--- title: "MarsadLab at AraGenEval Shared Task — Biswas 2025" description: "The paper describes the MarsadLab entry to the AraGenEval shared task at ArabicNLP 2025." authors: ["Md. Rafiul Biswas", "Mabrouka Bessghaier", "Firoj Alam", "Wajdi Zaghouani"] year: 2025 venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks" source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.15/" doi: "10.18653/v1/2025.arabicnlp-sharedtasks.15" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/biswas-2025-marsadlab-arageneval-shared-task-llm/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/biswas-2025-marsadlab-arageneval-shared-task-llm/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/2/) / MarsadLab at AraGenEval Shared Task: LLM-Based Approaches to Arabic Authorship Style Transfer and Identification # MarsadLab at AraGenEval Shared Task: LLM-Based Approaches to Arabic Authorship Style Transfer and Identification **Authors:** [Md. Rafiul Biswas](https://md.research.elamur.ai/free/en/authors/md-rafiul-biswas/), [Mabrouka Bessghaier](https://md.research.elamur.ai/free/en/authors/mabrouka-bessghaier/), [Firoj Alam](https://md.research.elamur.ai/free/en/authors/firoj-alam/), [Wajdi Zaghouani](https://md.research.elamur.ai/free/en/authors/wajdi-zaghouani/) · **Year:** 2025 · **Venue:** Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.15/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.15.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.15](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.15) · [OpenAlex](https://openalex.org/W4416037798) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) · [1](#cited-by) in our library, of the 140884 papers whose bibliographies we read ## Summary The paper describes the MarsadLab entry to the AraGenEval shared task at ArabicNLP 2025. The task has two parts: rewriting a Modern Standard Arabic passage so it reads like a particular author while keeping its meaning, and recognising which of 21 authors wrote a given excerpt. The authors note that Arabic work in this area has been held back by small, scattered datasets and a lack of common benchmarks. For generation, they adapted instruction-tuned LLMs with LoRA and Arabic-specific text normalisation. Qwen2.5-7B-Instruct gave their best output and placed third on the leaderboard. For attribution, they fine-tuned the compact encoder AraBERTv2 with a classification head. It beat MARBERT and a Qwen2.5 classifier, though it still trailed the winning system. The work shows that open models with adapter-based tuning on modest GPUs can be competitive on both tasks. It also suggests that encoder models remain the stronger choice for discrimination, while instruction-tuned decoders suit stylistic rewriting. The scripts are released on GitHub. ## Key points - The data has 47,692 parallel pairs of neutral MSA text and author-styled text, split 35,122 / 4,157 / 8,413 for train / validation / test, across 21 authors. One author has over 4,000 samples, so the classes are skewed. - Style transfer used Qwen2.5-7B-Instruct with LoRA (r = 16, α = 32, dropout 0.1) on the attention projections via HuggingFace peft. Decoding used top_p 0.9 and temperature 0.7. - The style transfer result was BLEU 20.30 and chrF 52.56, placing 3rd. The top system scored 24.58 / 59.01. Fanar-1.9B (18.39 / 48.32) and Jais-13B (15.17 / 47.32) scored lower. - Identification used AraBERTv2 (aubmindlab/bert-base-arabertv2) with 512-token inputs and 4 epochs. Other settings were learning rate 2e-5, batch 8 and gradient accumulation 4, with early stopping on validation macro-F1. - AraBERTv2 reached 86.49% accuracy and 82.82% macro-F1, against the winner's 92.42% / 89.89%. MARBERT and Qwen2.5-7B lagged behind, and less-represented authors were the weak spot. ## Interesting ideas - If the generator returns nothing or fails, the system outputs the unchanged neutral source text, so every input still gets a sensible answer. — For Mova (back): Graph-to-text generation could emit the original sentence whenever realisation fails, keeping the compressor's output complete and the failure visible in logs. - Text is normalised by merging spelling variants of the same letter and removing script noise, while features that carry an author's style are deliberately kept. — For Mova (tokens): The tokenizer