Research On Implicit Feature Identification Of Product Reviews

The explosive development of e-commerce has attracted a lot of people to shoponline, Most of the online merchants provide functions that allow users to postreviews about products. The massive amounts of reviews on the web can providevaluable resources for both potential customers and e-commerce companies.The task of opinion mining is extracting opinions and sentiments fromcustomer reviews by automatically analyzing and processing them. In order to getdetails of each review, fine-grained feature-level opinion mining has received moreand more attention. Most of the existing researches on feature-level opinion miningare dedicated to extract explicitly appeared features and opinion words on reviewsentences. However, among the numerous kinds of reviews on the web, there are asignicant number of reviews that contain only opinion words and these opinionwords implies product features. The identication of such implicit features is still oneof the most dicult problems in opinion mining, and it’s a even harder task onChinese reviews due to complexity of Chinese language.In this paper, we have tried to explore the fine-grained opinion mining inChinese product reviews, and focused on solving the problem of implicit featureidentication from review sentences. We propose a novel classication-based approachto deal with the problem of implicit feature identication. Firstly, By exploiting theword segmentation, part-of-speech(POS) tagging and dependency parsing, wepropose a rule based method to extract the explicit feature-opinion pairs fromcustomer reviews. Secondly, we cluster the feature-opinion pairs for each opinionword, and then construct the training document for each clustered feature-opinionpair from customer reviews. Finally, we formulate the identication of implicitfeature into a classication-based feature selection problem, so we can identify theimplicit feature by exploiting the text classication approach. Empirical evaluationdemonstrates that our approach outperforms existing state-of-the-art methodssignicantly.