Aiming at the three problems of strong professionalism, semantic confusion and data imbalance in the text classification data of "three violation" in coal mine, the cBert-GCN model for the classification of "three violation" in coal mine is proposed. Considering the existence of professional field information, the text data of "three violation" in coal mine is short, and it has close context correlation and inherent ambiguity. Pinyin and glyph vectors are introduced to enhance the expression of “three violation” text data in coal mine. GCN is used in the text classification of "three violation" in coal mine, and a text co-occurrence graph is constructed to capture the structural information and dependency relationships in the text. The Chinese pre-training model and graph convolutional neural network are combined for feature learning, and character-level and word-level are fused, and the weights of the two are set to achieve accurate classification of the text data of "three violation" in coal mine. The results show that the accuracy of the cBert-GCN model on the training samples is higher than that of other models, reaching 97.03%, and on the test samples it reaches 93.17%, with good generalization ability. Therefore, the cBert-GCN model has obvious application advantages in the text data of "three violation" in coal mine.