Résumé: Methods that detect hate speech in memes have become vital in our connected society, especially in the context of many social media companies. Memes are a quick way to transfer ideas, events, or other content from the real world to the digital one. Massively created, they reproduce like viruses and aim to get people's attention. They are powerful tools, that, when used to spread hate speech, are able to have global reach. Meme has a broad definition and different formats, such as short videos, GIFS, challenges, among others. In this paper, we follow the classical format of an image with superimposed text. In this context, the hateful meme detection task is extremely challenging, especially due to memes' multimodal nature, i.e., they have two different sources: image and text. Consequently, when dealing with memes, a classification model needs to tackle both components in order to classify them as hateful or not-hateful. This work contributes to the effort to solve this task. We list the most recent research, synthesize and discuss the approaches proposed in the current literature by providing a critical analysis of these methods, highlighting their strengths and points to improve. We also introduce a taxonomy to allow grouping similar approaches. Our conclusion indicates that, despite the few studies currently available and the few public datasets specially designed for this topic, there is an evolution in the methodologies used, which is reflected in the evolution of the results attained.

Langue: Anglais