Opinion retrieval based on mutual reinforcement between opinon analysis and relavence estimation

Different from most existing opinion retrieval systems separately process opinion analysis and relevance estimation as two one-step classification, this paper proposes a coarse-fine multi-pass opinion retrieval system incorporating mutual reinforcement between opinion analysis and relevance estimation. Based on linguistic observation on the opinion expression, some inner-and inter-sentence features are discovered. A multi-pass opinion retrieval system is then designed. Firstly, by using inner-sentence features, two base classifiers corresponding to opinion analysis and relevance estimation tasks, respectively, analyze the opinion and relevance of each sentence in the document. The inter-sentence features, including neighboring sentence-level, paragraph-level and document-level features, are obtained based on coarse analysis results. Secondly, both inner-sentence and inter-sentence features are incorporated the improved classifiers to refine the sentence analysis results and then update the inter-sentence features. Considering the strong association between opinionated sentences and topic-relevance sentences, the individual analysis results are refined following a mutual reinforcement mechanism. The updated features are then feed back to the improved classifier to further refine the sentence analysis results. Such circles terminate until the analysis results converge. Evaluations on NTCIR-7 MOAT dataset show that the proposed system achieved promising results. It shows that the proposed opinion retrieval system integrating coarse-fine analysis strategy and mutual reinforcement mechanism between opinion analysis and relevance estimation are effective.