main.dvi

Size: px

Start display at page:

Download "main.dvi"

てるえありの
5 years ago
Views:

1 LDA Estimating and Analyzing a Domain Topic Model of Entries Kensaku Makita 1 Hiroko Suzuki 1 Daichi Koike 1 Takehito Utsuro 2 Yasuhide Kawada 3 Abstract: In order to address the issue of quickly overviewing the distribution of the contents of the collection of blog posts, this paper proposes a framework of estimating a topic model, namely a domain topic model, which is a topic model estimated with the texts of entries extracted from the collection of blog posts. In this domain topic model of entries, each topic represents domains such as meteorology, astronomy, biology, energy, and industry, that are closely related to a query term, e.g., global warming. We compare the proposed approach of topic modeling with knowledge source and the standard topic modeling without knowledge source. Both topic modeling results have quite different nature and contribute to quickly overviewing the search result of blog posts in a quite complementary fashion. Keywords: blog,, topic model, LDA, topic analysis 1. 1 Graduate School of Systems and Information Engineering, University of Tsukuba, Tsukuba, , Japan 2 Faculty of Engineering, Information and Systems, University of Tsukuba, Tsukuba, , Japan 3 ( ) LOG WORKS Co., Ltd., Tokyo , Japan c 2012 Information Processing Society of Japan 1

2 (a) 1 (b) [15] [17] (LDA; Latent Dirichlet Allocation) [3] c 2012 Information Processing Society of Japan 2

3 LDA 1(b) 6 BT1 BT16 LDA LDA 1(b) 1(a) ( 1(a) ) 6 WT1 WT8 6 BT1 BT [13,16] (b) 2. TREC-2009 [11] 3 [6] () [9] LDA Web [7] Web [1,5,14] [2] c 2012 Information Processing Society of Japan 3

4 ( [4]) LDA [16] plsi [8] [10] LDA 3. t 0 t 0 BP(t 0 ) t 0 Yahoo!Japan API *1 t 0 8 * ,000 URL 200 t 0 t 0 BP(t 0 ) (LDA; Latent Dirichlet Allocation) [3] LDA w V w( V ) K z k (k =1,...,K) w P (w z k )(w V ) d z k P (z k d) (k =1,...,K) GibbsLDA++ *3 [12] LDA α β GibbsLDA++ α =50/K β =0.1 LDA K *1 *2 fc2.com yahoo.co.jp yaplog.jp ameblo.jp goo.ne.jp livedoor.jp Seesaa.net hatena.ne.jp * BP(t 0 ) BP(t 0 ) E t(e) BP(t 0 ) df 10 BP(t 0 ) E(BP(t 0 )) { } E(BP(t 0 )) = E df( BP(t 0 ), t(e) ) 10 E(BP(t 0 )) LDA w V *4 BP(t 0 ) zk e (k =1,...,K) zk e (k =1,...,K) w P (w zk e )(w V ) E zk e P (ze k E) (k =1,...,K) 4.3 [13] q d P (d q) d P (q) d d P (d) P (d q) = P (q d)p (d) P (q) P (q d)p (d) P (q d) d q P (q d) q d B( BP(t 0 )) B t(e) E B B * ,000 c 2012 Information Processing Society of Japan 4

5 B E E(BP(t 0 )) (E E(BP(t 0 ))) *5 B w B E B P (B E) P (B E) = w B P (w E) [16] P (w E) LDA P lda (w E) E w P ML (w E) E(BP(t 0 )) w P ML (w E(BP(t 0 )) ) *6 { P (w E) = λ μp ML (w E) + (1) } (1 μ)p ML (w E(BP(t 0 )) ) +(1 λ)p lda (w E) LDA P lda (w E) zk e P lda (w E) = K P (w zk e )P (ze k E) k=1 4.4 E B P (B E) B E 4.2 E zk e P (zk e E) (k =1,...,K) B zk e score e (B,zk e) *5 B E t(e) B freq(b, t(e)) 3 *6 E w P ML (w E) (BP(t 0 )) w P ML (w (BP(t 0 )) ) [16] λ = μ =0.7 1 t 0 BP(t 0) 8,834 1,402 7,199 1,823 4,211 score e (B,z e k)= E E 30(B) P (B E)P (z e k E) (2) B score e (B,zk e ) E P (B E) 30 E 30 (B) * (4.1 K = 50) BP(t 0 ) 60 * % 60% % 70% score e(b, zk e) 30 c 2012 Information Processing Society of Japan 5

6 2 : score e(b, z e k ) (= t 0 zk e ) (%) 82.5 (429/520) 78.7 (365/464) 71.8 (348/485) 80.1 (370/462) 72.2 (372/515) 2 : (1) 4.3 P (B E) ( 2 LDA ) (2) (1) P (B E) (1) P lda (w E) E w P ML (w E) E(BP(t 0 )) w P ML (w E(BP(t 0 )) ) (μ =0.7) ( 2 ) (3) E B [17] B E ( 2 tf-idf) % 0% 10%... 90% 100% LDA tf-idf LDA LDA tf-idf 10%... 90% 9 1% LDA tf-idf P (B E) BP(t 0 ) E(BP(t 0 )) w B E B E w 5.2 t 0 BP(t 0 ) 500 *8 B 4.4 (2) score e (B,zk e) ( zk e) *9 *8 500 B score e(b, zk e ) ( zk e) scoree(b, ze k ) BP(t 0) 500 *9 zk e z e k =argmax P (z k E) (3) z k E z e k c 2012 Information Processing Society of Japan 6

