An Improved BERT-BiLSTM-CRF Model for Constructing Safety Field

To address the challenges of named entity recognition posed by the long text structure in construction safety accident investigation reports, an improved model combining convolutional neural networks and a smoothing mechanism is proposed to enhance the entity recognition performance in construction safety accident texts. Based on investigation reports from official government websites, a domain-specific dataset for construction safety was constructed after data preprocessing, with ten entity types closely related to accident risks defined. The proposed BERT-BiLSTM-CNN-CRF+ model enhances feature extraction capabilities, reduces over-reliance on specific boundaries and achieves in-depth analysis of construction safety texts. Additionally, introducing large models improves feature extraction efficiency and the use of pre-trained models for semantic encoding and context feature capture enables efficient automated annotation, reducing manual costs. Experimental results indicate that this model, based on BERT-BiLSTM-CRF, improves precision by 5.0%, recall by 1.0% and the F1 score by 4.0%. Although there is still room for improvement in recognizing entities related to direct and indirect causes, the improved model shows slight enhancements in their recognition, offering potential applications for accident causation analysis. Overall, this method provides an effective technical means for accident prevention in the field of construction safety and lays the foundation for constructing a knowledge graph of construction safety accidents.