面向经营类铁路信息资源管理的自然语言交互关键技术研究

Research on Key Natural Language Interaction Technology for Management of Railway Information Resources

摘要

随着行业数字化转型的不断深入，经营类铁路信息资源服务模式正经历从传统人工沟通与静态报表向智能化、实时化的深刻变革。本文针对该领域存在的查询效率低、语义理解难、系统适应性差等问题，提出了一种基于自然语言交互的智能技术体系。该体系通过融合意图识别、语义检索（RAG）、双通道查询生成等核心技术，构建了一个能够对多源异构数据进行快速语义理解与高效查询执行的闭环系统。实验结果表明，本系统在真实业务数据集上的意图识别准确率达到96.2%，SQL生成执行成功率达94.5%，平均响应时间低于2秒，显著优于传统模板匹配方案。本研究不仅为铁路信息资源管理提供了可行的技术路径，其“模板优先，LLM回退”的双通道架构与持续优化机制，对同类企业级应用也具有重要的参考价值。

Abstract

As the digital transformation of the railway industry deepens, operational information resource services are undergoing a profound shift from traditional manual communication and static reporting to intelligent, real-time systems. This paper addresses challenges in the field such as low query efficiency, semantic comprehension difficulties, and poor system adaptability by proposing an intelligent technology framework based on natural language interaction. The system integrates core technologies including intent recognition, semantic retrieval (RAG), and dual-channel query generation to create a closed-loop system capable of rapid semantic understanding and efficient query execution for multi-source heterogeneous data. Experimental results demonstrate that the system achieves 96.2% intent recognition accuracy, 94.5% successful SQL generation execution rate, and an average response time under 2 seconds on real business datasets, significantly outperforming traditional template matching solutions. This research not only provides a viable technical pathway for railway information resource management but also offers valuable reference for enterprise-level applications through its “template-first, LLM fallback” dual-channel architecture and continuous optimization mechanism.

关键词

自然语言交互 / 经营类铁路信息资源 / RAG / 意图识别 / LLM / SQL生成

Key words

natural language interaction / railway operational information resources / RAG / intent recognition / LLM / SQL generation

参考文献

[1] Devlin J, Chang M W, Lee K, et al. BERT: Pre—training of deep bidirectional transformers for language understanding[J]. arXiv preprint arXiv:1810.04805, 2018.

[2] Lewis P, Perez E, Piktus A, et al. Retrieval—augmented generation for knowledge—intensive nlp tasks[J]. Advances in Neural Information Processing Systems, 2020, 33: 9459-9474.

[3] Touvron H, Martin L, Stone K, et al. Llama 2: Open foundation and fine—tuned chat models[J]. arXiv preprint arXiv:2307.09288, 2023.

[4] Hu E J, Shen Y, Wallis P, et al. Lora: Low—rank adaptation of large language models[J]. arXiv preprint arXiv:2106.09685, 2021.

[5] Schick T, Schütze H. Few—shot text generation with natural language instructions[J]. arXiv preprint arXiv:2012.15723, 2020.

[6] Thoppilan R, De Freitas D, Hall J, et al. Lamda: Language models for dialog applications[J]. arXiv preprint arXiv:2201.08239, 2022.

[7] Li Y, Li Z, Zhang K, et al. A survey of text sql parsing: Concepts, methods and future directions[J]. arXiv preprint arXiv:2305.15036, 2023.

基金资助

中国铁路信息科技集团有限公司科技研究开发计划课题《信息公司经营类铁路信息资源共享及分析关键技术研究》研究成果(WJZG-CKY-2024022（2024N10）)