BOUNCE: Sentiment Classification in Twitter using Rich Feature Sets

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 554–561, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics BOUNCE: Sentiment Classification in Twitter using Rich Feature Sets Nadin K¨okciyan†, Arda C¸ elebi†, Arzucan ¨Ozg¨ur, Suzan ¨Usk¨udarlı Department of Computer Engineering Bogazici University Istanbul, Turkey {nadin.kokciyan,arda.celebi,arzucan.ozgur,suzan.uskudarli}@boun.edu.tr Abstract The widespread use of Twitter makes it very interesting to determine the opinions and the sentiments expressed by its users. The short- ness of the length and the highly informal na- ture of tweets render it very difficult to auto- matically detect such information. This paper reports the results to a challenge, set forth by SemEval-2013 Task 2, to determine the posi- tive, neutral, or negative sentiments of tweets. Two systems are explained: System A for de- termining the sentiment of a phrase within a tweet and System B for determining the senti- ment of a tweet. Both approaches rely on rich feature sets, which are explained in detail. 1 Introduction Twitter consists of a massive number of posts on a wide range of subjects, making it very interesting to extract information and sentiments from them. For example, answering questions like ‘What do Twitter users feel about the brand X?’ are quite interesting. The constrained length and highly informal nature of tweets presents a serious challenge for the auto- mated extraction of such sentiments. Twitter supports special tokens (i.e. mentions and hashtags), which have been utilized to determine the sentiment of tweets. In (Go et al., 2009), emoticons are used to label tweets. In (Davidov et al., 2010), Twitter emoticons as well as hashtags are used to la- bel tweets. O’Connor et al. (2010) demonstrated a correlation between sentiments identified in pub- lic opinion polls and those in tweets. A subjectivity † These authors contributed equally to this work lexicon was used to identify the positive and nega- tive words in a tweet. In (Barbosa and Feng, 2010), subjective tweets are used for sentiment classifica- tion. They propose the use of word specific (e.g. POS tags) and tweet specific (e.g. presence of a link) features. Most of these studies use their own anno- tated data sets for evaluation, which makes it diffi- cult to compare the performances of their proposed approaches. Sentiment Analysis in Twitter 2013 (SemEval 2013 Task 2) (Wilson et al., 2013) presented a chal- lenge for exploring different approaches examin- ing sentiments conveyed in tweets: interval-level (phrase-level) sentiment classification (TaskA) and message-level sentiment classification (TaskB). Sen- timent are considered as positive, negative, or neu- tral. For TaskA, the goal is to determine the sen- timent of an interval (consecutive word sequence) within a tweet. For TaskB, the goal is to determine sentiment of an entire tweet. For example, let’s con- sider a tweet like ‘Can’t wait until the DLC for ME3 comes out tomorrow. :-)’. For TaskA, the interval 0-1 (Can’t wait) is ‘positive’ and the interval 10-10 (:-)) is ‘positive’. For TaskB, this tweet is ‘positive’. In …

Content truncated to 500 words