基于Lora微调的阿尔兹海默症知识图谱构建

Construction of Alzheimer’s Disease Knowledge Graph Based on LoRa Fine-Tuning

摘要: 阿尔茨海默症是一种不可逆的神经退行性疾病，利用知识图谱对阿尔茨海默症和轻度认知障碍患者进行准确的辅助诊断具有重要意义。然而传统知识图谱构建方法通常依赖大量人工标注，成本高且领域适应性有限。近年来，人工智能技术特别是大语言模型的快速发展，为此提供了新的技术支撑。本文提出一种基于大语言模型与Lora微调的阿尔茨海默症知识图谱构建方法，旨在为低资源、低成本场景下高效构建知识图谱提供参考。该方法通过设计合理的信息抽取提示模板并构建指令数据集，分别采用中英文语料的5个大语言模型进行少样本的Lora微调，对比分析实体关系联合抽取的不同表现。实验结果表明，Llama-3.1-Tulu-3-8B在实体关系联合抽取方面表现最优，在60个训练轮次下精确率达到82.5%，并进一步实现了从相关文献中自动抽取阿尔茨海默症知识，并完成知识图谱的构建与可视化分析。

Abstract: Alzheimer’s disease is an irreversible neurodegenerative disorder. Utilizing knowledge graphs for accurate auxiliary diagnosis of patients with Alzheimer’s disease and mild cognitive impairment holds significant importance. However, traditional knowledge graph construction methods often rely on extensive manual annotation, which is costly and has limited domain adaptability. In recent years, the rapid development of artificial intelligence technology, especially large language models, has provided new technical support for this purpose. This paper proposes a method for constructing an Alzheimer’s disease knowledge graph based on large language models and Lora fine-tuning, aiming to provide a reference for efficiently constructing knowledge graphs in low-resource and low-cost scenarios. This method involves designing reasonable information extraction prompt templates and constructing an instruction dataset. Five large language models in both Chinese and English corpora are employed for few-shot Lora fine-tuning, and the different performances of joint entity relation extraction are comparatively analyzed. Experimental results show that Llama-3.1-Tulu-3-8B performs optimally in joint entity relation extraction, achieving an accuracy rate of 82.5% after 60 training epochs. Furthermore, it automatically extracts Alzheimer’s disease knowledge from relevant literature, completes the construction of the knowledge graph, and performs visual analysis.