This paper presents M-NLE, a compact knowledge-augmented multitask model for jointly detecting offensive memes and generating natural language explanations, and suggests that knowledge-augmented explanation generation is a practical direction for more interpretable offensive meme detection.
Abstract
Memes combine visual content and short textual cues to convey humor, sarcasm, and social commentary, but they can also encode offensive or harmful messages targeting individuals and social groups. While recent multimodal systems have improved offensive meme detection, many of them operate as opaque classifiers and do not explain why a meme is considered offensive. This limits their usefulness in moderation settings, where human reviewers and policy stakeholders often require an interpretable account of the harmful implication. In this paper, we present M-NLE, a compact knowledge-augmented multitask model for jointly detecting offensive memes and generating natural language explanations. M-NLE retrieves external contextual knowledge in the form of visual entities and their descriptions, incorporates this knowledge through a dedicated knowledge encoding module, and uses the resulting multimodal representation for both offensiveness classification and explanation generation. Experiments on the extended Facebook Hateful Memes and MultiOFF datasets show that M-NLE achieves competitive classification and reference-based generation performance across multiple data partitions while remaining substantially smaller than general-purpose LLM/VLM systems. Human evaluation shows that M-NLE improves explanation adequacy and fluency over the evaluated baselines while also highlighting areas for further improvement. Overall, 38% of the generated explanations were rated as not related to the input, whereas 15% were judged to fully justify the reference interpretation. These findings suggest that M-NLE produces more useful and fluent explanations than existing baselines, while leaving room to further improve explanation relevance and justification. Ablation studies further show the contribution of knowledge encoding and multitask learning, while detailed qualitative analyses reveal both the benefits and limitations of retrieved knowledge, including cases where it introduces noise or bias. Overall, our results suggest that knowledge-augmented explanation generation is a practical direction for more interpretable offensive meme detection. To facilitate reproducibility, we publicly release the complete implementation, experimental scripts, and detailed documentation at https://github.com/newcodevelop/M-NLE
This work introduces MemeCF, a cue-focused benchmark of 9,895 memes across harm, hate, and sarcasm, with annotations identifying the modality and rationale of the pivotal evidence.
Akshit Sharma, Prashant W. Patil· Conference on Empirical Meth...· 1 citation
The findings validate that LLMs can effectively detect strategic narratives in a zero-shot setting and when combined with simple ensembling and human-written descriptions, they can rival supervised systems, offering a scalable solution for narrative detection, specially when there is no training data for the vast major...
Jesús M. Fraile-Hernández, Anselmo Peńas, Patrick Giedemann· 0 citations
The results demonstrate that multilevel evaluation improves the reliability that the proposed framework effectively balances performance and efficiency, which makes the framework suitable for practical hate speech moderation systems that require accurate, generalizable, and explainable decisions.
Rameesha Zia, Muhammad Shahid Iqbal Malik· 0 citations
A multi-agent framework in which a story agent, an image agent, and a critic agent collaborate to produce fake social media posts that plausibly counter true news is introduced, providing a foundation for developing robust defenses against social media fake news.
Ji-Yao Yang, Yang Liu, Zhen Qin et al.· Conference on Empirical Meth...· 0 citations
Hateful memes spread abusive content through implicit interactions between images and text, posing serious threats to the safety of online communities. In recent years, multimodal large language models have been widely used for hateful meme detection and are increasingly adopted to generate explainable detection result...
Bo Xu, Chen-Yuan Wang, Xin-Yu Chen et al.· 0 citations
Sarcasm detection remains a challenging task in Natural Language
Processing (NLP) because sarcasm often conveys a negative meaning with positive words.
Furthermore, the extensive use of social media has significantly transformed the way people communicate,
making automatic sarcasm detection gradually important. Tra...