体育健身知识智能问答系统构建

宋薇; 李丹阳; 蔡宾瑶; 毛宇辉

doi:10.14036/j.cnki.cn11-4513.2026.03.010

PDF(2299 KB)

首都体育学院学报 ›› 2026, Vol. 38 ›› Issue (3) : 339-352. DOI: 10.14036/j.cnki.cn11-4513.2026.03.010

体育治理与体育产业

体育健身知识智能问答系统构建

作者信息 +

Construction of an Intelligent Question Answering System for Sports and Fitness Knowledge

Author information +

文章历史 +

摘要

针对通用大语言模型在体育健身知识生成过程中存在领域适配精度不足、规则响应时效性偏弱等问题，构建了一种面向体育健身领域的知识精准响应与智能生成机制。通过构建体育健身领域专属知识库(Fitness-KB)，运用Chroma向量数据库与BGE嵌入模型，结合LangChain框架与量化因果关系的语言模型（DeepSeek-R1-Distill），实现毫秒级精准语义检索与高效推理。为提升生成内容的专业性与用户匹配度，通过人类反馈强化学习（RLHF）与近端策略优化（PPO）的双策略协同，使损失函数在知识准确性与用户偏好之间实现动态平衡。实验结果表明，检索准确率达到91.3%，生成准确率达到86.7%，偏好得分提高至4.23，BLEU-4指标提高至78.6%，能生成包含个性化参数与风险提示的专业建议。为体育健身方面的垂直领域知识问答的量化模型应用提供实践参考，为体育科技创新与人工智能深度融合提供可复用的量化技术路径，促进体育健身知识智能问答系统生成内容的精准化。

Abstract

To address the problems of general large language models existing in the generation of sports and fitness knowledge, such as low domain adaptation accuracy and poor timeliness of rule-based responses, this paper constructs an accurate knowledge response and intelligent generation mechanism for the sports and fitness domain. A specialized knowledge base for sports and fitness is built. By leveraging the Chroma vector database and the BGE embedding model, and integrating the LangChain framework with aquantized casual language model (DeepSeek-R1-Distill), millisecond-level semantic retrieval and efficient inference are achieved. To enhance the professionalism of generated content and its alignment with user preferences, a dual-strategy optimization framework combining reinforcement learning from human feedback and proximal policy optimization (RLHF+PPO) is designed. A joint loss function is employed to dynamically balance knowledge accuracy and user preference alignment. Experimental results demonstrate that the proposed system achieves a retrieval accuracy of 91.3% and a generation accuracy of 86.7%, with the preference score increasing to 4.23 and the BLEU-4 metric improving to 78.6%. The system is capable of generating professional fitness recommendations that incorporate personalized parameters and risk warnings. The proposed approach provides a practical paradigm for deploying quantizedmodels in vertical sports and fitness domains and offers a reusable quantization-based technical pathway for the deep integration of sports science and technology and artificial intelligence, promoting more precise content generation by intelligent question-answering systems for sports and fitness

导出引用

宋薇 , 李丹阳 , 蔡宾瑶 , 等. 体育健身知识智能问答系统构建[J]. 首都体育学院学报. 2026, 38(3): 339-352 https://doi.org/10.14036/j.cnki.cn11-4513.2026.03.010

SONG Wei , LI Danyang , CAI Binyao , et al. Construction of an Intelligent Question Answering System for Sports and Fitness Knowledge[J]. Journal of Capital University of Physical Education and Sports. 2026, 38(3): 339-352 https://doi.org/10.14036/j.cnki.cn11-4513.2026.03.010

中图分类号： G806 (体育锻炼) TP3 (计算技术、计算机技术)

参考文献

列表( 原文顺序 | 文献年度倒序 | 文中引用次数倒序 ) 可视化分析

[1]	张新苗, 陈日生. 基于体育强国背景下科学健身知识与方法的普及与推广研究[C]// 中国体育科学学会体育社会科学分会. 2023年体育社会科学分会年会论文集. 北京: 首都体育学院, 2023:865-866. 本文引用 [1]

[2]	邵梦霓, 张阳, 游永豪, 等. 健康中国背景下安徽省老年人群科学健身知识普及策略研究[J]. 安徽师范大学学报(自然科学版), 2024, 47(3):280. 本文引用 [1]

[3]	谢一博, 梁智杰, 张正萌. 人工智能技术在科学健身知识与方法普及与推广中的应用研究——以ChatGPT为例[C]// 中国体育科学学会体育社会科学分会. 2023年体育社会科学分会年会论文集. 郑州: 河南大学武术学院, 2023:475-478. 本文引用 [1]

[4]	Merilehto J. From PDFs to structured data: Utilizing LLM analysis in sports database management[Z/OL].(2024-10-23)[2026-04-18]. https://arxiv.org/abs/2410.17619. https://arxiv.org/abs/2410.17619 本文引用 [1]

[5]	张旭峰. 大语言模型在体育教育及社会体育中的创新应用探索[C] // 国际班迪联合会(FIB),国际体能协会(ISCA),中国班迪协会(CBF). 第五届国际体育科学大会论文集(下). 北京: 国家开放大学实验学院, 2024:218-221. 本文引用 [2]

[6]	徐凯, 赵亚楠, 戴剑松, 等. 大语言模型辅助制定运动处方的可行性研究[C] // 中国体育科学学会体质与健康分会. 2024年全国运动增强体质与健康学术会议论文摘要集. 南京: 南京体育学院运动健康学院,南京师范大学体育学院, 国家体育总局体育科学研究所,2024:107-108. 本文引用 [1]

[7]	Qiu Y. The impact of LLM hallucinations on motor skill learning: A case study in badminton[J]. IEEE Access, 2024, 12:139669-139682. https://doi.org/10.1109/ACCESS.2024.3444783 https://ieeexplore.ieee.org/document/10638024/ 本文引用 [1]

