A Method for Chinese Sarcasm Detection Based on Enhanced Cross-Entropy and Regularization

In the context of social media, the automatic identification of sarcastic discourse presents a complex challenge within the field of natural language processing. Due to the significant disparity between the literal and implied meanings of sarcastic statements, traditional text analysis algorithms struggle to accurately detect them. This paper introduces a method for detecting Chinese sarcastic comments on social media, which integrates an improved cross-entropy loss function with regularization techniques. The method employs a dynamic adjustment factor to optimize weighted cross-entropy loss function, calibrating the model's penalty intensity for high-confidence predictions, and enhances the consistency of the model's output through batch-averaged Kullback-Leibler divergence. Experiments were conducted using the Chinese-RoBERTa model as a baseline, validated on full data and small sample sizes (50%, 10%). The results demonstrate that, compared to the baseline model, the proposed method achieved accuracy improvements of 3.7%, 2.5%, and 2.3% across different data volumes, confirming its effectiveness and applicability in the detection of Chinese sarcastic discourse.