| 刘俏亮,刘东亮,张洁.基于DeepSeek的参考文献AI编校研究——以《吉林大学学报(信息科学版)》为例.编辑学报,2025,37(4):431-436 |
| 基于DeepSeek的参考文献AI编校研究——以《吉林大学学报(信息科学版)》为例 |
| Research on AI-Based reference editing using DeepSeek:a case study of Journal of Jilin University(Information Science Edition) |
| |
| DOI：10.16811/j.cnki.1001-4314.2025.04.015 |
| 中文关键词: 参考文献 AI编校 DeepSeek 差错率 |
| 英文关键词: references AI editing and proofreading DeepSeek error rate |
| 基金项目: |
|
| 摘要点击次数: 0 |
| 全文下载次数: 0 |
| 中文摘要: |
| 为探究AI大模型在参考文献编校中的实际效果,明确其可用性与局限性,本研究以《吉林大学学报(信息科学版)》的参考文献为样本,采用DeepSeek大模型进行编校。依据国标GB/T7714—2015及本刊自身格式要求设计著录规则表和提示词,构建典型文献(10篇)和大样本文献(443篇)测试集,通过对比原稿文献、AI编校后文献与正确文献的差异,统计差错率等指标进行评估。结果显示DeepSeek在标点修正、格式规范等简单任务中表现尚可,但在广泛存在的内容性差错(如题名翻译错误、出版项错漏)修正上能力薄弱,且存在引入新差错的风险。虽然经过DeepSeek编校后整体差错率有所下降,但差错率的绝对数值仍数十倍于期刊出版质量控制阈值(0.000 2)。因此,AI大模型在帮助编辑完成基础性文献格式规范工作上具备一定可行性,但对于提升最终的参考文献出版质量,解决大量存在于已出版期刊中的文献内容差错问题,作用非常有限。 |
| 英文摘要: |
| To investigate the practical efficacy of large AI models in reference editing and proofreading tasks, and to systematically analyze their strengths and limitations, this study employs the reference lists from the Journal of Jilin University(Information Science Edition) as research samples, utilizing the DeepSeek large model for automated editing and proofreading. Based on the national standard GB/T 7714-2015 and the journal’s specific formatting requirements, we developed a structured cataloging rule matrix and tailored prompt templates, constructing a typical document set(10 papers)and a large-scale sample set(443 papers)for validation. The evaluation methodology involved comparative analysis of discrepancies among original documents, AI-edited versions, and manually corrected references, with quantitative assessment of error rates and other key metrics. Results indicate that DeepSeek achieves satisfactory performance in basic tasks such as punctuation normalization and format standardization, but exhibits significant limitations in addressing systemic content inaccuracies(e.g., title translation errors, omissions or inaccuracies in publication details) while carrying inherent risks of introducing new errors. Although the overall error rate decreased post-AI editing, the absolute error rate remained dozens of times higher than the journal’s stringent quality control threshold(≤0.0002). These findings suggest that while large AI models demonstrate certain utility as supplementary tools for foundational reference formatting tasks, their effectiveness remains highly constrained in addressing final publication quality requirements and resolving pervasive content errors in academic references. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
