A semantically enhanced two-stage framework for few-shot named entity recognition

Named entity recognition (NER) underpins information extraction, yet its performance often degrades in specialized domains where annotations are scarce and entity inventories change frequently. Few-shot NER targets such settings, but existing two-stage pipelines that separate span detection from type classification remain vulnerable: domain shift blurs span boundaries, and type representations estimated from tiny support sets are unstable. To address both issues, we propose a semantically enhanced two-stage framework. We first train a boundary-aware span detector with a contrastive objective on a high-resource source domain; then, for each target-domain episode, we construct label-guided hybrid prototypes that fuse label-text semantics with support-set mentions. Results on cross-domain benchmarks, Few-NERD, and a real-world power equipment defect dataset show that our framework consistently outperforms strong two-stage baselines, yielding about 1–3 F1 gains in the most challenging low-resource scenarios.