Harnessing Ensemble of Data Preprocessing and Hand-crafted Features for Irony Detection in Tweets

In the decade of microblogging, people are more inclined to use figurative languages such as irony to express their opinions towards various issues. Determining the ironic orientation of tweets is crucial for various natural language processing (NLP) tasks especially in opinion mining due to its twisted nature. However, the noisy and informal characteristics of tweets demand a formidable system to address the challenges of detecting ironic tweets. In this paper, we propose an approach for irony detection in tweets. We focus on an ensemble of preprocessing techniques to address the noisy tweet characteristics and idiosyncratic natures. Besides, we utilize an effective combination of n-gram features and various hand-crafted features in a unified supervised classification model. We leverage the lexical syntax, parts-of-speech (POS), and opinionated lexicons to extract the hand-crafted features and employ a feature selection technique to select the best combination. We present a thorough analysis of the contribution of various features and preprocessing techniques using the SemEva1-2018 benchmark irony detection dataset in tweets. Experimental findings demonstrate the efficacy of our method over several state-of-the-art methods.