Towards Resource Inequities in Catching the “Dark Side” of Social Media: a Hateful Meme Classification Framework for Low-resource Scenarios

Abstract

The increasing prevalence of social media platforms has led to the emergence of multimodal information such as memes. Hateful memes poses a risk by perpetuating discrimination, reinforcing stereotypes, and causing online harassment, thereby marginalising certain groups and impeding efforts towards inclusivity and social justice. Detecting hateful memes is crucial for creating a safe and equitable online environment. However, existing research heavily relies on complex and large deep learning models, requiring substantial computational resources for training. This creates a barrier for under-resourced researchers and small companies, limiting their participation in hateful information detection and exacerbating inequalities in the field of artificial intelligence. This paper attempts to tackle the problem by proposing a low-resource- oriented framework of hateful meme classification to address limitations in training data, computing power, and modality integration. Our approach achieves faster performance with reduced computational requirements, while maintaining a 94.7% accuracy comparable to the existing highest-scoring model.