Deep Learning Multi-label Classification of News Text Based on BERT-GAC Model

With the continued development of internet media, the volume of various types of media information, particularly news texts, has grown rapidly. The traditional classification methods, which mix hot topics and content themes, can no longer meet the requirements for effective information storage, utilization, or user needs in the face of such vast amounts of news information from various domains. A well-structured news tagging system plays a crucial role in news classification and subsequent retrieval and utilization. This study fine-tunes the Graphormer model to model the structure and information of patent tags. The modeled tag information is then used to enhance the text representation of the BERT model, enabling automatic annotation of raw news content. By incorporating adversarial training, perturbations are added to the word embeddings, enhancing the model's robustness against adversarial samples and improving generalization performance. Finally, Softmax-CorNet is utilized to learn tag correlations, thereby improving prediction and classification performance. Experiments demonstrate that the improved BERT-GAC model achieves significant performance gains, with an accuracy of 95.3%, representing a 6.2% improvement over the baseline BERT model. Comparative experiments show that the improved model effectively enhances performance on multi-label text classification tasks.