--- title: "ANLPers at AraGenEval Shared Task — Nacar 2025" description: "The paper describes the winning system for Subtask 1 of the AraGenEval 2025 shared task: rewriting Modern Standard Arabic text so it reads like one of 21…" authors: ["Omer Nacar", "Mahmoud Reda", "Serry Sibaee", "Yasser Alhabashi", "Adel Ammar", "Wadii Boulila"] year: 2025 venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks" source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.8/" doi: "10.18653/v1/2025.arabicnlp-sharedtasks.8" 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/nacar-2025-anlpers-arageneval-shared-task-descriptive/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/nacar-2025-anlpers-arageneval-shared-task-descriptive/) [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/) / ANLPers at AraGenEval Shared Task: Descriptive Author Tokens for Transparent Arabic Authorship Style Transfer # ANLPers at AraGenEval Shared Task: Descriptive Author Tokens for Transparent Arabic Authorship Style Transfer **Authors:** [Omer Nacar](https://md.research.elamur.ai/free/en/authors/omer-nacar/), [Mahmoud Reda](https://md.research.elamur.ai/free/en/authors/mahmoud-reda/), [Serry Sibaee](https://md.research.elamur.ai/free/en/authors/serry-sibaee/), [Yasser Alhabashi](https://md.research.elamur.ai/free/en/authors/yasser-alhabashi/), [Adel Ammar](https://md.research.elamur.ai/free/en/authors/adel-ammar/), [Wadii Boulila](https://md.research.elamur.ai/free/en/authors/wadii-boulila/) · **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.8/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.8.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.8](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.8) · [OpenAlex](https://openalex.org/W4416037681) **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 winning system for Subtask 1 of the AraGenEval 2025 shared task: rewriting Modern Standard Arabic text so it reads like one of 21 specific Arabic authors. The team fine-tuned UBC-NLP/AraT5v2-base-1024, an Arabic T5 encoder–decoder. They compared four ways of telling the model which author to imitate: - no conditioning at all (plain input–output pairs); - an opaque numeric author ID token; - a readable token carrying the author's name; - a short Arabic instruction placed before the source text. All conditioned variants clearly beat the plain baseline. The readable name token scored about the same as the numeric ID, so making the signal transparent costs nothing. The Arabic instruction gave the best BLEU and took first place overall. A stylometric profiling of the authors showed they are very similar on surface features. The authors use this to argue that the model's pretrained knowledge, reached through the prompt, matters more than shallow cues. The team also reports mid-table results on authorship identification and on detecting AI-generated Arabic text. ## Key points - Data: 21 authors, with 35,122 training, 4,157 validation and 8,413 test samples. Preprocessing normalized whitespace, punctuation forms and diacritics. - Test results (BLEU / chrF): - baseline: 20.50 / 58.50 - numeric tokens: 24.04 / 59.15 - descriptive tokens: 24.00 / 59.00 - Arabic prompt: 24.58 / 59.01 The prompt variant ranked first, although the numeric tokens have a marginally higher chrF. - Training setup: - AdamW with weight decay 0.01, learning rate 5e-5; - OneCycleLR with 1000 warmup steps; - batch 64 across 4 A100 GPUs; - up to 10 epochs, with early stopping on validation loss. - Decoding: beams=2 with sampling (temperature 0.6, top_k 20, top_p 0.8) and a 512-token output cap. - Other subtasks: - Authorship identification used bert-base-arabic-camelbert-mix-sentiment with weighted loss. It reached 