OBP-LLM: Optimizing Boundary Perception of Large Language Model for Few-shot NER

ACL ARR 2025 May Submission6324 Authors

Abstract: Few-shot Named Entity Recognition (NER) enables models to learn effectively from limited annotated samples and perform robustly, even in resource-rich domains, addressing the challenge of scarce labeled data in many fields. Recently, Large Language Models (LLMs) have demonstrated strong adaptability and generalization capabilities in few-shot learning, offering new solutions for few-shot NER tasks. In this paper, we propose OBP-LLM, a novel large language model-based method that integrates contrastive learning and Direct Preference Optimization (DPO) to address attention mismatch and generation fallacy in LLM-based NER, by refining internal attention and generation preferences. Experimental results demonstrate that our method significantly outperforms existing approaches on multiple Few-shot NER benchmarks, including Few-NERD and CrossNER, particularly in cross-domain and extremely low-resource scenarios. This study validates the potential of contrastive learning and DPO in optimizing LLMs and provides new directions and practical solutions for NER tasks in low-resource domains.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: Information Extraction,Information Retrieval and Text Mining

Contribution Types: Approaches to low-resource settings, Publicly available software and/or pre-trained models

Languages Studied: Python

Submission Number: 6324