Multi-Modal Sarcasm Detection and Humor Classification in Code-Mixed Conversations

Sarcasm detection and humor classification are inherently subtle problems, primarily due to their dependence on the contextual non-verbal information. Furthermore, existing studies in these two topics usually constrained non-English languages such as Hindi, unavailability of qualitative annotated datasets. In this work, we make major contributions considering above limitations: (1) develop a Hindi-English code-mixed dataset, MaSaC , 1 for multi-modal sarcasm conversational dialog, which our knowledge is first dataset its kind; (2) propose MSH-COMICS 2 novel attention-rich neural architecture utterance classification. We learn efficient representation utilizing hierarchical attention mechanism that attends small portion input sentence at time. Further, incorporate dialog-level leverage dialog history perform extensive experiments both tasks by varying inputs various submodules . also conduct comparative analysis against approaches. observe attains superior performance over models $>$ 1 F1-score point 10 points diagnose model thorough results understand superiority pitfalls.