95.3% accuracy on dev and test F1 0.831, ranking 6th. - AI-text detection (ARATECT) used XLM-RoBERTa. It reached about 99.4% accuracy on dev but only F1 0.76 on test, also 6th. ## Interesting ideas - Swapping opaque numeric class IDs for name-bearing tokens kept scores essentially unchanged, so readable control signals can be had at no cost. — For Mova (back): condition the graph→text generator on readable tags (e.g. register or speaker names) instead of indices, so every generation choice stays explainable in logs. - The comparison is a ladder from implicit to explicit conditioning, with model, data and decoding all held fixed, so each rung's gain is attributable to the conditioning alone. — For Mova (learn): structure ablations of new features (e.g. a VerbNet hint added to parser input) as a controlled ladder alongside the negative-control runs. - A post-hoc stylometric profile showed that the target classes barely differ on measurable surface features, which explains where the difficulty lies rather than blaming the model. — For Mova (learn): before training on silver labels, profile whether the label categories are actually separable by simple features, to flag ill-posed distinctions early. - Reporting word-level BLEU next to character-level chrF exposed a disagreement about which variant is best, a useful signal for morphology-heavy output. — For Mova (back): score the text→graph→text compressor with both a word and a character metric, and treat ranking disagreements as cases to inspect. - Phrasing the task as a natural-language instruction in the model's own language let fine-tuning reuse pretrained knowledge better than special tokens did. — For Mova (learn): when querying LLM teacher tiers, state the annotation task as a plain-English instruction rather than symbolic codes, to get more faithful silver labels. *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/) Topics: [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) ## 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{nacar2025anlpers, title = {{ANLPers at AraGenEval Shared Task: Descriptive Author Tokens for Transparent Arabic Authorship Style Transfer}}, author = {Nacar, Omer and Reda, Mahmoud and Sibaee, Serry and Alhabashi, Yasser and Ammar, Adel and Boulila, Wadii}, booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks}, year = {2025}, doi = {10.18653/v1/2025.arabicnlp-sharedtasks.8}, url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.8} } ``` - [← previous Ancient and Rare Readings in the Syriac Harklean Version of the Gospel of John and their Significance for its Interpretation](https://md.research.elamur.ai/free/en/papers/jutkiewicz-2025-ancient-rare-readings-syriac-harklean/) - [next → Annotating Second Language in Universal Dependencies: a Review of Current Practices and Directions for Harmonized Guidelines](https://md.research.elamur.ai/free/en/papers/masciolini-2025-annotating-second-language-universal-dependencies/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/) ## References 12 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Sebastian Ruder, Anders Søgaard, and Ivan Vulić. 2019. Unsupervised cross-lingual representation learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Tutorial Abstracts, pages 31–38. [find](https://scholar.google.com/scholar?q=Sebastian+Ruder%2C+Anders+S%C3%B8gaard%2C+and+Ivan+Vulic%CC%81%2E+2019%2E+Unsupervised+cross%2Dlingual+representation+learning%2E+In+Proceedings+of+the+57th+Annual+Meeting+of+the+Association+for+Computational+Linguistics%3A+) 2. 