Enhancing Multimodal Sarcasm Detection Via Global and Local Prompt Mechanisms

With the rapid development of social media, the volume and complexity of multimodal data have significantly increased, posing new challenges for tasks such as multimodal sarcasm detection. This paper proposes a novel framework, Global and Local Prompt Learning (GLPL), to address these challenges by jointly capturing both global contextual and local fine-grained features in multimodal data. The global prompts align semantic information across modalities to preserve the overall meaning, while the local prompts extract detailed, modalityspecific interactions from text-image pairs, enhancing sensitivity to subtle sarcastic cues. To further improve robustness and generalization, the global channel incorporates a random prompt loss strategy to avoid over-reliance on specific prompt features, and the local channel introduces a multi-scale information mining mechanism to capture sarcasm cues at different levels of granularity. This paper also conducts comprehensive ablation studies to validate the effectiveness of each module. Experimental results on the benchmark dataset demonstrate that GLPL achieves strong and consistent performance, significantly outperforming most existing approaches in identifying nuanced multimodal sarcasm.