Introduction: The coal industry in China plays a vital role in the national economy, yet safety production remains a significant challenge. Although coal mine accidents have decreased in number, their complexity and the diversity of accident types have increased. The government has implemented regulations to improve safety, and leveraging big data and intelligent technologies for digital transformation is seen as a key path for enhancing safety production. This study introduces a novel deep learning model, Tri-Input-BERT-MLP, to enhance relation extraction in coal mine accident reports, which is also being developed into a patented system for improving safety monitoring. Methods: The paper proposes a hybrid architecture that combines Tri-Input-BERT with Multilayer Perceptron (MLP) to extract relations between entities in coal mine safety reports. The model integrates semantic features from BERT, entity pair representations, and positional encodings. The study also builds a specialized dataset from publicly available accident reports from China between 2010 and 2020, focusing on handling complex and unstructured data. Results: Experimental results show that the Tri-Input-BERT-MLP model significantly outperforms baseline models like BERT-BiLSTM, BERT-CNN, and others. It achieves the highest precision (91. 29%), recall (91. 32%), and F1 score (91. 30%) in relation extraction tasks. The model handles class imbalance by augmenting data and adjusting loss weights, improving performance, especially on minority relations like “R-belongₜo” and “R-locatedᵢn. ” Discussion: The Tri-Input-BERT-MLP model demonstrates substantial improvements in relation extraction for coal mine accident reports. It excels in handling industry-specific terminology and complex entity relationships. However, challenges remain in dealing with long-tail categories due to limited data. Future research should expand the dataset, apply data augmentation, and explore advanced techniques like Graph Neural Networks to further enhance model performance. Conclusion: This study presents the Tri-Input-BERT-MLP model, which effectively extracts relationships in coal mine accident reports, outperforms traditional methods, and offers significant improvements in precision, recall, and F1 score. The model's success provides valuable support for the development of intelligent supervision systems in coal mine safety, helping to reduce accidents and improve safety standards. Future research will focus on expanding the model's capabilities to handle long-tail relations and improving its generalization to diverse real-world scenarios.
Li et al. (Thu,) studied this question.