Elamur Research⌕
☰
home / papers / 2025 / ARWI: Arabic Write and Improve

ARWI: Arabic Write and Improve

Paper · 2025 · Kirill Chirkunov, Bashar Alhafni, Chatrine Qwaider et al.

Summary

Arabic has hundreds of millions of speakers, yet tools that help people write better Modern Standard Arabic (MSA) are scarce. Existing tools mostly fix frequent spelling slips and do not assess overall writing quality. ARWI is a free web-based writing assistant built to close this gap. It combines several parts:

  • a bank of essay prompts graded by CEFR level;
  • an Arabic editor that supports diacritics;
  • grammatical error detection and correction (GED/C) built on CAMeLBERT-MSA and trained on QALB-2014 and ZAEBUC;
  • an automated essay scoring (AES) module that estimates CEFR levels A1–C2;
  • progress tracking and optional user profiles (native language, dialect, self-rated level).

Each submitted essay is automatically annotated, so the system also builds a growing Arabic learner corpus for future GED/C and AES research. A small pilot with non-native university learners found the interface moderately well received. Most users who revised their essays reduced their error counts, though no CEFR level changed within the short sessions. The work matters because Arabic GED/C and AES resources lag far behind English, and a deployed tool can help close that data gap.

Key points

  • GED works as a cascade of two token classifiers built on CAMeLBERT-MSA. The first makes a binary erroneous/correct decision, and the second assigns ARETA error categories. Running them in sequence is meant to keep precision high.
  • GEC is a text-editing model rather than a text generator: it predicts character-level edit operations for each token. Both GED/C models were fine-tuned on QALB-2014 plus ZAEBUC.
  • AES fine-tunes CAMeLBERT-MSA on ZAEBUC combined with a larger synthetic essay set that varies topic, genre and level (Qwaider et al., 2025).
  • Prompts follow exam formats such as CIMA and ALPT and set minimum lengths: 50 words for A1–A2, 100 for B1–B2, and 200+ for C1–C2.
  • The pilot study details:
    • 34 non-native undergraduates at levels A1–B1 produced 112 checked submissions.
    • 67 iterative submissions from 12 users were analysed, and 8 of those users reduced their errors.
    • Average Likert ratings ranged from 3.03 (visual appeal) to 3.71, and about 85% of users would recommend the tool.

Interesting ideas

  • I1. First decide whether something is wrong at all, and only then classify what kind of error it is; gating the fine-grained step behind a binary detector trades recall for trustworthy flags.
  • I2. Correction is framed as predicting small character edit operations per token, so every change can be traced to a specific operation rather than to free-form generated text.
  • I3. A small gold set is enlarged with synthetic data deliberately spread across topics, genres and levels, so the model covers the space rather than overfitting to a few essay types.
  • I4. The deployed tool doubles as a corpus builder: every use produces auto-annotated text tagged with user metadata, which later allows profiling by subgroup.
  • I5. The evaluation admits that a coarse score (CEFR level) cannot move within a short session, so progress is measured with a finer signal, namely error counts across revisions.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tree): A binary detector could first flag suspicious attachments, and only flagged arcs would get a typed label (wrong head, wrong deprel). This keeps the error reports explainable and lowers false alarms.
  • I2 (tokens): Lemmas and spelling normalisation could be expressed as per-token edit scripts. Each lemma would then carry an inspectable transformation instead of an opaque lookup.
  • I3 (learn): When building silver data, generate examples to fill explicit stratification gaps (genre, sentence type, construction). Coverage gaps then become visible and fixable.
  • I4 (learn): Logging each processed text with its source metadata and pipeline outputs would allow error profiles by genre or source. Those profiles could then feed repair-rule induction.
  • I5 (answer): Besides pass/fail on school tests, track finer-grained intermediate measures such as correct frames or entities per question. Small improvements would then show up before the final scores change.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓New York University Abu DhabiAbu Dhabi · United Arab Emirates124 works · 4,709 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSocial Sciences · 1,954 worksEducation · 212 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 3,066 works

Cite

Cite as Kirill Chirkunov, Bashar Alhafni, Chatrine Qwaider, Nizar Habash and Ted Briscoe. 2025. ARWI: Arabic Write and Improve. Proceedings of the Fourth Workshop on Intelligent and Interactive Writing Assistants (In2Writing 2025).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.in2writing-1.2

@inproceedings{chirkunov2025arwi,
  title = {{ARWI: Arabic Write and Improve}},
  author = {Kirill Chirkunov, Bashar Alhafni, Chatrine Qwaider, Nizar Habash and Ted Briscoe},
  booktitle = {Proceedings of the Fourth Workshop on Intelligent and Interactive Writing Assistants (In2Writing 2025)},
  year = {2025},
  doi = {10.18653/v1/2025.in2writing-1.2},
  url = {https://doi.org/10.18653/v1/2025.in2writing-1.2}
}

