Abstract: Internet memes,characterized by high metaphoricity and multimodal features,have emerged as a new and covert me-dium for disseminating online hate speech.Current end-to-end detection methods lack the capability for deep metaphorical decoding,while generative reasoning based on large language models is prone to reasoning hallucinations detached from visual facts and faces bottlenecks in high-parameter deployment.To address these issues,this paper proposes a hateful meme detection framework enhanced by large models and knowledge graphs.The framework achieves the modal decoupling of perception and rea-soning.Firstly,a visual structure prior dictionary is introduced to accurately extract image elements.Secondly,deep multi-hop logical evidence chains are constructed through knowledge graph augmentation to establish profound semantic associations.Subsequently,a “draft-audit-refine” reflective reasoning mechanism is designed to anchor reasoning in objective facts through cross-modal self-audit,curbing hallucinations at the algorithmic level.Finally,a progressive multimodal knowledge distillation strategy is proposed to efficiently transfer complex reasoning capabilities to lightweight target models.Experimental results demonstrate that the proposed framework achieves a detection accuracy of 91.92% on the Harm-P dataset,outperforming existing baseline methods.Notably,the faithfulness score of the generated explanations reaches 4.63(out of 5),significantly suppressing reasoning hallucinations.In cross-architecture experiments,the framework improves the accuracy of base models by up to 33.82 percentage points.

Key words: Hateful meme detection, Knowledge graph augmentation, Interpretability, Hallucination suppression, Knowledge distillation, Multimodal large language models