Stylometric Analysis and Detection of Large Language Models Text

View PDF file from the paper entitled AI Arabic: Stylometric analysis and discovering the text of large language models, written by Maged S. Al-Shaibani and 1 other authors

PDF view

a summary:LLMS models have achieved unprecedented capabilities in generating a human -like text, which constitutes hidden but important challenges to integrate information through critical fields, including education, social media and academic circles, allowing advanced misinformation campaigns, clarifying health care directions, and facilitating targeted publication. This challenge becomes severe, especially in unstable languages ​​and low resources such as Arabic. This paper provides a comprehensive investigation of the text created by Arabic machine guns, and the study of multiple generation strategies (generating the title only, generating content, and refining the text) through various typical structures (Allam, Jais, Lama, and GPT-4) in the academic and social fields. Our analysis reveals the distinctive linguistic patterns that were characterized by the writing of man from the Arabic text that was created by machine guns through these various contexts. Despite its human -like characteristics, we prove that LLMS produces signatures that can be discovered in their Arab outputs, with characteristics of the field that differ greatly between the different contexts. Based on these ideas, we have developed the BERT-based detection models that have achieved exceptional performance in official contexts (up to 99.9 % F1-SCORE) with strong accuracy through the typical structure. Our analysis across the field confirms the challenges of generalization previously reported in the literature. To the extent of our knowledge, this work represents the most comprehensive investigation in the text that was created by Arabic machine guns so far, and uniquely combines the methods of generating multiple waves, the buildings of various models, and in -depth systems systems across the various textual fields, and establishing a basis for the development of the basic basic basic detection and linguistic detection systems to maintain the integration of Arab information.