Mining Opinion Words and Targets from Online Reviews in a Hybrid Framework

Abstract

Extracting opinion words and targets is a main task in opinion mining. This paper proposes a novel approach with a dynamic process of joint propagation and refinement. In the propagation process, two initial datasets of opinion words and targets are separately obtained by given seed words and seed dependency patterns under the pre-defined extraction rules, and meanwhile new dependency patterns are found and added into seed dependency patterns. In the following refinement process, an Opinion Relation Graph (ORG) is modeled to represent relations between opinion words and targets, which is employed to measure the confidence of each candidate from opinion words and targets datasets. The words or targets with high confidence are kept in their respective datasets and the rest are removed as false results which are used to refine extraction rules with an Automatic Rule Refinement (ARR) method. Update ORG model and repeat the joint process of propagation and refinement until ORG model reaches stable. Experimental results on both English and Chinese datasets demonstrate the effectiveness of proposed method comparing with the-state-of-the-art methods.