Research on news text classification based on improved BERT-UNet model

The classification of news text is crucial for various applications such as information retrieval, sentiment analysis, and intelligence gathering. In light of the limitations of conventional convolutional neural networks in news text classification, this study introduces an enhanced BERT-UNet model for improved long-distance text feature extraction. Initially, the model leverages BERT for pre-training the text word vectors, followed by embedding and mapping them onto the UNet architecture to extract contextual key features. The Softmax function is then utilized for news opinion text categorization. To validate the model's performance, comparative experiments are conducted on the THUCNews dataset. The results indicate that the BERT-UNET model outperforms the standard TextCNN model and standalone BERT approach with a 3.11% and 0.29% increase in macro average F1 value, respectively. These findings demonstrate the effectiveness of the enhanced BERT-UNet model in capturing textual relationships, offering a fresh perspective on enhancing traditional news text classification methods.