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

Problem
Zero-shot Named Entity Recognition (ZS-NER) is prone to performance degradation when faced with domain and schema shifts, particularly when the target label definitions are unseen or semantically similar to existing labels. This paper addresses the gap in existing literature regarding the systematic semantic drift that occurs during the direct mapping of entity mentions to fine-grained target labels, which can lead to misalignment with a large language model’s (LLM’s) intrinsic semantic organization. The authors propose a novel framework, SAM-NER, to mitigate these issues. This work is presented as a preprint and has not yet undergone peer review.

Method
SAM-NER introduces a three-stage framework based on Semantic Archetype Mediation. The stages are as follows:

1. Entity Discovery: This stage employs cooperative extraction and consensus-based denoising techniques to identify high-coverage and high-fidelity entity spans from the input text.
2. Abstract Mediation: Entities are projected into a compact set of universal semantic archetypes, which are derived from high-level ontological abstractions. This step aims to create a domain-invariant representation of entities.
3. Semantic Calibration: The final stage resolves predictions at the archetype level into target-domain types using constrained, definition-aligned inference with a frozen LLM. This approach ensures that the predictions are aligned with the definitions of the target labels, reducing the risk of semantic drift.

Results
The authors evaluate SAM-NER on the CrossNER benchmark, demonstrating consistent performance improvements over strong prior ZS-NER baselines. While specific numerical results are not disclosed in the abstract, the paper claims that SAM-NER outperforms existing methods across various cross-domain scenarios, indicating a significant effect size in terms of accuracy and robustness against domain shifts.

Limitations
The authors acknowledge that while SAM-NER improves ZS-NER performance, it may still struggle with highly ambiguous entities or those that lack clear archetype representations. Additionally, the reliance on a frozen LLM may limit adaptability to rapidly evolving language or domain-specific terminologies. The paper does not address potential computational overhead introduced by the three-stage process, which could impact real-time applications.

Why it matters
The implications of SAM-NER extend to various applications in natural language processing, particularly in scenarios where labeled data is scarce or unavailable. By providing a robust framework for ZS-NER that stabilizes cross-domain transfer, this work paves the way for more reliable entity recognition systems in diverse and dynamic environments. The open-sourcing of the implementation further encourages community engagement and potential enhancements, fostering advancements in zero-shot learning methodologies.

Authors: Ruichu Cai, Juntao Gan, Miao Mai, Zhifeng Hao, Boyan Xu
Source: arXiv:2605.03706
URL: https://arxiv.org/abs/2605.03706v1