Multi-Perspective Enhanced Sarcasm Detection Models

This paper proposes the Multi-perspective Enhanced Sarcasm Detection Model (MESDM), which improves sarcasm detection by integrating three perspectives: textual, image-based, and inconsistency-based views. The core innovation of the model lies in enhancing the processing of multimodal information through interaction and reinforcement learning across these perspectives. The model uses a Mixture of Experts (MoE) self-attention mechanism in both the textual and image-based modules, combined with contrastive learning to extract meaningful features. A Graph Neural Network (GNN) is employed in the inconsistency-based module to capture structural relationships between text and images. Additionally, a cross-modal attention mechanism is used to fuse information and detect inconsistencies, improving detection accuracy. Experimental results on multiple datasets show that MESDM outperforms traditional methods, achieving significant improvements in precision (85.36%), recall (84.97%), F1 score (84.66%), and accuracy (87.08%), demonstrating the effectiveness of the multi-perspective enhancement in sarcasm detection.