PURE: A Novel Tripartite Model for Review Sentiment Analysis and Recommendation

Nowadays, more and more users like to leave online reviews. These reviews, which are based on their experiences on a set of service or products, often express different opinions and sentiments. Correlated topic model (CTM), an effective text mining model, can reduce the dimension without losing important information. However, traditional analyses based on CTM still have some problems. In this paper, we propose the Product-User-Review tripartite sEntiment model (PURE), which is based on content-based clustering to optimize CTM, to select topic number, extract feature, estimate the reviews’ utility. Moreover, our model analyzes the reviews from the user’s preferences, review content and product properties in three dimensions. Based on the five indexes, such as informative attributes and sentiment attributes, the feature vector of the review data is constructed. We found that after adding user’s preference feature in sentiment analysis and utility estimation, PURE achieves high accuracy and classification speed in the review-mixing Chinese and English processing, and the quality of selection is improved significantly by 21%. To the best of our knowledge, this is the first work to incorporate users’ preference feature in optimized CTM to do the study of sentiment analysis, review selection and recommendation.