Research and Implementation of Production Safety Accident Description Element Extraction

Accident description element extraction is an important basis for building accident knowledge base, and its purpose is to extract the key accident description features from the text. Previous element extraction usually uses multi-label classification methods for modeling, and the model mainly relies on text features, ignoring domain-specific semantic information. At the same time, due to the lack of large-scale annotated data corpus in production safety accident texts, it is difficult to solve this problem by traditional supervised learning methods. To address the above problems, we propose an accident description element extraction method based on a few-shot Named Entity Recognition (NER) framework, define seven description element categories reflecting basic accident information, and build a few-shot element extraction model based on transfer learning — RPBMC model, which introduces a RoBERTa-WWM pre-trained language model adapted to Chinese datasets; we further design a Partial Least Squares (PLS) domain word mapping layer based on feature transformation to enhance the word vector generalization ability. And we combine the Multi-Head Self-Attention (MHSA) mechanism to learn more fine-grained semantic information. The experimental results show that the model effectively completes the element extraction with an overall F1 value of 91.34%.