Improving Hateful Meme Detection through Retrieval-Guided Contrastive Learning

TL;DR

This work demonstrates a retrieval-based hateful memes detection system, which is capable of identifying hatefulness based on data unseen in training, and allows developers to update the hateful memes detection system by simply adding new examples 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.

Figures (2)

• Figure 1: Illustrative examples from KielaFBHMC2020. The meme on the left is hateful, the middle one is a benign image confounder, and the right one is a benign text confounder. We show HateCLIPper's KumarHateClip2022prediction below each meme. HateCLIPper misclassifies the hateful meme on the left as benign.
• Figure 2: Model overview. Using VL Encoder $\mathcal{F}$ to extract the joint vision-language representation for a training example $i$. Additionally, the VL Encoder encodes the training memes into a retrieval database $\mathbf{G}$. During training, pseudo-gold and hard negative examples are obtained using the Faiss nearest neighbour search. During inference, $K$ nearest neighbours are obtained using the same querying process to perform the KNN-based inference. During training, we optimise the joint loss function $\mathcal{L}$. For inference, we use conventional logistic classifier and our proposed retrieval-based KNN majority voting. For a test meme $j$, we denote the prediction from logistic regression and KNN classifier as $\hat{y}_j$ and $\hat{y}'_j$, respectively.