Detecting Harmful Memes with Decoupled Understanding and Guided CoT Reasoning

Summary

This paper talks about U-CoT+, a new method designed to detect harmful memes by first turning the images and text of memes into written descriptions. It then uses human-created guidelines combined with a special reasoning approach called zero-shot Chain-of-Thought (CoT) prompting to help small language models understand and explain why a meme might be harmful.

What's the problem?

The problem is that memes combine pictures and words in ways that are often tricky for AI to understand, making it hard to spot if a meme is harmful or offensive. Many existing AI models struggle because they either can't explain their decisions well or need lots of training data and large models, which can be expensive and less flexible.

What's the solution?

The solution was to create U-CoT+, which breaks down the problem by converting memes into detailed text descriptions first, making it easier for models to analyze. It then guides the AI with clear, expert-created rules and uses zero-shot Chain-of-Thought prompting to let the model think step-by-step without needing additional training. This approach allows small language models to better understand and explain harmful content in memes without extra data or big models.

Why it matters?

This matters because harmful memes can spread quickly online and cause real damage, so having smart, explainable AI that can detect them is very important. U-CoT+ makes it possible to use smaller, more efficient AI systems to spot harmful content while also providing clear reasons behind their decisions, helping build safer social platforms and making AI tools more trustworthy and practical.