A Comparative Study of Different Sentiment Lexica for Sentiment Analysis of Tweets

Proceedings of Recent Advances in Natural Language Processing, pages 488–496, Hissar, Bulgaria, Sep 7–9 2015. A Comparative Study of Different Sentiment Lexica for Sentiment Analysis of Tweets Canberk ¨Ozdemir and Sabine Bergler CLaC Labs, Concordia University 1455 de Maisonneuve Blvd West Montreal, Quebec, Canada, H3G 1M8 ozdemir.berkin@gmail.com, bergler@cse.concordia.ca Abstract We report on interoperability of different sentiment lexica with each other and with the linguistic notions negation and modal- ity for sentiment analysis of tweets in a comprehensive ablation study and in com- petition results for SemEval 2015. Our ap- proach performed well at the tweet level, but excelled in the presence of figurative language. 1 Introduction Increasing interest in social media is reflected in SemEval competitions on sentiment analysis of tweets. Sentiment analysis categorizes text into positive or negative sentiment, possibly with an additional neutral category (Pang and Lee, 2008). Tweets use more informal and non-standard lan- guage than other text forms posing additional chal- lenges. The winners of the past two years made heavy use of their specially designed NRC lex- icon (Mohammad et al., 2013; Kiritchenko et al., 2014), a large lexical resource extracted from tweets with hashtags that are unmistakably pos- itive or negative. This leads to the question we address here: is a bigger lexicon (proportionally) more useful? Is there something special about the NRC lexicon? Is a lexicon that is designed like the NRC lexicon but ten times its size more useful? And finally, can linguistic contexts negation and modality improve the lexicon and the final classi- fication? We compiled a NRC-inspired lexical resource, Gezi, of seven times the size of the NRC lexi- con. We used several lexica in various combina- tions: Gezi, NRC, Bing Liu’s lexicon (Hu and Liu, 2004), MPQA (Wilson et al., 2005), and aFinn (Nielsen, 2011), and add negation and modality sensitive features, performing comprehensive ab- lation experiments. The system competed in Se- mEval 2015 and ranked 9/40 in Task 10B, senti- ment classification of tweets, and 1/35 in Task 11, tweets featuring figurative language. 2 Previous Work Since Pang and Lees pioneering work on movie re- view classification into thumbs up—thumbs down (Pang et al., 2002), the major resource for senti- ment determination was a sentiment lexicon, mod- elled after the independently and previously cre- ated Harvard General Inquirer (Stone et al., 1966), a list of words labelled as positive or negative sen- timent carriers. Rule-based approaches yielded strong baselines that depended mainly on the cov- erage of the lexicon used, leading to various ef- forts to compile dedicated sentiment lexica (Esuli and Sebastiani, 2006; Wilson et al., 2005). The growing number of sentiment laden text on so- cial media led to more efforts to annotate corpora, enabling machine learning approaches which mo- nopolize the current exercises on sentiment an- notation of tweets at SemEval (Rosenthal et al., 2015; Rosenthal et al., 2014; Nakov et al., 2013). The lexicon is still the major tool used, and for the non-standard use of language encountered in tweets, special resources have been …

Content truncated to 500 words