Irony Detection Using Transformers

With the ever-expanding social net, the use cases of irony detection and classification is also exponentially increasing. With this work, we take Task3 of SemEva1-2018 as our problem statement which further has two tasks. We intend to first determine if a given tweet is ironic or not (Task A) and then classify the tweets into four classes viz. non-ironic, verbal irony with contrast, verbal irony without contrast and situational irony (Task B). Existing papers have mainly exploited the lexical features of tweets using supervised machine learning. Here, we have proposed two NLP Transformer models viz. BERT (Bidirectional Encoder Representations from Transformers) and XLNets to classify tweets and have also compared our results to that of past papers. Using BERT, we have achieved F1 scores of 0.70 and 0.75 and using XLNets 0.74 and 0.59 for Task A and Task B respectively.