Accidents Named Entity Recognition with Rules of Chinese Fine-Grained Complex Production Processes based on improved Convolution Attention Network

Accident named entity recognition (NER) in complex production process is an important prerequisite for chemical accident information extraction, and also an effective way for unstructured chemical safety data. An effective NER provides data-driven support for accident cause analysis, preventive action development, and emergency response. Existing methods for chemical accident NER primarily focus on coarse-grained entity. However, chemical accident entities are typically fine-grained. To address this, a Chinese fine-grained chemical accident entity corpus has been developed, incorporating 32 types of fine-grained labels. Consequently, a specialized fine-grained chemical accident entity recognition model is proposed using an improved rules based convolution attention network (RB-CAN) is proposed, which leverages the embedding layer to extract both word level information and character level information and adds chemical accident naming rules to correct the proposed model outputs for improving complicated chemical accident entities recognition. The method for fine-grained recognition of entities in chemical accidents is assessed using investigation reports of chemical incidents as the test dataset. Respectively, the experimental results show that the RB-CAN model achieves precision, recall, and F1-score values of 87.84%, 89.32%, and 90.13%. Meanwhile, the proposed method provides theoretical and practical support for intelligent perception and optimal control of complex production processes.