A Joint Segmentation and Classification Framework for Sentiment Analysis Duyu Tang ♮ ∗ , Furu Wei ‡ , Bing Qin ♮ , Li Dong ♯ ∗ , Ting Liu ♮ , Ming Zhou ‡ ♮ Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, China ‡ Microsoft Research, Beijing, China ♯ Beihang University, Beijing, China ♮ { dytang, qinb, tliu } @ir.hit.edu.cn ‡ { fuwei, mingzhou } @microsoft.com ♯ donglixp@gmail.com Abstract timent classification as a special case of text cate- gorization task. Under this perspective, previous In this paper, we propose a joint segmenta- studies typically use pipelined methods with two tion and classification framework for sen- steps. They first produce sentence segmentation- timent analysis. Existing sentiment clas- s with separate text analyzers (Choi and Cardie, sification algorithms typically split a sen- 2008; Nakagawa et al., 2010; Socher et al., 2013b) tence as a word sequence, which does not or bag-of-words (Paltoglou and Thelwall, 2010; effectively handle the inconsistent senti- Maas et al., 2011). Then, feature learning and sen- ment polarity between a phrase and the timent classification algorithms take the segmenta- words it contains, such as “ not bad ” and tion results as inputs to build the sentiment classi- “ a great deal of ”. We address this issue fier (Socher et al., 2011; Kalchbrenner et al., 2014; by developing a joint segmentation and Dong et al., 2014). classification framework ( JSC ), which si- The major disadvantage of a pipelined method multaneously conducts sentence segmen- is the problem of error propagation, since sen- tation and sentence-level sentiment classi- tence segmentation errors cannot be corrected by fication. Specifically, we use a log-linear the sentiment classification model. A typical kind model to score each segmentation candi- of error is caused by the polarity inconsistency be- date, and exploit the phrasal information tween a phrase and the words it contains, such of top-ranked segmentations as features to as � not bad , bad � and � a great deal of , great � . build the sentiment classifier. A marginal The segmentations based on bag-of-words or syn- log-likelihood objective function is de- tactic chunkers are not effective enough to han- vised for the segmentation model, which dle the polarity inconsistency phenomenons. The is optimized for enhancing the sentiment reason lies in that bag-of-words segmentations re- classification performance. The joint mod- gard each word as a separate unit, which losses el is trained only based on the annotat- the word order and does not capture the phrasal ed sentiment polarity of sentences, with- information. The segmentations based on syntac- out any segmentation annotations. Experi- tic chunkers typically aim to identify noun group- ments on a benchmark Twitter sentimen- s, verb groups or named entities from a sentence. t classification dataset in SemEval 2013 However, many sentiment indicators are phrases show that, our joint model performs com- constituted of adjectives, negations, adverbs or id- parably with the state-of-the-art methods. ioms (Liu, 2012; Mohammad et al., 2013a), which are splitted by syntactic chunkers. Besides, a bet- 1 Introduction ter approach would be to utilize the sentiment in- formation to improve the segmentor. Accordingly, Sentiment classification, which classifies the senti- the sentiment-specific segmentor will enhance the ment polarity of a sentence (or document) as posi- performance of sentiment classification in turn. tive or negative, is a major research direction in the field of sentiment analysis (Pang and Lee, 2008; In this paper, we propose a joint segmentation Liu, 2012; Feldman, 2013). Majority of existing and classification framework ( JSC ) for sentimen- approaches follow Pang et al. (2002) and treat sen- t analysis, which simultaneous conducts sentence segmentation and sentence-level sentiment clas- ∗ This work was partly done when the first and fourth sification. The framework is illustrated in Fig- authors were visiting Microsoft Research. 477 Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 477–487, October 25-29, 2014, Doha, Qatar. c � 2014 Association for Computational Linguistics
CG SC SEG SEG SC that is not bad -1 <+1,-1> NO 0.6 0.6 <+1,-1> NO that is not bad -1 0.4 0.4 that is not bad that is not bad <+1,+1> YES 2.3 2.3 +1 Polarity: +1 Top K 1.6 1.6 that is not bad +1 <+1,+1> YES Input Segmentations Polarity Update Rank Update Figure 1: The joint segmentation and classification framework ( JSC ) for sentiment classification. CG represents the candidate generation model, SC means the sentiment classification model and SEG stands for the segmentation ranking model. Down Arrow means the use of a specified model, and Up Arrow indicates the update of a model. 2 Related Work ure 1. We develop (1) a candidate generation mod- el to generate the segmentation candidates of a sentence, (2) a segmentation ranking model to s- Existing approaches for sentiment classification core each segmentation candidate of a given sen- are dominated by two mainstream directions. tence, and (3) a classification model to predic- Lexicon-based approaches (Turney, 2002; Ding t the sentiment polarity of each segmentation. The et al., 2008; Taboada et al., 2011; Thelwall et phrasal information of top-ranked candidates from al., 2012) typically utilize a lexicon of sentiment the segmentation model are utilized as features to words, each of which is annotated with the sen- build the sentiment classifier. In turn, the predict- timent polarity or sentiment strength. Linguis- ed sentiment polarity of segmentation candidates tic rules such as intensifications and negations are from classification model are leveraged to update usually incorporated to aggregate the sentimen- the segmentor. We score each segmentation can- t polarity of sentences (or documents). Corpus- didate with a log-linear model, and optimize the based methods treat sentiment classification as a segmentor with a marginal log-likelihood objec- special case of text categorization task (Pang et al., tive. We train the joint model from sentences an- 2002). They mostly build the sentiment classifier notated only with sentiment polarity, without any from sentences (or documents) with manually an- segmentation annotations. notated sentiment polarity or distantly-supervised We evaluate the effectiveness of our joint mod- corpora collected by sentiment signals like emoti- el on a benchmark Twitter sentiment classifica- cons (Go et al., 2009; Pak and Paroubek, 2010; tion dataset in SemEval 2013. Results show that Kouloumpis et al., 2011; Zhao et al., 2012). the joint model performs comparably with state- Majority of existing approaches follow Pang et of-the-art methods, and consistently outperforms al. (2002) and employ corpus-based method for pipeline methods in various experiment settings. sentiment classification. Pang et al. (2002) pi- The main contributions of the work presented in oneer to treat the sentiment classification of re- this paper are as follows. views as a special case of text categorization prob- lem and first investigate machine learning meth- • To our knowledge, this is the first work that ods. They employ Naive Bayes, Maximum En- automatically produces sentence segmenta- tropy and Support Vector Machines (SVM) with a tion for sentiment classification within a joint diverse set of features. In their experiments, the framework. best performance is achieved by SVM with bag- of-words feature. Under this perspective, many s- • We show that the joint model yields com- tudies focus on designing or learning effective fea- parable performance with the state-of-the-art tures to obtain better classification performance. methods on the benchmark Twitter sentiment On movie or product reviews, Wang and Man- classification datasets in SemEval 2013. ning (2012) present NBSVM, which trades-off 478
Recommend
More recommend
