基于心力衰竭领域数据增强的问答模型优化与应用
施雪斐 , 郭奇 , 马琳 , 秦志英 , 石兆峰 , 王肖龙
中国医学教育技术 ›› 2026, Vol. 40 ›› Issue (1) : 72 -79.
基于心力衰竭领域数据增强的问答模型优化与应用
Optimization and application of question-answering model based on heart failure data enhancement
心力衰竭是心血管领域的一种复杂疾病,对精准诊断、治疗和管理有着很高的要求。针对心力衰竭领域高精度可溯源的问答场景,本文提出了一种心力衰竭领域数据增强的问答模型优化方法。首先收集并整理大量心力衰竭相关的内容,并构建包含170余万个令牌的心力衰竭领域语料库,用于BGE-M3模型的增量预训练;其次构建超过3 200个心力衰竭专业问答的数据集,对预训练后的模型进一步进行细粒度的微调;最后将优化后的模型应用于检索增强生成(retrieval-augmented generation,RAG)中,实现了最终的问答系统。通过实验对比,较BGE-M3模型,微调后的模型与增量预训练并微调后的模型准确度分别提升了48%和52%,且在回答的精准性和内容全面性上均优于DeepSeek和通用RAG,验证了基于领域数据驱动的模型优化的有效性。本文方法证明了针对心力衰竭领域的智能化知识服务方案是实际可行的,尤其在医学教育场景中能显著提升教学效果,对于其他垂直领域的建模工作同样具有重要的参考价值。
Heart failure (HF) is a complex disease in the cardiovascular field, imposing high requirements for precise diagnosis, treatment, and management. Aiming at high-precision and traceable question-answering scenarios in the heart failure domain, this paper proposes a method for optimizing question-answering models with HF domain data enhancement. First, a large amount of heart failure-related content was collected and sorted to construct an HF domain corpus containing over 1.7 million tokens, which was used for incremental pre-training of the BGE-M3 model. Second, a dataset of over 3,200 professional HF Q&A pairs was built to further fine-tune the pre-trained model at a fine-grained level. Finally, the optimized model was applied to Retrieval-Augmented Generation (RAG) to implement the final question-answering system. Experimental comparisons show that compared with the general RAG, the accuracy of the fine-tuned model and the model after incremental pre-training and fine-tuning increased by 48% and 52% respectively. Moreover, it outperforms DeepSeek in both answer precision and comprehensiveness, verifying the effectiveness of domain data-driven model optimization. The method in this paper proves that the intelligent knowledge service scheme for the HF domain is practical, especially in improving teaching effectiveness in medical education, and provides important reference value for modeling work in other vertical domains.
| [1] |
|
| [2] |
|
| [3] |
黄子扬, 董超, 姜会珍, |
| [4] |
蔡文涛, 丁增辉, 李梦, |
| [5] |
刘泓泽, 王耀国, 唐圣晟, |
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
陈冬雷, 秦薇. 基于检索增强生成的智能问答系统关键技术研究[J]. 办公自动化, 2024, 29(19): 82-86. |
| [11] |
丁宁, 宋雨欣, 单泽田, |
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
上海市科委医学创新研究专项“基于‘心肾同治’理论指导的中医药干预治疗射血分数保留型心力衰竭的随机对照、双盲、安慰剂、多中心临床研究”(23Y31920200)
国家自然科学基金青年科学基金项目“基于深度学习与多模态构建中西医射血分数保留型心衰复发风险预测模型研究”(82405185)
/
| 〈 |
|
〉 |