References

24 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. A. Alfaifi, E. Atwell, and G. Abuhakema. 2013. Error annotation of the Arabic learner corpus. In Language Processing and Knowledge in the Web, volume 8105 of Lecture Notes in Computer Science. Springer, Berlin, Heidelberg. find
  2. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The interplay of variant, size, and task type in Arabic pre-trained language models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 92– 104, Kyiv, Ukraine (Virtual). Association for Computational Linguistics.
  3. Abdullah Alfaifi and Eric Atwell. 2012. Arabic learner corpora (alc): A taxonomy of coding errors. In The 8th International Computing Conference in Arabic. ↗
  4. Bashar Alhafni and Nizar Habash. 2025. Enhancing text editing for grammatical error correction: Arabic as a case study. Preprint, arXiv:2503.00985. ↗
  5. Minsun Kim, SeonGyeom Kim, Suyoun Lee, Yoosang Yoon, Junho Myung, Haneul Yoo, Hyunseung Lim, Jieun Han, Yoonsu Kim, So-Yeon Ahn, Juho Kim, Alice Oh, Hwajung Hong, and Tak Yeon Lee. 2024. Designing prompt analytics dashboards to analyze student-chatgpt interactions in efl writing. Preprint, arXiv:2405.19691. ↗
  6. Bashar Alhafni, Go Inoue, Christian Khairallah, and Nizar Habash. 2023. Advancements in Arabic grammatical error detection and correction: An empirical investigation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6430–6448, Singapore. Association for Computational Linguistics.
  7. Riadh Belkebir and Nizar Habash. 2021. Automatic error type annotation for Arabic. In Proceedings of the 25th Conference on Computational Natural Language Learning, pages 596–606, Online. Association for Computational Linguistics. ↗
  8. Behrang Mohit, Alla Rozovskaya, Nizar Habash, Wajdi Zaghouani, and Ossama Obeid. 2014. The first QALB shared task on automatic text correction for Arabic. In Proceedings of the EMNLP 2014 Workshop on Arabic Natural Language Processing (ANLP), pages 39–47, Doha, Qatar. Association for Computational Linguistics.
  9. Christopher Bryant, Mariano Felice, Øistein E. Andersen, and Ted Briscoe. 2019. The BEA-2019 shared task on grammatical error correction. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 52–75, Courtney Napoles, Maria Nădejde, and Joel Tetreault. 2019. Enabling robust grammatical error correction in new domains: Data sets, metrics, and analyses. Transactions of the Association for Computational Linguistics, 7:551–566.
  10. Courtney Napoles, Keisuke Sakaguchi, and Joel Tetreault. 2017. Jfleg: A fluency corpus and benchmark for grammatical error correction. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 229–234, Valencia, Spain. Association for Computational Linguistics. multimodal generative ai platform. arXiv preprint arXiv:2501.13944. ↗
  11. United Nations. 2024. Arabic language and AI: Advancing innovation while preserving cultural heritage. Accessed: 2025-02-28. find
  12. Haoran Wu, Wenxuan Wang, Yuxuan Wan, Wenxiang Jiao, and Michael Lyu. 2023. Chatgpt or grammarly? evaluating chatgpt on grammatical error correction benchmark. Preprint, arXiv:2303.13648. ↗
  13. Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. 2014. The CoNLL-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, pages 1–14, Baltimore, Maryland. Association for Computational Linguistics. find
  14. K. Yancey, Geoffrey T. LaFlair, Anthony Verardi, and Jill Burstein. 2023. Rating short l2 essays on the cefr scale with gpt-4. In Workshop on Innovative Use of NLP for Building Educational Applications. find
  15. Diane Nicholls, Andrew Caines, and Paula Buttery. 2024. find
  16. The Write & Improve Corpus 2024: Error-annotated and CEFR-labelled essays by learners of English. Research Outputs. find
  17. Kostiantyn Omelianchuk, Andrii Liubonko, Oleksandr Skurzhanskyi, Artem Chernodub, Oleksandr Korni-ienko, and Igor Samokhin. 2024. Pillars of grammatical error correction: Comprehensive inspection of contemporary approaches in the era of large language models. Preprint, arXiv:2404.14914. ↗
  18. Austin Pack, Alex Barrett, and Juan Escalante. 2024. Large language models and automated essay scoring of English language learner writing: Insights into validity and reliability. Computers and Education: Artificial Intelligence, 6:100234. ↗
  19. Chatrine Qwaider, Bashar Alhafni, Kirill Chirkunov, Nizar Habash, and Ted Briscoe. 2025. Enhancing arabic automated essay scoring with synthetic data and error injection. Preprint, arXiv:2503.17739. ↗
  20. Alla Rozovskaya, Houda Bouamor, Nizar Habash, Wajdi Zaghouani, Ossama Obeid, and Behrang Mohit. 2015. The second QALB shared task on automatic text correction for Arabic. In Proceedings of the Second Workshop on Arabic Natural Language Processing, pages 26–35, Beijing, China. Association for Computational Linguistics.
  21. Karin Ryding and David Wilmsen, editors. 2021. The Cambridge Handbook of Arabic Linguistics. Cambridge Handbooks in Language and Linguistics. Cambridge University Press. find
  22. Lucía Sanz-Valdivieso. 2024. Technology-powered multilingual professional and technical writing: An integrative literature review of landmark and the latest writing assistance tools. IEEE Transactions on Professional Communication, 67(3):301–315. ↗
  23. Neha Sengupta, Sunil Kumar Sahu, Bokang Jia, and et al. 2023. Jais and jais-chat: Arabic-centric foundation and instruction-tuned open generative large language models. Preprint, arXiv:2308.16149. ↗
  24. Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad, and et al. 2025. Fanar: An arabic-centric find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/chirkunov-2025-arwi-arabic-write-improve/