| 吴诺曼,赖伟,胡琳.基于“人在回路”的生成式人工智能在图书编目中的应用——以Kimi和DeepSeek技术适配性对比为例[J].数字图书馆论坛,2025,21(10):23~29 |
| 基于“人在回路”的生成式人工智能在图书编目中的应用——以Kimi和DeepSeek技术适配性对比为例 |
| Application of Generative Artificial Intelligence Based on “Human-in-the-Loop” in Bibliographic Cataloging: Comparative Study on Technical Adaptability of Kimi and DeepSeek |
| 投稿时间：2025-09-26 |
| DOI：10.3772/j.issn.1673-2286.2025.10.003 |
| 中文关键词: 生成式人工智能；人在回路；智能化编目；Kimi；DeepSeek |
| 英文关键词: Generative Artificial Intelligence; Human-in-the-Loop; Intelligent Cataloging; Kimi; DeepSeek |
| 基金项目: |
| 作者 | 单位 | | 吴诺曼 | 四川大学图书馆 | | 赖伟 | 四川大学图书馆 | | 胡琳 | 四川大学图书馆 |
|
| 摘要点击次数: 5 |
| 全文下载次数: 7 |
| 中文摘要: |
| 生成式人工智能为图书馆编目业务提供了新的技术路径，但其专业应用面临技术适配性挑战。本研究基于“人在回路”（HITL）机制设计了“基线测试-模型对比-定向优化”三阶段实验框架，系统评估了Kimi与DeepSeek两款国产大模型在西文图书MARC21编目中的表现。实验结果表明：在无专业干预条件下，两款模型原生能力存在显著差异（Kimi F1=7.41%，DeepSeek F1=51.30%）；经统一提示词引导后，DeepSeek（F1=83.00%）综合表现优异，显著优于Kimi（F1=63.50%）；进一步实施精细化提示工程后，DeepSeek（F1=95.16%）性能实现跃升。通过人类编目员的动态校验与反馈，生成式人工智能可突破初始技术限制，实现从通用对话到专业编目任务的适配。研究验证了“人在回路”机制在模型选型与性能优化中的有效性，提出了建立提示词知识库、实施字段分级管控等实践建议，为图书馆构建人机协同的智能编目体系提供了参考方案。 |
| 英文摘要: |
| Generative artificial intelligence offers new technological pathways for library cataloging, yet its professional application faces technical adaptability challenges. Based on the “human-in-the-loop” (HITL) mechanism, this study designs a three-phase experimental framework—baseline testing, model comparison, and targeted optimization—to systematically evaluate the performance of two domestic large language models, Kimi and DeepSeek,in MARC21 cataloging of Western-language books. Experimental results indicate that without professional intervention, the native capabilities of the two models differ significantly (Kimi F1=7.41%, DeepSeek F1=51.30%). Under unified prompt guidance, DeepSeek demonstrates superior overall performance (F1=83.00%), significantly outperforming Kimi (F1=63.50%). After implementing refined prompt engineering, DeepSeek’s performance achieves a notable improvement (F1=95.16%). Through dynamic calibration and feedback from human catalogers, generative artificial intelligence can overcome in tial technical limitations and adapt from general dialogue to professional cataloging tasks. This study validates the effectiveness of the HITL mechanism in model selection and performance optimization, and proposes practical recommendations, such as establishing a prompt knowledge base and implementing hierarchical field management, providing a reference solution for building a human-machine collaborative intelligent cataloging system in libraries. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
