Hovy, 2003)",
"ref_id": null
},
{
"start": 1247,
"end": 1267,
"text": "(Lin and Hovy, 2003)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 525,
"end": 532,
"text": "Table 1",
"ref_id": null
},
{
"start": 675,
"end": 682,
"text": "Table 1",
"ref_id": null
},
{
"start": 1849,
"end": 1856,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "5"
},
{
"text": "In this paper we have argued that choosing sentences from multiple topics makes a better generic summary. It is especially true if we compare our method to graph based ranking methods like HITS and PageRank. Richardson and Domingos (2002) have mentioned that both HITS and PageRank suffer from the topic drift. This not only makes these algorithms susceptible for exclusion of important sentences outside the main topic but miss the sentences from main topic as well. Cohn and Chang (2001) also have shown similar results for HITS. They (Cohn and Chang) have shown that the central topic identified by HITS (principal eigenvector) may not always correspond to the most authoritative topic. The main topic in fact may be represented by smaller eigenvectors rather than the principal one. They also show that the topic segre- gation in HITS is quite extreme so if we just use principal eigenvector, first there is a chance of being drifted away from the main topic hence producing low quality summary and there is also a chance of missing out other important topics due to the extreme segregation of communities. In PLSI the segregation of topics is not as extreme. If a sentence is related to several topics the sentence can attain high rank in many topics. We can see from the scores that the performance of graph based algorithms like LexPageRank and HITS are not as good as our method. This can be attributed to the fact that the graph based summarizers take only a central topic under consideration. The method that proved most successful in our summarization was the one where we extracted the sentences that had the most influence in the document. We used the tempered version of EM-algorithm (4) in our summarization task. We evaluated the effect of tempering factor B in performance of summarization for PROC3. We found that that the tempering factor did not influence the results by a big margin. We conducted our experiment using values of B from 0.1 through 1.0 incrementing each step by 0.1. The results are shown in Figure 2 . In the results shown in Table 1 the value for tempering factor was set to 0.75.",
"cite_spans": [
{
"start": 208,
"end": 238,
"text": "Richardson and Domingos (2002)",
"ref_id": "BIBREF12"
},
{
"start": 468,
"end": 489,
"text": "Cohn and Chang (2001)",
"ref_id": "BIBREF3"
},
{
"start": 537,
"end": 553,
"text": "(Cohn and Chang)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 2028,
"end": 2036,
"text": "Figure 2",
"ref_id": "FIGREF1"
},
{
"start": 2063,
"end": 2070,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "In this paper we presented a method for creating generic summaries of the documents using PLSI. PLSI allowed us classify the sentences present in the document into several topics. Our summary included sentences from all the topics, which made the generation of generic summary possible. Our experiments showed that the results we obtained in summarization tasks were better than some other methods we compared with. LSA can also be used to summarize documents in similar manner by extracting sentences from several topics, but our experiments showed that PLSI performs better than LSA. In the future we plan to investigate how more recent methods such as LDA (Blei et al) perform in document summarization tasks. We also plan to apply our methods to multiple document summarization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "7"
},
{
"text": "We pay our special gratitude to the reviewers who have taken their time to give very useful comments on our work. The comments were very useful for us to as we were able to provide wider perspective on our work with the help of those comments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": "8"
},
{
"text": "ROUGE:http://openrouge.com/default.aspx 2 http://duc.nist.gov",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF1": {
"ref_id": "b1",
"title": "The Anatomy of a Large-Scale Hypertextual Web Search Engine",
"authors": [
{
"first": "S",
"middle": [],
"last": "Brin",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Page",
"suffix": ""
}
],
"year": 1998,
"venue": "Computer Networks",
"volume": "30",
"issue": "1-7",
"pages": "107--117",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Brin S and Page L.1998. The Anatomy of a Large- Scale Hypertextual Web Search Engine. Computer Networks 30(1-7): 107-117.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries",
"authors": [
{
"first": "Carbonell J",
"middle": [],
"last": "Goldstein",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 1998,
"venue": "Proc. ACM SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Carbonell J and Goldstein J.1998. The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries. Proc. ACM SIGIR",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Learning to probabilistivally identify authoritative documents",
"authors": [
{
"first": "D",
