A Comparative Study on the Performance of BERT Sentiment Classification Models Based on Data Augmentation

With the development of social media and e-commerce platforms, text sentiment analysis is of great value in public opinion monitoring, market analysis, and user experience research. Pre-trained language models such as BERT significantly improve sentiment classification performance, but are still prone to overfitting in small sample scenarios 1. In order to improve the robustness and generalization ability of the model, this paper compares the performance differences of three lightweight data enhancement methods, namely synonym substitution, random masking and backtranslation, in Chinese sentiment classification tasks 2,3. Based on the ChnSentiCorp dataset, the enhancement strategy is analyzed from three dimensions: semantic retention, computational cost and performance improvement. The results show that both synonym substitution and random masking can stably improve the model performance, while the effect of unfiltered backtranslation is significantly reduced due to semantic shift. After adding semantic similarity and sentiment consistency screening, the backtranslation performance is restored. This paper provides suggestions for the selection of enhancement strategies for emotion classification in small sample scenarios.