A Method for Tibetan Offensive Language Detection Based on Prompt Learning and Information Theory

Offensive language on social media is a serious social problem, which affects people’s mental health and social harmony. However, there is a lack of effective detection methods for Tibetan offensive language, which makes the supervision and management of Tibetan social media platforms challenging. To effectively detect Tibetan offensive language, this paper proposes a method based on prompt learning and information theory, which has the following three characteristics: (1) It uses prompt learning technique to transform the text classification problem into a masked language modeling problem, and uses multi-label words based on external knowledge as the output of the verbalizer function, thus improving the model’s transferability and semantic expression; (2) It uses weighted cross-entropy loss and batch-averaged Kullback-Leibler divergence loss to balance the loss of different categories, and constrains the model output distribution under different dropout to be consistent, thus solving the overfitting problem and enhancing robustness; (3) It uses prior distribution to adjust the model output logits, making the model output closer to the true label distribution, increasing mutual information, compensating for the loss caused by class imbalance and magnitude mismatch, thus solving the class imbalance problem in the dataset. This paper uses current best mainstream model XLM-RoBerta (CINO) as the baseline, and conducts experiments on the newly collected dataset TOLD (Tibetan Offensive Language Dataset), which shows that this paper’s method can significantly improve the performance and generalization ability of Tibetan offensive language detection model, achieving 82.2% and 79.9% in accuracy and F1 score respectively, which are 3.5% and 2.5% higher than the baseline. This paper provides an effective offensive language detection tool for Tibetan social media platforms, which helps to maintain the order and security of cyberspace.