Abstract: Construction of a hydraulic project often features a complex environment, difficult operation, and frequent occurrence of accidents. The accident report, as a document for accident analysis, usually includes a summary of the accident and its cause analysis and is used as a basis for accident prevention. However, most of the current analyses on hydraulic construction accidents rely on manual analysis by on-site managers, which is not only of more mistakes but time-consuming and labor-intensive; previous numerical models cannot directly analyze or predict the accident texts with high accuracy. This paper develops a hybrid deep learning model combining the bidirectional encoding representation of transformers (BERT) and the bidirectional long short-term memory model (BiLSTM) for in-depth analysis of the causes of hydraulic construction accidents. This hybrid model uses the BERT model in its upstream to generate a character-level dynamic feature vector of the accident text, and mines the semantic features of the accident report text in its downstream using an improved BiLSTM model, so as to realize an intelligent analysis of accident report texts. Its efficacy is compared with those of seven state-of-the-art deep learning models, and the results show it is superior and useful for analysis and decision-making of hydraulic construction accidents.

Key words: construction accident text, intelligent analysis, deep learning, BERT, BiLSTM, natural language processing