BiRDy: Bullying Role Detection in Multi-Party Chats

Description

Recent studies have highlighted that private instant messaging platforms and channels are major media of cyber aggression, especially among teens. Due to the private nature of the verbal exchanges on these media, few studies have addressed the task of hate speech detection in this context. Moreover, the recent release of resources mimicking online aggression situations that may occur among teens on private instant messaging platforms is encouraging the development of solutions aiming at dealing with diversity in digital harassment. In this study, we present BiRDy: a fully Web-based platform performing participant role detection in multi-party chats. Leveraging the pre-trained language model mBERT (multilingual BERT), we release fine-tuned models relying on various contextual window strategies to classify exchanged messages according to the role of involvement in cyberbullying of the authors. Integrating a role scoring function, the proposed pipeline predicts a unique role for each chat participant. In addition, detailed confidence scoring are displayed. Currently, BiRDy publicly releases models for French and Italian.

Abstract

International audience

Additional details