Year: 2025
Type: conference-paper
Abstract: This paper presents our approach and findings for Subtask 1 (Authorship Style Transfer) of the AraGenEval2025 shared task.We explore methods to transform neutral Arabic text into the distinctive style of a specified author while preserving its original meaning.Our work details a twophase development: an initial baseline model leveraging few-shot prompting with Gemini and K-means clustering, followed by fine-tuning of pre-trained seq2seq models that support Arabic, including representatives from the mT5 and mBART model families.We evaluated our models using BLEU and chrF metrics, demonstrating significant improvements in fine-tuning, particularly in capturing Arabic-specific stylistic nuances.To complement these surface-level overlap metrics, we incorporate BERTScore to assess semantic preservation across style transfer.Additionally, we introduce a style classifier to quantify author-specific style transfer strength.We discuss the challenges encountered, including Arabic linguistic complexities, handling long Arabic text, and hardware constraints, and outline future directions for enhancing Arabic Authorship Style Transfer.
Authors: Hafsa Kara Achira, Mourad Bouache, Mourad Dahmane
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4416037650
FWCI:
Cites:
Cited by:
Related to: