Abstract:

To solve the problems existing in the traditional NER methods in the domain of tunnel construction safety, such as fuzzy entity boundary, difficulty in small-sample learning, and insufficiently comprehensive extraction of feature information, an entity recognition method for tunnel construction accident text based on the BERT-BiLSTM-CRF model was proposed. Firstly, the BERT model was used to encode the tunnel construction accident text to obtain word vectors containing semantic features. Then, the word vectors output after the training of the BERT model were input into the BiLSTM model to further obtain the context feature of the tunnel construction accident text and conduct label probability prediction. Finally, by utilizing the constraints of the annotation rules of the CRF layer, the output result of the BiLSTM model was corrected, and the maximum probability sequence annotation result was obtained, so as to realize the intelligent classification of the labels of the tunnel construction accident texts. Comparative experiments were conducted between this model and other four commonly used traditional NER models on the tunnel construction safety accident corpus dataset. The results show that the recognition accuracy rate, recall rate and F₁ value of the BERT-BiLSTM-CRF model are 88%, 89% and 88% respectively, and the entity recognition effect is better than other benchmark models. By using the established NER model to recognize the entities in the actual tunnel construction accident texts, its application effect in the domain of tunnel construction safety is verified.

Key words: bidirectional encoder representations from transformers (BERT), bidirectional long short-term memory(BiLSTM), conditional random fields (CRF), tunnel construction, safety field, named entity recognition (NER), deep learning