Microsoft Word - 5 魏志生.doc

Size: px

Start display at page:

Download "Microsoft Word - 5 魏志生.doc"

瑜润奚
5 years ago
Views:

1 ISSN CODEN JKYTA8 E-mal: Journal of Fronters of Computer Scence and Technology /2011/05(12) Tel: DOI: /j.ssn * 魏志生, 吉阳生 +, 罗春勇, 陈家骏南京大学计算机软件新技术国家重点实验室, 南京 Generatve Sentment Classfcaton Model Afflatng Doman-Specfc Sentment Lexcons * WEI Zhsheng, JI Yangsheng +, LUO Chunyong, CHEN Jajun State Key Laboratory for Novel Software Technology, Nanjng Unversty, Nanjng , Chna + Correspondng author: E-mal: jys@nlp.nju.edu.cn WEI Zhsheng, JI Yangsheng, LUO Chunyong, et al. Generatve sentment classfcaton model afflatng doman-specfc sentment lexcons. Journal of Fronters of Computer Scence and Technology, 2011, 5(12): Abstract: Sentment classfcaton focuses on learnng a classfcaton model from labeled opnon text, whch can predct the sentment polarty n other opnon text. In ths feld, one mportant research topc s to combne the pror knowledge wth a generatve classfer to construct a new model. By studyng the doman character and weght of sentment lexcons, ths paper proposes an approach to automatcally construct doman-specfc sentment lexcons whch can be reformulated as a generatve pror, and then combne t wth a generatve model. Expermental results show that the proposed generatve model performs sgnfcantly and consstently better than some of the-state-ofthe-art doman-ndependent methods. Key words: sentment classfcaton; pror knowledge; doman-specfc sentment lexcon; sentment lexcon 情感分类是通过分析数据中的情感信息, 来预测数据所传递的情感倾向其中结合语言学词典与产生式分类器构造带有先验知识的分类模型, 是一类重要的研究课题通过研究情感词的领域性和不同权重的特性, 提出了一种新的融入情感先验知识的情感分类方法通过自动分析构造领域相关的情感词及其权重信息, 将其作为情感先验知识, 融入到产生式分类模型中, 得到更适合特定领域的分类模型实验结果 *The Natonal Grand Basc Research 973 Program of Chna under Grant No. 2010CB ((973)). Receved , Accepted

2 1106 Journal of Fronters of Computer Scence and Technology 2011, 5(12) 表明, 该方法在分类性能上, 显著并一致地优于其他结合了领域无关先验知识的产生式分类模型情感分类 ; 先验知识 ; 领域情感词 ; 情感词 A TP Web 2.0,,, ;, () [1 4],, ( ) (),,,,,,,,, good bad ;,,, long_hour,,,,,, (opnon retreval) [5 6] (expresson level sentment classfcaton) [7] Melvlle [8], Naïve Bayes,,,,,,,,, (Naïve Bayes),, 2 ; 3 ; 4 ; 5, ;,,, 2,, [9 10], a P( + D) =, a b a+ b P( + D) > t, ;, t,, t = 0.5,

3 1107,, Lexcal Classfer, (baselne), Pang [2], (bag of words),, Pang [11],,,,,, [12 13], Wu [14],,, Weghted Margn Support Vector Machne Melvlle [8] (poolng multnomals classfer), Naïve Bayes,,, Melvlle [8], ;,,, [15 17], (Naïve Bayes) LnearDomanPoolng, Melvlle [8] 3, Pang [2],,, Melvlle [8],,, (multnomal Naïve Bayes classfer), [18] (1) ; (2) ; (3),, P( D) = P( D cj ) P( cj ) j, P( c j ) c j ; P( D cj ) D c j C, { +, },, ( w ), D c j P( D cj ) = P( w cj), D P( cj ) P( w cj) Pc ( j D)= PD ( ) D Classfy(D)=arg max P( c ) P( w c ) cj j j [8]

