Detecting Sarcasm in Multiple Modalities by Leveraging Sarcasm Type Correlations

Sarcasm detection is essential for accurately interpreting communication in applications such as dialogue systems. However, most existing approaches treat sarcasm as a single phenomenon and ignore the linguistic distinction between illocutionary, embedded, and propositional sarcasm, each of which relies on different contextual and emotional cues. In this work, we propose a type-aware multimodal framework that explicitly models these three sarcasm types by training dedicated representations for each and combining them within a unified architecture. To encourage the learning of emotionally informative features, we incorporate explicit and implicit emotion prediction as auxiliary regularization tasks in a multitask learning setting. Experiments on the multimodal sarcasm dataset MUStARD++ show that our system achieves an F1 score of 82.65% in sarcasm detection. These results demonstrate that type-specific modeling improves multimodal sarcasm detection.