Adaptação do algoritmo Ferret/Webferret para detecção de plágio em textos em português: desafios com textos gerados por IA

Abstract

Plagiarism detection in academic texts is a growing challenge, especially with the advancement of artificial intelligence (AI) technologies, such as ChatGPT, which enable the generation of cohesive and complex texts. This study proposes the reimplementation of the WebFerret algorithm, originally designed for textual similarity detection, with specific adaptations for the Portuguese language and phrase-by-phrase analysis, aiming to enhance plagiarism detection in AI-generated texts. The methodology included modifying the algorithm to consider rare trigrams in phrases and applying similarity metrics based on the Jaccard coefficient. The experiment was conducted with three categories of texts: direct plagiarism, mosaic plagiarism, and AIgenerated texts, using a database created for evaluation purposes. The results showed high algorithm effectiveness in identifying direct and mosaic plagiarism but significant limitations in detecting plagiarism in AI-generated texts, highlighting the need for new approaches to address this type of content. It is concluded that the proposed adaptation represents progress in plagiarism detection in Portuguese but still requires improvements to tackle the challenges posed by AI.