Exploring Emotional Cues Within Conversational Contexts for Sarcasm Detection

Abstract

Sarcasm, widely used in social media and everyday conversations, presents a substantial challenge for natural language understanding. Enhancing the ability to accurately interpret sarcasm is crucial for advancing natural language processing capabilities and enhancing communication comprehension. Traditional approaches to sarcasm detection often overlook the subtle emotional cues embedded within conversational contexts, which play a crucial role in its interpretation. In this paper, we investigate the impact of integrating emotions into contextual data and analyze how excluding neutral emotion influences sarcasm detection. This analysis aims to provide insights into factors that affect sarcasm comprehension. By using Mustard dataset, we employ the DistilBERT model to extract emotional features for each speaker in conversational exchanges. Subsequently, we use transformer architectures like BERT, DistilBERT, Electra, and RoBERTa in training and evaluating sarcasm detection. Our findings highlight the critical role of emotions in sarcasm interpretation, with RoBERTa achieving the highest accuracy of 0.937 and F1-score of 0.943, outperforming other models by up to 4.6% in F1-score. It also surpasses the existing work by 1.6% in F1-score when neutral emotion is excluded from the conversation context and by 4.3 % when neutral emotion is excluded from both the conversation context and the implicit emotions of the last utterance.