TensorGCN model is one of the state-of-the-art (SOTA) models applied by graph neural networks in the field of text classification. However, in terms of processing text semantic information, the long short-term memory (LSTM) used by the model has difficulty in completely extracting the semantic features of short text and performs poorly in handling complex semantic information. At the same time, due to the large number of semantic and syntactic features contained in long texts, feature sharing is incomplete when heterogeneous information is shared among graphs, which affects the accuracy of text classification. To solve these two problems, the TensorGCN model is improved, and a text classification method based on the tensor graph convolutional network fusing BERT and the self-attention mechanism (BTSGCN) is proposed. Firstly, BERT is used to replace the LSTM module in the TensorGCN architecture for semantic feature extraction. It captures the dependencies between words by considering the surrounding words on both sides of a given word, thus extracting the semantic features of short texts more accurately. Then, the self-attention mechanism is added during the propagation among graphs to help the model better capture the features among different graphs and complete the feature fusion. Experimental results on MR, R8, R52, and 20NG datasets show that BTSGCN has higher classification accuracy than other comparison methods.