Assembling the optimal sentiment classifiers

Abstract

Sentiment classification aims to classify documents according to their overall sentiment orientation, which plays an important role in many web applications, such as electronic commerce. Machine learning is an effective method for such tasks. In general, a classifier is determined by a feature type, a weighting function and a classification algorithm for a given training set. Thus, users are required to predetermine which ones should be applied, that is a troublesome problem for them, because each classifier always achieves different performance for different domains. To deal with this problem, we develop a three phase framework based on assembling multiple classifiers. In order to choose the optimal combination of classifiers, we propose a criterion for estimating the quality of the combination based on sentiment classification accuracy and diversity of the results generated by these classifiers. Moreover, we study the effect of the number of classifiers selected experimentally. With our solution, users can achieve a good performance without making a choice among plentiful combinations of different classifiers. We perform extensive experiments to demonstrate the effectiveness of our solution for different domains. © 2012 Springer-Verlag.