Sarcasm Through the Looking Glass: Multi-Domain Analysis for Improved Detection

Sarcasm plays a crucial role in communication today, especially on Social Media. Typically, sarcasm is conveyed through verbal cues (intonation) and non-verbal gestures (eye-roll, shoulder shrug). However, detecting sarcasm in text is complex compared to in-person communication. The challenge lies in recognizing sarcasm without the aid of verbal and non-verbal cues. Existing solutions involve creating Models that can predict if a text is Sarcastic in nature for a very specific text corpora pertaining to a single domain or source. However, these solutions fail when tested on different text from other domains, as the use of sarcasm varies within each of the social media domain. Our paper emphasizes on challenges that arise while attempting to gauge sarcasm using machine learning models and proves that when we combine text from multiple domains, we increase a models ability to correctly classify sarcasm. Specifically, we find increases of up to 27% in accuracy when a model incorporates multiple domains compared to single domain. We believe that by narrowing down the definition of sarcasm and excluding explicit text from its definition on various social media platforms, we can strengthen and aid the future of sarcasm detection within natural language processing. Our research conducts experiments that support this hypothesis.