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), 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/) 3. Vicky Shao, Xinyi Chen, Saizhu Zong, Yuerou Yang, Ao Peng, and He He. 2024. Authorship style transfer with inverse transfer data augmentation. AI Open, 5:14–22. [find](https://scholar.google.com/scholar?q=Vicky+Shao%2C+Xinyi+Chen%2C+Saizhu+Zong%2C+Yuerou+Yang%2C+Ao+Peng%2C+and+He+He%2E+2024%2E+Authorship+style+transfer+with+inverse+transfer+data+augmentation%2E+AI+Open%2C+5%3A14%E2%80%9322%2E) 4. Reem Alghamdi, Areej Al-Wabil, and Muna Al-Razgan. 2024. Distinguishing arabic genai-generated tweets and human tweets utilizing machine learning. Engineering, Technology & Applied Science Research, 14(5):16720–16726. [find](https://scholar.google.com/scholar?q=Reem+Alghamdi%2C+Areej+Al%2DWabil%2C+and+Muna+Al%2DRazgan%2E+2024%2E+Distinguishing+arabic+genai%2Dgenerated+tweets+and+human+tweets+utilizing+machine+learning%2E+Engineering%2C+Technology+%26+Applied+Science+Research%2C+1) 5. AbdelRahim Elmadany, Muhammad Abdul-Mageed, and 1 others. 2022. Arat5: Text-to-text transformers for arabic language generation. In Proceedings of the 60th annual meeting of the association for computational linguistics (Volume 1: Long papers), pages 628–647. [find](https://scholar.google.com/scholar?q=AbdelRahim+Elmadany%2C+Muhammad+Abdul%2DMageed%2C+and+1+others%2E+2022%2E+Arat5%3A+Text%2Dto%2Dtext+transformers+for+arabic+language+generation%2E+In+Proceedings+of+the+60th+annual+meeting+of+the+association+for+comput) [our card](https://md.research.elamur.ai/free/en/papers/nagoudi-2022-arat5-text-text-transformers-arabic/) 6. Helena Gómez-Adorno, Juan-Pablo Posadas-Duran, Germán Ríos-Toledo, Grigori Sidorov, and Gerardo Sierra. 2018. Stylometry-based approach for detecting writing style changes in literary texts. Computación y Sistemas, 22(1):47–53. [find](https://scholar.google.com/scholar?q=Helena+G%C3%B3mez%2DAdorno%2C+Juan%2DPablo+Posadas%2DDuran%2C+Germ%C3%A1n+R%C3%ADos%2DToledo%2C+Grigori+Sidorov%2C+and+Gerardo+Sierra%2E+2018%2E+Stylometry%2Dbased+approach+for+detecting+writing+style+changes+in+literary+texts%2E+Computaci) 7. Sol Horvitz, Luis Ortiz, Anjali Sharan, and Alexandra Getman. 2024. Tinystyler: Efficient few-shot text style transfer with authorship embeddings. arXiv preprint arXiv:2406.15586. [↗](https://arxiv.org/abs/2406.15586) 8. Zhumin Hu, Zhaofeng Tu, Zur G-BETH, Jdn Lrec, and Vıctor T-BI. 2022. Text style transfer: A review and experimental evaluation. ACM SIGKDD Explorations Newsletter, 24(1):1–21. [find](https://scholar.google.com/scholar?q=Zhumin+Hu%2C+Zhaofeng+Tu%2C+Zur+G%2DBETH%2C+Jdn+Lrec%2C+and+V%C4%B1ctor+T%2DBI%2E+2022%2E+Text+style+transfer%3A+A+review+and+experimental+evaluation%2E+ACM+SIGKDD+Explorations+Newsletter%2C+24%281%29%3A1%E2%80%9321%2E) 9. G Inoue, B Alhafni, N Baimukan, H Bouamor, and N Habash. The interplay of variant, size, and task type in arabic pre-trained language models. arxiv 2021. arXiv preprint arXiv:2103.06678. [↗](https://arxiv.org/abs/2103.06678) [our card](https://md.research.elamur.ai/free/en/papers/inoue-2021-interplay-variant-size-task-type/) 10. AraGenEval Organizers. 2024. Overview of the arageneval 2024 shared task. Shared Task Website. (Placeholder citation). [find](https://scholar.google.com/scholar?q=AraGenEval+Organizers%2E+2024%2E+Overview+of+the+arageneval+2024+shared+task%2E+Shared+Task+Website%2E+%28Placeholder+citation%29%2E) 11. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318. [↗](https://doi.org/10.3115/1073083.1073135) [our card](https://md.research.elamur.ai/free/en/papers/papineni-2002-bleu-method-automatic-evaluation-machine/) 12. Krish Patel, Saizhu Zong, Vicky Shao, Ao Peng, and He He. 2022. Low-resource authorship style transfer: [find](https://scholar.google.com/scholar?q=Krish+Patel%2C+Saizhu+Zong%2C+Vicky+Shao%2C+Ao+Peng%2C+and+He+He%2E+2022%2E+Low%2Dresource+authorship+style+transfer%3A)