Kea: Expression-level Sentiment Analysis from Twitter Data

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 530–534, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics Kea: Expression-level Sentiment Analysis from Twitter Data Ameeta Agrawal Computer Science and Engineering York University Toronto, Canada ameeta@cse.yorku.ca Aijun An Computer Science and Engineering York University Toronto, Canada aan@cse.yorku.ca Abstract This paper describes an expression-level senti- ment detection system that participated in the subtask A of SemEval-2013 Task 2: Senti- ment Analysis in Twitter. Our system uses a supervised approach to learn the features from the training data to classify expressions in new tweets as positive, negative or neutral. The proposed approach helps to understand the rel- evant features that contribute most in this clas- sification task. 1 Introduction In recent years, Twitter has emerged as an ubiquitous and an opportune platform for social activity. Ana- lyzing the sentiments of the tweets expressed by an international user-base can provide an approximate view of how people feel. One of the biggest chal- lenges of working with tweets is their short length. Additionally, the language used in tweets is very informal, with creative spellings and punctuation, misspellings, slang, new words, URLs, and genre- specific terminology and abbreviations, such as, RT for “re-tweet” and #hashtags, which are a type of tagging for tweets. Although several systems tackle the task of analyzing sentiments from tweets, the task of analyzing sentiments at term or phrase-level within a tweet has remained largely unexplored. This paper describes the details of our expression- level sentiment detection system that participated in the subtask A of SemEval-2013 Task 2: Sentiment Analysis in Twitter (Wilson et al., 2013). The goal is to mark expressions (a term or short phrases) in a tweet with their contextual polarity. This is chal- lenging given the fact that the entire length of a tweet is restricted to just 140 characters. We describe the creation of an SVM classifier that is used to classify the contextual polarity of expressions within tweets. A feature set derived from various linguistic fea- tures, parts-of-speech tagging and prior sentiment lexicons was used to train the classifier. 2 Related Work Sentiment detection from Twitter data has attracted much attention from the research community in re- cent times (Go et al., 2009; Pang et al., 2002; Pang and Lee, 2004; Wilson et al., 2005; T. et al., 2012). However, most of these approaches classify entire tweets by their overall sentiment (positive, negative or neutral). The task at hand is to classify expressions with their contextual sentiment. Most of these expres- sions can be found in sentiment lexicons already an- notated with their general polarity, but the focus of this task is to detect the polarity of that expression within the context of the tweet it appears in, and therefore, given the context, the polarity of the ex- pression might differ from that found in any lexicon. One of the primary goals of this task is to facilitate the creation of a …

Content truncated to 500 words