Evidential Fusion for Sentiment Polarity Classification

Abstract

This paper presents an evidential fusion approach for sentiment classification tasks and a comparative study with linear sum combination. It involves the formulation of sentiment classifier output in the triplet evidence structure and adaptation of combination formulas for combining simple support functions derived from triplet functions by using Smets’s rule, the cautious conjunctive rules and linear sum rule. Empirical comparisons on the performance have been made in individuals and in combinations by using these rules, the results demonstrate that the best ensemble classifiers constructed by the four combination rules outperform the best individual classifiers over two public datasets of MP3 and Movie-Review.