could split normalisation into "noise to remove" and "signal to keep" layers, so lemmatisation sees canonical forms while the original surface stays available for irony or register cues. - Checkpoints are chosen by macro-F1 instead of accuracy, so rare classes count as much as frequent ones on a skewed label set. — For Mova (learn): Model selection for taggers or frame classifiers could use macro-averaged scores, so rare dependency relations or VerbNet classes are not hidden by common ones. - The related work they build on uses inverse transfer: a strong LLM strips style from texts to create synthetic neutral–styled pairs, which then train a smaller model. — For Mova (learn): An LLM teacher tier could produce paired paraphrases of the same meaning, giving silver data to check that the parser and frames yield identical graphs for both versions. - On the same data, a compact encoder clearly beat a 7B generative model at classification, while the decoder won at rewriting. — For Mova (tree): This supports keeping a dedicated encoder like gparse for structural decisions rather than asking a general LLM to label, saving the LLM for teacher or generation roles. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/) · [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) ## Cited by Cited by 1 paper of our library cites it (of the 140884 whose bibliographies we read); each has a page here. - [The AraGenEval Shared Task on Arabic Authorship Style Transfer and AI Generated Text Detection](https://md.research.elamur.ai/free/en/papers/abudalfa-2025-arageneval-shared-task-arabic-authorship/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{biswas2025marsadlab, title = {{MarsadLab at AraGenEval Shared Task: LLM-Based Approaches to Arabic Authorship Style Transfer and Identification}}, author = {Biswas, Md. Rafiul and Bessghaier, Mabrouka and Alam, Firoj and Zaghouani, Wajdi}, booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks}, year = {2025}, doi = {10.18653/v1/2025.arabicnlp-sharedtasks.15}, url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.15} } ``` - [← previous Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets](https://md.research.elamur.ai/free/en/papers/moghe-2025-machine-translation-meta-evaluation-through/) - [next → Mechanisms vs. Outcomes: Probing for Syntax Fails to Explain Performance on Targeted Syntactic Evaluations](https://md.research.elamur.ai/free/en/papers/agarwal-2025-mechanisms-vs-outcomes-probing-syntax/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/2/) ## References 14 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Muhammad Abdul-Mageed, Abdelrahim Elmadany, and 1 others. 2021. ARBERT & MARBERT: Deep bidirectional transformers for arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7088–7105. [find](https://scholar.google.com/scholar?q=Muhammad+Abdul%2DMageed%2C+Abdelrahim+Elmadany%2C+and+1+others%2E+2021%2E+ARBERT+%26+MARBERT%3A+Deep+bidirectional+transformers+for+arabic%2E+In+Proceedings+of+the+59th+Annual+Meeting+of+the+Association+for+Computati) [our card](https://md.research.elamur.ai/free/en/papers/abdul-2021-mageed-arbert-marbert-deep-bidirectional-transformers/) 2. Imane Guellil, Houda Saâdane, Faical Azouaou, Billel Gueni, and Damien Nouvel. 2021. Arabic natural language processing: An overview. Journal of King Saud University-Computer and Information Sciences, 33(5):497–507. [find](https://scholar.google.com/scholar?q=Imane+Guellil%2C+Houda+Sa%C3%A2dane%2C+Faical+Azouaou%2C+Billel+Gueni%2C+and+Damien+Nouvel%2E+2021%2E+Arabic+natural+language+processing%3A+An+overview%2E+Journal+of+King+Saud+University%2DComputer+and+Information+Sciences%2C) [our card](https://md.research.elamur.ai/free/en/papers/guellil-2019-arabic-natural-language-processing-overview/) 3. Shadi Abudalfa, Saad Ezzini, Ahmed Abdelali, Hamza Alami, Abdessamad Benlahbib, Salmane Chafik, Mo El-Haj, Abdelkader El Mahdaouy, Mustafa Jarrar, Salima Lamsiyah, and Hamzah Luqman. 2025. The arageneval shared task on arabic authorship style transfer and ai-generated text detection. In Proceedings of the Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou, China. