RBA: A Detection Method for Offensive Language in Chinese

As social media continues to gain popularity, offensive language on the Internet is on the rise, and how to effectively detect it has become a hot spot in the academic community. Due to the strong context-dependence of Chinese offensive text, the feature information is challenging to extract, and the training is susceptible to noise interference. Therefore, we propose a Chinese offensive language detection model integrating RoBERTa-base-cold, BiLSTM and adversarial training. The model generates dynamic word vectors containing offensive features by segmenting the text with RoBERTa-base-cold, then adds noise perturbations to these output vectors to generate adversarial samples and sends these adversarial samples to BiLSTM to enhance model robustness through adversarial training in the face of noise interference. Experiments are conducted on the COLDataset dataset, and the results show that this method is more effective than other modelling methods in detecting Chinese offensive speech.