---
title: "REGLAT at AbjadGenEval — paper by Labib et al. (2026)"
description: "The authors tackle the AbjadGenEval shared task, a binary classification problem that asks whether an Arabic text… 0.624 recall. 9 references."
authors: ["Mariam Labib", "Nsrin Ashraf", "Ahmed M. Fetouh", "Hamada Nayel"]
year: 2026
venue: "Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script"
source: "https://aclanthology.org/2026.abjadnlp-1.62/"
doi: "10.18653/v1/2026.abjadnlp-1.62"
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/labib-2026-reglat-abjadgeneval-multi-model-ensemble/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/labib-2026-reglat-abjadgeneval-multi-model-ensemble/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/) / REGLAT at AbjadGenEval: Multi-Model Ensemble Approach for Arabic AI-Generated Text Detection

# REGLAT at AbjadGenEval: Multi-Model Ensemble Approach for Arabic AI-Generated Text Detection

**Authors:** [Mariam Labib](https://md.research.elamur.ai/free/en/authors/mariam-labib/), [Nsrin Ashraf](https://md.research.elamur.ai/free/en/authors/nsrin-ashraf/), [Ahmed M. Fetouh](https://md.research.elamur.ai/free/en/authors/ahmed-m-fetouh/) and [Hamada Nayel](https://md.research.elamur.ai/free/en/authors/hamada-nayel/) · **Year:** 2026 · **Venue:** [Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script](https://md.research.elamur.ai/free/en/venues/abjadnlp/) ([Workshop on NLP for Languages Using Arabic Script 2026](https://md.research.elamur.ai/free/en/venues/abjadnlp/2026/)) · **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/2026.abjadnlp-1.62/) · [PDF](https://aclanthology.org/2026.abjadnlp-1.62.pdf) · [DOI 10.18653/v1/2026.abjadnlp-1.62](https://doi.org/10.18653/v1/2026.abjadnlp-1.62) · [OpenAlex](https://openalex.org/W7140099281)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The authors tackle the AbjadGenEval shared task, a binary classification problem that asks whether an Arabic text was written by a person or produced by a language model. Their system averages the predictions of two fine-tuned Arabic encoders, AraBERTv2 and BERT-base-arabic. Each encoder is trained with 5-fold stratified cross-validation and a class-weighted cross-entropy loss.

Before training, the text goes through orthographic cleanup: diacritics are stripped, Alef, Taa Marbuta and Alef Maksura variants are unified, and Tatweel is removed. The training data is also extended with the KFUPM-JRCAI Arabic Generated Abstracts corpus. The decision threshold is then tuned on pooled validation predictions.

On the 200-item official test set, the system reaches F1 0.763 with very high recall (0.980) and only moderate precision (0.624). In other words, it rarely misses machine text but often flags human text as machine-written. Cross-validation scores were far higher, so the paper's main practical lesson is the size of the gap between in-domain validation and the real test distribution in Arabic AI-text detection.

## Key points
- The ensemble is P = 0.6·AraBERTv2 + 0.4·BERT-base-arabic (aubmindlab/bert-base-arabertv2, asafaya/bert-base-arabic). Oddly, the weaker model gets the larger weight, and the authors themselves concede the weighting was probably suboptimal.
- Cross-validation F1 was 0.978 ± 0.005 for BERT-base-arabic (precision 0.957) and 0.904 for AraBERTv2 (precision 0.826). Both models had recall of at least 0.998.
- Test-set results: precision 0.624, recall 0.980, F1 0.763, accuracy 0.695. That corresponds to roughly a 38% false-positive rate on human texts.
- Augmentation brought the training pool to 16,678 samples. The added data covers three generation scenarios (polishing a human abstract, generating from a title, generating from title plus content), with machine text produced by Allam and GPT-4.
- The threshold search covered 0.35–0.70 in steps of 0.01 and selected 0.69. Class weights were 1.478 for human and 0.756 for machine. These numbers seem to contradict the paper's own claim that the machine class was weighted more heavily.

## Interesting ideas
- Weight ensemble members by their measured validation quality, and then check that the final weights actually reflect those measurements; here a mismatch between the two likely cost precision.
- Tune the decision threshold on pooled out-of-fold predictions, but treat an optimum at the edge of the search range (0.69 out of 0.35–0.70) as a warning that the range is too narrow or the objective is biased.
- Compare in-domain cross-validation against a distribution-shifted test set; a drop from about 0.97 to 0.76 F1 shows that CV alone can badly overstate readiness.
- Use external data that pairs the same content under graded generation conditions (polish, title-only, title plus content), so the model learns a property rather than the quirks of one source.
- Normalize orthographic variants that writers apply inconsistently before modelling, which reduces sparsity without changing meaning.

## For Mova

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

- **I1** (learn): When several witness parsers vote on silver trees, derive each witness's vote weight from its dev-set LAS, and add an automatic check that flags weights that contradict the measured ranking.
- **I2** (frames): Coreference-link and implicature confidence cutoffs could be chosen by a sweep over held-out data, with a rule that widens the search whenever the best value lands on a boundary.
- **I3** (answer): Score the readers on fables, tests and word problems drawn from sources not used in development, and report that number alongside the in-domain score as the honest measure.
- **I4** (back): Graded paired data could serve as negative controls for the text → graph → text compressor, testing whether the recovered graph stays stable when the surface varies but the content does not.
- **I5** (tokens): Fold quote and dash variants, ligatures and spelling variants into a canonical form for the tagger, while keeping the original surface token attached so every step remains explainable.

