RT-GPNet: a cascaded contextual encoding and rotary position-aware framework for named entity recognition in Chinese emergency texts

Chinese emergency event domain Named Entity Recognition faces key challenges: ambiguous long-entity boundaries, complex nesting, and poor deep contextual modeling. To address these, we propose RT-GPNet, an end-to-end span-based framework. Its core includes a Dual-Stage Contextual Encoder (RoBERTa cascaded with Multi-Layer Custom Transformer Blocks) for task-specific feature refinement, boosting long-range dependency modeling. The critical innovation is the Rotary Position-aware Span Localization Head, integrating Dynamic Frequency Rotary Position Embedding into GlobalPointer—infusing fine-grained relative position awareness to enhance boundary detection precision and enable nested entity extraction. Additionally, a boundary-sensitive learning objective ensures robust convergence with sparse domain data. Extensive experiments demonstrate the framework’s superior performance and robustness: RT-GPNet not only achieves a competitive 78.96% F1-score on the target Chinese emergency corpus, significantly outperforming mainstream baselines, but also demonstrates exceptional cross-domain and cross-lingual universality. Specifically, generalization tests on the ACE2005 English nested NER dataset (F1-score = 87.62%) and the CMeEE Chinese medical dataset (F1-score = 74.39%) verify that the proposed architecture is fundamentally domain-agnostic and language-independent.