Toxicity in Texts and Images on the Internet

In this paper we studied the most typical characteristics of toxic images on the web. To get a set of toxic images we collected a set of 8800 images from 4chan.org. Then we trained a BERT-based classifier to find toxic texts with accompanying images. We manually labelled approximately 2000 images accompanying these texts. This revealed that toxic content in images does not correlate with toxic content in texts. On top of manually annotated images there was trained a neural network that inferred labels for unannotated pictures. Neural network layer activations for these images were clustered and manually classified to find the most typical ways of expressing aggression in images. We find that racial stereotypes are the main cause of toxicity in images (https://github.com/denis-gordeev/specom20).