---
title: "ZAI at BAREC Shared Task 2025 — paper by Ahmad (2025)"
description: "The paper describes the ZAI entry to the BAREC 2025 shared task. 12 references, cited by 1 paper here."
authors: ["M. Nazzal, Ahmad"]
year: 2025
venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks"
source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.37/"
doi: "10.18653/v1/2025.arabicnlp-sharedtasks.37"
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/nazzal-2025-zai-barec-shared-task-2025/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/nazzal-2025-zai-barec-shared-task-2025/)

[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/page/16/) / ZAI at BAREC Shared Task 2025: AraBERT CORAL for Fine Grained Arabic Readability

# ZAI at BAREC Shared Task 2025: AraBERT CORAL for Fine Grained Arabic Readability

**Authors:** M. Nazzal, Ahmad · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.37/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.37.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.37](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.37) · [OpenAlex](https://openalex.org/W4416037814)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/nazzal-2025-zai-barec-shared-task-2025/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper describes the ZAI entry to the BAREC 2025 shared task. The task is to grade the readability of single Arabic sentences on a 19-level ordered scale, and the author competed in the Sentence-level (Strict) track. Arabic is hard here because of its rich morphology and the diglossia between Modern Standard Arabic and dialects. The system is deliberately small. It fine-tunes AraBERTv2 (aubmindlab/bert-base-arabertv2) on text cleaned with the AraBERTPreprocessor and replaces the usual softmax classifier with a CORAL ordinal head. That head predicts 18 cumulative "harder than level k" probabilities. Checkpoints are chosen by Quadratic-Weighted Kappa (QWK), the official metric, with early stopping. At inference, a single threshold tuned on dev turns the probabilities into a level. Training used only the official BAREC splits, with no external data and no ensembling. The model reached QWK 85.5 on the Blind Test. Swapping CORAL for a nominal softmax head lowered QWK. The main lesson is that respecting label order and tuning every choice to the official metric are cheap but strong design decisions.

## Key points
- Setup: AraBERTv2 with a CORAL head (K = 19, so 18 binary thresholds), trained with summed BCEWithLogits over the cumulative targets. Inputs were capped at 128 tokens with no dynamic padding.
- Training settings:
  - AdamW with lr 2×10⁻⁵ and weight decay 0.01, plus linear decay with warmup and grad-norm clipping at 1.0.
  - Batch size 16 with gradient accumulation of 2, FP16 on a T4 GPU.
  - Early stopping with patience 2; the best checkpoint was chosen by eval_qwk.
- QWK scores:

  | Split | QWK |
  |---|---|
  | Open Dev | 66.6 |
  | Open Test | 72.9 |
  | Blind Test | 85.5 |

- Other metrics:
  - Exact accuracy stayed modest at 35–47%. Within-one-level accuracy was much higher at 56–74%.
  - Collapsed-scale accuracies (7, 5 and 3 levels) were also reported.
  - The distance score was 1.6 for Dev and decreased to 1.0 for Blind.
  - On Blind, exact accuracy was the lowest of the three splits even though QWK was the highest.
- Ablations that hurt dev QWK:
  - a nominal softmax head instead of CORAL;
  - label smoothing;
  - temperature scaling.

  The final system uses none of them, and no error analysis was carried out.

## Interesting ideas
- Model ordered labels as a chain of cumulative "greater than k" binary decisions instead of independent classes, so the model cannot ignore that neighbouring levels are close.
- Decode a level by counting how many cumulative probabilities pass one shared threshold, and tune that single threshold on dev. This gives a transparent decision rule with only one knob.
- Choose checkpoints and post-processing by the official metric (QWK) itself, not by loss or plain accuracy, since the paper treats this alignment as a high-leverage choice.
- Report several evaluation granularities side by side: exact accuracy, accuracy within one level, accuracy after collapsing to 7, 5 and 3 levels, and a distance score. Together they show whether errors are near misses or wild guesses.
- Treat seemingly harmless tricks such as label smoothing and temperature scaling as hypotheses, and keep them only if they improve the target metric on dev; here both made results worse.

## For Mova

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

- **I1** (answer): Grading tasks with ordered outcomes, such as text difficulty, answer partial-credit levels or confidence bands, could use a CORAL-style head so that a wrong prediction tends to land near the true level.
- **I2** (learn): Any graded output Mova produces could be decoded this way, which keeps the step explainable: the system can state how many level boundaries the item crossed.
- **I3** (learn): When training gparse or the perceptron parser, early stopping and model selection could be driven directly by LAS or by the downstream frame or answer score that we actually report.
- **I4** (learn): Evaluations of answer readers or attachment decisions could add similar "near-miss" tolerance views to separate small slips from real breakdowns.
- **I5** (learn): Regularization and calibration defaults in our training pipelines should be run as explicit ablations against a baseline, like the negative controls we already use, rather than assumed to help.