4 1108 Journal of Fronters of Computer Scence and Technology 2011, 5(12), ;, ( ) (lnear opnon pool) [8] K P( w c ) = α P ( w c ) j k k j k = 1, K ; Pk( w cj) k w c ; α 1 1 errk αk = lg err k, errk k α k P( c j ) P( w cj) P( c j ),, ; Pg ( w cj), Pd( w cj), P( w c ) P( w c ) = βp ( w c ) + (1 β) P ( w c ) j j d j g j, β, β,,, 0.6 j k 4.2 Pg ( w c j), [8], 1 Pg( w+ + ) = Pg( w ) = p+ n 1 1 Pg( w + ) = Pg( w+ ) = p+ n r n(1 1/ r) Pg( wu + ) = Pg( wu ) = ( p+ n)( m p n), p ; n ; m ; r, 0< 1/ r 1,, r = P ( w c ) d j,,,,, (mutual nformaton, MI), MI Pwordc (, j ) MI( word) = P( word, c j )lb c ( ) ( ) j C P word P c j, P( word, c j ) word c j 1 (1) 10 (2) MI

5 1109 (3) MI 10% (4) (pontwse mutual nformaton, PMI), PMI (5) MI, weght( word) = MI ( word) 4.4, [19], Pwords ; Nwords ; A( word1, word2) word1 word2 SO- A( word) = A( word, pword) Awordnword (, ), A( word1, word2), word1 word2 SO- A( word), word, PMI PMI( word1, word2) lb p( word, word ) 1 2 = pword ( 1) pword ( 2), p( word1, word2) word1 word2,, PMI( word, word ) SO-A 1 2 A( word1, word 2), SO-A SO- PMI ( word) = PMI ( word, pword) SO-PMI > 0 PMI( word, nword), word ; SO-PMI < 0, word, ( good)( badly) 2, (1) SO- PMI ( word) (2) max ( SO- PMI( word )) > 0, Pwords, ; mn ( SO- PMI ( word )) < 0, Nwords, (3) (1) (2), SO-PMI > 0 SO-PMI < 0, word, weght( word) = weght( pword) PMI ( word, pword) PMI( word, pword) word, weght( word) = weght( nword) PMI( word, nword) PMI( word, nword), weght( w+ ) Pd ( w+ + ) =, w+ Pwords weght( pword) weght( w ) Pd ( w ) =, w Nwords weght( nword) P ( w ) = P ( w + ) = P ( w + ) = P ( w ) = 0 d + d d u d u

6 1110 Journal of Fronters of Computer Scence and Technology 2011, 5(12) Mult-Doman Sentment Dataset [20] amazon Book DVD Electronc Ktchen 1~5 3, 3, 3,, DF 5 (Book, DVD, Electronc, Ktchen),, 1 Table 1 Data nformaton 1 Book DVD Electronc Ktchen MPQA(multperspectve queston answerng) [21] Lu [4] 2 Table 2 Global sentment lexcon nformaton 2 MPQA Lu Naïve Bayes Lexcal Classfer LnearPoolng [8] LnearDomanPoolng Naïve Bayes, Pw ( c) = j t t j j + ε + ε V, tj w c j ; ε = 1 LnearPoolng LnearDomanPoolng, r = 100, 100 LnearDomanPoolng, β = Table 3 Classfcaton performance usng MPQA lexcon 3 MPQA (%) Book DVD Electronc Ktchen Lexcal Naïve Bayes LnearPoolng LnearDomanPoolng Table 4 Classfcaton perfomance usng Lu lexcon 4 Lu (%) Book DVD Electronc Ktchen Lexcal Naïve Bayes LnearPoolng LnearDomanPoolng ,, LnearPoolng (Naïve Bayes) ; LnearDomanPoolng,, Lexcal Classfer,,

