Beyond a Joke: Multi-Angle Reasoning for Detecting and Explaining Harmful Humor in Memes

Short summary

MAR-12 is a framework using Vision Language Models to detect and explain harmful humor in internet memes, where humorous and hateful elements coexist. It interprets each meme through twelve structured perspectives derived from humor and hate theories, applies role-aware soft-gated attention, and uses a prototype-based classifier. It achieves up to 80.3% accuracy for humor detection and 75.9% for hate detection, outperforming prior approaches while producing coherent explanations validated by human and GPT-4 evaluations.

• •MAR-12 framework uses 12 structured perspectives from humor and hate theories for meme analysis
• •Achieves 80.3% humor detection and 75.9% hate detection accuracy, beating SOTA
• •Human and GPT-4 evaluations confirm coherent, persuasive explanations for memes with co-occurring humor and harm

Generated with AI, which can make mistakes.