7 3 ID (= zk e ) ( : ) (1) WT1: WT2: WT3: WT4: 30 (B1) () B1 CO2 30 (B2) () B2 30 (B3) () B3 4 (= zk e ) ( : ) (2) WT5: ID B4 30 (B4) () WT6: WT7: WT8: 30 (B5) () B5 30 (B6) () B % * 10 (3) E *10 t 0 6 B1 B6 B score e (B,zk e ) 2 c 2012 Information Processing Society of Japan 7

8 5 t 0 (= z e k ) 3 4 ( ) * 11 zk e (3) E B E 30 (Bi) ( ) (i =1,...,6) B3 B4 B3 WT2 WT4 B4 WT5 WT8 B1 B2 B5 B6 BP(t 0 ) 6. 1(b) * LDA 1(b) w V 4.2 zk b (k =1,...,K) (4.1 K = 50) zk b P (z k B) z k zb k B B b (z b k ) B b (zk b )={ B z b k =argmaxp (z k B)} z k zk e score e (B,z k ) z k ze k B B e(zk e) B e (zk)= { e B z e k =argmaxscore e (B,z k) } z k z e i zj b Dice Dice Dice ( B e (zi e ), B b (zj) ) b = 2 Be (zi e) B b (zj b) Be (zi e) + Bb (zj b) 6 t WT1 WT8 zj b Dice 0.05 Be (zi e) B b (zj b) zj b ID BT1 BT16 B b (zk b) zj b zi e zj b c 2012 Information Processing Society of Japan 8

9 6 ( : ) zi e Dice ID e (zi e) zj b ID B e (z e i ) B b(z b j ) Dice B e (z e i ), B b(z b j ) WT1: e (z e i ) = 172 BT1: BT1: BT2: WT2: BT3: BT4: e (zi e ) =1, 974 BT5: BT6: BT7: BT8: WT3: BT7: e (zi e ) = 192 BT8: WT4: BT5: e (zi e ) = 255 BT9: BT10: WT5: e (z e i ) = 219 BT11: BT12: CO WT6: BT13: e (zi e ) = 588 BT14: BT15: WT7: BT1: e (zi e ) = 128 BT16: BT5: WT8: BT6: e (zi e ) = 396 BT9: BT12: CO (a) ( 7(a) ) ( ) ( ) ( c 2012 Information Processing Society of Japan 9

10 7 ( 50 ) (a) t (b) t ) ( 7(a) ) ( 7(a) ) 7(b) ( 7(b) ) ( 7(b) ) ( 7(b) ) [1] Vol. 50, No. 4, pp (2009). [2] Blei, D. M., Griffiths, T. L., Jordan, M. I. and Tenenbaum, J. B.: Hierarchical Topic Models and the Nested Chinese Restaurant Process, NIPS 03 (2003). [3] Blei, D. M., Ng, A. Y. and Jordan, M. I.: Latent Dirichlet Allocation, Journal of Machine Learning Research, Vol. 3, pp (2003). [4] Carmel, D., Roitman, H. and Zwerdling, N.: Enhancing Cluster Labeling Using, Proc. 32nd SIGIR, pp (2009). [5] de Winter, W. and de Rijke, M.: Identifying Facets in Query-Biased Sets of Blog Posts, Proc. ICWSM, pp (2007). [6] BLO- GRANGER, OIS pp (2006). [7] PLSI 16 pp (2010). [8] Hoffman, T.: Probabilistic Latent Semantic Indexing, Proc. 22nd SIGIR, pp (1999). c 2012 Information Processing Society of Japan 10

11 [9] Li, C., Yan, N., Roy, S. B., Lisham, L. and Das, G.: Facetedpedia: Dynamic Generation of Query-Dependent Faceted Interfaces for, Proc. 19th WWW, pp (2010). [10] Liu, X. and Croft, W. B.: Cluster-based Retrieval using Language Models, Proc. 27th SIGIR, pp (2004). [11] Macdonald, C., Ounis, I. and Soboroff, I.: Overview of the TREC-2009 Blog Track, Proc. TREC-2009 (2009). [12] Phan, X.-H. and Nguyen, C.-T.: GibbsLDA++: A C/C++ implementation of latent Dirichlet allocation (LDA) (2007). [13] Ponte, J. M. and Croft, W. B.: A Language Modeling Approach to Information Retrieval, Proc. 21st SIGIR, pp (1998). [14] Vol. 46, No. SIG 13(TOD 27), pp (2005). [15] Tunkelang, D.: Faceted Search, Synthesis Lectures on Information Concepts, Retrieval, and Services, Morgan & Claypool Publishers (2009). [16] Wei, X. and Croft, W. B.: LDA-Based Document Models for Ad-hoc Retrieval, Proc. 29th SIGIR, pp (2006). [17] Yokomoto, D., Makita, K., Utsuro, T., Kawada, Y. and Fukuhara, T.: Utilizing in Categorizing Topic related Blogs into Facets, Procedia - Social and Behavioral Sciences, Vol. 27, pp (2011). c 2012 Information Processing Society of Japan 11

main.dvi

main.dvi DEIM Forum 2012 E2-4 1 2 2 2 3 4 5 6 7 1 305-8573 1-1-1 2 305-8573 1-1-1 3 305-8573 1-1-1 4 ( ) 141-0031 8-3-6 5 060-0808 8 5 6 101-8430 2-1-2 7 135-0064. 2-3-26 113-0033 7-3-1 305-8550 1-2 Analyzing Correlation