Elamur Research⌕
☰
home / papers / 2025 / ArbESC+: Arabic Enhanced Edit Selection System Combination for Grammatical Error Correction Resolving conflict and improving system combination in Arabic GEC

ArbESC+: Arabic Enhanced Edit Selection System Combination for Grammatical Error Correction Resolving conflict and improving system combination in Arabic GEC

Summary

Arabic grammatical error correction (GEC) is difficult because of rich inflection, inconsistent spelling conventions, complex syntax and scarce annotated data. Earlier Arabic work mostly relied on single models or simple voting. The authors present ArbESC+, which they describe as the first system-combination framework for Arabic GEC. It pools corrections from nine systems. AraT5, ByT5, mT5 and AraBART were fine-tuned by the authors. Three AraBART+Morph+GEC variants and two text-editing variants were used off-the-shelf.

Every proposed micro-edit becomes a vector of numerical features. A classifier then decides, edit by edit, whether that edit is applied, which turns combination into binary classification. Extra steps filter overlapping or conflicting edits and estimate how reliable each decision is.

The combined system beats every individual model, reaching F0.5 of 82.63% on QALB-14, 84.64% on QALB-15 L1 and 65.55% on QALB-15 L2. An ARETA error-type analysis shows sharp drops in overcorrection. The approach is presented as transferable to other low-resource, morphologically complex languages.

Key points

  • Nine base systems are combined:

    • AraT5, ByT5, mT5 and AraBART, fine-tuned for this work;
    • AraBART+Morph+GEC trained on QALB-14, QALB-15 and ZAEBUC;
    • text-editing models trained on QALB-14 and ZAEBUC.

  • Combination works at the micro-edit level: each candidate edit is featurized and receives an independent accept/reject decision from a classifier, followed by conflict filtering and a reliability estimate.

  • Reported F0.5 is 82.63% (QALB-14), 84.64% (QALB-15 L1) and 65.55% (QALB-15 L2). The previous voting ensemble of Alhafni & Habash scored 81.7% on QALB-2014 and 82.9% on QALB-2015.

  • In the ARETA per-tag comparison, ArbESC+ nearly matches the best single model on surface tags:

    • spelling: OA, OG, OW;
    • merge/split: MG, SP, XT.

    It lags on semantic and syntactic tags such as XG, MT and OR, most visibly on L2 learner data.

  • Overcorrection-prone tags are cut sharply. PT and XT fall to very low counts, and PC and XM vanish entirely.

Interesting ideas

  • I1. Rather than voting on whole output sentences, the method splits every system's output into atomic edits and lets a classifier accept or reject each one using features such as agreement among systems, which keeps the best fragments of each system. — For Mova (learn): Silver data built from several witness parsers could be decided arc by arc, with a small classifier over agreement features instead of a plain majority vote, and each arc decision stays explainable.
  • I2. Edits that overlap or contradict each other are resolved by an explicit filtering step before the final output is assembled, so incompatible choices from different systems never mix. — For Mova (tree): When arcs from different parsers are merged, a dedicated conflict-resolution pass could enforce tree well-formedness (a single head per token, no cycles) before the silver tree is accepted.
  • I3. The ensemble deliberately mixes system types with different biases: full-sentence seq2seq generators, byte-level ByT5, and edit-tagging models. Diversity of errors is what makes the combination pay off. — For Mova (learn): Witnesses should be chosen for differing error profiles, for example pairing the perceptron parser with gparse, rather than as several near-copies of one model.
  • I4. Overcorrection is measured directly, by counting errors of change-prone tags before and after correction, instead of relying only on aggregate F0.5. — For Mova (learn): Evaluation suites could include an "unnecessary change" counter as a negative control, for instance how often repair rules alter tokens that were already correct.
  • I5. Results are broken down by linguistic error category and by data source (native vs. learner), which shows where combination stops helping, namely deep semantic and syntactic errors. — For Mova (answer): Per-category reporting of reader failures, such as surface-cue questions versus ones needing deep inference, would show which stage to strengthen next.

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

Topics: Semitic languages · Grammatical error correction & learner language

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2511.14230 · arXiv 2511.14230

@misc{alrehili2025arbesc,
  title = {{ArbESC+: Arabic Enhanced Edit Selection System Combination for Grammatical Error Correction Resolving conflict and improving system combination in Arabic GEC}},
  author = {Alrehili, Ahlam and Alhothali, Areej},
  year = {2025},
  doi = {10.48550/arxiv.2511.14230},
  eprint = {2511.14230},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2511.14230}
}

