Elamur Research⌕
☰
home / papers / 2026 / SyriSign: A Parallel Corpus for Arabic Text to Syrian Arabic Sign Language Translation

SyriSign: A Parallel Corpus for Arabic Text to Syrian Arabic Sign Language Translation

Summary

Deaf and hard-of-hearing people in Syria have little access to local news, which mostly arrives as spoken or written Arabic. Only about ten sworn interpreters are registered nationwide, and before this work no public dataset existed for Syrian Arabic Sign Language (SyArSL). The authors built SyriSign, a word-level parallel resource that pairs Arabic words with recorded signs. They chose the vocabulary by counting how often words appeared in a regional newspaper over six months, so it fits news and public announcements. Seven signers with varied ages and hearing status recorded the clips, and MediaPipe Holistic turned the videos into body, face and hand landmarks. Three text-to-sign approaches were fine-tuned on the data: MotionCLIP (a hybrid that places motion in CLIP space), T2M-GPT (generative, based on discrete motion tokens) and SignCLIP (retrieval through contrastive text–pose embeddings). Generative methods looked promising, but the small corpus limited how well any model generalized. Evaluation used only proxy metrics. The main value is a first open SyArSL benchmark and baseline.

Key points

  • SyriSign contains 1,500 RGB clips spanning 150 distinct lexical items. Two signers performed every item, repeating it five times apiece, which yields roughly three hours of video at 60 FPS, with no clip longer than seven seconds.
  • There are seven signers (three male, four female) with differing hearing abilities, and one of them is a sworn translator. The data is released on Hugging Face and GitHub.
  • Preprocessing uses MediaPipe Holistic, which tracks 543 landmarks across pose, face and hands.
  • The adapted MotionCLIP feeds Arabic text through a translation layer into the CLIP text encoder and takes truncated MediaPipe motion as input. Its loss combines reconstruction, velocity smoothness and latent regularization terms.
  • Stated limitations:
    • the scope is word-level only;
    • non-manual features such as facial expressions are weakly covered;
    • there is no human evaluation or standardized production metric.

Interesting ideas

  • I1. Choosing the lexicon by word frequency in a target-domain corpus focuses a small annotation budget on the items that actually occur. — For Mova (learn): Rank annotation and silver-data targets by their frequency in the texts Mova must read, such as tales and school tests, rather than sampling words uniformly.
  • I2. A velocity smoothness term penalizes abrupt changes between consecutive frames, which builds temporal consistency directly into the training objective. — For Mova (world): A similar consistency check over successive event states could flag implausible jumps in entity states or holdings between adjacent clauses.
  • I3. A translation layer lets a low-resource input reuse a strong pretrained encoder built for another language instead of training one from scratch. — For Mova (back): Translation through the meta-world could rely on English-side resources as the shared pivot, with the bridge step kept as a separate, inspectable module.
  • I4. Retrieval, generative and hybrid methods are benchmarked side by side on identical data, which shows their trade-offs instead of crowning one paradigm. — For Mova (tree): Compare the perceptron parser and gparse on the same small treebank slices to see where each fails when data is scarce.
  • I5. The authors openly treat proxy metrics as weak and plan back-translation and human checks. — For Mova (back): Score the text → graph → text compressor by round-trip back-translation, so fidelity is measured against meaning rather than surface overlap alone.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2603.29219 · arXiv 2603.29219

@misc{khalil2026syrisign,
  title = {{SyriSign: A Parallel Corpus for Arabic Text to Syrian Arabic Sign Language Translation}},
  author = {Khalil, Mohammad Amer and Nahas, Raghad and Nassar, Ahmad and Jallad, Khloud Al},
  year = {2026},
  doi = {10.48550/arxiv.2603.29219},
  eprint = {2603.29219},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2603.29219}
}

