KELabTeam: A Statistical Approach on Figurative Language Sentiment Analysis in Twitter

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 679–683, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics KELabTeam: A Statistical Approach on Figurative Language Sentiment Analysis in Twitter Hoang Long Nguyen, Trung Duc Nguyen and Dosam Hwang Department of Computer Engineering Yeungnam University, Korea {longnh238, duc.nguyentrung, dosamhwang}@gmail.com Jason J. Jung∗ Department of Computer Engineering Chung-Ang University, Korea j2jung@gmail.com Abstract In this paper, we propose a new statistical method for sentiment analysis of figurative language within short texts collected from Twitter (called tweets) as a part of SemEval- 2015 Task 11. Particularly, the proposed model focuses on classifying the tweets into three categories (i.e., sarcastic, ironic, and metaphorical tweet) by extracting two main features (i.e., term features and emotion pat- terns). Our experiments have been conducted with two datasets, which are Trial set (1000 tweets) and Test set (4000 tweets). Perfor- mance is evaluated by cosine similarity to gold annotations. Using this evaluation methodol- ogy, the proposed method achieves 0.74 on the Trial set. On the Test set, we achieve 0.90 on sarcastic tweets and 0.89 on ironic tweets. 1 Introduction Sentiment analysis in computer science is a diffi- cult task which aims to identify the emotion from a given data source. The goal of sentiment analysis is to dissect a given document and determine whether its opinion represent positive, negative, or neutral. There have been many studies (which use lexicon- based methods and machine learning-based meth- ods) to extract and identify the sentiment (Medhat et al., 2014). In case of figurative language, the task becomes more challenging because the document can have secondary or extended meanings. Hence, exactly finding the truth meaning of figurative lan- guage is an interesting problem for researchers due to its importance. ∗Corresponding author The first work that we want to mention here is contributed by Reyes and Rosso (2013a). The authors captured ironic sentences from low-level to high-level of irony according to three con- ceptual layers and their eight textual features. With customer reviews on Amazon, Reyes and Rosso (2012a) contributed an approach for distin- guishing irony and non-irony based on six mod- els. Also focusing on detecting irony, Hao and Veale (2010) classifies irony and non-irony by ana- lyzing the large quantity of simile forms with 9-steps sequence. By considering short texts with case- study is Twitter, Reyes et al. (2013b) introduced a model to detect verbal irony by combining four types of conceptual features and their dimensions. Focusing on comprehending metaphor, Shutova et al. (2010) used unsupervised methods to find the associate from a small set of metaphorical expres- sions by verb and noun clustering processing to de- tect similarity structure of metaphor. Finally, Reyes et al. (2012b) analyzed humor and irony by adding more features to express the favorable and unfavor- able ironic contexts using the theory of textual. These above studies tried to solve the problem by focusing on lexical level. Therefore, the goal of our research is to find a new way to identify …

Content truncated to 500 words