Research on Entity Recognition for Coal Mine Safety Production

To address the challenges in annotation arising from highly specialized terminology and ambiguous entity boundaries in the field of coal mine safety production, we propose a hybrid entity recognition model. Based on Chinese RoBERTa, this model resolves the representation balance issue between the general features of the lower layers and the task-specific features of the upper layers in the pre-trained model by freezing the parameters of the first six layers and weighting the fusion of the hidden states from the last four layers. During the feature extraction phase, the model employs a hybrid structure combining parallel and serial processing: the context enhancement module and the attention-based feature pyramid network extract multi-scale local features in parallel. The former directly participates in subsequent fusion after optimizing RoBERTa features in situ; the latter’s multi-scale features are further processed through Transformer layers to model global dependencies; Subsequently, a two-layer gating module and a three-layer bidirectional LSTM achieve dynamic fusion of multi-source features, enhancing the accuracy of entity boundary recognition. The loss function is composed of three weighted components: entity-aware Focal Loss, enhanced boundary loss, and transfer loss, which are used to mitigate entity category imbalance, strengthen entity boundary recognition, and constrain the validity of B-I label sequences, respectively. In experiments on our self-built dataset, the model achieved a precision of 87.73%, a recall of 86.79%, and an F1 score of 87.26%. Compared to mainstream models, it demonstrates a superior balance between precision and recall, significantly alleviating the challenges of ambiguous entity boundaries and the difficulty in identifying low-resource entities in mine safety texts.