Skip to content
Book Open access

Deconstructing Harmful Memes: Fine-Grained Analysis of Subtle Semantics

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 69 references

Abstract

Memes, widely shared on the internet as a form of cultural expression, present a challenge in identifying harmful content due to their compact format and subtle meanings. Current detection methods employ pre-trained models to extract image and text features from memes, subsequently combining these features through deep neural networks for classification. However, these approaches often overlook the underlying meanings embedded in memes, focusing solely on shallow semantics and surface-level harmful signals. This makes it difficult to capture the cultural connotations and subtle expressions inherent in memes. Moreover, they lack explanations for their judgments. To address these challenges, we propose a novel framework for explainable harmful meme detection. This framework combines the advantages of Multimodal Large Language Models (MLLMs) and fine-grained heterogeneous semantic graphs. By leveraging multi-hop reasoning, we utilize MLLMs' extensive knowledge to provide comprehensive meme analyses, addressing meme literacy gaps. We extract key objects, relationships, and attributes from images to produce detailed scene graphs, which support a deeper understanding of the meme. The fine-grained heterogeneous semantic graphs filter and integrate multiple information sources to uncover hidden harmful elements. Additionally, through multi-hop reasoning and scene graphs, our approach generates reliable explanations via MLLMs. Experimental results demonstrate that our method outperforms existing state-of-the-art approaches on widely used datasets.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.