基于深度学习的AI生成文本判别模型研究  

Study of text discriminative model of AI generation based on deep learning

作　　者：徐璐 唐大卫 XU Lu;David Tang(Jinling University of Science and Technology,Nanjing 211100,China;Jiangsu Sumeida Group Co.,LTD.,Nanjing 210018,China)

机构地区：[1]金陵科技学院,南京211100 [2]江苏苏美达集团有限公司,南京210018

出　　处：《高科技与产业化》2024年第8期43-45,共3页High-Technology & Commercialization

摘　　要：本文针对识别大型语言模型(LLMs)生成文本与人类创作文本的核心难题展开研究,通过在多样化的数据集上检验模型性能,验证升级后的鉴别策略的有效性。本研究重点评估GPT-3.5-Turbo模型,并将其性能与多种主流分类模型进行了对比。研究结果显示,模型鉴别准确率显著受文本序列长度的影响,揭示了长度作为影响鉴别效能关键因素的新视角。这些发现不仅加深了对AI生成文本特性的理解,也为开发更精准的鉴别算法提供了方向。This paper focuses on addressing the core challenge of distinguishing text generated by Large Language Models(LLMs)from human-written content.Through testing model performance on a diversified dataset,the effectiveness of an upgraded discrimination strategy is substantiated.The study particularly evaluates the GPT-3.5-Turbo model and compares its performance against various mainstream classification models.The findings indicate that the accuracy of model discrimination is significantly influenced by the length of text sequences,unveiling a new perspective on length as a critical factor impacting discrimination efficacy.These insights not only deepen the understanding of characteristics unique to AI-generated text but also provide direction for the development of more precise discrimination algorithms.

关 键 词：深度学习 文本判别 鉴别准确率 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.1[自动化与计算机技术—控制科学与工程]