[8]	Qi G, Nguyen C T. Mixed preference optimization: Reinforcement learning withdata selection and better reference model[Z/OL]. (2024-03-28)[2026-04-18]. https://arxiv.org/abs/2403.19443. https://arxiv.org/abs/2403.19443 本文引用 [1]

[9]	Xu R, Shi W, Zhuang Y, et al. Collab-RAG: Boosting retrieval-augmented generation for complex question answering via white-Box and black-box LLM collaboration[Z/OL]. (2025-04-07)[2026-04-18]. https://arxiv.org/abs/2504.04915. https://arxiv.org/abs/2504.04915 本文引用 [1]

[10]	刘洋. “大模型+RAG”技术在档案工作中的应用探析[J]. 中国档案, 2025(3):64. 本文引用 [1]

[11]	Lin J, Ma Z, Gomez R, et al. A review on interactive reinforcement learning from human social feedback[J]. IEEE Access, 2020, 8:120757-120765. https://doi.org/10.1109/Access.6287639 https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=6287639 本文引用 [1]

[12]

Pokhrel

, Bina

K C

, Shah

P B

. A Practical application of retrieval-augmented generation for website-based chatbots: Combining web scraping, vectorization, and semantic search[J]. Journal of Trends in Computer Science and Smart Technology, 2025, 6(4):424.

https://doi.org/10.36548/jtcsst.2024.4

https://irojournals.com/tcsst/issue/view/273

本文引用 [1]

[13]	Kurniawan D, Triloka J. Penerapan teknologi langchain dan LLM pada sistem question answering berbasis chatbot telegram: Literature review[G]// Institut Informatika dan Bisnis Darmajaya. Seminar Nasional Hasional Hasil Penelitian dan Pengabdian Masyarakat. [S.I.]:[s.n.] 2025(1):95-104. 本文引用 [1]

[14]	窦凤岐, 胡珊, 李佳隆, 等. 基于LangChain的RAG问答系统设计与实现——以C语言课程问答系统为例[J]. 信息与电脑(理论版), 2024, 36(6):101. 本文引用 [1]

[15]	李全兴, 吴国林. 论人工智能的理解——以DeepSeek-R1等新兴人工智能为例[J]. 华南理工大学学报(社会科学版), 2025, 27(2):9. 本文引用 [1]

[16]	Ren T, Li R, Zhao M M, et al. Separate source channel coding is still what you need: An LLM-based rethinking[J]. ZTE Communications, 2025, 23(1):30-44. 本文引用 [1]

[17]	Deepseek A, Guo D, Yang D, et al. DeepSeek-R1:Incentivizing reasoning capability in LLMs via reinforcement learning[Z/OL]. (2025-01-22)[2026-04-18]. https://arxiv.org/abs/2501.12948. https://arxiv.org/abs/2501.12948 本文引用 [1]

[18]

Pternea

, Singh

, Chakraborty

, et al. The RL/LLM taxonomy tree: Reviewing synergies between reinforcement learning and large language models[J]. Journal of Artificial Intelligence Research, 2024, 80:1525-1573.

https://doi.org/10.1613/jair.1.15960

https://www.jair.org/index.php/jair/article/view/15960

本文引用 [1] 摘要

In this work, we review research studies that combine Reinforcement Learning (RL) and Large Language Models (LLMs), two areas that owe their momentum to the development of Deep Neural Networks (DNNs). We propose a novel taxonomy of three main classes based on the way that the two model types interact with each other. The first class, RL4LLM, includes studies where RL is leveraged to improve the performance of LLMs on tasks related to Natural Language Processing (NLP). RL4LLM is divided into two sub-categories depending on whether RL is used to directly fine-tune an existing LLM or to improve the prompt of the LLM. In the second class, LLM4RL, an LLM assists the training of an RL model that performs a task that is not inherently related to natural language. We further break down LLM4RL based on the component of the RL training framework that the LLM assists or replaces, namely reward shaping, goal generation, and policy function. Finally, in the third class, RL+LLM, an LLM and an RL agent are embedded in a common planning framework without either of them contributing to training or fine-tuning of the other. We further branch this class to distinguish between studies with and without natural language feedback. We use this taxonomy to explore the motivations behind the synergy of LLMs and RL and explain the reasons for its success, while pinpointing potential shortcomings and areas where further research is needed, as well as alternative methodologies that serve the same goal.

[19]	Yu Q, Zhang Z, Zhu R, et al. DAPO: An open-source LLM reinforcement learning system at scale[Z/OL]. (2025-03-18)[2026-04-18]. https://arxiv.org/abs/2503.14476. https://arxiv.org/abs/2503.14476 本文引用 [1]

[20]	Xu B, Zubiaga A. Understanding the effects of RLHF on the quality and detectability of LLM-generated texts[EB/OL]. (2025-03-23)[2026-04-18]. https://arxiv.org/abs/2503.17965. https://arxiv.org/abs/2503.17965 本文引用 [1]

[21]	Wu Z Q, Hu Y S, Shi W J, et al. Fine-grained human feedback gives better rewards for language model training[J]. Advances in Neural Information Processing Systems, 2023, 36:59008-59022. 本文引用 [1]

[22]	Zheng R, Dou S, Gao S, et al. Secrets of RLHF in large language models part I: PPO[Z/OL]. (2023-07-11)[2026-04-18]. https://arxiv.org/abs/2307.04964. https://arxiv.org/abs/2307.04964 本文引用 [1]

[23]	CRUD-RAG: A comprehensive chinese benchmark for retrieval-Augmented generation of large language models[J]. ACM Transactions on Information Systems, 2025, 43(2):1-32. 本文引用 [1]