LLM-NER: Advancing Named Entity Recognition with LoRA+ Fine-Tuned Large Language Models
Yi Zhu, Yunan Liu
Abstract
Named Entity Recognition (NER) is a crucial task in natural language processing, but it faces persistent challenges in deep learning approaches, such as ambiguity, entity overlap, and domain variability. This study explores the potential of large language models (LLMs) to address these challenges by leveraging LoRA (Low-Rank Adaptation) and LoRA+ fine-tuning techniques. We present a specialized fine-tuned LLM based on Meta-Llama-3-8B-Instruct, optimized for NER tasks. Our model is evaluated against three benchmarks—BERT, RoBERTa, and DeBERTa—and demonstrates competitive performance using prompting alone. Moreover, the fine-tuned model with LoRA+ surpasses benchmark models by $10 \%$ in F1 score, showcasing its ability to distinguish nuanced entities effectively. These findings highlight the potential of LLMs in redefining the state of the art for NER tasks, enabling more robust and adaptable entity recognition solutions.