A Cyclic Cascaded CRFs Model for Opinion Targets Identification Based on Rules and Statistics

摘要

Opinion sentences on e-commerce platform, microblog and forum contain lots of emotional information. And opinion targets identification plays an import role in huge potential commercial value mining, especially in sales decision making and development trend forecasting. Traditional CRFs-based method has achieved a pretty good result to a certain extent. However, its discovery ability of out-of-vocabulary words and optimization of the mining model are both insufficient. We propose a novel cyclic cascaded CRFs model for opinion targets identification which incorporates rule-based and statistic-based methods. The approach acquires candidate opinion targets through part-of-speech, syntactic and semantic rules, and integrates them in a cyclic cascaded CRFs model for the accurate opinion targets identification. Experimental results on COAE2014 dataset show the outperformance of this method.