An information theoretic approach to sentiment polarity classification

Sentiment classification is a task of classifying documents according to their overall sentiment inclination. It is very important and popular in many web applications, such as credibility analysis of news sites on the Web, recommendation system and mining online discussion. Vector space model is widely applied on modeling documents in supervised sentiment classification, in which the feature presentation (including features type and weight function) is crucial for classification accuracy. The traditional feature presentation methods of text categorization do not perform well in sentiment classification, because the expressing manners of sentiment are more subtle. We analyze the relationships of terms with sentiment labels based on information theory, and propose a method by applying information theoretic approach on sentiment classification of documents. In this paper, we adopt mutual information on quantifying the sentiment polarities of terms in a document firstly. Then the terms are weighted in vector space based on both sentiment scores and contribution to the document. We perform extensive experiments with SVM on the sets of multiple product reviews, and the experimental results show our approach is more effective than the traditional ones.