2007/2 Vol. J90 D No Web 2. 1 [3] [2], [11] [18] [14] YELLOW [16] [8] tfidf [19] 2. 2 / 30% 90% [24] 2. 3 [4], [21] 428

Size: px

Start display at page:

Download "2007/2 Vol. J90 D No Web 2. 1 [3] [2], [11] [18] [14] YELLOW [16] [8] tfidf [19] 2. 2 / 30% 90% [24] 2. 3 [4], [21] 428"

ひろといさやま
5 years ago
Views:

1 Informative Summarization Method by Key Sentences Extraction Considering Sub-Topics Naoki SAGARA, Wataru SUNAYAMA, and Masahiko YACHIDA WWW World Wide Web Web [15] Graduate School of Engineering Science, Osaka University, Toyonaka-shi, Japan Faculty of Information Sciences, Hiroshima City University, Hiroshima-shi, Japan indicative informative [7] D Vol. J90 D No. 2 pp c

2 2007/2 Vol. J90 D No Web 2. 1 [3] [2], [11] [18] [14] YELLOW [16] [8] tfidf [19] 2. 2 / 30% 90% [24] 2. 3 [4], [21] 428

3 1 Fig. 1 Story model. MMR Maximal Marginal Relevance [5] [20] [23]

4 2007/2 Vol. J90 D No [25] [26] % 36/ Fig. 2 Informative summarization system considering sub-topics. 2 / 3. 1 Chasen [6] STEP1 1 n S n S (1) freq p(n) mean p std p p n freq p(n) mean p + std p (1) 2 430

5 3 (2) freq p(n) mean p (2) STEP2 S 1 S 2 n (3) (4) freq text(n) subjectf req text(n) n val1(n) = val2(n) = subjectf req text(n) max w S1 {subjectf req text(w)} (3) freq text(n) max w S2 {freq text(w)} (4) STEP3 n (5) (6) key(n) =val1(n)+val2(n) (5) implevel(n) = key(n) max w S1 S 2 {key(w)} (6) STEP4 implevel(n) <implevel(n) < 0.7 n STEP1 1 2 (1) 1 3 (6) Fig. 3 Key sentences extraction module [22] S w (7) n(w) w (8) F-measure 431

6 2007/2 Vol. J90 D No. 2 n(w s) 2 view(w) = (7) n(w)n(s) s S feat(w) = s S n(w s) n(w) (8) s k rank text(s, k) p p s k rank p(s, k) k s rank text(s, k) s p rank p(s, k) [0, 1] (9) rank text(s, k) p rank p(s, k) s k story(s, k) 4 story(s, k) =rank text(s, k)+rank p(s, k) (9) story(s, k) rate k (9) p (10) N p N p = linenumber rate 1 2 linep(k) mainline text (10) linenumber mainline text line p(k) p k k 1/2 1/ /2 432

7 (6) Seg. Sent. Web SitePlots A 1 30% Plots 2 Label A 1 A Table 1 Texts for the experiments. Text Seg. Sent. Web Site Plots 4 37 Logos [10] J-TEXTS [9] [1] [12] 3 50 Rieti [17] Rieti Rieti [13] Label A B C D E F G H I J K L M N 2 Table 2 Plots of Urashima Taro story. Plots A

8 2007/2 Vol. J90 D No. 2 3 Table 3 Keywords for Urashima Taro story. Keywords Key value % Table 4 Keywords for summarization. (Summarization Rate = 30%) Text Main keywords [Sub keywords] [ ] [] [ ] [ ] [] [ ] [] [] [ ] [ ] 4. [1] STEP STEP % 80% Proposal 434

9 5 6 Table 5 Correction set of Main keywords and sub keywords. (Underline means the words not included in Table 6) Text Main keywords Sub keywords 6 5 Table 6 Main keywords and sub keywords by the proposed system. (Underline means the words not included in Table 5) Text Main keywords Sub keywords Main PVS [22] Freq MMI MMR Importance [20] Naive 2 2 MMI (11) 7 S A Imp(S i) S i 1 Sim(S i,s j) S i S j 8 MMI = max S i S Ā (Imp(S i) α max Sim(S i,s j)) S j A (11) 1 20% 25% 30% α =2.0 8 S i S j S i S j % 2 54%