References

15 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Akkazeh. (2025, January). مصلا عامس نع نوناقلا زجعي نيح :ةمجرت لاب قوقح . Retrieved January 18, 2026, from Akkazeh ةزاكع: https://akkazeh.net/3173/ عامس - نع - نوناقلا - زجعي - نيح - ةمجرت - لاب - قوقح Alethary, A. A., Aliwy, A. A., & Ali, N. S. (2022). Automated Arabic-Arabic sign language translation system based on 3D avatar technology. International Journal of Advances in Applied Sciences (IJAAS), 11, 383–396. Aulamo, M., Tiedemann, J., Bakshandaeva, D., Boggia, M., Grönroos, S.-A., Nieminen, T., . . . Virpioja, S. (2023). Democratizing Neural Machine Translation with OPUS-MT. Language Resources and Evaluation. ↗
  2. Camgöz, N. C., Hadfield, S., & Koller, O. (2018). Neural Sign Language Translation. IEEE Conf. on Computer Vision and Pattern Recognition. Salt Lake City, UT. find
  3. Grishchenko , I., & Bazarevsky, V. (2020, December 10). MediaPipe Holistic — Simultaneous Face, Hand and Pose Prediction, on Device. Retrieved from https://research.google/blog/mediapipe-holistic-simultaneous-face-hand-and-pose-prediction-on-device/ Humanitarian Needs Assessment Programme (HNAP). (2021). Disability in Syria: Investigation on the Intersectional Impacts of Gender, Age and a Decade of Conflict on Persons with Disabilities. Retrieved from https://www.hi-deutschland-projekte.de/lnob/wp-content/uploads/sites/2/2021/09/hnap-disability-in-syria-investigation-on-intersectional-impacts-2021.pdf ↗
  4. Jedlička, P., Krňoul, Z., Željezný, M., & Müller, L. (2022). MC-TRISLAN: A Large 3D Motion Capture Sign Language Dataset (Czech Sign Language). Proceedings of the Twelfth Language Resources and Evaluation Conference (LREC 2022) (pp. 88–95). Marseille, France: European Language Resources Association (ELRA). Retrieved from https://aclanthology.org/2022.signlang-1.14/ ↗
  5. Jiang, Z., Sennrich, R., Moryossef, A., Sant, G., Ebling, S., & M, M. (2024). SignCLIP: Connecting Text and Sign Language by Contrastive Learning. Zurich. Retrieved 7 Jan, 2026, from https://arxiv.org/pdf/2407.01264 ↗
  6. Koller, O., Zargaran, S., Ney, H., & Bowden, R. (2018). Deep Sign: Enabling Robust Statistical Continuous Sign Language Recognition via Hybrid CNN-HMMs. International Journal of Computer Vision, 126(12), 1311–1325. doi:10.1007/s11263-018-1121-3 ↗
  7. Latif, G., Mohammad, N., Alghazo, J., AlKhalaf, R., & AlKhalaf, R. (2019). ArASL: Arabic Alphabets Sign Language Dataset. Data in Brief, 23. doi:10.1016/j.dib.2019.103777 ↗
  8. Li, D., Rodriguez Opazo, C., Yu, X., & Li, H. (2020). Word-level Deep Sign Language Recognition from Video: A New Large-scale Dataset and Methods Comparison. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), (pp. 1459–1468). doi:10.1109/WACV45572.2020.9093305 Lugaresi, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., . . . Grundmann, M. (2019). MediaPipe: A Framework for Building Perception Pipelines. arXiv. ↗
  9. Luqman, H., Mohandes, M., Mahmoud, S., & I. Sidig, A. A. (2021). KArSL: Arabic Sign Language Database. ACM Transactions on Asian and Low-Resource Language Information Processing, 1 - 19. Retrieved 1 19, 2026, from https://doi.org/10.1145/3423420 ↗
  10. Nagoudi, E. M., Elmadany, A., & Abdul-Mageed, M. (2022). AraT5: Text-to-Text Transformers for Arabic Language Generation. ACL 2022. Retrieved from https://arxiv.org/abs/2109.12068 ↗ our card
  11. Saunders, B., Camgoz, N. C., & Bowden, R. (2020). Progressive Transformers for End-to-End Sign Language Production. ECCV 2020. Retrieved from https://arxiv.org/abs/2004.14874 ↗
  12. Tevet, G., Gordon, B., Hertz, A., Bermano, A. H., & Cohen-Or, D. (2022). MotionCLIP: Exposing Human Motion Generation to CLIP Space. find
  13. Vaezi Joze, H. R., & Koller, O. (2019). MS-ASL: A Large-Scale Data Set and Benchmark for Understanding American Sign Language. British Machine Vision Conference (BMVC). doi:10.48550/arXiv.1812.01053 ↗
  14. Zelinka, J., & Kanis, J. (2020). Neural Sign Language Synthesis: Words Are Our Glosses. Conference: 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), (pp. 3395–3403). doi:10.1109/WACV45572.2020.9093516 Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., . . . Shen, X. (2023). T2M-GPT: Generating Human Motion from Textual Descriptions with Discrete Representations. arXiv. CVPR 2023. ↗
  15. Zifan Jiang, G. S. (2024). SignCLIP: Connecting Text and Sign Language by Contrastive Learning. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/khalil-2026-syrisign-parallel-corpus-arabic-text/