A Joint Model of Product Properties, Aspects and Ratings for Online Reviews

Proceedings of Recent Advances in Natural Language Processing, pages 131–137, Hissar, Bulgaria, Sep 7–9 2015. A Joint Model of Product Properties, Aspects and Ratings for Online Reviews Ying Ding School of Information Systems Singapore Management University ying.ding.2011@smu.edu.sg Jing Jiang School of Information Systems Singapore Management University jingjiang@smu.edu.sg Abstract Product review mining is an important task that can benefit both businesses and con- sumers. Lately a number of models com- bining collaborative filtering and content analysis to model reviews have been pro- posed, among which the Hidden Factors as Topics (HFT) model is a notable one. In this work, we propose a new model on top of HFT to separate product prop- erties and aspects. Product properties are intrinsic to certain products (e.g. types of cuisines of restaurants) whereas aspects are dimensions along which products in the same category can be compared (e.g. service quality of restaurants). Our pro- posed model explicitly separates the two types of latent factors but links both to product ratings. Experiments show that our proposed model is effective in separat- ing product properties from aspects. 1 Introduction Online product reviews and the numerical ratings that come with them have attracted much attention in recent years. During the early years of research on product review mining, there were two separate lines of work. One focused on content analysis using review texts but ignored users, and the other focused on collaborative filtering-based rating pre- diction using user-item matrices but ignored texts. However, these studies do not consider the identi- fies of reviewers, and thus cannot incorporate user preferences into the models. In contrast, the ob- jective of collaborative filtering-based rating pre- diction is to predict a target user’s overall rating on a target product without referring to any review text (e.g. Salakhutdinov and Mnih (2007)). Col- laborative filtering makes use of past ratings of the target user, the target item and other user-item rat- ings to predict the target user’s rating on the target item. Presumably if review texts, numerical ratings, user identities and product identities are analyzed together, we may achieve better results in rat- ing prediction and feature/aspect identification. This is the idea explored in a recent work by McAuley and Leskovec (2013), where they pro- posed a model called Hidden Factors as Topics (HFT) to combine collaborative filtering with con- tent analysis. HFT combines latent factor models for recommendation with Latent Dirichlet Alloca- tion (LDA). In the joint model, the latent factors play dual roles: They contribute to the overall rat- ings, and they control the topic distributions of in- dividual reviews. While HFT is shown to be effective in both predicting ratings and discovering meaningful la- tent factors, we observe that the discovered la- tent factors are oftentimes not “aspects” in which products can be evaluated and compared. In fact, the authors themselves also pointed out that the topics discovered by HFT “are not similar to as- pects” (McAuley and Leskovec, 2013). Here we use “aspects” to refer to criteria that can be used …

Content truncated to 500 words