PANER: A Paraphrase-Augmented Framework for Low-Resource Named Entity Recognition

TL;DR

PANER addresses low-resource NER by combining a paraphrase-augmented data strategy with an instruction-tuning template that uses a word/tag output format and negative instances. It leverages long-context LLMs with efficient LoRA fine-tuning to maintain compute efficiency while enabling broader domain coverage ($128K$ and $32K$ context windows). Empirical results on CrossNER, MIT, and BUSTER show competitive zero-shot performance with strong few-shot gains attributable to paraphrase augmentation, and substantial reductions in training resources compared to full fine-tuning. These findings suggest PANER as a practical approach for domain-specific NER under limited data and computational constraints, with potential extensions to multilingual settings and selective guideline application.

Abstract

Named Entity Recognition (NER) is a critical task that requires substantial annotated data, making it challenging in low-resource scenarios where label acquisition is expensive. While zero-shot and instruction-tuned approaches have made progress, they often fail to generalize to domain-specific entities and do not effectively utilize limited available data. We present a lightweight few-shot NER framework that addresses these challenges through two key innovations: (1) a new instruction tuning template with a simplified output format that combines principles from prior IT approaches to leverage the large context window of recent state-of-the-art LLMs; (2) introducing a strategic data augmentation technique that preserves entity information while paraphrasing the surrounding context, thereby expanding our training data without compromising semantic relationships. Experiments on benchmark datasets show that our method achieves performance comparable to state-of-the-art models on few-shot and zero-shot tasks, with our few-shot approach attaining an average F1 score of 80.1 on the CrossNER datasets. Models trained with our paraphrasing approach show consistent improvements in F1 scores of up to 17 points over baseline versions, offering a promising solution for groups with limited NER training data and compute power.

Figures (4)

• Figure 1: Illustration of paraphrasing-based data augmentation process.
• Figure 2: Prompt used for generating paraphrases.
• Figure 3: Prompt used for Instruction-tuning LLMs.
• Figure 4: Impact of augmented sample size on model performance (F1 score, in %) for CrossNER dataset b20.