Collective Sentiment Classification Based on User Leniency and Product Popularity

Collective Sentiment Classification based on User Leniency and Product Popularity 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 We propose a method of collective senti- ment classification that assumes dependen- cies among labels of an input set of reviews. The key observation behind our method is that the distribution of polarity labels over reviews written by each user or written on each product is often skewed in the real world; intolerant users tend to report com- plaints while popular products are likely to receive praise. We encode these char- acteristics of users and products (referred to as user leniency and product popular- ity) by introducing global features in su- pervised learning. To resolve dependencies among labels of a given set of reviews, we explore two approximated decoding algo- rithms, “easiest-first decoding” and “two- stage decoding”. Experimental results on two real-world datasets with product and user/product information confirmed that our method contributed greatly to the classifica- tion accuracy. 1 Introduction In document-level sentiment classification, early studies have exploited language-based clues (e.g., n-grams) extracted from the textual content (Tur- ney, 2002; Pang et al., 2002), followed by recent studies which adapt the classifier to the reviews written by a specific user or written on a specific product (Tan et al., 2011; Seroussi et al., 2010; Speriosu et al., 2011; Li et al., 2011). Although the user- and product-aware methods exhibited bet- ter performance over the methods based on purely textual clues, most of them use only the user in- formation (Tan et al., 2011; Seroussi et al., 2010; Speriosu et al., 2011), or they assume that the user and the product of a test review is known in ad- vance (Li et al., 2011). These assumptions heav- ily limit their applicability in a real-world scenario where new users and new products are ceaselessly emerging. This paper proposes a method of collective sen- timent classification that is aware of the user and the product of the target review, which benefits from the skewed distributions of polarity labels: intolerant users tend to report complaints while popular products are likely to receive praise. We introduce global features to encode these charac- teristics of a user and a product (referred to as user leniency and product popularity), and then com- pute the values of global features along with test- ing. Our method is therefore applicable to reviews written by users and on products that are not ob- served in the training data. Because global features depend on labels of test reviews while the labels reversely depend on the global features, we need to globally optimize a la- bel configuration for a given set of reviews. In this study, we resort to approximate algorithms, easiest-first (Tsuruoka and Tsujii, 2005) and two- stage strategies (Krishnan and Manning, 2006), in decoding labels, and empirically compare their speed and accuracy. We …

Content truncated to 500 words