Named Entity Recognition in the Coal Mine Accident Domain Based on ALBERT-IDCDA-SCRF (ALBERT–Iterative Dilated Convolutional Dropout Adam–Semi-Conditional Random Fields)

Introduction: Named Entity Recognition (NER) plays a crucial role in the coal mine safety accident domain, being key to constructing a knowledge graph for coal mine safety production. Traditional methods face challenges such as long training cycles and insufficient recognition accuracy when processing coal mine accident cases. Methods: This study proposes an enhanced deep learning framework, ALBERT-IDCDA-SCRF, to address the issues of long training cycles and low recognition accuracy in traditional methods. The framework uses the ALBERT pre-trained language model to generate word embeddings, which are then processed by a convolutional neural network (CNN) incorporating the Iterative Dilated Convolution Dropout Adam (IDCDA) module. This framework integrates a semi-conditional random field (SCRF) model to ensure annotation consistency. The model was trained using a self-constructed coal mine corpus, a domain-specific annotated dataset for coal mine accidents. Additionally, the NER method presented in this study was applied to related patents, significantly improving model performance through technological innovations. Results: Experimental results show that the ALBERT-IDCDA-SCRF model has significantly improved precision (90.21%), recall (88.69%), and F-measure (88.95%), while greatly reducing training time. The model demonstrates high efficiency in coal mine safety applications and exhibits strong applicability in coal mine accident management tasks. Discussion: The ALBERT-IDCDA-SCRF model achieved significant improvements in NER performance for coal mine accidents, providing a solid technical foundation for accident analysis and knowledge graph construction. However, further optimization of the model is a key focus for future work, including expanding the coal mine corpus and incorporating both word-level and character- level features in the word embeddings. Conclusion: The ALBERT-IDCDA-SCRF model has achieved significant improvements in NER performance for coal mine accidents, providing effective technical support for accident analysis and safety production management. Future work will focus on expanding the corpus and further optimizing the model.