BERT-DXLMA: Enhanced representation learning and generalization model for english text classification
Xingliang Mao, Zhuhao Li, Qingxi Li, Shichao Zhang · Neurocomputing · 2025
In recent years, the rapid development of artificial intelligence technology has significantly improved the performance of BERT-based deep learning models in text classification tasks . However, these models still have shortcomings in capturing the deep semantics of text, emerging text classification , and handling minority class samples in imbalanced data . To address these challenges, we proposed an innovative BERT-DXLMA hybrid model. This model enhances the semantic feature extraction and representation learning capabilities by integrating xLSTM, and uses semantic fusion technology to reduce the information loss caused by the increase in the number of network layers. In addition, we re-improved the loss function based on focal loss to improve the model’s attention to minority class samples and generalization ability . We conducted extensive experiments and in-depth analysis on six public datasets, and the results showed that BERT-DXLMA outperformed multiple baseline methods in terms of precision and overall accuracy, demonstrating its superiority in text classification tasks .