7 1111 Fg.1 Classfcaton performance wth respect to dfferent numbers of tranng data on Book doman 1 Book Fg.2 Classfcaton performance wth respect to dfferent numbers of tranng data on DVD doman 2 DVD Fg.3 Classfcaton performance wth respect to dfferent numbers of tranng data on Electronc doman 3 Electronc,,,, 3 4, MPQA [21] Lu [4],, Lu [4] 1~ 4, Lu [4],, Fg.4 Classfcaton performance wth respect to dfferent numbers of tranng data on Ktchen doman 4 Ktchen,,,, LnearPoolng LnearDomanPoolng,,, 6 (1),,

8 1112 Journal of Fronters of Computer Scence and Technology 2011, 5(12) (2),,,,,, References: [1] Hatzvassloglou V, Mckeown K R. Predctng the semantc orentaton of adjectve[c]//proceedngs of the 8th Conference on European Chapter of the Assocaton for Computatonal Lngustcs (EACL 97), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 1997: [2] Pang Bo, Lee L, Vathyanathan S. Thumbs up?: sentment classfcaton usng machne learnng technques[c]//proceedngs of the ACL-02 Conference on Emprcal Methods n Natural Language Processng (EMNLP 02), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2002: [3] Turney P. Thumbs up or thumbs down? semantc orentaton appled to unsupervsed classfcaton of revews[c]// Proceedngs of the 40th Annual Meetng of the Assocaton for Computatonal Lngustcs (ACL 02), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2002: [4] Hu Mngqng, Lu Bng. Mnng and summarzng customer revews[c]//proceedngs of the 10th ACM SIGKDD Internatonal Conference on Knowledge Dscovery and Data Mnng (KDD 04). New York, NY, USA: ACM, 2004: [5] Jjkoun V, Rjke M D, Weerkamp W. Generatng focused topc-specfc sentment lexcons[c]//proceedngs of the 48th Annual Meetng of the Assocaton for Computatonal Lngustcs (ACL 10), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2010: [6] Na S H, Lee Y, Nam S H, et al. Improvng opnon retreval based on query-specfc sentment lexcon[c]//proceedngs of the 31st European Conference on IR Research on Advances n Informaton Retreval (ECIR 09), Toulouse, France, Berln, Hedelberg: Sprnger-Verlag, 2009: [7] Cho Y, Carde C. Adaptng a polarty lexcon usng nteger lnear programmng for doman-specfc sentment classfcaton[c]//proceedngs of the 2009 Conference on Emprcal Methods n Natural Language Processng (EMNLP 09), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2009: [8] Melvlle P, Gryc W, Lawrence R D. Sentment analyss of blogs by combnng lexcal knowledge wth text classfcaton[c]//proceedngs of the 15th ACM SIGKDD Internatonal Conference on Knowledge Dscovery and Data Mnng (KDD 09). New York, USA: ACM, 2009: [9] Ng V, Dasgupta S, Arfn S M N. Examnng the role of lngustc knowledge sources n the automatc dentfcaton and classfcaton of revews[c]//proceedngs of the COLING/ACL on Mal Conference Poster Sessons (COLING-ACL 06), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2006: [10] Kamps J, Marx M J, Mokken R J, et al. Usng wordnet to measure semantc orentaton of adjectves[c]//proceedngs of the 4th Internatonal Conference on Language Resources and Evaluaton (LREC 04), Pars, France, 2004: [11] Pang Bo, Lee L. A sentmental educaton: sentment analyss usng subjectvty summarzaton based on mnmum cuts[c]//proceedngs of the 42nd Annual Meetng on Assocaton for Computatonal Lngustc (ACL 04), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2004: [12] Schapre R E, Rochery M, Rahm M, et al. Incorporatng pror knowledge nto boostng[c]//proceedngs of the 19th Internatonal Conference on Machne Learnng (ICML 02), Sydney, Australa, San Francsco, CA, USA: Morgan Kaufmann Publshers Inc, 2002: [13] Druck G, Mann G, McCallum A. Learnng from labeled features usng generalzed expectaton crtera[c]//proceedngs of the 31st Annual Internatonal ACM SIGIR Conference on Research and Development n Informaton Retreval (SgIR 08). New York, NY, USA: ACM, 2008: [14] Wu Xaoyun, Srhar R. Incorporatng pror knowledge

1113 wth weghted margn support vector machnes[c]//proceedngs of the 10th ACM SIGKDD Internatonal Conference on Knowledge Dscovery and Data Mnng (KDD 04). New York, NY, USA: ACM, 2004: 326 333.