10 2007/2 Vol. J90 D No. 2 7 Table 7 Averaged recall value of plots. Sum.rate20% Sum.rate25% Sum.rate30% AVG STD AVG STD AVG STD Proposal Main PVS Freq MMI Naive Plots 10 k n 7 recall(n, k) 11 Output outlinerecall (12) K outlinerecall = 1 K K max n Output k=1 {recall(n, k)} (12) outlinep recision (13) N outlinep recision = 1 N 5. 2 K N recall(n, k) k=1 n=1 (13) % % 10% 4 Fig. 4 Comparison between summarization rate and recall value of plots. 8 30% Table 8 Recall value of plots for each text. (summarization rate = 30%) Text Prop. Main PVS Freq MMI Naive % 8 30% 1020% 25% 100% 30% 100% 25% 97.4% 20% 88.5%

11 20% % 7 30% 82%9.4% 8 MMI MMI 8 30% MMI MMI MMI MMI MMI 5 Fig. 5 Comparison between summarization rate and precision value of plots. 5 (13) 20% MMI Main MMI Naive

12 2007/2 Vol. J90 D No % 80% 3 2 [1] [2] vol.20, no.3, pp , [3] R. Barzilay and M. Elhadad, Using lexical chains for text summarization, Advances in Automatic Text Summarization, pp.1 12, The MIT Press, London, [4] E. Boros, P.B. Kantor, and D.J. New, A clustering based approach to creating multi-document summaries, Proc. ACM SIGIR Workshop on Text Summarization, pp.1 4, [5] J. Carbonell and J. Goldstein, The use of MMR, diversity-based reranking for reordering documents and producing summaries, Proc. 21st ACM-SIGIR Conference on Research and Development in Information Retrieval, pp , [6] version2.0 NAIST-IS-TR99012, [7] T. Hand, Proposal for task-based evaluation of text summarization systems, Proc. ACL Workshop on Intelligent Scalable Text Summarization, pp.31 38, [8] 160 pp.43 48, [9] J-TEXTS, [10] Logos, [11] D. Marcu, Discourse trees are good indicators of importance in text, in Advances in Automatic Text Summarization, pp , MIT Press, [12] [13] [14] 120 pp.71 76, [15] vol.9, no.4, pp , [16] YELLOW vol.43, no.sig2 (TOD 13), pp.37 47, [17] Rieti, [18] G. Salton, A. Singhal, M. Mitra, and C. Buckley, Automatic text structuring and summarization, Inf. Process. Manage., vol.33, no.2, pp , [19] G. Salton and C. Buckey, Term-weighting approaches in automatic text retrieval, in Readings in Information Retrieval, ed. K.S. Jones and P. Willett, pp , Morgan Kaufmann Publishers, San Francisco, [20] MMR 438

13 9 pp , [21] Y. Seki, K. Eguchi, and N. Kando, User-focused multi-document summarization with paragraph clustering and sentence-type filtering, Proc. Fourth NT- CIR Workshop on Research in Information Access Technologies Information Retrieval, Question Answering and Summarization, pp , [22] vol.17, no.1, pp.14 22, [23] vol.18, no.2, pp , [24] vol.12, no.1, pp.51 78, [25] Y. Nakao, An algorithm for one-page summarization of a long text based on thematic hierarchy detection Full text, Proc. 38th Annual Meeting on Association for Computational Linguistics, pp , [26] T. Nomoto and Y. Matsumoto, A new approach to unsupervised text summarization, Proc. 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp.26 34, Table A 1 A 1 Summarization for Urashima Taro story. Plots Extracted Key Sentences A(7/7) [ ] A(7/7) B(6/7) B(7/7) [ ] C(7/7) D(7/7) [] Plots E(5/7) E(7/7) F(7/7) F(3/7) G(7/7) G(5/7) H(3/7) H(7/7) I(7/7) I(2/7) Extracted Key Sentences [] [ ] [ ] [ ] [ ] [ ] [] [ ] 439

14 2007/2 Vol. J90 D No. 2 Plots J(7/7) J(2/7) K(4/7) L(7/7) L(7/7) M(7/7) N(4/7) Extracted Key Sentences [ ] [] [ ] [ ]

1 1 tf-idf tf-idf i

14 A Method of Article Retrieval Utilizing Characteristics in Newspaper Articles 1055104 2003 1 31 1 1 tf-idf tf-idf i Abstract A Method of Article Retrieval Utilizing Characteristics in Newspaper Articles