Mei Jingbiao ; Chen Jinghong ; Lin Weizhe ; Byrne Bill ; Tomalin Marcus

Abstract

Hateful memes have emerged as a significant concern on the Internet.Detecting hateful memes requires the system to jointly understand the visualand textual modalities. Our investigation reveals that the embedding space ofexisting CLIP-based systems lacks sensitivity to subtle differences in memesthat are vital for correct hatefulness classification. We propose constructinga hatefulness-aware embedding space through retrieval-guided contrastivetraining. Our approach achieves state-of-the-art performance on theHatefulMemes dataset with an AUROC of 87.0, outperforming much largerfine-tuned large multimodal models. We demonstrate a retrieval-based hatefulmemes detection system, which is capable of identifying hatefulness based ondata unseen in training. This allows developers to update the hateful memesdetection system by simply adding new examples without retraining, a desirablefeature for real services in the constantly evolving landscape of hateful memeson the Internet.