HCIG: A Hierarchical Cross-Modal Incongruity Graph Network for Multimodal Sarcasm and Cyberbullying Detection

This paper introduces HCIG, a graph-based framework for multimodal sarcasm and cyberbullying detection. It models text-image incongruity at token, phrase, and global levels with graph attention networks, then combines these representations using learned hierarchical attention. The authors also present GCCN, which uses contradiction-aware pooling for efficient multimodal interaction. On the MMSD benchmark, HCIG reports 85.74% accuracy and 85.29% macro-F1. On MultiBully, GCCN achieves the highest macro-F1 at 68.66%, while HCIG reaches 69.62% accuracy and 74.90% bullying-class F1.