Detection and Recognition of Network Offensive Language Based on BERT Model

While enjoying the many conveniences brought by the internet, people are also deeply affected by the interference of harmful information, including offensive language. The detection of offensive language and the identification of attack targets are particularly important in the era of the Internet. This article proposes a multi task attack speech detection model based on BERT. The main idea is to introduce more useful information for detection tasks while effectively utilizing the text context. The BERT pre-trained model can provide contextual word representations for text and obtain additional language information learned from large-scale corpora. In response to the focus of recognition tasks, this paper proposes a multi task attack object recognition model that integrates attention. The main idea is to use attention mechanisms to focus on the parts of aggressive text that reflect attacks and the objects they modify. The experiment shows that this model has better detection and recognition performance compared to other common models.