main.dvi

Size: px

Start display at page:

Download "main.dvi"

ゆりなやぶき
5 years ago
Views:

1 THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEERS TECHNICAL REPORT OF IEICE. ( ) Estmaton and Analyss of Topc Models n Tme Seres Japanese / Chnese News and Blogs Shuo HU,LyZHENG, Yusuke TAKAHASHI,DachKOIKE,KensakuMAKITA,Takehto UTSURO, and Masaharu YOSHIOKA Grad. Sch. of Systems and Informaton Engneerng, Unversty of Tsukuba Faculty of Engneerng, Informaton and Systems, Unversty of Tsukuba Grad. Sch. of Informaton Scence and Technology, Hokkado Unversty Abstract Ths paper studes ssues regardng topc modelng of nformaton flow n multlngual news and blog streams. If someone wants to fnd dfferences n the topcs of Japanese and Chnese news / blog, t s usually necessary for hm/her to carefully watch every artcle n Japanese and Chnese news streams at every moment. In such a stuaton, topc models such as LDA (Latent Drchlet Allocaton) and DTM (dynamc topc model) are qute effectve n estmatng dstrbuton of topcs over a document collecton such as artcles n a news / blog stream. Based on the results of estmatng dstrbuton of topcs n Japanese / Chnese news streams, ths paper proposes how to analyze cross-lngual algnment of topcs n tme seres Japanese / Chnese news streams. Ths paper also proposes how to analyze dfferences of concerns and opnons among varous Chnese news stes as well as blog hosts. Key words news, blog, topc model, tme seres analyss 1. (LDA, Latent Drchlet Allocaton) [1] DTM(dynamc topc model) [2] 1

2 1 2 ( ) 3. 1 DTM Wkpeda 2

3 3 ( ) 4 ( ) LDA LDA 3

4 (LDA, Latent Drchlet Allocaton) DTM(dynamc topc model) w K z n(n =1,...,K) w p(w z n)(w V ) b z n p(z n b)(n =1,...,K) V DTM LDA p(w z n)(w V ) p(z n b)(n =1,...,K) LDA GbbsLDA++ 1 DTM 2 α β K LDA α =2.5 β =0.1 K =20. DTM α =0.01 β =0.1 K = D K d(d D) z n(n =1,...,K) D(z n) D(z n)= d D z n = argmax p(z u d) z u(u=1,...,k) d d θ JC ( θ JC 10 ) Wkpeda Wkpeda , ,595 78,519 N JC(d J,d C) N JC(d J,d C)= J, S ( JS W ) J d J S d C d J d C θ JC DD JC(θ JC) DD CJ(θ JC) DD JC(θ JC) = DD CJ(θ JC) = d J,d C N JC(d J,d C) > = θ JC, d C =argmaxn JC(d J,d C) d C d J,d C N JC(d J,d C) > = θ JC, d J =argmaxn JC(d J,d C) d J 3. 2 DD JC DD CJ d J d C DTM TTJ TTC TTJ TTC t J( TTJ) P (t J d J) > = θ t (θ t =0.6) d J t C( TTC) P (t C d C) > = θ t (θ t =0.6) d C t J t C P (t J d J) > = θ t P (t C d C) > = θ t d J,d C DD JC(θ JC) DD CJ(θ JC) M JC(t J,t C,θ t,θ JC) TTJ t C TTC t J TA J (t C,TT J,θt,θ JC)= ( max M JC (t J,t C,θ t,θ JC )=1) t J TT J argmax M JC (t J,t C,θ t,θ JC ) t J TT J ( max M JC (t J,t C,θ t,θ JC ) > = 2) t J TT J TA C (t J,TT C,θt,θ JC)= ( max M JC (t J,t C,θ t,θ JC )=1) t C TT C argmax M JC (t J,t C,θ t,θ JC ) t C TT C ( max M JC (t J,t C,θ t,θ JC ) > = 2) t C TT C

5 ( IMF ( 10 1 ) ( ) ( ) : 80.0% (4/5) 66.7% (4/6) DTM Q J = TTJ 1,TTJ 2,...,TTJ n Q C = TTC,TT 1 C,...,TT 2 J n TTJ 1 1 t J TT J TA C(t J,TTC,θ t,θ JC) t C TTC TA J(t C,TTJ,θ t,θ JC) =1,...,n , , , ,278 6, % ( ) 53.3% ( ) % ( ) 86.6% ( ) [3] [4] 791 1,361 DTM ( )

6 8 Sna ,937 1,446 LDA Yahoo! Search BOSS API 10 Sna ,937 1,446 LDA Sna 159 Sna Sna [5] LDA On-lne LDA [6] [7] DTM [8] [1] D. M. Ble, A. Y. Ng, and M. I. Jordan. Latent Drchlet allocaton. Journal of Machne Learnng Research, Vol.3, pp , [2] D. M. Ble and J. D. Lafferty. Dynamc topc models. In Proc. 23rd ICML, pp , [3] J. Klenberg. Bursty and herarchcal structure n streams. In Proc. 8th SIGKDD, pp , [4],,,,,,,,.. 4 DEIM, [5],.. 18, pp , [6] L. AlSumat, D. Bardara, and C. Domencon. On-Lne LDA: Adaptve topc models for mnng text streams wth applcatons to topc detecton and trackng. In Proc. 8th ICDM, pp. 3 12, [7],,,,,,,,,,,.. 4 DEIM, [8],,,,. Wkpeda., Vol. 25, No. 5, pp ,

main.dvi

main.dvi DEIM Forum 2015 D3-1 305-8573 1-1-1 305-8573 1-1-1 ( ) 151-0051 5-13-18 101-8430 2-1-2.com,,,, Market Share Estimation based on Statistics of Search Engine Suggests Takakazu IMADA,IchiroMORIYA, Yusuke