Abstract

The automatic identification of harmful content is a crucial concern for social media platforms, policymakers, and society. While researchers have analyzed textual, visual, and audio content separately, harmful media often integrates multiple modalities, such as memes. This study provides a comprehensive survey on harmful meme detection, introducing a new typology based on recent literature. Findings indicate that certain harmful meme categories, such as those related to self-harm and extremism, are underrepresented in existing research due to dataset limitations. The study explores challenges related to multimodal analysis, technological barriers, and social implications, proposing future research directions.