Generating High-Coverage Semantic Orientation Lexicons From Overtly Marked Words and a Thesaurus

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 599–608, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Generating High-Coverage Semantic Orientation Lexicons From Overtly Marked Words and a Thesaurus Saif Mohammadφ†∗, Cody Dunne⋄‡, and Bonnie Dorrφ†‡∗ Laboratory for Computational Linguistics and Information Processingφ Human-Computer Interaction Lab⋄ Institute for Advanced Computer Studies† Department of Computer Science‡, University of Maryland. Human Language Technology Center of Excellence∗ {saif,bonnie}@umiacs.umd.edu and {cdunne}@cs.umd.edu Abstract Sentiment analysis often relies on a se- mantic orientation lexicon of positive and negative words. A number of approaches have been proposed for creating such lex- icons, but they tend to be computation- ally expensive, and usually rely on signifi- cant manual annotation and large corpora. Most of these methods use WordNet. In contrast, we propose a simple approach to generate a high-coverage semantic orien- tation lexicon, which includes both indi- vidual words and multi-word expressions, using only a Roget-like thesaurus and a handful of affixes. Further, the lexicon has properties that support the Polyanna Hypothesis. Using the General Inquirer as gold standard, we show that our lexi- con has 14 percentage points more correct entries than the leading WordNet-based high-coverage lexicon (SentiWordNet). In an extrinsic evaluation, we obtain signifi- cantly higher performance in determining phrase polarity using our thesaurus-based lexicon than with any other. Additionally, we explore the use of visualization tech- niques to gain insight into the our algo- rithm beyond the evaluations mentioned above. 1 Introduction Sentiment analysis involves determining the opin- ions and private states (beliefs, emotions, specu- lations, and so on) of the speaker (Wiebe, 1994). It has received significant attention in recent years due to increasing online opinion content and ap- plications in tasks such as automatic product rec- ommendation systems (Tatemura, 2000; Terveen et al., 1997), question answering (Somasundaran et al., 2007; Lita et al., 2005), and summarizing multiple view points (Seki et al., 2004) and opin- ions (Mohammad et al., 2008a). A crucial sub-problem is to determine whether positive or negative sentiment is expressed. Auto- matic methods for this often make use of lexicons of words tagged with positive and negative seman- tic orientation (Turney, 2002; Wilson et al., 2005; Pang and Lee, 2008). A word is said to have a positive semantic orientation (SO) (or polarity) if it is often used to convey favorable sentiment or evaluation of the topic under discussion. Some example words that have positive semantic orien- tation are excellent, happy, honest, and so on. Sim- ilarly, a word is said to have negative semantic ori- entation if it is often used to convey unfavorable sentiment or evaluation of the target. Examples include poor, sad, and dishonest. Certain semantic orientation lexicons have been manually compiled for English—the most notable being the General Inquirer (GI) (Stone et al., 1966).1 However, the GI lexicon has orientation labels for only about 3,600 entries. The Pitts- burgh subjectivity lexicon (PSL) (Wilson et al., 2005), which draws from the General Inquirer and other sources, also has semantic orientation labels, but …

Content truncated to 500 words