Columbia NLP: Sentiment Detection of Subjective Phrases in Social Media

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 478–482, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics Columbia NLP: Sentiment Detection of Subjective Phrases in Social Media Sara Rosenthal Department of Computer Science Columbia University New York, NY 10027, USA sara@cs.columbia.edu Kathleen McKeown Department of Computer Science Columbia University New York, NY 10027, USA kathy@cs.columbia.edu Abstract We present a supervised sentiment detection system that classifies the polarity of subjec- tive phrases as positive, negative, or neutral. It is tailored towards online genres, specifically Twitter, through the inclusion of dictionaries developed to capture vocabulary used in on- line conversations (e.g., slang and emoticons) as well as stylistic features common to social media. We show how to incorporate these new features within a state of the art system and evaluate it on subtask A in SemEval-2013 Task 2: Sentiment Analysis in Twitter. 1 Introduction People use social media to write openly about their personal experiences, likes and dislikes. The follow- ing sentence from Twitter is a typical example: “To- morrow I’m coming back from Barcelona...I don’t want! :(((”. The ability to detect the sentiment ex- pressed in social media can be useful for understand- ing what people think about the restaurants they visit, the political viewpoints of the day, and the products they buy. These sentiments can be used to provided targeted advertising, automatically gen- erate reviews, and make various predictions, such as political outcomes. In this paper we develop a sentiment detection al- gorithm for social media that classifies the polarity of sentence phrases as positive, negative, or neutral and test its performance in Twitter through the par- ticipation in the expression level task (subtask A) of the SemEval-2013 Task 2: Sentiment Analysis in Twitter (Wilson et al., 2013) which the authors helped organize. To do so, we build on previous work on sentiment detection algorithms for the more formal news genre, notably the work of Agarwal et al (2009), but adapt it for the language of social me- dia, in particular Twitter. We show that exploiting lexical-stylistic features and dictionaries geared to- ward social media are useful in detecting sentiment. In this rest of this paper, we discuss related work, including the state of the art sentiment system (Agar- wal et al., 2009) our method is based on, the lexicons we used, our method, and experiments and results. 2 Related Work Several recent papers have explored sentiment anal- ysis in Twitter. Go et al (2009) and Pak and Paroubek (2010) classify the sentiment of tweets containing emoticons using n-grams and POS. Bar- bosa and Feng (2010) detect sentiment using a po- larity dictionary that includes web vocabulary and tweet-specific social media features. Bermingham and Smeaton (2010) compare polarity detection in twitter to blogs and movie reviews using lexical fea- tures. Agarwal et al (2011) perform polarity senti- ment detection on the entire tweet using features that are somewhat similar to ours: the DAL, lexical fea- tures …

Content truncated to 500 words