Statistical Mechanical Analysis of Semantic Orientations on Lexical Network

Proceedings of COLING 2012: Technical Papers, pages 977–994, COLING 2012, Mumbai, December 2012. Statistical Mechanical Analysis of Semantic Orientations on Lexical Network Takuma GOTO∗ Yoshi yuki KABASHIMA Hiroya TAKAMURA Tokyo Institute of Technology, 4259 Nagatsuta-cho, Midori-ku, Yokohama, Japan takuma510@gmail.com, kaba@dis.titech.ac.jp, takamura@pi.titech.ac.jp ABSTRACT Many of the state-of-the-art methods for constructing a polarity lexicon rely on the propagation of polarity on the lexical network. In one of those methods, where the Ising spin model is em- ployed as a probabilistic model, it is reported that the system exhibits the phase transition in the vicinity of the optimal temperature parameter. We provide an analysis of this phenomenon from the viewpoint of statistical mechanics and clarify the underlying mechanism. On the basis of this analysis, we propose a scheme for improving the extraction performance, i.e., by remov- ing the largest eigenvalue component from the weight matrix. Experimental results show that the scheme significantly improves the accuracy of the extraction of the semantic orientations at negligible additional computational cost, outperforming the state-of-the-art algorithms. We also explore the origin of the high classification performance by analyzing eigenvalues of the weight matrix and a linearized model. KEYWORDS: sentiment analysis, polarity lexicon, spin model, label propagation. ∗ Takuma Goto now works for ACCESS CO., LTD. 977 1 Introduction A huge amount of semantic information is constantly being produced and accumulated on the Internet by the activities of individuals through, for example, their blog, Twitter, and Facebook postings. The information tends to focus on personal interests but can include generally useful information such as opinions about fashion and comments about new products. This means that extracting and structuralizing such information can be beneficial for both producers and consumers, which led us to focus on the development of methods for handling semantic infor- mation. In general, each word constituting sentences possesses its specific orientation. For ex- ample, we usually receive positive impressions for words such as “good”, “excellent” and “enjoyable”, while “bad”, “poor” and “boring” sound negative. tation of impression is termed polarity (or semantic orientation).Such word-specific orien- A polarity lexicon is a list of words and phrases that are labeled by their polarity, and is an important re- source in extracting semantic information from natural language data. Accordingly, the construction of such lists under various conditions has been a major focus in sentiment analysis research (Hatzivassiloglou and McKeown, 1997; Choi and Cardie, 2009). Many construction methods have been developed so far (Hatzivassiloglou and McKeown, 1997; Takamura et al., 2005; Turney and Littman, 2003; Velikovich et al., 2010; Kamps et al., 2004; Rao and Ravichandran, 2009). Among those methods for polarity lexicon construction, the method proposed by Takamura et al. (2005) is distinctive in terms of emphasizing the utility of probabilities. In their method, the construction of a polarity lexicon is mapped to the Ising spin model of mag- netism at a finite temperature. This mapping, in conjunction with the formalism of equilibrium statistical mechanics, yields a probabilistic model for assigning a polarity to each word. The optimal assignment of …

Content truncated to 500 words