Cross-domain Classification Based Sentiment Analysis For Product Reviews

With the fast growing of e-business, e-Shopping becomes more and more popular. Also, users would like to publish product reviews on e-business sites, product forums, or personal blogs. Information mining from product reviews is very useful to both potential customers and product suppliers. For a potential customer, this can help him/her decide which product to buy. For a product supplier, this enables it to easily gather marketing intelligence and product benchmark information. So review mining and analysis becomes very meaningful.The amount of reviews is very large and growing fast. Collecting and analyzing reviews by human is unpractical. So an automatic method to mine and analyze reviews from web is needed. This paper proposes a product review mining and analyzing system framework. This framework includes review mining, review sentiment orientation classification, and review search. Review mining module crawls product related web pages from web and identifies the review contents. Review sentiment orientation classification classifies product reviews to positive and negative. Review search module is a platform of online product review search. Users can learn the overall reputation of products and also the detail of each product feature.Product sentiment classification is a very important task in review analysis. We consider it as a binary classification problem and supervised learning techniques are applied. But we find that sentiment classification is a very domain-specific problem. It means we have to collect large amount of labeled training data for each product domain if we use traditional classification techniques. Labeled data collection is very costly. Therefore, how to use training data from a domain to do sentiment classification of other domains becomes a popular research problem.This paper proposes a cross-domain text classification algorithm named Reinforcement Iterative Transferring Classification (RITC) to solve cross-domain text classification problem. We apply RITC and other two cross-domain text classification algorithms to cross-domain review sentiment classification problem. Empirically, three cross-domain classification algorithms perform better than traditional classification algorithms and RITC performs best in most cases.