Hate Speech in Pixels: Detection of Offensive Memes towards Automatic Moderation (Meta)

Unlike any other study to our knowledge, this one aims to automatically detect hate speech by utilizing visual information. It tackles the problem of hate speech identification in Internet memes. Memes are multimedia documents made of pixels that include sentences that, when put together, take on a humorous meaning. However, automated identification of hate memes would lessen their detrimental societal influence because they are also used to disseminate hatred through social networks. The model can learn to identify some of the memes, according to Metas’s research, but a straightforward design is still far from solving the problem. A dataset of 5,020 memes was constructed for Meta’s studies in order to train and assess a multi-layer perceptron across the verbal and visual representations, either alone or in combination.

https://ai.meta.com/research/publications/hate-speech-in-pixels-detection-of-offensive-memes-towards-automatic-moderation