Detecting Sarcasm Using Commonsense Reasoning Capabilities of Comet and Representation Power of Bert

Abstract

Sarcasm poses significant challenges for natural language processing systems due to its reliance on implicit meaning and contextual cues. While models like BERT have advanced sarcasm detection, they still face difficulties with non-literal interpretations. This paper introduces a hybrid model that combines BERT's contextual representations with COMET's commonsense reasoning. The approach involved fine-tuning BERT on sarcasm datasets and integrating commonsense inferences via a feature integration layer. Evaluated on three key datasets, the BERT–COMET model demonstrated superior performance in recognizing subtle sarcastic expressions and contradictions, highlighting that incorporating commonsense knowledge can enhance detection accuracy and reduce errors.