A Statistical NLP Approach for Feature and Sentiment Identification from Chinese Reviews

Existing methods for extracting features from Chinese reviews only use simplistic syntactic knowledge, while those for identifying sentiments rely heavily on a semantic dictionary. In this paper, we present a systematic technique for identifying features and sentiments, using both syntactic and statistical analysis. We firstly identify candidate features using a proposed set of common syntactic rules. We then prune irrelevant candidates with topical relevance scores below a cut-off point. We also propose an association analysis method based on likelihood ratio test to infer the polarity of opinion word. The sentiment of a feature is finally adjusted by analyzing the negative modifiers in the local context of the opinion word. Experimental results show that our system performs significantly better than a well-known opinion mining system.