人工智能生成中文学术论文文本检测研究  

Research on Detecting AI⁃generated Chinese Academic Text

作　　者：薛德军 孔祥煜 耿崇 吴晨 安彤 XUE Dejun;KONG Xiangyu;GENG Chong;WU Chen;AN Tong(Tongfang CNKI Digital Publishing Technology Co.,Ltd,Beijing 100192,P.R.China)

机构地区：[1]同方知网数字出版技术股份有限公司,北京100192

出　　处：《北京电子科技学院学报》2024年第3期104-112,共9页Journal of Beijing Electronic Science And Technology Institute

基　　金：国家重点研发计划“面向办案的检察机关法律监督知识融合与智能交互关键技术研究”(2020YFC0833003);国家卓越行动计划“科技期刊数字化运营国际平台服务项目”(WKZB1911BJM501173/02)。

摘　　要：以ChatGPT为代表的AIGC技术的普及应用,为AI自动生成论文和报告等文本内容提供了极大便利,出版和教育等管理部门迫切需要利用智能技术对提交的文本进行AIGC检测。本文对AI续写和AI扩写这两种常用人机协作方式生成中文学术论文文本检测方法进行研究和实验,设计了一种面向多学科学术论文文本检测数据的人机交互生成及检测方法,采用不同大模型生成的四个检测数据集,使用AIGC⁃CHECK检测工具进行检测,验证了在实际场景中的检测效果。检测结果显示,检测工具最优检测标注准确率达到98%。相较于内容全部由AI生成的中文学术论文检测,针对AI续写和AI改写这两种难度更大、检测粒度更细的人机协作生成文本检测依然可取得较高准确率,说明辅助开展AI生成学术论文检测具有可行性。Widespread application of the AIGC technology such as ChatGPT provides great convenience for AI automatically generating text content including papers and reports,urging the administrative de⁃partments in fields of publication and education to perform AIGC detection on submitted texts utilizing intelligent technology.In this paper,text detection methods for Chinese academic paper texts generated by the AI continuation and the AI expansion respectively are explored and experimented,which are two familiar human⁃computer collaboration methods.A human⁃computer interaction⁃based approach for gen⁃erating and detecting multidisciplinary academic paper text data is designed,where four detection data⁃sets generated by different large language models are adopted.Detection is performed by using the tool AIGC⁃CHECK to assess the effectiveness of the approach in actual scenarios.Detection results indicate that AI annotation accuracy using the tool AIGC⁃CHECK optimally reaches 98%.Compared with detec⁃ting text of Chinese academic papers completely generated by AI,our approach could still achieve high accuracy for the AI continuation and the AI expansion which demand greater difficulty and finer detec⁃tion granularity,demonstrating the feasibility of applying such methods to assist in detecting AI⁃genera⁃ted academic papers.

关 键 词：中文学术论文 AIGC检测 AI扩写检测 AI续写检测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]