Sentiment Polarity Classification Using Statistical Data Compression Models

With growing availability and popularity of user generated content, the discipline of sentiment analysis has come to the attention of many researchers. Existing work has mainly focused on either knowledge based methods or standard machine learning techniques. In this paper we investigate sentiment polarity classification based on adaptive statistical data compression models. We evaluate the classification performance of the loss less compression algorithm Prediction by Partial Matching (PPM) as well as compression based measures using PPM-like character n-gram frequency statistics. Comprehensive experiments on three corpora show that compression based methods are efficient, easy to apply and can compete with the accuracy of sophisticated classifiers such as support vector machines.