LLM-Augmented Prototype Representation for Few-Shot Named-Entity Recognition

Abstract

Named Entity Recognition (NER) models face challenges in adapting to data distribution shifts, especially with unseen entity types and limited data. Few-shot learning is used to address long-tailed distributions and unseen classes, but struggles with few labeled examples, leading to inaccurate prototypes. Recently, Large Language Models (LLMs) excel at Few-Shot Named Entity Recognition (FS-NER) via in-context learning, but their substantially higher computational cost and latency compared to prototype-based models limit their practicality for resource-constrained or real-time applications. In this paper, we propose a new way to use LLMs to augment the FS-NER system and a way for the prototype network to utilize the augmented dataset. We aim to examine whether LLM In-context learning (LLM-ICL) can effectively replace the encoder model in the FS-NER task. We study an LLM-based data augmentation framework that generates new mentions and descriptions for specific entity types when fine-tuning, while maintaining inference-time efficiency, improving the robustness of embeddings and prototype representations. Our method shows consistent improvements across FS-NER setups, demonstrating the potential of LLMs as external knowledge sources for NER.