Augmented BERT for Online Sarcasm Detection

Sarcasm is a complex and subtle linguistic construct, which often conveys a meaning opposite to its literal interpre-tation. Detecting sarcasm from social media conversations is challenging yet critical to correctly understand the underlying sentiment and opinion, particularly during socially volatile events like the COVID-19 pandemic. This paper proposes a multi-modal approach that integrates non-textual, user- and post-specific metadata features into BERT, a state-of-the-art transformer model that traditionally operates only on textual input. Although our approach was motivated by the politically charged social media conversations during the COVID-19 pandemic, we explore its performance on three other public domain data sets that express sarcasm in a more structured and formal way in contrast to the noisy and informal social media posts. The results show that our approach offers a higher F1-score across all the data sets, when compared with baseline transformers as well as conventional machine learning and deep learning models, as it provides a better balance between precision and recall. This paper thus advances sarcasm detection, and looking beyond, provides a framework to detect misinformation and hostility from online conversations.