A Unified Framework for Fine-Grained Opinion Mining from Online Reviews

Abstract

Extracting opinion words and opinion targets from online reviews is an important task for fine-grained opinion mining. Usually, traditional extraction methods under the pipeline-based framework have higher precision but lower recall, while methods in the propagation-based framework possess greater recall but poorer precision. To achieve better performance both in precision and recall, this paper proposes a unified framework for fine-grained opinion mining, combining propagation with refinement in a dynamic and iterative process. In the propagation process, syntactic patterns are chosen as opinion relations to extract new opinion words and targets. Besides, syntactic patterns are further generalized to make them more flexible and scalable. In the refinement process, a three-layer opinion relations graph (ORG) model is constructed based on three types of candidates: opinion word candidates, opinion target candidates and syntactic pattern candidates. A sorting algorithm based on ORG model is proposed to rank all the candidates in their own type, and low-rank candidates are removed from candidate datasets. Repeat propagation and refinement until the syntactic pattern candidate set reaches stable. Experimental results on both English and Chinese online reviews demonstrate the effectiveness of proposed framework and its methods, comparing with the-state-of-the-art methods.