Mining Product Reviews in a Comprehensive View
Hongning Wang, Xiaoyan Zhu · 2009
Previous work in product review mining mainly focuses on flne-grained opinions, i.e. feature-opinion analysis. However, since one product usually associates with various reviews concerning difierent aspects, these approaches could not provide an inclusive depiction about the given product. Besides, only identifying the sentiment as positive/negative could not preserve su‐cient information about the public preference on the products. Focusing on these problems, in this paper we propose to perform product review mining in a comprehensive view: we process the reviews of a given product as a whole and formulate the sentiment as a distribution rather than a determined label over the product. In particular, we present a supervised generative topic model, the Word-Sentiment Model (WSModel), to reveal the latent structure between the content of reviews and related sentiments. Difierent from the traditional approach of review mining, which relies on a list of sentiment-oriented words, in this paper we propose to exploit the correlation between the content of reviews and labeled ratings to infer the sentiment distribution. To evaluate our model, we collected 6 categories of products consisting of 3,741 products and 134,772 reviews with labeled ratings from amazon.com. Extensive experiments were conducted on this collection and the encouraging performance conflrms the efiectiveness of our proposed approach.