SafeSpeak: A Real-Time Message Filtering System for Safe Guarding Against Toxicity using BERT

A huge number of user-generated messages are created each day on social media platforms. Messages may contain abusive or toxic messages which could cause cyberbullying, mental harassment, and emotional disturbance. The keyword-based approach fails to understand the context and is incapable of handling different languages such as english, hindi, marathi. This paper showcases SafeSpeak, which is a real-time message filtering system based on a Bidirectional Encoder Representations from Transformers (BERT) based transformer approach which would prevent toxic messages from being posted. SafeSpeak handles multiple languages as well as mixed languages such as english, hindi, and marathi, and tests messages at the time of submission to prevent the posting of abusive messages. We are using BERT to create our system. Using BERT gives us better results because the system is able to comprehend the context as well as be more accurate. The system have a good accuracy rate as well as an adequate response time which will make it effective for real-life application.