Sarcasm Detection on News Headline Dataset Using Language Models

Mechanized detection of sarcasm in textual data is a challenging task in NLP, as it involves recognizing the opposite of what is actually meant. This has become important today since many artificial intelligence tasks require understanding the user and answering appropriately. In this paper, we investigate multiple deep learning models for sarcasm detection on a news headline dataset, including BiLSTM and BERT. For the first time, hybrid BERT models have been used in our study to identify sarcasm in non-social media writings. The results show that the BERT and GBDT strategy combined performs better than all the other strategies listed.