Effective Sentiment Analysis based on Term Evaluation by Bayesian Model Selection Criteria

Sentiment analysis (or opinion mining) becomes more and more important in many industrial applications because there have been huge needs in extracting writers' intention in their text documents. Previous approaches have adopted classical techniques used in text categorization and information retrieval. For example, for the analysis of terms in a document, information theoretic measures such as pointwise mutual information (PM!) have been widely used. Although these approaches are sometimes shown to be effective, it often takes a long time and expensive computation to find a right pair of words with maximum mutual information. Thus, it will be more beneficial if we can use more powerful measures such as Bayesian model selection criteria (i.e. likelihood and Bayesian information criteria) for the term analysis. In this paper, we estimate likelihood and Bayesian information criteria (BIC) for each term conditional to class labels. After we estimate model selection criteria, we calculate a ratio of the model selection criteria for each class label. Based on the obtained ratios, we can sort and identify highly influential terms for each class label. We have performed extensive experiments on several public benchmark data sets. Experimental results on those data sets indicate that our proposed approach is effective in term evaluation for sentiment classification, and exhibits scalability comparable to that of classical Bayesian classifiers.