Reproduce “Show Less, Instruct More: Enriching Prompts with Definitions and Guidelines for Zero-Shot NER” on OpenResearch

About this paper

Andrew Zamai, Andrea Zugarini, Leonardo Rigutini, Marco Ernandes, Marco Maggini

An instruction-tuned Large Language Model (LLM) named SLIMER enhances zero-shot Named Entity Recognition (NER), particularly for genuinely novel entity types, by employing richly defined prompts with explicit annotation guidelines and a significantly reduced training data footprint. This methodology achieves superior performance on datasets featuring completely unseen entity classes compared to existing state-of-the-art models.

Key insights

• Integrating explicit Definitions and Guidelines into prompts is critical for improving true zero-shot NER performance and model stability, especially when identifying novel entity types.
• High-quality, detailed instructions are more impactful than large volumes of training data or numerous examples per class for effective instruction-tuning, leading to greater efficiency and generalization.
• The proposed instruction-tuning method is less dependent on specific LLM backbones, making it more adaptable and broadly applicable compared to some prior approaches that benefit significantly from specialized models (e.g., CodeLLaMA).

Summary by alphaXiv.