Named Entity Recognition for Safety Risks in Electrical Work

The texts of safety risks in eletrical work describe the safety risks involved and management measures. The text records information on potential risks, risk levels, and preventive measures within the electrical work, which is crucial for extracting and mining safety risks. Named Entity Recognition for Safety Risks in Electrical Work (EWSR-NER) is a key technique for information extraction and knowledge discovery. It identifies objects related to safety risks in electrical work, such as risk level, risk factors, risk categories and protective measures, which convey critical meanings within the text. Current Named Entity Recognition (NER) is used to identify generic entities. However, their accuracy is limited by inherent linguistic irregularities in natural language texts, such as nested entities, lengthy entities, and the scarcity of domain-specific annotated corpora. To address this issue, the established standards and principles of safety risk texts in electrical work were followed to analyze the structure and characteristics of the text, developing a dataset of safety risk in electrical work. A new model, RoBERTa-BiLSTM-Efficient GlobalPointer (RoBEGP), was proposed in this paper. It uses the Robertly Optimized Bidirectional Encoder Representation from Transformers Pretraining Approach(RoBERTa) pretraining model to learn the characteristics of electrical work content and generate representation vectors. These representation vectors are input into a dynamic fusion layer, which adjusts the weights of each layer's representation vectors based on the specific requirements of the task, reducing the dimensionality from 1024 to 512. The fused vectors are input into a Bidirectional Long Short-Term Memory(BiLSTM) network to generate feature representations. Efficient GlobalPointer decodes these representations to identify nested entities within the text. Experimental results show that the proposed model achieves an F1 score of 92.38% on a self-constructed dataset of safety risk in electrical work, identifying 11 types of entities, totaling 2408 entity nodes.