A novel hybrid approach based on self-supervised transformers and explainable AI for multilingual fake text detection

Abstract

With the rapid development and advancement of artificial intelligence, similarities in language structure and grammatical rules are making it increasingly difficult for AI-generated text to distinguish itself from human-written text. To overcome this problem, this study proposes a novel approach for multilingual AI-generated text detection using self-supervised learning and explainable AI techniques. Starting with an unlabeled dataset, self-supervised training is used to enable the model to deeply learn all structural and semantic features of a language. This robust language representation is then subjected to a supervised fine-tuning process to customize the model's capability for detecting synthetic text. Following multi-stage training, the proposed framework is evaluated on three rigorously compiled datasets encompassing English, Turkish, and Arabic texts from forums, academic publications, and social media platforms. To enhance transparency and trust in the model's decision-making process, we integrate Local Interpretable Model-agnostic Explanations (LIME) and Shapley Additive exPlanations (SHAP), providing interpretable insights into the model's predictions. The experimental results validate the robustness of the proposed architecture, demonstrating superior performance with 99.9% accuracy in English, followed by 96.1% and 96.6% in Turkish and Arabic, respectively. This study presents a pivotal explainable AI framework for detecting synthetic text while simultaneously holding great promise for bolstering digital content security.