Dependency Tree-based Sentiment Classification using CRFs with Hidden Variables

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 786–794, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Dependency Tree-based Sentiment Classification using CRFs with Hidden Variables Tetsuji Nakagawa∗, Kentaro Inui∗† and Sadao Kurohashi∗‡ ∗National Institute of Information and Communications Technology †Tohoku University ‡Kyoto University tnaka@nict.go.jp, inui@ecei.tohoku.ac.jp, kuro@i.kyoto-u.ac.jp Abstract In this paper, we present a dependency tree- based method for sentiment classification of Japanese and English subjective sentences us- ing conditional random fields with hidden variables. Subjective sentences often con- tain words which reverse the sentiment po- larities of other words. Therefore, interac- tions between words need to be considered in sentiment classification, which is difficult to be handled with simple bag-of-words ap- proaches, and the syntactic dependency struc- tures of subjective sentences are exploited in our method. In the method, the sentiment po- larity of each dependency subtree in a sen- tence, which is not observable in training data, is represented by a hidden variable. The po- larity of the whole sentence is calculated in consideration of interactions between the hid- den variables. Sum-product belief propaga- tion is used for inference. Experimental re- sults of sentiment classification for Japanese and English subjective sentences showed that the method performs better than other meth- ods based on bag-of-features. 1 Introduction Sentiment classification is a useful technique for an- alyzing subjective information in a large number of texts, and many studies have been conducted (Pang and Lee, 2008). A typical approach for sentiment classification is to use supervised machine learning algorithms with bag-of-words as features (Pang et al., 2002), which is widely used in topic-based text classification. In the approach, a subjective sen- tence is represented as a set of words in the sen- tence, ignoring word order and head-modifier rela- tion between words. However, sentiment classifi- cation is different from traditional topic-based text classification. Topic-based text classification is gen- erally a linearly separable problem ((Chakrabarti, 2002), p.168). For example, when a document con- tains some domain-specific words, the document will probably belong to the domain. However, in sentiment classification, sentiment polarities can be reversed. For example, let us consider the sentence “The medicine kills cancer cells.” While the phrase cancer cells has negative polarity, the word kills re- verses the polarity, and the whole sentence has pos- itive polarity. Thus, in sentiment classification, a sentence which contains positive (or negative) polar- ity words does not necessarily have the same polar- ity as a whole, and we need to consider interactions between words instead of handling words indepen- dently. Recently, several methods have been proposed to cope with the problem (Zaenen, 2004; Ikeda et al., 2008). However, these methods are based on flat bag-of-features representation, and do not consider syntactic structures which seem essential to infer the polarity of a whole sentence. Other methods have been proposed which utilize composition of sentences (Moilanen and Pulman, 2007; Choi and Cardie, 2008; Jia et al., 2009), but these methods use rules to handle polarity reversal, …

Content truncated to 500 words