DGLF: A Dual Graph-based Learning Framework for Multi-modal Sarcasm Detection

Capturing inter-modal incongruities within the text-image pair is a critical challenge in multimodal sarcasm detection (MSD). Fortunately, graph neural networks (GNNs) have made promising advancements in MSD, which show advantages in explicitly capturing data relationships. Nevertheless, current GNN-based MSD methods do not effectively address some of the inherent limitations of GNNs, which include: 1) neglecting high-order relationships, and 2) underestimating high-frequency messages. In this paper, we propose a Dual Graph-based Learning Framework (DGLF) to address the above two issues. Specifically, we construct a hypergraph to perform high-order aware propagation and a vanilla graph to perform highfrequency enhanced propagation, respectively. We empower GNNs to 1) better capture the inherent and complicated relationships based on the hypergraph and 2) deliver sufficient modeling through high-frequency enhanced messages on the vanilla graph. Besides, we introduce multi-modal fusion information bottleneck to effectively fuse the two learned graph features. Experimental results on two benchmark datasets show that the proposed model outperforms previous state-of-the-art methods.