Joint Propagation and Refinement for Mining Opinion Words and Targets

Abstract

This paper proposes a novel Joint Propagation and Refinement (JPR) method to extract opinion words and targets. We adopt a growing heuristic method to extract new opinion words and targets in two parallel processes: propagation and refinement. In the propagation process, we generate the candidate sets of opinion words and targets and construct Sentiment Graph Model (SGM) to evaluate the relations between opinion words and targets. We employ statistical word co-occurrence and dependency patterns to identify these relations. In addition, we discover new patterns by the newly extracted opinion words and targets, which can capture opinion relations more precisely in the case of informal texts. In the refinement process, we prune false results and update model iteratively. We employ Automatic Rule Refinement (ARR) to refine the rules of extraction, which means to refine the rule to extract false results. By using false results pruning and ARR process, we can efficiently alleviate the error propagation problem in traditional bootstrapping based methods. Experimental results on both English and Chinese datasets demonstrate the effectiveness of our method.