In the domain of natural language processing, BERT has emerged as a highly effective model due to its ability to capture rich contextual information from text, while graph neural networks (GNNs), such as GraphSAGE, excel at learning representations from graph-structured data by efficiently aggregating neighborhood information. In this article, we integrate the strengths of BERT and GraphSAGE to enhance document classification tasks in citation networks. Specifically, BERT is fine-tuned to encode textual information, representing documents as nodes with enriched node features, while GraphSAGE is employed to learn from the graph structure. Our approach combines these models by using BERT for feature extraction and GraphSAGE for graph-based learning. The model with the highest validation accuracy from BERT is used to encode the node features, which are then fed into the GraphSAGE model for further training and evaluation. Joint training of BERT and GraphSAGE allows the model to leverage the linguistic features captured by BERT and the structural information captured by GraphSAGE. In our experiments, in addition to testing on the Open Graph Benchmark dataset ogbn-arxiv, we also evaluated the model on four text classification datasets: 20 Newsgroups (20NG), R8, Ohsumed, and Movie Review (MR). We used accuracy as the primary evaluation metric, and the experimental results show that the BERT-GraphSAGE model achieved an average improvement of approximately 2.1% in accuracy compared to traditional models across the datasets. These findings further demonstrate the broad applicability and effectiveness of the hybrid model across various tasks and datasets.