Advancing Sarcasm Detection: The Case for Naturalistic Conversational Datasets

Abstract-Sarcasm detection is a critical research area in Natural Language Sarcasm detection plays a pivotal role in advancing Natural Language Processing (NLP), influencing sentiment analysis, emotion recognition, and conversational AI. Effective models rely on diverse, well-annotated datasets that capture subtle linguistic and contextual cues. This review analyzes 46 research papers and categorizes sarcasm detection datasets into three types: text-based, visual-textual, and audio-visual. The findings highlight the growing importance of multimodal datasets for improving recognition in domains such as social media, news, and dialogues. However, current resources—particularly in audio sarcasm—often overlook critical aspects like prosody, emotional tone, and speaker variability. To address these gaps, the review emphasizes the need for naturalistic conversational data that integrates varied accents, emotional nuances, and dynamic contexts. By advancing dataset design toward real-world dialogue, sarcasm detection models can achieve greater accuracy and robustness, enhancing practical applications including healthcare chatbots, e-commerce reviews, and virtual assistants.