Abstract:
Sarcasm is a nuanced and context-dependent communication that poses a significant challenge for natural language processing (NLP) systems. This study proposes a novel approach to improving sarcasm understanding by generating explanations and targets for sarcastic input using a multi-modal encoder-decoder transformer model. Our approach builds on an existing dataset called MORE by adding 2000 more instances and annotating the target of ridicule for all the sarcastic instances. The key idea behind our approach is to provide the model with more contextual information and common-sense knowledge to better understand sarcasm’s complex and subtle nature. The model can give users a more accurate and adequate understanding of the underlying sarcasm by generating explanations and targets for sarcastic input. Our experimental results show that our proposed approach gives better results than the existing benchmark on the task of sarcasm explanation. Furthermore, our approach is highly interpretable, as the generated explanations and targets provide valuable insights into the model’s decisionmaking process. Overall, our study demonstrates the effectiveness of our proposed approach for improving sarcasm understanding in NLP systems. Our approach has critical practical applications in sentiment analysis, opinion mining, and social media monitoring, where sarcasm understanding is crucial.