Named Entity Recognition in Track Circuits Based on Multi-Granularity Fusion and Multi-Scale Retention Mechanism

To enhance the efficiency of reusing massive unstructured operation and maintenance (O&M) data generated during routine railway maintenance inspections, this paper proposes a Named Entity Recognition (NER) method that integrates multi-granularity semantics and a Multi-Scale Retention (MSR) mechanism. The proposed approach effectively transforms expert knowledge extracted from manually processed fault data into structured triplet information, enabling the in-depth mining of track circuit O&M text data. Given the specific characteristics of railway domain texts, which include a high prevalence of technical terms, ambiguous entity boundaries, and complex semantics, we first construct a domain-specific lexicon stored in a Trie tree structure. A lexicon adapter is then introduced to incorporate these terms as external knowledge into the base encoding process of RoBERTa-wwm-ext, forming the lexicon-enhanced LE-RoBERTa-wwm model. Subsequently, a hidden feature extractor captures semantic representations from all 12 output layers of LE-RoBERTa-wwm, performing weighted fusion to fully leverage multi-granularity semantic information across encoding layers. Furthermore, in the downstream processing stage, two computational paradigms are designed based on the MSR mechanism and the Regularized Dropout (R-Drop) mechanism, enabling low-cost inference and efficient parallel training. Comparative experiments conducted on the public Resume and Weibo datasets demonstrate that the model achieves F1 scores of 96.75% and 72.06%, respectively. Additional experiments on a track circuit dataset further validate the model’s superior recognition performance and generalization capability.