Artificial Intelligence-Based Data-Driven Approach for Classifying AI-Generated Text from Human-Generated Text

Le résumé fourni par la source

The rapid advancement of artificial intelligence (AI) language models, such as ChatGPT, has raised growing concerns about distinguishing AI-generated text from human-written content. Accurately identifying the origin of textual content has become a challenging research problem because AI-generated and human-written texts often exhibit high linguistic and contextual similarity. This challenge has important implications for content authenticity, spam detection, misinformation prevention, and the security of online information systems. Although conventional machine learning (ML) and natural language processing (NLP) techniques have been applied to this task, they often struggle to capture complex semantic and contextual relationships. To address this issue, this study proposes a Hierarchical Deep Ensemble Learning Approach for Distinguishing AI-Generated Text from Human-Generated Text (HDEL-AITHG) using contextual representations. Initially, DeBERTa is employed to extract rich contextual embeddings from the input text. These representations are then processed through a weighted soft-voting ensemble comprising Kolmogorov-Arnold Networks, a Hierarchical Attention Network, and a Temporal Convolutional Network, which capture complementary nonlinear, hierarchical, and sequential characteristics of textual data. Furthermore, the Lookahead optimizer is employed to improve training stability and convergence, while SHAP analysis enhances the interpretability of the framework. Extensive experiments conducted on the benchmark ChatGPT Classification Dataset demonstrate the effectiveness and reliability of HDEL-AITHG in distinguishing AI-generated from human-written text. The proposed approach can support applications in content authenticity verification and automated text analysis.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.