*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

- [Prince Sattam Bin Abdulaziz University](https://md.research.elamur.ai/free/en/universities/prince-sattam-bin-abdulaziz-university/) — University, Al Kharj, Saudi Arabia · 17 works · 489 citations
- [Benha University](https://md.research.elamur.ai/free/en/universities/benha-university/) — University, Banhā, Egypt · 14 works · 89 citations
- [ElSewedy University of Technology](https://md.research.elamur.ai/free/en/universities/elsewedy-university-of-technology/) — University, Cairo, Egypt · 5 works · 5 citations

📚 Shelves (5): [🖥️ 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 · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 2,869 works

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

## Cite

Cite as: Mariam Labib, Nsrin Ashraf, Ahmed M. Fetouh and Hamada Nayel. 2026. REGLAT at AbjadGenEval: Multi-Model Ensemble Approach for Arabic AI-Generated Text Detection. Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script.

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

```bibtex
@inproceedings{labib2026reglat,
  title = {{REGLAT at AbjadGenEval: Multi-Model Ensemble Approach for Arabic AI-Generated Text Detection}},
  author = {Mariam Labib, Nsrin Ashraf, Ahmed M. Fetouh and Hamada Nayel},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.62},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.62}
}
```

- [← previous Regional Variation in the Performance of ASR Models on Croatian and Serbian](https://md.research.elamur.ai/free/en/papers/samardzic-2026-regional-variation-performance-asr-models/)
- [next → REGLAT at AbjadMed: Handling Imbalanced Arabic Medical Text Classification via Hierarchical KNN-MLP Architecture](https://md.research.elamur.ai/free/en/papers/fetouh-2026-reglat-abjadmed-handling-imbalanced-arabic/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/)

## References

9 references, 3 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](https://md.research.elamur.ai/free/en/papers/abudalfa-2025-arageneval-shared-task-arabic-authorship/). In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 1– 13, Suzhou, China. Association for Computational Linguistics.
2. Souha Al Katat, Chamseddine Zaki, Hussein Hazimeh, Ibrahim El Bitar, Rafael Angarita, and Lionel Trojman. 2024. Natural language processing for arabic sentiment analysis: A systematic literature review. IEEE Transactions on Big Data, 10(5):576– 594. [↗](https://doi.org/10.1109/tbdata.2024.3366083)
3. Abdulaziz M. Alayba. 2025. Arabic natural language processing (nlp): A comprehensive review of challenges, techniques, and emerging trends. Computers, 14(11). [↗](https://doi.org/10.3390/computers14110497)
4. Adeeba. 2026. [Abjadgeneval: Abjad ai generated text detection shared task for languages using arabic script at abjadnlp 2026](https://md.research.elamur.ai/free/en/papers/ezzini-2026-abjadgeneval-abjad-ai-generated-text/). 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.
5. Ola Karajeh, Mohammed N. Al-Kabi, and Edward A. Fox. 2023. Fusing arabert and graph neural networks for enhanced arabic text classification. In 2023 24th International Arab Conference on Information Technology (ACIT), pages 1–8. [find](https://scholar.google.com/scholar?q=Ola+Karajeh%2C+Mohammed+N%2E+Al%2DKabi%2C+and+Edward+A%2E+Fox%2E+2023%2E+Fusing+arabert+and+graph+neural+networks+for+enhanced+arabic+text+classification%2E+In+2023+24th+International+Arab+Conference+on+Information+T)
6. Mariam Labib, Nsrin Ashraf, Mohammed Aldawsari, and Hamada Nayel. 2025. [REGLAT at AraGenEval shared task: Morphology-aware AraBERT for detecting Arabic AI-generated text](https://md.research.elamur.ai/free/en/papers/labib-2025-reglat-arageneval-shared-task-morphology/). In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 94–98, Suzhou, China. Association for Computational Linguistics.
7. Yu Wang and Wen Qu. 2025. A tutorial on fine-tuning pretrained language models: Applications in social and behavioral science research. Behavior Research Methods, 57(12):336. [↗](https://doi.org/10.3758/s13428-025-02868-z)
8. Saad Yaquine, Amine Hmimou, and Paolo Rosso. 2025. Magenta: Generating and detecting arabic machine-generated text in multiple domains. In Arabic Language Processing: From Theory to Practice, volume 2339, pages 151–159, Cham. Springer Nature Switzerland. [find](https://scholar.google.com/scholar?q=Saad+Yaquine%2C+Amine+Hmimou%2C+and+Paolo+Rosso%2E+2025%2E+Magenta%3A+Generating+and+detecting+arabic+machine%2Dgenerated+text+in+multiple+domains%2E+In+Arabic+Language+Processing%3A+From+Theory+to+Practice%2C+volume+2)
9. Hongzhi Zhang and M. Omair Shafiq. 2024. Survey of transformers and towards ensemble learning using transformers for natural language processing. Journal of Big Data, 11(1):25. [↗](https://doi.org/10.1186/s40537-023-00842-0)

*Updated 1 November 2026*