References

49 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mohamed Abdelrehim, Marwan Torki, and Nagwa El-Makky. Hybrid llm and rule-based synthetic data generation for arabic grammatical error correction. In 2025 International Conference on Machine Intelligence and Smart Innovation (ICMISI), pages 280–285. IEEE, 2025. find
  2. Ahmad Aiman, Rehab Duwairi, and Mahmoud Al-Ayyoub. Deep learning-based arabic grammatical error correction. International Journal of Arabic Linguistics, 2021. find
  3. Abdullah Alfaifi and Eric Atwell. Arabic learner corpora (alc): a taxonomy of coding errors. In The 8th International Computing Conference in Arabic, 2012. *Figure 9: Raw (gray) vs. ARBESC+ Fixed (blue) counts for frequent over-correction tags. PT/XT values are dataset averages; PC/XM use range midpoints (with error bars). PC and XM are not observed in the latest ARBESC+ run.* find
  4. Bashar Alhafni and Nizar Habash. Enhancing text editing for grammatical error correction: Arabic as a case study, 2025. URL https://arxiv.org/abs/2503.00985. ↗
  5. Bashar Alhafni and Nizar Habash. Enhancing text editing for grammatical error correction: Arabic as a case study. arXiv preprint arXiv:2503.00985, 2025. ↗
  6. Bashar Alhafni, Go Inoue, Christian Khairallah, and Nizar Habash. Advancements in arabic grammatical error detection and correction: An empirical investigation. arXiv preprint arXiv:2305.14734, 2023. ↗ our card
  7. Ahlam Alrehili and Areej Alhothali. Tibyan corpus: balanced and comprehensive error coverage corpus using chatgpt for arabic grammatical error correction. PeerJ Computer Science, 11:e2724, 2025. find our card
  8. Ahlam Alrehili and Areej Alhothali. Towards the development of balanced synthetic data for correcting grammatical errors in arabic: An approach based on error tagging model and synthetic data generating model. arXiv preprint arXiv:2502.05312, 2025. ↗ our card
  9. Ahlam Alrehili and Areej Alhothaly. Balanced synthetic data generation for correcting grammatical errors in arabic. arXiv preprint arXiv:2502.05312, 2025. ↗ our card
  10. Wissam Antoun, Fady Baly, and Hazem Hajj. Arabert: Transformer-based model for arabic language understanding. In LREC 2020 Workshop Language Resources and Evaluation Conference 11–16 May 2020, page 9, 2020. find our card
  11. Riadh Belkebir and Nizar Habash. Automatic error type annotation for arabic. arXiv preprint arXiv:2109.08068, 2021. ↗ our card
  12. Christopher Bryant, Zheng Yuan, Muhammad Reza Qorib, Hannan Cao, Hwee Tou Ng, and Ted Briscoe. Grammatical error correction: A survey of the state of the art. Computational Linguistics, 49(3):643–701, 2023. find
  13. CJ Bryant, Mariano Felice, and Edward Briscoe. Automatic annotation and evaluation of error types for grammatical error correction. In Association for Computational Linguistics, 2017. find our card
  14. Daniel Dahlmeier and Hwee Tou Ng. Better evaluation for grammatical error correction. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 568–572, 2012. find
  15. Nizar Habash and Owen Rambow. Arabic tokenization, part-of-speech tagging and morphological disambiguation in one fell swoop. In Proceedings of the 43rd annual meeting of the association for computational linguistics (ACL’05), pages 573–580, 2005. find our card
  16. Nizar Habash, Owen Rambow, and Ryan Roth. Mada+ tokan: A toolkit for arabic tokenization, diacritization, morphological disambiguation, pos tagging, stemming and lemmatization. In Proceedings of the 2nd international conference on Arabic language resources and tools (MEDAR), Cairo, Egypt, volume 41, page 62, 2009. find
  17. Bashar Hefny and Nizar Habash. Enhancing text editing for grammatical error correction: Arabic as a case study. In Proceedings of ACL 2025, 2025. find
  18. Svanhvít Lilja Ingólfsdóttir, Pétur Orri Ragnarsson, Haukur Páll Jónsson, Haukur Barri Símonarson, Vilhjálmur Þorsteinsson, and Vésteinn Snæbjarnarson. Byte-level grammatical error correction using synthetic and curated corpora. arXiv preprint arXiv:2305.17906, 2023. ↗
  19. Karim Ismail, Sherif Abdou, Mohamed Farouk, and Ahmed Salem. Transformers to the rescue: alleviating data scarcity in arabic grammatical error correction with pre-trained models. Neural Computing and Applications, 37 (18):13011–13038, 2025. find
  20. Shyamsundar Jayaraman and Alon Lavie. Multi-engine machine translation guided by explicit word matching. In Proceedings of the 10th EAMT Conference: Practical applications of machine translation, 2005. find
  21. Yoav Kantor, Yoav Katz, Leshem Choshen, Edo Cohen-Karlik, Naftali Liberman, Assaf Toledo, Amir Menczel, and Noam Slonim. Learning to combine grammatical error corrections. arXiv preprint arXiv:1906.03897, 2019. ↗
