Osint at AraGenEval shared task: Fine-Tuned Modeling for Tracking Style Signatures and AI Generation in Arabic Texts
Year: 2025
Type: conference-paper
Abstract: The increasing complexity of large language models (LLMs) has made human-written and machine-translated text difficult to distinguish, reinforcing the requirement for effective stylistic modeling and authorship analysis in Arabic.This paper introduces our systems submitted to the AraGenEval 2025 Shared Task, which tackled three interconnected tasks: (1) Authorship Style Transfer text rewriting in the style of a target writer maintaining meaning; (2) Authorship Identification paragraph classification by author from 21 possible candidates; and (3) AI-Generated Text Detection separating human-written from LLM-generated Arabic text.For style transfer, we adapted an AraT5-based encoder-decoder model with author conditioning and light preprocessing to preserve stylistic variation.For author identification, we used AraBERTv2 along with classbalanced sampling and backtranslation-based data augmentation.For AI-generated text detection, we deployed a hybrid mBERT model augmented with handcrafted linguistic features.Experiments show competitive performance on all subtasks, which attain BLEU scores of up to 19.87 in style transfer, an F1-score of 0.79673 in identifying the author, and an F1-score of 0.75 in detecting AI-generated text.Ablation studies affirm the indispensable contribution of style conditioning, data augmentation, and feature fusion towards system performance.
Authors: Shifali Agrahari, Hemanth Prakash Simhadri, Ashutosh Kumar Verma, Ranbir Singh Sanasam
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037675