Measuring Term Specificity Information for Assessing Sentiment Orientation of Documents in a Bayesian Learning Framework

Abstract

The assessment of document sentiment orientation using term specificity information is advocated in this study. An interpretation of the mathematical meaning of term specificity information is given based on Shannon’s entropy. A general form of a specificity measure is introduced in terms of the interpretation. Sentiment classification using the specificity measures is proposed within a Bayesian learning framework, and some potential problems are clarified and solutions are suggested when the specificity measures are applied to estimation of posterior probabilities for the NB classifier. A novel method is proposed which allows each document to have multiple representations, each of which corresponds to a sentiment class. Our experimental results show, while both the proposed method and IR techniques can produce high performance for sentiment classification, that our method outperforms the IR techniques.