Modeling User Leniency and Product Popularity for Sentiment Classification

International Joint Conference on Natural Language Processing, pages 1107–1111, Nagoya, Japan, 14-18 October 2013. Modeling User Leniency and Product Popularity for Sentiment Classification Wenliang Gao∗, Naoki Yoshinaga†, Nobuhiro Kaji† and Masaru Kitsuregawa†‡ ∗Graduate School of Information Science and Technology, The University of Tokyo †Institute of Industrial Science, The University of Tokyo ‡National Institute of Informatics {wl-gao, ynaga, kaji, kitsure}@tkl.iis.u-tokyo.ac.jp Abstract Classical approaches to sentiment classi- fication exploit only textual features in a given review and are not aware of the per- sonality of the user or the public sentiment toward the target product. In this paper, we propose a model that can accurately es- timate the sentiment polarity by referring to the user leniency and product popular- ity computed during testing. For decoding with this model, we adopt an approximate strategy called “two-stage decoding.” Pre- liminary experimental results on two real- world datasets show that our method sig- nificantly improves classification accuracy over existing state-of-the-art methods. 1 Introduction Document-level sentiment classification estimates the sentiment polarity for a given subjective text (hereafter, review). Traditionally, researchers have tried to estimate the sentiment polarity from only the textual content of the review (Pang and Lee, 2004; Li et al., 2011). However, since reviews are written by a user to express his/her emotion toward a particular product, taking the users and products into consideration would play an important role in solving this task. Recently, the increase of opinionated text within social media, e.g., Twitter, has motivated re- searchers to exploit the user or product informa- tion in the sentiment classification task. Some re- searchers take advantages of the friend relation in a social network because friends are likely to hold common tastes (Tan et al., 2011; Seroussi et al., 2010; Speriosu et al., 2011). Others incorporate user- or product-specific n-gram features (Li et al., 2011; Seroussi et al., 2010). Although these stud- ies have showed that user or product information is useful for sentiment classification, they implicitly assume that the same users or products appear in both training and testing data. Thus, to train such a model, a large amount of the reviews should be labeled for each user and each product. In a real- world scenario, however, this is unrealistic since new users and products are ceaselessly emerging and labeling reviews written by such users (or on such products) is impractical. In the real world, different users have different rating standards, while different products receive different rating tendencies. For example, a crit- ical person is likely to point out flaws and gives negative ratings, while a popular product receives more praise than negative feedback. We refer to these user- or product-specific polarity biases as user leniency and product popularity, respectively. A sentiment classifier would resort to these biases when textual features are not reliable enough to estimate the sentiment polarity. In this study, we build a model that automati- cally computes and uses user leniency and product popularity for sentiment classification. We rep- resent these biases with two types of real-valued global features. …

Content truncated to 500 words