DsUniPi: An SVM-based Approach for Sentiment Analysis of Figurative Language on Twitter

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 709–713, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics DsUniPi: An SVM-based Approach for Sentiment Analysis of Figurative Language on Twitter Maria Karanasou Dept. of Digital Systems University of Piraeus Greece karanasou@gmail.com Christos Doulkeridis Dept. of Digital Systems University of Piraeus Greece cdoulk@unipi.gr Maria Halkidi Dept. of Digital Systems University of Piraeus Greece mhalk@unipi.gr Abstract The DsUniPi team participated in the SemEval 2015 Task#11: Sentiment Analysis of Figura- tive Language in Twitter. The proposed ap- proach employs syntactical and morphological features, which indicate sentiment polarity in both figurative and non-figurative tweets. These features were combined with others that indi- cate presence of figurative language in order to predict a fine-grained sentiment score. The method is supervised and makes use of struc- tured knowledge resources, such as Senti- WordNet sentiment lexicon for assigning sentiment score to words and WordNet for cal- culating word similarity. We have experiment- ed with different classification algorithms (Naïve Bayes, Decision trees, and SVM), and the best results were achieved by an SVM clas- sifier with linear kernel. 1 Introduction Sentiment analysis on figurative speech is a chal- lenging task that becomes even more difficult on short social-media related text. Tweet text can be rich in irony that is either stated with hashtags ex- plicitly (such as #irony) or implied. Identifying the underlying sentiment of such text is challenging due to its restricted size and features such as use of abbreviations and slang. Consequently, assigning positive or negative polarity is quite a difficult task. The actual meaning can be very different than what is stated, since, for example, in ironic lan- guage what is said can be the opposite of what it is meant. To address this challenge, we propose a system for sentiment analysis of figurative lan- guage, which relies on feature selection and trains a classifier to predict the label of a tweet. Given a labelled trial set, the objective of the system is to correctly determine how positive, negative or neu- tral a tweet is considered to be on a scale of [-5, 5]. 2 Related Work Tweets have unique characteristics compared to other text corpora, such as emoticons, abbrevia- tions, and hashtags. Use of emoticons is considered a reasonably effective way to conveying emotion (Derks et al. 2008, Thelwall et al.). Go et al. (2009) show that machine learning algorithms achieve accuracy above 80% when trained with emoticon data. It is also indicated that the use of hashtags and presence of intensifiers, such as capitalization and punctuation, can affect sentiment identification (Kouloumpis et al., 2010). According to Agarwal et al. (2011) such features can add value to a clas- sifier, but only marginally. Additionally, natural language related features, such as part-of-speech tagging and use of lexicon resources, can signifi- cantly contribute to detecting the sentiment of a tweet. Moreover, features that combine the prior polarity of words and their parts-of-speech tags are considered most useful. The problem of sentiment …

Content truncated to 500 words