Improving Hateful Meme Detection through Retrieval-Guided Contrastive Learning

TL;DR

This paper introduces a retrieval-guided contrastive learning method to improve hateful meme detection by creating a hatefulness-sensitive embedding space, achieving state-of-the-art results and enabling easy updates without retraining.

Contribution

The paper proposes a novel retrieval-guided contrastive training approach that enhances the sensitivity of multimodal embeddings for hatefulness detection, outperforming larger models.

Findings

Achieved an AUROC of 87.0 on the HatefulMemes dataset.

Demonstrated effective detection of unseen hateful memes.

Enabled system updates through data addition without retraining.

Abstract

Hateful memes have emerged as a significant concern on the Internet. Detecting hateful memes requires the system to jointly understand the visual and textual modalities. Our investigation reveals that the embedding space of existing CLIP-based systems lacks sensitivity to subtle differences in memes that are vital for correct hatefulness classification. We propose constructing a hatefulness-aware embedding space through retrieval-guided contrastive training. Our approach achieves state-of-the-art performance on the HatefulMemes dataset with an AUROC of 87.0, outperforming much larger fine-tuned large multimodal models. We demonstrate a retrieval-based hateful memes detection system, which is capable of identifying hatefulness based on data unseen in training. This allows developers to update the hateful memes detection system by simply adding new examples without retraining, a desirable feature for real services in the constantly evolving landscape of hateful memes on the Internet.