The viability of web-derived polarity lexicons

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 777–785, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics The viability of web-derived polarity lexicons Leonid Velikovich Sasha Blair-Goldensohn Kerry Hannan Ryan McDonald Google Inc., New York, NY {leonidv|sasha|khannan|ryanmcd}@google.com Abstract We examine the viability of building large polarity lexicons semi-automatically from the web. We begin by describing a graph propa- gation framework inspired by previous work on constructing polarity lexicons from lexi- cal graphs (Kim and Hovy, 2004; Hu and Liu, 2004; Esuli and Sabastiani, 2009; Blair- Goldensohn et al., 2008; Rao and Ravichan- dran, 2009). We then apply this technique to build an English lexicon that is signifi- cantly larger than those previously studied. Crucially, this web-derived lexicon does not require WordNet, part-of-speech taggers, or other language-dependent resources typical of sentiment analysis systems. As a result, the lexicon is not limited to specific word classes – e.g., adjectives that occur in WordNet – and in fact contains slang, misspellings, multi- word expressions, etc. We evaluate a lexicon derived from English documents, both qual- itatively and quantitatively, and show that it provides superior performance to previously studied lexicons, including one derived from WordNet. 1 Introduction Polarity lexicons are large lists of phrases that en- code the polarity of each phrase within it – either positive or negative – often with some score rep- resenting the magnitude of the polarity (Hatzivas- siloglou and McKeown, 1997; Wiebe, 2000; Turney, 2002). Though classifiers built with machine learn- ing algorithms have become commonplace in the sentiment analysis literature, e.g., Pang et al. (2002), the core of many academic and commercial senti- ment analysis systems remains the polarity lexicon, which can be constructed manually (Das and Chen, 2007), through heuristics (Kim and Hovy, 2004; Esuli and Sabastiani, 2009) or using machine learn- ing (Turney, 2002; Rao and Ravichandran, 2009). Often lexicons are combined with machine learning for improved results (Wilson et al., 2005). The per- vasiveness and sustained use of lexicons can be as- cribed to a number of reasons, including their inter- pretability in large-scale systems as well as the gran- ularity of their analysis. In this work we investigate the viability of polar- ity lexicons that are derived solely from unlabeled web documents. We propose a method based on graph propagation algorithms inspired by previous work on constructing polarity lexicons from lexical graphs (Kim and Hovy, 2004; Hu and Liu, 2004; Esuli and Sabastiani, 2009; Blair-Goldensohn et al., 2008; Rao and Ravichandran, 2009). Whereas past efforts have used linguistic resources – e.g., Word- Net – to construct the lexical graph over which prop- agation runs, our lexicons are constructed using a graph built from co-occurrence statistics from the entire web. Thus, the method we investigate can be seen as a combination of methods for propagat- ing sentiment across lexical graphs and methods for building sentiment lexicons based on distributional characteristics of phrases in raw data (Turney, 2002). The advantage of breaking the dependence on Word- …

Content truncated to 500 words