MarsadLab at AraGenEval Shared Task: LLM-Based Approaches to Arabic Authorship Style Transfer and Identification
Year: 2025
Type: conference-paper
Abstract: We present our system submitted to the Ara-GenEval Shared Task at ArabicNLP 2025, which addresses the tasks of Authorship Style Transfer and Authorship Identification.For Subtask 1 (Style Transfer), we fine-tuned instruction-following Arabic large language models using Low-Rank Adaptation (LoRA).Among the evaluated models, Qwen2.5-7B-Instructachieved a BLEU score of 20.30 and a chrF score of 52.56, ranking 3 rd on the official leaderboard.For Subtask 2 (Authorship Identification), AraBERTv2 attained an accuracy of 86.49% and a macro-F1 score of 82.82%, demonstrating robust performance in multiclass author classification across 21 categories.Our approach integrates Arabic-specific preprocessing, task-oriented prompt design, and transformer-based architectures, which enables effective handling of both generative and discriminative aspects of authorship analysis.We have made experimental scripts publicly available for the community.1
Institution: Northwestern University in Qatar
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4416037798
FWCI:
Cites:
Cited by:
Related to: