LSP: Empowering Few-Shot NER with Demonstration Augmentation via Label Subset Partition

ACL ARR 2025 May Submission2968 Authors

Abstract: Leveraging the strong generalization capabilities of Large Language Models (LLMs) for data augmentation is an effective means to address the data sparsity of few-shot named entity recognition (FS-NER). Typically, existing methods manage to select appropriate demonstrations from a large amount of labeled data to be filled into the context of LLMs, thereby significantly enhancing the ability for in-context learning (ICL) in FS-NER. However, on the one hand, we have not yet figured out how demonstrations affect ICL in FS-NER so that we cannot do targeted optimization. On the other hand, labeled data is not abundant to select demonstrations from in real low-resource scenarios. In this study, we first systematically explore the impact of demonstrations on the ICL for FS-NER from 5 perspectives: sentence inclusion, number of demonstrations, label accuracy, label diversity, and label coverage. We find that label diversity and label coverage are important factors for ICL in FS-NER. So, we propose three metrics to quantify them: Label Space Per Instance (LSPI), Label Coverage (LC), and Label Measure(LM). Second, focusing on improving LSPI, LC, and LM, we devise a method named label subset partition (LSP) to augment demonstrations. It's an out-of-the-box augmentation method which is training-free, prompt-agnostic, and model-agnostic. Experiments on extensive NER datasets have demonstrated that LSP can effectively improve the performance of ICL for FS-NER.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: Efficient/Low-Resource Methods for NLP, Generation, Information Extraction

Contribution Types: NLP engineering experiment, Approaches to low-resource settings

Languages Studied: English, Chinese

Submission Number: 2968