The investigation of hidden dangers of coal mine is an important basis for the construction of the “Three in One” safety production standardization system. Most coal mines have established and used the safety production information system to carry out the investigation and management of hidden dangers, but the relevant data of hidden dangers of accidents has not been fully utilized. Taking the new edition of Coal Mine Safety Regulations as the classification standard, the classification system of 17 hidden danger large classes and 109 hidden danger small classes is constructed as the sample labels of the hidden danger data of coal mine, BiLSTM model combined with Attention mechanism is used to conduct text classification of coal mine accident hidden danger data in a two-layer classification system, and BERT model is used as the baseline for comparative study. In the hidden danger large class classification experiment, for the whole classification results, it shows that BiLSTM+Attention model has 2 percentage points higher in accuracy, precision,recall and F₁ value than BERT model in the hidden danger classification experiment. For the classification results of each hidden dangers, with F₁ value as the main measurement standard, the classification performance of BiLSTM+Attention model is up to 91%, which is generally 1% to 4% higher than BERT model. In the hidden danger small class classification experiment, the classification performance of BiLSTM+Attention model is up to 99%, which is also generally 1% to 10% higher than BERT model. It can be seen that the coal mine accident hidden danger classification algorithm based on BiLSTM+Attention model has a significant classification effect, which can provide a convenient application for fast entry for the information system of coal mine accident hidden danger investigation.