中国铁道科学2025，Vol.46Issue(6)：44-54,11.DOI:10.3969/j.issn.1001-4632.2025.06.05
铁路路基压实施工知识图谱构建与检索增强方法
Knowledge Graph Establishment and Retrieval-Augmented Method for Railway Subgrade Compaction Construction
摘要
Abstract
To address the issues such as delays of knowledge timeliness,insufficient understanding of professional knowledge,and lack of output stability in the application of general models to railway subgrade compaction construction,a knowledge graph establishment and retrieval-augmented method specifically for the railway subgrade compaction construction domain was proposed.First,a specialized corpus encompassing standards and specifications,professional books,academic papers,and engineering documents was built.Then,a Bert-BiLSTM-CRF model was designed for entity extraction from this corpus.By integrating hierarchical clustering and DeepSeek-Prompt techniques,relation extraction and reasoning were accomplished,resulting in a knowledge graph comprising 1 315 entities and 1 003 relations within the railway subgrade compaction domain.Finally,this knowledge graph was integrated into the DeepSeek-R1 model through Retrieval-Augmented Generation(RAG)technology to enable retrieval-augmented in the field of railway subgrade compaction.The results show that the model integrated with the knowledge graph significantly outperforms traditional general models in key performance metrics,achieving ROUGE-1,ROUGE-L,and accuracy rate of 0.321 2,0.383 4,and 0.822 9,respectively.This approach not only successfully achieves systematic extraction and structured representation of knowledge,but also effectively improves the model's precision in understanding complex professional knowledge and the reliability of its generation.It opens up a new technical approach for enhancing the level of intelligence in engineering practice.关键词
铁路路基/知识图谱/检索增强/路基施工/DeepSeekKey words
Railway subgrade/Knowledge graph/Retrieval-augmented/Subgrade construction/DeepSeek分类
交通工程引用本文复制引用
史小萌,徐功博,蔡国庆,王子南,邓志云,刘保国..铁路路基压实施工知识图谱构建与检索增强方法[J].中国铁道科学,2025,46(6):44-54,11.基金项目
国家重点研发计划项目(2023YFB2603902) （2023YFB2603902）
国家自然科学基金青年项目(42307246) （42307246）