*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

- [Zayed University](https://md.research.elamur.ai/free/en/universities/zayed-university/) — University, Abu Dhabi, United Arab Emirates · 6 works · 67 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works

## Cite

Cite as: M. Nazzal, Ahmad. 2025. ZAI at BAREC Shared Task 2025: AraBERT CORAL for Fine Grained Arabic Readability. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{ahmad2025zai,
  title = {{ZAI at BAREC Shared Task 2025: AraBERT CORAL for Fine Grained Arabic Readability}},
  author = {M. Nazzal, Ahmad},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.37},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.37}
}
```

- [← previous YassirEA at MAHED 2025: Fusion-Based Multimodal Models for Arabic Hate Meme Detection](https://md.research.elamur.ai/free/en/papers/attar-2025-yassirea-mahed-2025-fusion-based/)
- [next → «Are you Afraid of Ghosts?» A Proposal for Busting Predicate Ellipsis in Universal Dependencies](https://md.research.elamur.ai/free/en/papers/corbetta-2025-guillemotleftare-you-afraid-ghosts-guillemotright/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/16/)

## References

12 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Annalena Aicher, Alisa Gazizullina, Aleksei Gusev, Yuri Matveev, and Wolfgang Minker. 2022. Towards speech-only opinion-level sentiment analysis. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2000–2006, Marseille, France. European Language Resources Association. [find](https://scholar.google.com/scholar?q=Annalena+Aicher%2C+Alisa+Gazizullina%2C+Aleksei+Gusev%2C+Yuri+Matveev%2C+and+Wolfgang+Minker%2E+2022%2E+Towards+speech%2Donly+opinion%2Dlevel+sentiment+analysis%2E+In+Proceedings+of+the+Thirteenth+Language+Resources+an)
2. Nizar Habash, Hanada Taha-Thomure, Khalid N. Elmadani, Zeina Zeino, and Abdallah Abushmaes. 2025. [Guidelines for fine-grained sentence-level Arabic readability annotation](https://md.research.elamur.ai/free/en/papers/habash-2025-guidelines-fine-grained-sentence-level/). In Proceedings of the 19th Linguistic Annotation Workshop (LAW-XIX-2025), pages 359–376, Vienna, Austria. Association for Computational Linguistics.
3. Muhamed Al Khalil, Nizar Habash, and Zhengyang Jiang. 2020. [A large-scale leveled readability lexicon for Standard Arabic](https://md.research.elamur.ai/free/en/papers/khalil-2020-large-scale-leveled-readability-lexicon/). In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 3053–3062, Marseille, France. European Language Resources Association.
4. Reem Hazim, Hind Saddiki, Bashar Alhafni, Muhamed Al Khalil, and Nizar Habash. 2022. [Arabic word-level readability visualization for assisted text simplification](https://md.research.elamur.ai/free/en/papers/hazim-2022-arabic-word-level-readability-visualization/). In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 242–249, Abu Dhabi, UAE. Association for Computational Linguistics.
5. Bashar Alhafni, Reem Hazim, Juan David Pineros Liberato, Muhamed Al Khalil, and Nizar Habash. 2024. [The SAMER Arabic text simplification corpus](https://md.research.elamur.ai/free/en/papers/alhafni-2024-samer-arabic-text-simplification-corpus/). In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 16079–16093, Torino, Italia. ELRA and ICCL.
6. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. [AraBERT: Transformer-based model for Arabic language understanding](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/). In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 9–15, Marseille, France. European Language Resource Association. [↗](https://doi.org/10.48550/arxiv.2003.00104)
7. Ibrahim A. Asadi and Salim Abu-Rabia. 2019. The impact of the position of phonemes and lexical status on phonological awareness in the diglossic arabic language. Journal of Psycholinguistic Research, 48(5):1051–1062. [↗](https://doi.org/10.1007/s10936-019-09646-x)
8. Juan Liberato, Bashar Alhafni, Muhamed Khalil, and Nizar Habash. 2024. [Strategies for Arabic readability modeling](https://md.research.elamur.ai/free/en/papers/liberato-2024-strategies-arabic-readability-modeling/). In Proceedings of the Second Arabic Natural Language Processing Conference, pages 55–66, Bangkok, Thailand. Association for Computational Linguistics.
9. Elinor Saiegh-Haddad and Ola Ghawi-Dakwar. 2017. Impact of diglossia on word and non-word repetition among language impaired and typically developing arabic native speaking children. Frontiers in Psychology, 8:2010. [↗](https://doi.org/10.3389/fpsyg.2017.02010)
10. Haitham Taha and Elinor Saiegh-Haddad. 2016. The role of phonological versus morphological skills in the development of arabic spelling: An intervention study. Journal of Psycholinguistic Research, 45(3):507–535. [↗](https://doi.org/10.1007/s10936-015-9362-6)
11. Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. 2020. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140:325–331. [↗](https://doi.org/10.1016/j.patrec.2020.11.008)
12. Edgar Dale and Jeanne S. Chall. 1948. A formula for predicting readability. Educational Research Bulletin, 27(1):11–28. [find](https://scholar.google.com/scholar?q=Edgar+Dale+and+Jeanne+S%2E+Chall%2E+1948%2E+A+formula+for+predicting+readability%2E+Educational+Research+Bulletin%2C+27%281%29%3A11%E2%80%9328%2E)

*Updated 1 November 2026*