Sarc-M: Sarcasm Detection in Typo-graphic Memes

Abstract

Detecting sarcastic tone, which conveys a sharp, bitter, or cutting expression, remark or taunt in natural language is tricky even for humans, making its automated detection more arduous. The growing use of typo-graphic images, that is text represented as an image further characterizes the power of expressiveness in online social data. This research proffers a model Sarc-M, a sarcastic meme predictor, for sarcasm detection in typo-graphic memes using supervised learning based on lexical, pragmatic and semantic features. The learning model is evaluated using five different classifiers and the results are evaluated using a balanced dataset of typo-graphic images, called MemeBank, scrapped from Instagram. The contribution of the research is two-fold, firstly, typo-graphic text is extracted using optical character recognizer and then analyzed for sarcasm and secondly for detecting sarcasm the need of contextual information is explored, that is, contextual cues such as frequency of punctuations and sentiment words are considered as features. The best sarcasm prediction model for typo-graphic memes is built using Multi Layer perceptron which achieves an accuracy of approximately 88%.