ProFine：基于弱监督学习的关系抽取渐进式微调框架
ProFine: A Progressive Fine-Tuning Framework for Relation Extraction under Weak Supervision
吴英晔 1张笑燕1
作者信息
- 1. 北京邮电大学计算机学院（国家示范性软件学院），北京 100876
- 折叠
摘要
实体关系抽取是医学知识图谱构建中的一个重要步骤，而当前LLM的大规模应用也在许多领域引起了关注，为了使LLM适用于特定领域的任务，通常需要对其进行微调。然而，在对关系抽取任务进行LLM微调时事先需要领域专家标注的高质量数据，存在成本高，效率低的问题。为了解决这一问题，本文提出了一个基于弱监督学习的关系抽取渐进式微调框架ProFine，该框架结合了LLM本身的推理能力和用于参数高效微调的低秩自适应（LoRA）的优点。为了缓解LLM生成训练数据可能会引入噪声的问题，本文还设计了用于衡量标签质量的指标，通过剔除质量较差的标签，来提升训练数据的质量。在医学文本数据集上的实验表明，ProFine要优于现有的参数高效微调方法。
Abstract
Entity relation extraction is an important step in the construction of medical knowledge graph, and the current large-scale application of LLM has also attracted attention in many fields. In order to make LLM suitable for domain-specific tasks, it usually needs to be fine-tuned. However, LLM fine-tuning for relation extraction tasks requires high-quality data annotated by domain experts in advance, which has the problems of high cost and low efficiency. To address this issue, we propose a Progressive Fine-Tuning Framework for Relation Extraction under Weak Supervision (ProFine), which combines the advantages of LLM\'s own inference capabilities and Low-Rank Adaptation (LoRA) for efficient parameter fine-tuning. In order to alleviate the problem that the training data generated by LLM may introduce noise, we design an indicator to measure the quality of labels, which can improve the quality of training data by eliminating poor quality labels. Experiments on medical text datasets show that ProFine outperforms existing parameter-efficient fine-tuning methods.关键词
知识图谱/实体关系抽取/大语言模型Key words
Knowledge Graph/Entity Relation Extraction/Large Language Models/引用本文复制引用
吴英晔,张笑燕.ProFine：基于弱监督学习的关系抽取渐进式微调框架[EB/OL].(2026-03-02)[2026-03-04].http://www.paper.edu.cn/releasepaper/content/202603-17.学科分类
医学研究方法
评论