Graph-Based Methods to Detect Hate Speech Diffusion on Twitter

In this paper, we investigate models to detect the spread of hate speech on Twitter based on its diffusion in the network graph. We experiment with a dataset of 10,000 tweets manually labelled as hate speech or not and show that classification based solely on the sharing graph yields strong F1 scores for our task and high hate speech detection precision. We also highlight the vulnerability of existing textual hate speech detection methods to adversarial attacks and demonstrate that while our methods do not outperform state-of-the-art text models, graph-based models provide robust detection mechanisms and are able to detect instances of hate speech that fool text classifiers. We find that graph convolutional networks produce the strongest hate speech F1 score of 0.58 and find other success with kernel methods. Finally, we also consider the effects of automated bots in the sharing of hate speech content and find they are insignificant in our experiments.