Sarcasm detection of tweets: A comparative study

Abstract

Sarcasm is a nuanced form of communication where the individual states opposite of what is implied. One of the major challenges of sarcasm detection is its ambiguous nature. There is no prescribed definition of sarcasm. Another major challenge is the growing size of the languages. Every day hundreds of new slang words are being created and used on these sites. Hence, the existing corpus of positive and negative sentiments may not prove to be accurate in detecting sarcasm. Also, the recent developments in online social networks allow its users to use varied kind of emoticons with the text. These emoticons may change the polarity of the text and make it sarcastic. Due to these difficulties and the inherently tricky nature of sarcasm it is generally ignored during social network analysis. As a result the results of such analysis are affected adversely. Thus, sarcasm detection poses to be one of the most critical problems which we need to overcome. Detection of sarcastic content is vital to various NLP based systems such as text summarization and sentiment analysis. In this paper we address the problem of sarcasm detection by leveraging the most common expression of sarcasm — “positive sentiment attached to a negative situation”. Our work uses two ensemble based approaches — voted ensemble classifier and random forest classifier. Unlike current approaches to sarcasm detection which rely on existing corpus of positive and negative sentiments for training the classifiers, we use a seeding algorithm to generate training corpus. The proposed model also uses a pragmatic classifier to detect emoticon based sarcasm.