Extractng semantc orentatons of words usng spn model[c]//proceedngs of the 43rd Annual Meetng on Assocaton for Computatonal Lngustcs (ACL 05), Phladelpha, USA, 2005.

Journal of Computer Scence and Technology, 2011, 26(1): 81 98. [18] Ngam K P. Usng unlabeled data to mprove text classfcaton[d]. Carnege Mellon Unversty, 2001. [19] Turney P, Lttman M L.

Bographes, bollywood, boom-boxes and blenders: doman adaptaton for sentment classfcaton[c]//proceedngs of the 45th Annual Meetng of the Assocaton of Computatonal Lngustcs (ACL 07), Prague, Czech

9 1113 wth weghted margn support vector machnes[c]//proceedngs of the 10th ACM SIGKDD Internatonal Conference on Knowledge Dscovery and Data Mnng (KDD 04). New York, NY, USA: ACM, 2004: [15] Webe J. Learnng subjectve adjectves from corpora[c]// Proceedngs of the 17th Natonal Conference on Artfcal Intellgence (AAAI 00), Texas, USA, 2000: [16] Takamura H, Inu T, Okumura M. Extractng semantc orentatons of words usng spn model[c]//proceedngs of the 43rd Annual Meetng on Assocaton for Computatonal Lngustcs (ACL 05), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2005: [17] J Yangsheng, Chen Jajun, Nu Gang, et al. Transfer learnng va mult-vew prncpal component analyss[j]. Journal of Computer Scence and Technology, 2011, 26(1): [18] Ngam K P. Usng unlabeled data to mprove text classfcaton[d]. Carnege Mellon Unversty, [19] Turney P, Lttman M L. Measurng prase and crtcsm: nference of semantc orentaton from assocaton[j]. ACM Transactons on Informaton Systems (TOIS), 2003, 21(4): [20] Bltzer J, Dredze M, Perera F. Bographes, bollywood, boom-boxes and blenders: doman adaptaton for sentment classfcaton[c]//proceedngs of the 45th Annual Meetng of the Assocaton of Computatonal Lngustcs (ACL 07), Prague, Czech Republc, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2007: [21] Wlson T, Webe J, Hoffmann P. Recognzng contextual polarty n phrase-level sentment analyss[c]//proceedngs of the Conference on Human Language Technology and Emprcal Methods n Natural Language Processng (HLT 05), Phladelpha, USA, Stroudsburg, PA, USA: Assocaton for Computatonal Lngustcs, 2005: WEI Zhsheng was born n He s a master canddate at Nanjng Unversty. Hs research nterests nclude text mnng and natural language processng. (1986 ),,,,, JI Yangsheng was born n He s a Ph.D. canddate at Nanjng Unversty. Hs research nterests nclude sentment analyss, revew summarzaton, transfer learnng and natural language processng. (1984 ),,,,,,, LUO Chunyong was born n He s a master canddate at Nanjng Unversty. Hs research nterests nclude machne learnng and natural language processng. (1987 ),,,,, CHEN Jajun was born n He receved hs Ph.D. degree from Nanjng Unversty n Now he s a professor and doctoral supervsor at Department of Computer Scence and Technology, Nanjng Unversty. Hs research nterests nclude natural language processng and software engneerng. (1963 ),,, 1998,,,

基于词语关联度的查询缩略*

基于词语关联度的查询缩略* * 基于词语关联度的查询缩略陈炜鹏 1, 付瑞吉 1, 胡熠 2, 秦兵 1, 刘挺 (1. 哈尔滨工业大学计算机科学与技术学院社会计算与信息检索研究中心, 黑龙江省哈尔滨市,150001; 2. 腾讯公司搜索平台部, 广东省深圳市,518057) 摘要 : 冗