Improving Blog Polarity Classification via Topic Analysis and Adaptive Methods

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 309–312, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Improving Blog Polarity Classification via Topic Analysis and Adaptive Methods Feifan Liu University of Wisconsin, Milwaukee liuf@uwm.edu Dong Wang, Bin Li, Yang Liu The University of Texas at Dallas dongwang,yangl@hlt.utdallas.edu Abstract In this paper we examine different linguistic features for sentimental polarity classification, and perform a comparative study on this task between blog and re- view data. We found that results on blog are much worse than reviews and investigated two methods to improve the performance on blogs. First we ex- plored information retrieval based topic analysis to extract relevant sentences to the given topics for po- larity classification. Second, we adopted an adaptive method where we train classifiers from review data and incorporate their hypothesis as features. Both methods yielded performance gain for polarity clas- sification on blog data. 1 Introduction Sentimental analysis is a task of text categorization that focuses on recognizing and classifying opinionated text towards a given subject. Different levels of sentimental analysis has been performed in prior work, from binary classes to more fine grained categories. Pang et al. (2002) defined this task as a binary classification task and ap- plied it to movie reviews. More sentiment classes, such as document objectivity and subjectivity as well as dif- ferent rating scales on the subjectivity, have also been taken into consideration (Pang and Lee, 2005; Boiy et al., 2007). In terms of granularity, this task has been investigated from building word level sentiment lexicon (Turney, 2002; Moilanen and Pulman, 2008) to detecting phrase-level (Wilson et al., 2005; Agarwal et al., 2009) and sentence-level (Riloff and Wiebe, 2003; Hu and Liu, 2004) sentiment orientation. However, most previous work has mainly focused on reviews (Pang et al., 2002; Hu and Liu, 2004), news resources (Wilson et al., 2005), and multi-domain adaptation (Blitzer et al., 2007; Man- sour et al., 2008). Sentiment analysis on blogs (Chesley et al., 2005; Kim et al., 2009) is still at its early stage. In this paper we investigate binary polarity classifica- tion (positive vs. negative). We evaluate the genre effect between blogs and review data and show the difference of feature effectiveness. We demonstrate improved polarity classification performance in blogs using two methods: (a) integrating topic relevance analysis to perform topic specific polarity classification; (b) adopting an adaptive method by incorporating multiple classifiers’ hypotheses from different review domains as features. Our manual analysis also points out some challenges and directions for further study in blog domain. 2 Features for Polarity Classification For the binary polarity classification task, we use a super- vised learning framework to determine whether a docu- ment is positive or negative. We used a subjective lex- icon, containing 2304 positive words and 4145 negative words respectively, based on (Wilson et al., 2005). The features we explored are listed below. (i) Lexical features (LF) We use the bag of words for the lexical features …

Content truncated to 500 words