| 摘 要: 针对高速公路施工标准复杂、人工查阅效率低、专业术语易误解等问题,围绕基于大语言模型的高速公路施工标准知识问答系统展开研究。通过整合5册高速公路施工标准化指南,构建结构化数据集,并采用多级清洗、动态标注及质量验证流程提升数据一致性。以 Qwen2.5-3B为基座模型,系统对比 LoRA、QLoRA 等6种参数高效微调方法。实验结果表明:结合数据增强的 LoRA 策略综合性能最优,ROUGE-L指标提升至29.0%。领域模型在专业问答中展现高准确性,较通用模型更贴合工程规范,能精准推荐施工参数,有效降低施工风险。研究验证了轻量化微调框架在垂直领域的实用性,为施工智能化管理提供技术支撑。 |
| 关键词: 大语言模型 参数高效微调 施工标准 知识问答 |
|
中图分类号: TH16
文献标识码: A
|
|
| Research on Knowledge Q&A System for Highway Construction Standards Based on Large Language Models |
|
GAO Qian, ZHU Xiaohong
|
(School of Automotive and Transportation Engineering, Wuhan University of Science and Technology, Wuhan 430070, China)
2907976159@qq.com; 83223740@qq.com
|
| Abstract: Addressing the challenges of complex highway construction standards, low efficiency in manual retrieval, and frequent misinterpretation of specialized terminology, this study focuses on knowledge Q&A for highway construction standards based on large language models. By integrating five volumes of highway construction standardization guides, a structured dataset was constructed, and data consistency was enhanced through mult-i level cleaning, dynamic annotation, and quality validation processes. Using Qwen2.5-3B as the base model, six paramete-refficient fine-tuning methods, including LoRA ( Low-Rank Adaptation ) and QLoRA ( Quantized LoRA), were systematically compared. Experimental results show that the LoRA strategy combined with data augmentation achieves optimal comprehensive performance, with the ROUGE-L metric improving to 29. 0% . The domain-specific model demonstrates high accuracy in professional Q&A, aligning more closely with engineering specifications compared to genera-l purpose models, accurately recommending construction parameters and effectively reducing construction risks.This study validates the practicality of lightweight fine-tuning frameworks in vertical domains, providing technical support for intelligent construction management. |
| Keywords: large language models paramete-r efficient fine-tuning construction standard knowledge Q&A |
