BGNN-TC: BERT GNN Based Text Classification Algorithm

The traditional Text-Level-GNN effectively captures the structural information of texts but fails to fully extract the semantic information of words during text-level composition. To address this limitation, we employ BERT to initialize the word nodes within text-level graphs, constructing a graph network that integrates both spatial structure and semantic information. Based on BERT optimization, we proposes a BGNN-TC based text classification algorithm to enhance the performance of existing methods, particularly on long-text classification tasks. This approach leverages the strengths of large-scale pre-trained models and text-level graph neural networks. We perform text category prediction using both BERT+GNN and BERT models based on word features from text-level compositions and train the model with a mixed loss of the two prediction results. Experiments on the R8, R52, and Ohsumed datasets demonstrate that the proposed method achieves outstanding performance in handling long texts, effectively extracting both semantic and structural information. The results highlight the robustness and generalization capabilities of the model.