An emoji centric approach to sarcasm detection in online discourse.

Sarcasm detection has gained significance in sentiment analysis, especially when social media is rife with cyberbullying and trolling. Emojis have garnered researchers' interest as they are polysemic. But they are usually employed in combination with other modalities in sarcasm detection. This work proposes an emoji-focused approach to study their role in sarcasm detection. This structured approach studies how emojis, sentiment of text and emojis, most frequent emoji in text, and its position helps with sarcasm classification task. Experiments include various machine learning classifiers and BERT fine-tuned for emojis to reveal the decisive role of emojis in sarcasm detection, even in the absence of any other modality. Novel sarcasm-aware GloVe-based emoji embeddings are presented that outperform other available emoji embeddings to achieve highest F1, MCC, and RoC_AuC scores on two unseen datasets. Emoji embeddings, BERT fine-tuned with emojis, and emoji-focused models presented in this work can be used by researchers as baseline for sarcasm classification when situational or conversational context and other modalities like visuals, audio etc. are absent. This emoji-focused approach can be useful in identifying bullying or hateful content on public platforms and even on private chat-based platforms where users may more frequently imply sarcasm under the guise of emojis.