Sentiment Analyzer with Rich Features for Ironic and Sarcastic Tweets

Sentiment Analyzer with Rich Features for Ironic and Sarcastic Tweets Piyoros Tungthamthiti1, Enrico Santus2, Hongzhi Xu2, Chu-Ren Huang2, and Kiyoaki Shirai1 1Japan Advanced Institute of Science and Technology 1-1, Asahidai, Nomi City, Ishikawa, Japan 923-1292 {s1320204,kshirai}@jaist.ac.jp 2Dept. of Chinese and Bilingual Studies The Hong Kong Polytechnic University, Hong Kong {e.santus,hongzhi.xu}@connect.polyu.hk {churen.huang}@polyu.edu.hk Abstract Sentiment Analysis of tweets is a complex task, because these short messages employ unconventional language to increase the ex- pressiveness. This task becomes even more difficult when people use figurative language (e.g. irony, sarcasm and metaphors) because it causes a mismatch between the literal mean- ing and the actual expressed sentiment. In this paper, we describe a sentiment analysis sys- tem designed for handling ironic and sarcastic tweets. Features grounded on several linguis- tic levels are proposed and used to classify the tweets in a 11-scale range, using a decision tree. The system is evaluated on the dataset re- leased by the organizers of the SemEval 2015, task 11. The results show that our method largely outperforms the systems proposed by the participants of the task on ironic and sar- castic tweets. 1 Introduction Whenever a message is encoded into linguistic form for being communicated – either in a spoken or writ- ten text1 – information revealing judgments, evalua- tions, attitudes and emotions is also encoded (Mar- tin and White, 2005). This is true for both infor- mal and formal texts, independently of how much attention the writer pays in cleaning such informa- tion out. This is also true for texts posted on social networks (i.e. Facebook, Twitter, etc.), where judg- ments, evaluations, attitudes and emotions constitute an important part of the message (Pak and Paroubek, 2010). Sentiment analysis (also known as opinion mining and subjectivity analysis) is a Natural Language Pro- cessing (NLP) task that focuses on identification of 1In this paper we will mainly refer to written texts, but most of what is said is also applicable to spoken ones. such judgments, evaluations, attitudes and emotions. It can be compared to other classification tasks, as it consists in associating the analyzed texts with a la- bel that represents the sentiment of the message or the affective state of the writer (Hart, 2013). In its earliest incarnations, sentiment analysis was limited to the identification of the polarity of the texts, and the classification label was either positive or negative. Later on, the task was extended to ad- dress more challenging and complex goals, such as the identification of the sentiment of the messages or the writer’s affective state in a more fine scale, with labels including anger, happiness or depression. Such extension could not avoid considering one of the most pervasive tools used in communication, namely figurative language. In fact, this expressive tool is not only very frequent in various kinds of texts, but it also strongly affects the sentiment ex- pressed in the text, often completely reversing its polarity (Xu et al., 2015; Ghosh et al., 2015). Because figurative language is used in unpre- dictable ways …

Content truncated to 500 words