Aspect and sentiment unification model

Abstract

User-generated reviews on the Web contain reviewers’ sentiments about detailed aspects of the products and services reviewed. However, most of the reviews are plain text and thus a user must read through many of them to obtain information about relevant details. This thesis addresses the problem of automatically discovering what aspects are evaluated in reviews and how sentiments are expressed for each of those aspects. As an approach to this problem, this thesis proposes the Aspect and Sentiment Unification Model (ASUM). This probabilistic generative model incorporates aspect and sentiment together to discover from reviews the aspects that are evaluated positively and the ones evaluated negatively. Given a collection of reviews, ASUM outputs language models, i.e., probability distributions over words, for positive aspects and negative aspects. ASUM was applied to reviews of electronic devices and restaurants, and the results show that the aspects discovered by ASUM match evaluative details of the reviews and capture important aspects that are closely coupled with a sentiment. The results of sentiment classification show that ASUM outperforms other generative models and comes close to supervised classification methods. One important advantage of ASUM is that it does not require any sentiment labels of the reviews, which are often expensive to obtain.