Elamur Research⌕
☰
home / papers / 2026 / HCMUS_The Fangs at AbjadStyleTransfer Shared Task: Learning to Query Style, Contrastive Representations for Zero-Shot Arabic Authorship Style Transfer

HCMUS_The Fangs at AbjadStyleTransfer Shared Task: Learning to Query Style, Contrastive Representations for Zero-Shot Arabic Authorship Style Transfer

Summary

The paper presents the top-ranked system in the AbjadStyleTransfer shared task (ArabicNLP 2026). The task is to rewrite formal Modern Standard Arabic so that it reads like a given literary author. The team's data analysis showed that none of the 21 test authors, among them Naguib Mahfouz, Taha Hussein and Khalil Gibran, appear in the training set of 32,784 authors. This makes the problem purely zero-shot.

Because author-ID or classifier-based conditioning cannot generalize under these conditions, the authors built a dual-encoder model. An AraBERT encoder is trained contrastively to place texts in a normalized style space. The resulting style vector is then added to the hidden states of an AraT5 generator.

The system reaches 19.77 BLEU and 55.74 chrF, ahead of retrieval-augmented and multi-task baselines. More elaborate style-injection designs performed markedly worse. The authors conclude that pre-trained Arabic models already hold much stylistic knowledge about famous writers. Light conditioning that leaves this knowledge intact works better than heavy architectural changes. This gives practical guidance for low-resource and zero-shot style transfer.

Key points

  • Data:
    • Training: 35,122 samples from 32,784 authors, most of whom contribute a single text.
    • Validation: 4,157 samples.
    • Test: 8,413 samples from 21 famous authors.
    • Author overlap between train and test is 0%.
  • Architecture:
    • Style encoder: AraBERT-base [CLS] output passed through a two-layer MLP, giving a 256-dimensional L2-normalized style embedding.
    • Generator: AraT5, with the style embedding added through a shared projection to all encoder layers.
  • Training objective:
    • Generation loss, plus a SimCSE-style contrastive loss with dropout-based positives (τ = 0.07), plus 0.1 × an auxiliary classification loss.
    • Trained on a single T4 GPU with an effective batch size of 32 for 3 epochs.
  • Results on the private test set (BLEU / chrF):
    • Contrastive system: 19.77 / 55.74.
    • Retrieval-augmented baseline (FAISS, k = 3): 16.68 / 49.63.
    • Multi-task baseline: 15.07 / 47.45.
  • Style injection variants underperform badly:
    • Cross-attention injection scores 9.63 BLEU.
    • Prompting with the author's name scores 6.54 BLEU.
    • Evaluation used only automatic metrics, with no human or stylometric assessment.

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 · Natural language processing

Topics: Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.abjadnlp-1.52

@inproceedings{sy2026hcmus,
  title = {{HCMUS\_The Fangs at AbjadStyleTransfer Shared Task: Learning to Query Style, Contrastive Representations for Zero-Shot Arabic Authorship Style Transfer}},
  author = {Sy, Duy Minh Dao and Huynh, Trung Kiet and Tran, Nguyen Chi and Quy, Nguyen Lam Phu and Hoa, Pham Phu and Duong, Nguyen Dinh Ha},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.52},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.52}
}

References

18 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 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 at EMNLP 2025, Suzhou, China. Association for Computational Linguistics. find
  2. Shadi Abudalfa, Saad Ezzini, Ahmed Abdelali, Mustafa Jarrar, Mo El-Haj, Nadir Durrani, Hassan Sajjad, and Farah Adeeba. 2026. Abjadstyletransfer: Authorship style transfer for arabic-script languages at abjadnlp 2026. In Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script (AbjadNLP 2026), co-located with the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), Rabat, Morocco. find
  3. Alec Radford, Jong Wook Kim, Chris Hallacy, et al. 2021. Learning transferable visual models from natural language supervision. In ICML. ↗ our card
  4. Yujin Shao, Yue Zhang, and Ruifeng He. 2024. Authorship style transfer with inverse transfer learning. In ACL. find
  5. Efstathios Stamatatos. 2009. A survey of modern authorship attribution methods. Journal of the American Society for Information Science and Technology, 60(3):538–556. ↗
  6. Yixuan Su, Tian Lan, Yan Wang, Dani Yogatama, Lingpeng Kong, and Nigel Collier. 2022. A contrastive framework for neural text generation. In NeurIPS. find
  7. Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. In NeurIPS. ↗ our card
  8. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: Transformer-based model for arabic language understanding. In LREC Workshop on Open-Source Arabic Corpora and Processing Tools. ↗ our card
  9. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In ICML. ↗ our card
  10. Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In EMNLP. find our card
  11. Beliz Gunel, Jingfei Du, Alexis Conneau, and Ves Stoyanov. 2021. Supervised contrastive learning for pre-trained language model fine-tuning. In ICLR. ↗ our card
  12. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In ICLR. ↗
  13. Zachary Horvitz, Ajay Agarwal, Yufei Xie, Vidhisha Balachandran, and Yulia Tsvetkov. 2024. TinyStyler: Efficient few-shot text style transfer with authorship embeddings. In EMNLP. find
  14. Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In ICML. ↗ our card
  15. Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data. ↗ our card
  16. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR. find
  17. El Moatez Billah Nagoudi, Abdul Rahman Elmadany, and Muhammad Abdul-Mageed. 2022. AraT5: Text-to-text transformers for arabic language generation. In ACL. find
  18. Ajay Patel, Sudhanshu Bhattamishra, Piyush Mitra, and Manish Gupta. 2022. Low-resource authorship style transfer with in-context learning. In EMNLP. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/sy-2026-hcmus-fangs-abjadstyletransfer-shared-task/