“In this work we provide the first large scale analysis of hate-speech on Parler. We experiment with an array of algorithms for hate-speech detection, demonstrating limitations of transfer learning in that domain, given the illusive and ever changing nature of the ways hate-speech is delivered. In order to improve classification accuracy we annotated 10K Parler posts, which we use to fine-tune a BERT classifier. Classification of individual posts is then leveraged for the classification of millions of users via label propagation over the social network. Classifying users by their propensity to disseminate hate, we find that hate mongers make 16.1% of Parler active users, and that they have distinct characteristics comparing to other user groups.”


By author

Leave a Reply

Your email address will not be published. Required fields are marked *