Research on News Text Classification Based on Graph Convolutional Neural Networks

This paper proposes a new approach for classifying news text. It utilizes Graph Convolutional Networks (GCNs) as the foundation for this method. The approach starts with building a document-word graph. This graph is crafted to capture the content of the text. It also captures the structural information inherent in the text. Then, it integrates the features of nodes along with their neighboring nodes to improve the representation of each node. Furthermore, contrastive learning is utilized to boost the GCN's ability to represent data, thereby improving the modeling of graph structures. The contrastive loss, derived from the contrastive learning process, is combined with the classification loss from the Softmax classifier. Together, these losses jointly optimize the overall learning objective, enhancing model performance through both representation learning and label prediction. The experimental results have been analyzed. They show that the proposed method surpasses traditional sequence-based deep learning models. Additionally, it outperforms other graph neural network models in news text classification tasks across several public datasets. This confirms the efficacy of the suggested approach in capturing and utilizing the intricate relationships present in textual data for classification purposes.