Scholarlink

While Transformer-based Chinese Named Entity Recognition (NER) models have achieved remarkable performance on standard benchmarks, their robustness under real-world conditions remains underexplored. Existing research often assesses robustness using artificial perturbations, which may not accurately represent the natural variations encountered in practical applications. This study aims to bridge this gap by investigating the sentence-level non-adversarial robustness of Chinese NER models built on large language models (LLMs). In this work, we propose a robustness enhancement framework for LLMs, specifically designed to improve model stability in response to natural perturbations. Through comprehensive experiments conducted across three diverse Chinese NER datasets, we demonstrate that LLM-based NER models are vulnerable to sentence-level natural perturbations. Our method significantly mitigates these vulnerabilities, achieving an average robust F1 score improvement of up to 9.22%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

On Sentence-level Non-adversarial Robustness of Chinese Named Entity Recognition with Large Language Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chang Liu,
  • Libang Wang,
  • Peiyan Wang,
  • Feiyan Jiang

摘要

While Transformer-based Chinese Named Entity Recognition (NER) models have achieved remarkable performance on standard benchmarks, their robustness under real-world conditions remains underexplored. Existing research often assesses robustness using artificial perturbations, which may not accurately represent the natural variations encountered in practical applications. This study aims to bridge this gap by investigating the sentence-level non-adversarial robustness of Chinese NER models built on large language models (LLMs). In this work, we propose a robustness enhancement framework for LLMs, specifically designed to improve model stability in response to natural perturbations. Through comprehensive experiments conducted across three diverse Chinese NER datasets, we demonstrate that LLM-based NER models are vulnerable to sentence-level natural perturbations. Our method significantly mitigates these vulnerabilities, achieving an average robust F1 score improvement of up to 9.22%.