Sentiment Classification through Combining Classifiers with Multiple Feature Sets

A classifier combination approach to sentiment classification where different classifiers are generated through training the review data with different features: unigram and some POS features and the sum rule achieves the best performance.

Abstract

Sentiment classification aims at assigning a document to a predefined category according to the polarity of its subjective information (e.g. 'thumbs up' or 'thumbs down'). In this paper, we present a classifier combination approach to this task. First, different classifiers are generated through training the review data with different features: unigram and some POS features. Then, classifier selection method is used to select a part of the classifiers for the next-step combination. Finally, these selected classifiers are combined using several combining rules. The experimental results show that all the combination approaches with different combining rules outperform individual classifiers and the sum rule achieves the best performance with an improvement of 2.56% over the best individual classifier.