REGLAT at AraGenEval shared task: Morphology-Aware AraBERT for Detecting Arabic AI-Generated Text

Résumé

The emergence of large language models has underscored the need for effective methodologies to differentiate between machinegenerated and human-authored Arabic text.This study introduces a transformer-based classification system designed for the AraGenEval shared task focused on detecting AI-generated Arabic text.The proposed approach employs AraBERTv2 as the backbone architecture, augmented with a comprehensive preprocessing pipeline that addresses Arabic-specific orthographic variations through systematic diacritic removal and character normalization.Experimental results indicate that this preprocessingenhanced approach achieves a weighted F1 score of 0.63 on the test dataset, demonstrating particularly strong performance in modern standard Arabic texts.The results suggest that morphological normalization is crucial for the detection of AI-generated Arabic text, surpassing the significance of similar preprocessing techniques in other languages.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0