Work

LMSA at AraGenEval Shared Task: Ensemble-Based Detection of AI-Generated Arabic Text Using Multilingual and Arabic-Specific Models

Year: 2025

Type: conference-paper

Abstract: We address the problem of distinguishing between human-authored and AI-generated text in low-resource languages, particularly Arabic.We present the LMSA 1 team's participation in the ARATECT (Arabic AI-Generated Text Detection) subtask of the AraGenEval 2 shared task, which targets the detection of AI-generated Arabic texts.We propose an ensemble-based classification framework that integrates multilingual and Arabic-specific pre-trained language models, namely Fanar, AraBERT, and XLM-R, optimized through a dedicated fine-tuning pipeline.The approach is evaluated on the balanced Arabic text dataset provided by the shared task organizers.Our system achieved an F1-score of 0.864 and ranked first among all participating teams.

Source: Arabic Natural Language Processing Workshop (WANLP)

DOI: 10.18653/v1/2025.arabicnlp-sharedtasks.4

Authors: Kaoutar Zita, Attia Nehar, Abdelkader Khelil, Slimane Bellaouar, Hadda Cherroun

Institutions: University of Ghardaia, Ziane Achour University of Djelfa, Amar Telidji University of Laghouat

Language: English

Topic: Authorship Attribution and Profiling

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funder: Direction Générale de la Recherche Scientifique et du Développement Technologique

Award: G1874142762