Using transformer architecture and natural language processing for detecting offensive content and cyberbullying

The prevalence of offensive content and cyberbullying on the Internet has become an increasingly widespread issue. They can inflict emotional harm, instigate social isolation, and exacerbate mental health problems. Since content moderation is a labor-intensive task, machine learning might be helpful here. This research paper presents a comprehensive investigation of how Bidirectional Encoder Representations from Transformers (BERT) performs on the task of detecting offensive content and cyberbullying. The examination is done on how BERT suggests removing offensive content from a message while preserving the idea that a sender wants to express. Findings show that BERT requires fine-tuning to achieve high performance in detecting offensive content and cyberbullying. After fine-tuning, BERT gives useful suggestions on how to remove offensive content from messages while keeping the main idea of a person if offensive phrases are present in the context of a bigger main idea.