Sentiment Classification of Product Reviews Using SVM and Decision Tree Induction

This paper reports a study in automatic sentiment classification, i.e. automatically classifying documents as expressing positive or negative sentiments/opinions. The study investigates the effectiveness of using SVM (Support Vector Machine) and Decision Tree induction on various text features to classify product reviews into recommended (positive sentiment) and not recommended (negative sentiment). Compared with traditional topical classification, it was hypothesized that syntactic and semantic processing of text would be more important for sentiment classification. In this study, five different approaches, unigrams (individual words), part-of-speech tagging, association rules, use of negation, and use of WordNet synsets (identifying a set of synonyms) were investigated. A sample of 1,800 miscellaneous product reviews was retrieved from Review Centre (www.reviewcentre.com. 2003) for the study. 1,200 reviews were used for training, and 600 for validation. Using SVM, the baseline unigrams approach obtained an accuracy rate of 81.3%. The use of WordNet synsets obtained marginally better result of 81.7%. The other text features did not yield better results. Error analysis suggests 3 approaches for improving classification accuracy: making inference from superficial words, solving the problems of "comments on parts" and "negation". Finally, Decision Tree induction was used to generate a list of indicative words that can identify the polarity of articles.