"middle": [],
"last": "Cohn",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Chang",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of 18 th International Conference of Machine Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cohn D, Chang H.2001. Learning to probabilisti- vally identify authoritative documents. Proceedings of 18 th International Conference of Machine Learn- ing.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Indexing by Latent Semantic Analysis",
"authors": [
{
"first": "S",
"middle": [],
"last": "Deerwester",
"suffix": ""
},
{
"first": "S",
"middle": [
"T"
],
"last": "Dumais",
"suffix": ""
},
{
"first": "G",
"middle": [
"T"
],
"last": "Furnas",
"suffix": ""
},
{
"first": "T",
"middle": [
"K"
],
"last": "Landauer",
"suffix": ""
},
{
"first": "Harshman",
"middle": [
"R"
],
"last": "",
"suffix": ""
}
],
"year": 1990,
"venue": "Journal of the American Society of Information Science",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Deerwester S, Dumais ST, Furnas GT, Landauer TK, and Harshman R.1990. Indexing by Latent Semantic Analysis. Journal of the American Soci- ety of Information Science.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "LexPageRank: Prestige in Multi-Document Text Summarization",
"authors": [
{
"first": "G",
"middle": [],
"last": "Erkan",
"suffix": ""
},
{
"first": "D",
"middle": [
"R"
],
"last": "Radev",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Erkan G and Radev DR.2004. LexPageRank: Pres- tige in Multi-Document Text Summariza- tion.EMNLP.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Generic text Summarization using relevance measure and latent semantic analysis",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Gong",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2001,
"venue": "Proc ACM SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gong Y and Liu X.2001.Generic text Summariza- tion using relevance measure and latent semantic analysis.Proc ACM SIGIR.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Probabilistic Latent Semantic Indexing",
"authors": [
{
"first": "T",
"middle": [],
"last": "Hofmann",
"suffix": ""
}
],
"year": 1999,
"venue": "Twenty Second International ACM-SIGIR Conference on Informatioln Retrieval",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hofmann, T.1999.Probabilistic Latent Semantic Indexing. Twenty Second International ACM- SIGIR Conference on Informatioln Retrieval.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Unsupervised Learning from Dyadic Data",
"authors": [
{
"first": "",
"middle": [],
"last": "Hofmann",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hofmann, et al.1998. Unsupervised Learning from Dyadic Data. Technical Report TR-98-042, Inter- national Computer Science Insitute, Berkeley, CA.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Authoritative sources in a hyperlinked environment",
"authors": [
{
"first": "J",
"middle": [],
"last": "Kleinberg",
"suffix": ""
}
],
"year": 1998,
"venue": "Proc. 9th ACM-SIAM Symposium on Discrete Algorithms",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kleinberg J.1998. Authoritative sources in a hyper- linked environment. Proc. 9th ACM-SIAM Sym- posium on Discrete Algorithms.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Advances in Automatic Text Summarization",
"authors": [
{
"first": "I",
"middle": [],
"last": "Mani",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mani I. 1999. Advances in Automatic Text Sum- marization. MIT Press, Cambridge, MA, USA.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Language Independent Extractive Summarization",
"authors": [
{
"first": "R",
"middle": [],
"last": "Mihalcea",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mihalcea R.2005.Language Independent Extractive Summarization. AAAI",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "The Intelligent Surfer: Probabilistic Combination of Link and Content Information in PageRank",
"authors": [
{
"first": "M",
"middle": [],
"last": "Richardson",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Domingos",
"suffix": ""
}
],
"year": 2002,
"venue": "Advances in Neural Information Processing Systems",
"volume": "14",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Richardson M, Domingos P.2002. The Intelligent Surfer: Probabilistic Combination of Link and Con- tent Information in PageRank. Advances in Neural Information Processing Systems 14",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Aggregate and mixedorder Markov models for statistical language processing",
"authors": [
{
"first": "Saul",
"middle": [
"L"
],
"last": "Pereria",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 1997,
"venue": "Proc 21 st ACM-SIGIR International Conference on Research and Development in Information Retrieval",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Saul L and Pereria F.1997.Aggregate and mixed- order Markov models for statistical language proc- essing.Proc 21 st ACM-SIGIR International Confer- ence on Research and Development in Information Retrieval.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"text": "Effect of tempering factor B in the ROUGE-L score for PROC3.",
"num": null,
"type_str": "figure",
"uris": null
}
}
}
} |