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Shadi+Abudalfa%2C+Saad+Ezzini%2C+Ahmed+Abdelali%2C+Hamza+Alami%2C+Abdessamad+Benlahbib%2C+Salmane+Chafik%2C+Mo+El%2DHaj%2C+Abdelkader+El+Mahdaouy%2C+Mustafa+Jarrar%2C+Salima+Lamsiyah%2C+and+Hamzah+Luqman%2E+2025%2E+The+aragene) [our card](https://md.research.elamur.ai/free/en/papers/abudalfa-2025-arageneval-shared-task-arabic-authorship/) 4. Zachary Horvitz, Ajay Patel, Kanishk Singh, Chris Callison-Burch, Kathleen McKeown, and Zhou Yu. 2024. Tinystyler: Efficient few-shot text style transfer with authorship embeddings. arXiv preprint arXiv:2406.15586. [↗](https://arxiv.org/abs/2406.15586) 5. Fatimah Alqahtani and Mischa Dohler. 2023. Survey of authorship identification tasks on arabic texts. ACM Transactions on Asian and Low-Resource Language Information Processing, 22(4):1–24. [find](https://scholar.google.com/scholar?q=Fatimah+Alqahtani+and+Mischa+Dohler%2E+2023%2E+Survey+of+authorship+identification+tasks+on+arabic+texts%2E+ACM+Transactions+on+Asian+and+Low%2DResource+Language+Information+Processing%2C+22%284%29%3A1%E2%80%9324%2E) 6. Neha Sengupta, Sunil Kumar Sahu, Bokang Jia, Satheesh Katipomu, Haonan Li, Fajri Koto, Osama Mohammed Afzal, Samta Kamboj, Onkar Pandit, Rahul Pal, and 1 others. 2023. Jais and Jais-chat: Arabic-centric foundation and instruction-tuned open generative large language models. arXiv preprint arXiv:2308.16149. [↗](https://arxiv.org/abs/2308.16149) [our card](https://md.research.elamur.ai/free/en/papers/sengupta-2023-jais-jais-chat-arabic-centric/) 7. Shuai Liu, Shantanu Agarwal, and Jonathan May. 2024. Authorship style transfer with policy optimization. arXiv preprint arXiv:2403.08043. [↗](https://arxiv.org/abs/2403.08043) 8. Hamed Alshammari and Khaled Elleithy. 2024. Toward robust arabic AI-generated text detection: Tackling diacritics challenges. Information, 15(7):419. [find](https://scholar.google.com/scholar?q=Hamed+Alshammari+and+Khaled+Elleithy%2E+2024%2E+Toward+robust+arabic+AI%2Dgenerated+text+detection%3A+Tackling+diacritics+challenges%2E+Information%2C+15%287%29%3A419%2E) 9. Zhonghui Shao, Jing Zhang, Haoyang Li, Xinmei Huang, Chao Zhou, Yuanchun Wang, Jibing Gong, Cuiping Li, and Hong Chen. 2024. Authorship style transfer with inverse transfer data augmentation. AI Open, 5:94–103. [find](https://scholar.google.com/scholar?q=Zhonghui+Shao%2C+Jing+Zhang%2C+Haoyang+Li%2C+Xinmei+Huang%2C+Chao+Zhou%2C+Yuanchun+Wang%2C+Jibing+Gong%2C+Cuiping+Li%2C+and+Hong+Chen%2E+2024%2E+Authorship+style+transfer+with+inverse+transfer+data+augmentation%2E+AI+Open%2C) 10. Yousef Altaher, Ali Fadel, Mazen Alotaibi, Mazen Alyazidi, Mishari Al-Mutairi, Mutlaq Aldhbuiub, Abdulrahman Mosaibah, Abdelrahman Rezk, Abdulrazzaq Alhendi, Mazen Abo Shal, and 1 others. 2022. Masader plus: A new interface for exploring+ 500 arabic nlp datasets. arXiv preprint arXiv:2208.00932. Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad, Firoj Alam, Enes Altinisik, Ehsannedin Asgari, Yazan Boshmaf, Sabri Boughorbel, Sanjay Chawla, Shammur Chowdhury, Fahim Dalvi, Kareem Darwish, Nadir Durrani, Mohamed Elfeky, Ahmed Elmagarmid, Mohamed Eltabakh, Masoomali Fatehkia, Anastasios Fragkopoulos, Maram Hasanain, and 23 others. Fanar: An arabic-centric multimodal generative ai platform. [↗](https://arxiv.org/abs/2208.00932) [our card](https://md.research.elamur.ai/free/en/papers/altaher-2022-masader-plus-new-interface-exploring/) 11. Zaid Alyafeai, Maraim Masoud, Mustafa Ghaleb, and Maged S. Al-shaibani. 2021. Masader: Metadata sourcing for arabic text and speech data resources. Preprint, arXiv:2110.06744. [↗](https://arxiv.org/abs/2110.06744) [our card](https://md.research.elamur.ai/free/en/papers/alyafeai-2022-masader-metadata-sourcing-arabic-text/) 12. Fetoun Mansour AlZahrani and Maha Al-Yahya. 2023a. A transformer-based approach to authorship attribution in classical arabic texts. Applied Sciences, 13(12):7255. [find](https://scholar.google.com/scholar?q=Fetoun+Mansour+AlZahrani+and+Maha+Al%2DYahya%2E+2023a%2E+A+transformer%2Dbased+approach+to+authorship+attribution+in+classical+arabic+texts%2E+Applied+Sciences%2C+13%2812%29%3A7255%2E) 13. Qwen Team. 2024. Qwen2.5: A party of foundation models. [find](https://scholar.google.com/scholar?q=Qwen+Team%2E+2024%2E+Qwen2%2E5%3A+A+party+of+foundation+models%2E) 14. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, and 40 others. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671. [↗](https://arxiv.org/abs/2407.10671)