摘要
针对目前人工智能(Artificial Intelligence,AI)生成文本的滥用导致的学术不端、侵犯版权、隐私保护和舆情监控等问题,提出了一种基于自然语言处理的AI生成文本的识别和检测算法。该算法首先采用Word2vec方法中的连续词袋模型将文本词转换成词向量,并将词向量累加获得文本向量。随后利用softmax函数获取文本向量的概率分布,通过统计可视化分析AI生成文本的基本规律,并采用余弦相似性来判断文本类型。其次采用支持向量机递归特征消除算法判断文本是否由AI生成,通过K-近邻算法对文本重生成次数进行判断,进一步细化了文本检测的粒度。通过仿真实验验证了算法的有效性,结果显示算法识别准确率达80%及以上。
To address such issues as academic dishonesty,copyright infringement stemming,privacy protection and public opinion monitoring from the misuse of artificial intelligence(AI)-generated texts,an recognition and detection algorithm based on natural language processing(NLP)is proposed.This algorithm initially converts words into vectors using the continuous bag-of-words(CBOW)model within Word2vec,and accumulates them into text vectors.It then applies softmax to address their probability distribution,analyze the fundamental patterns of AI-generated texts with statistical visualization,and determin the type of text by using cosine similarity.Next,a support vector machine recursive feature elimination(SVM-RFE)is used to determine whether the text is generated by AI.For AI-generated texts,the K-nearest neighbor(KNN)algorithm estimates the extent of AI involvement,further refining the granularity of text detection.Finally,simulation experiments show the algorithm s effectiveness with recognition accuracy of 80%or above.
作者
王雨欣
刘柯飞
李雪莲
王红军
WANG Yuxin;LIU Kefei;LI Xuelian;WANG Hongjun(College of Information Science and Engineering,Hohai University,Changzhou 213200,China;Guangling College of Yangzhou University,Yangzhou 225000,China;College of Electronic Engineering,National University of Defense Technology,Hefei 230031,China)
出处
《电讯技术》
北大核心
2025年第3期378-384,共7页
Telecommunication Engineering
基金
国家自然科学基金面上项目(61971473)。
