A multimodal critical perspective on the challenges of automatic detection of hate speech in internet memes: The case of Simianization

In the digital ecosystem, memes are one of the most popular means of communication for Internet users worldwide. While some memes are harmless, others have become powerful means to disseminate hate messages against vulnerable categories and individuals. The ubiquity of the Internet and the potential for virality of malicious contents call for the improvement of Artificial Intelligence (AI) tools to support humans in the prevention of online discrimination and cyberbullying. Research on AI-driven hate speech detection is currently making great strides. Still, it proves problematic in several areas, including the automatic interpretation of text-image clusters (TICs), such as memes. Based on the qualitative analysis of two different datasets (the Hateful Memes Challenge dataset, created by the Facebook AI Research group to train machines, and the TIC dataset, including user-generated memes manually collected by the researcher), this paper investigates memes’ multifarious meaning-making processes, whose interpretation requires pragmatic and intercultural skills, as well as the ability to understand visual/verbal interplays in English multimodal texts. The tools of sociosemiotic and multimodal critical discourse studies are adopted to analyze a set of hateful memes and, in particular, racist forms of dehumanization via simianization (i.e., the portrayal of individuals as primates). In conclusion, this study provides empirical indications about how multimodal-informed research can assist computer science in the development of AI systems for the creation of inclusive digital spaces.