Few-Shot Named Entity Recognition Approach Based on Large Language Models

Abstract

Named Entity Recognition (NER), as a pivotal task in Natural Language Processing (NLP), holds significant research value. However, current NER approaches typically rely on large amounts of annotated data and exhibit severe overfitting in low-resource scenarios. In 2022, ChatGPT based on GPT-3.5 was introduced, marking the rapid advancement of large language models. Their robust semantic understanding capabilities demonstrate strong potential for few-shot learning. Consequently, this paper investigates few-shot named entity recognition based on large language models. Given that large models are generative and prone to hallucinations, this study proposes a novel framework—CoT-LoRA-NER—that integrates Chain-of-Thought (CoT) reasoning guidance with Low-Rank Adaptation (LoRA) fine-tuning. This approach enables large models to perform named entity recognition in a reasoning-enhanced manner, implemented through three stages: CoT-guided data augmentation: Chain-of-Thought samples with explicit entity-level reasoning are automatically generated via the Deepseek API, with quality controlled through prompt template design, label schema specification, and limited random inspection, without per-instance manual correction; prompt construction: a structured question-reasoning-entity input format is designed to enhance contextual awareness of entity boundaries; parameter-efficient fine-tuning: LoRA is employed to optimize attention weights for the NER task, thereby reducing the risk of overfitting.

Ähnliche Arbeiten