SAM-NER: Semantic Archetype Mediation for Zero-Shot Named Entity Recognition

2026-05-05

AI Summary

Zero-shot named entity recognition suffers from significant performance degradation under domain shifts and label schema variations due to semantic misalignment. To address this, this work proposes SAM-NER, a three-stage framework that leverages a domain-invariant semantic prototype space as an intermediary for stable transfer. The approach first jointly extracts high-fidelity entity spans and applies consensus-based denoising, then maps these spans into a universal semantic prototype space, and finally calibrates inference using a frozen large language model guided by definition-alignment constraints. Evaluated on the CrossNER benchmark, SAM-NER substantially outperforms existing zero-shot NER methods, demonstrating superior generalization and robustness, particularly in cross-domain settings.

Abstract

Zero-shot Named Entity Recognition (ZS-NER) remains brittle under domain and schema shifts, where unseen label definitions often misalign with a large language model's (LLM's) intrinsic semantic organization. As a result, directly mapping entity mentions to fine-grained target labels can induce systematic semantic drift, especially when target schemas are novel or semantically overlapping. We propose \textbf{SAM-NER}, a three-stage framework based on \emph{Semantic Archetype Mediation} that stabilizes cross-domain transfer through an intermediate, domain-invariant archetype space. SAM-NER: (i) performs \emph{Entity Discovery} via cooperative extraction and consensus-based denoising to obtain high-coverage, high-fidelity entity spans; (ii) conducts \emph{Abstract Mediation} by projecting entities into a compact set of universal semantic archetypes distilled from high-level ontological abstractions; and (iii) applies \emph{Semantic Calibration} to resolve archetype-level predictions into target-domain types through constrained, definition-aligned inference with a frozen LLM. Experiments on the CrossNER benchmark show that SAM-NER consistently outperforms strong prior ZS-NER baselines in cross-domain settings. Our implementation will be open-sourced at https://github.com/DMIRLAB-Group/SAM-NER.

Problem

• Zero-shot Named Entity Recognition
• domain shift
• schema shift
• semantic drift
• label misalignment

Innovation

• Zero-shot Named Entity Recognition
• Semantic Archetype Mediation
• Cross-domain Transfer
• Frozen LLM
• Ontological Abstraction