  22. Sang Yun Kwon, Gagan Bhatia, El Moatez Billah Nagoud, and Muhammad Abdul-Mageed. Chatgpt for arabic grammatical error correction. arXiv preprint arXiv:2308.04492, 2023. ↗
  23. Seunghyun Kwon, Young Park, and Joonho Lee. Prompting gpt-4 for low-resource grammatical correction: Arabic as a case study. In Findings of the Association for Computational Linguistics (ACL Findings), 2023. find
  24. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. ↗ our card
  25. Ruixi Lin and Hwee Tou Ng. System combination for grammatical error correction based on integer programming. arXiv preprint arXiv:2111.01465, 2021. ↗
  26. Wolfgang Macherey and Franz Josef Och. An empirical study on computing consensus translations from multiple machine translation systems. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 986–995, 2007. find
  27. Amr Mahmoud and Rawan Alghamdi. Error correction in arabic texts using hybrid deep learning. IEEE Access, 2024. find
  28. Amr Mahmoud, Hadeer El-Shazly, and Amal Alotaibi. Fine-tuning t5 for arabic grammatical error correction. Natural Language Engineering, 2024. find
  29. Behrang Mohit, Alla Rozovskaya, Nizar Habash, Wajdi Zaghouani, and Ossama Obeid. 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, 2014. find
  30. El Moatez Billah Nagoudi, AbdelRahim Elmadany, and Muhammad Abdul-Mageed. Arat5: Text-to-text transformers for arabic language generation. arXiv preprint arXiv:2109.12068, 2021. ↗ our card
  31. Nour Nawar. Automatic arabic grammar error correction. Master’s thesis, Cairo University, 2015. find
  32. Hwee Tou Ng, Siew Mei Wu, Yuanbin Wu, Christian Hadiwinoto, and Joel Tetreault. The conll-2013 shared task on grammatical error correction. arXiv preprint arXiv:2507.09474, 2025. ↗
  33. Piotr Pajak and Marta Pajak. Neural approaches to arabic grammar correction. Applied Computing Review, 2022. find
  34. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. find our card
  35. Muhammad Reza Qorib, Seung-Hoon Na, and Hwee Tou Ng. Frustratingly easy system combination for grammatical error correction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1964–1974, 2022. find
  36. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020. find
  37. Sascha Rothe, Jonathan Mallinson, Eric Malmi, Sebastian Krause, and Aliaksei Severyn. A simple recipe for multilingual grammatical error correction. arXiv preprint arXiv:2106.03830, 2021. ↗
  38. Alla Rozovskaya, Nizar Habash, and Dan Roth. The qalb-2014 shared task on automatic text correction for arabic. In Proceedings of EMNLP 2014 Workshop on Arabic Natural Language Processing, 2014. find
  39. Alla Rozovskaya, Houda Bouamor, Nizar Habash, Wajdi Zaghouani, Ossama Obeid, and Behrang Mohit. 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, 2015. find
  40. Ahmed Solyman, Mohamed Elsayed, and Mohamed Attia. Improving arabic grammatical error correction using hybrid neural models. Journal of King Saud University–Computer and Information Sciences, 2022. find
  41. Ahmed Solyman, Hossam Rashad, and Ahmed Elgohary. Arat5 for arabic grammatical error correction. In Proceedings of the International Conference on Language Resources and Evaluation (LREC), 2023. find
  42. Aiman Solyman, Wang Zhenyu, Tao Qian, Arafat Abdulgader Mohammed Elhag, Muhammad Toseef, and Zeinab Aleibeid. Synthetic data with neural machine translation for automatic correction in arabic grammar. Egyptian Informatics Journal, 22(3):303–315, 2021. find
  43. Aiman Solyman, Marco Zappatore, Wang Zhenyu, Zeinab Mahmoud, Ali Alfatemi, Ashraf Osman Ibrahim, and Lubna Abdelkareim Gabralla. Optimizing the impact of data augmentation for low-resource grammatical error correction. Journal of King Saud University-Computer and Information Sciences, 35(6):101572, 2023. find
  44. Felix Stahlberg and Shankar Kumar. Seq2edits: Sequence transduction using span-level edit operations. arXiv preprint arXiv:2009.11136, 2020. ↗
  45. Raymond Hendy Susanto, Peter Phandi, and Hwee Tou Ng. System combination for grammatical error correction. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 951–962, 2014. find
  46. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. find
  47. L Xue. mt5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934, 2020. ↗
  48. Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291–306, 2022. find
  49. Yukun Zhu. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. arXiv preprint arXiv:1506.06724, 2015. ↗ our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alrehili-2025-arbesc-arabic-enhanced-edit-selection/