Utilising Twitter Metadata for Hate Classification

Oliver Warke, Joemon M. Jose, Jan Breitsohl

Abstract: Social media has become an essential daily feature of people’s lives. Social media platforms provide individuals wishing to cause harm with an open, anonymous, and far-reaching channel. As a result, society is experiencing a crisis concerning hate and abuse on social media. This paper aims to provide a better method of identifying these instances of hate via a custom BERT classifier which leverages readily available metadata from Twitter alongside traditional text data. With Accuracy, F1, Recall and Precision scores of 0.85, 0.75, 0.76, and 0.74, the new model presents a competitive performance compared to similar state-of-the-art models. The increased performance of models within this domain can only benefit society as they provide more effective means to combat hate on social media.