Research on Generalized Named Entity Recognition Approach for Military Texts

Named Entity Recognition (NER) in military texts is critical for intelligence analysis but faces challenges including scarce labeled samples, complex terminology, and emerging new entities. Traditional methods, constrained by fixed classification systems and manual annotation, struggle to adapt to these demands. We propose GLiNER-Military, a lightweight and generalizable NER model tailored for military domain texts to address these issues. Our key contributions include:(1) Comprehensive Military Entity Dataset: Through systematic analysis of six major entity categories in military texts, we constructed and annotated a dataset comprising 6,021 open-source military documents (totaling 22,846 entities), effectively mitigating data scarcity.(2) GLiNER-Military Model Design: Utilizing a bi-encoder architecture, the model unifies entity-type and text-span embeddings into a latent space for dynamic matching. Optimized for Chinese syntax, it significantly enhances the recognition of nested entities.Experimental Results: Achieved an F1-score of 83.87% on our custom dataset, surpassing baseline models. Demonstrated feasibility in zero-shot transfer learning tasks, confirming robust generalization to unseen entities.