Elamur Research⌕
☰
home / papers / 2026 / AraDetox: A Multi-Dialect Arabic Detoxification Dataset

AraDetox: A Multi-Dialect Arabic Detoxification Dataset

Paper · 2026 · Mo El-Haj

Summary

Most Arabic work on harmful language focuses on detecting it. Rewriting toxic posts into acceptable text has received much less attention, and existing Arabic resources are small and offer only one rewrite per input. AraDetox addresses this gap. The author took 10,500 harmful posts from the Arabic Hate Speech Superset and had GPT-5 and Gemini 2.5 Flash rewrite each one into four varieties: MSA, Gulf, Levantine and Egyptian. This gives eight cleaned versions per post and 84,000 outputs in total. The prompts asked the models to keep target, stance, sentiment and intent while removing slurs, insults and threats. Generation was followed by automatic validity checks, regeneration of failed outputs, and review by a native speaker.

The outputs were scored on:

  • lexical change,
  • embedding similarity,
  • sentiment shift,
  • n-gram closeness to dialect corpora,
  • independent human judgement.

The rewrites change the wording heavily, yet semantic similarity stays high, so the task behaves like paraphrasing that preserves meaning rather than minimal word swapping. The resource shows that existing detection corpora can be turned cheaply into multi-dialect detoxification data.

Key points

  • Sources: Of the 16,000 harmful posts in Tonneau et al.'s (2024) Arabic Hate Speech Superset, 10,500 were used, spanning all ten member datasets (T-HSAB, JHSC, MLMA, L-HSAB, OSACT, Let-Mi, etc.) with no class balancing.
  • Scale: 84,000 rewrites (42,000 per model), 1,318,257 words in total, with a mean output length of 15.69 words and a median of 13.
  • Human check:
    • 300 stratified source posts, with all 8 variants judged by 3 native annotators (2,400 outputs, 7,200 records).
    • These annotators were different from the quality-control reviewer.
    • Labels cover offence removal, meaning preservation and added content.
  • Negative sentiment, using CAMeLBERT-DA/MSA/Mix classifiers:
    • MSA rewrites show the largest drops, roughly 15–23 points (e.g. 75.47 → 53.48 for GPT-MSA).
    • Dialect rewrites drop only about 2–7 points.
    • The MultiParaDetox-Ar 400-pair test split starts low and changes little.
  • Two contrasting styles emerge: MultiParaDetox-Ar favours minimal lexical edits, while AraDetox favours full reformulation. The comparison is descriptive only, and no downstream training benchmark is reported.

Interesting ideas

  • I1. The dialect outputs mix two transformations (detoxification and variety change), so the MSA outputs act as a within-variety baseline, and each effect is scored on its own axis instead of crediting all change to detox.
  • I2. Token-level F1 is rejected because many differently worded rewrites can all be correct; surface-distance measures are instead paired with meaning-similarity and human scores, each targeting one requirement.
  • I3. The pipeline names specific LLM failure types (refusals, generic moderation boilerplate, summaries, outsider-style descriptions, added meaning), rejects them, and regenerates.
  • I4. An insult is treated as carrying a real criticism that must survive the rewrite; only the abusive form is removed.
  • I5. The release separates the fully synthetic collection from the human-labelled subset, with per-output flags (offence removed, meaning kept, content added) that let users filter.

For Mova

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

  • I1 (tree): When converting treebanks between UD dialects, keep an unconverted control slice so parser score shifts caused by the conversion can be separated from those caused by the parser itself.
  • I2 (back): Evaluate the text → graph → text compressor by combining edit distance with re-parsed graph equivalence, not by reference token F1, so valid paraphrases are not penalised.
  • I3 (learn): Add these failure classes as explicit filters on LLM teacher tiers, so a teacher that describes or summarises a sentence instead of annotating it never enters the silver data.
  • I4 (frames): Represent insults in crate prag as an evaluative speech act with speaker, target and stance slots, so the frame stays stable when wording is softened and can be checked for preservation.
  • I5 (learn): Tag every silver item with its provenance and per-criterion verdicts so training and negative-control evaluation can select only verified examples.

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, 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 worksCommunication · 104 works

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

Cite

Cite as Mo El-Haj. 2026. AraDetox: A Multi-Dialect Arabic Detoxification Dataset. arXiv:2608.22894 · ArabicNLP 2026 at EMNLP 2026 Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2608.22894 · arXiv 2608.22894

@misc{elhaj2026aradetox,
  title = {{AraDetox: A Multi-Dialect Arabic Detoxification Dataset}},
  author = {Mo El-Haj},
  year = {2026},
  doi = {10.48550/arxiv.2608.22894},
  eprint = {2608.22894},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2608.22894}
}

References

22 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Muhammad Abdul-Mageed, AbdelRahim Elmadany, and El Moatez Billah Nagoudi. 2021. ARBERT & MARBERT: Deep bidirectional transformers for Arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7088–7105, Online. Association for Computational Linguistics. ↗
  2. Ashraf Ahmad, Mohammad Azzeh, Eman Alnagi, Qasem Abu Al-Haija, Dana Halabi, Abdullah Aref, and Yousef AbuHour. 2024. Hate speech detection in the arabic language: corpus design, construction, and evaluation. Frontiers in Artificial Intelligence, 7:1345445. ↗
  3. Salim Al Mandhari, Mo El-Haj, and Paul Rayson. 2024. Is it offensive or abusive? an empirical study of hateful language detection of arabic social media texts. In Proceedings of the First International Conference on Natural Language Processing and Artificial Intelligence for Cyber Security, pages 137–146.
  4. Nuha Albadi, Mohamed Kurdi, and Swapna Mishra. 2018. Are they our brothers? analysis and detection of religious hate speech in the arabic twittersphere. In Proceedings of AICS. find
  5. Seham Alghamdi, Youcef Benkhedda, Basma Alharbi, and Riza Batista-Navarro. 2024. AraTar: A corpus to support the fine-grained detection of hate speech targets in the Arabic language. In Proceedings of the 6th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT) with Shared Tasks on Arabic LLMs Hallucination and Dialect to MSA Machine Translation @ LREC-COLING 2024, pages 1–12, Torino, Italia. ELRA and ICCL.
  6. Safa Alsafari, Samira Sadaoui, and Malek Mouhoub. 2020. Hate and offensive speech detection on arabic social media. Online Social Networks and Media, 19:100096. ↗
  7. Raghad Alshaalan and Hend Al-Khalifa. 2020. Hate speech detection in saudi twittersphere: A deep learning approach. In Proceedings of the fifth Arabic natural language processing workshop, pages 12–23.
  8. Ghadah Alwakid, Taha Osman, Mahmoud El Haj, Saad Alanazi, Mamoona Humayun, and Najm Us Sama. 2022. Muldasa: Multifactor lexical sentiment analysis of social-media content in nonstandard arabic social media. Applied Sciences, 12(8):3806. ↗
  9. Mohamed Seghir Hadj Ameur and Hassina Aliane. 2021. Aracovid19-mfh: Arabic covid-19 multi-label fake news & hate speech detection dataset. Procedia Computer Science, 189:232–241. American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers), pages 124–140, Mexico City, Mexico. Association for Computational Linguistics. find
  10. Daryna Dementieva, Nikolay Babakov, Amit Ronen, Abinew Ali Ayele, Naquee Rizwan, Florian Schneider, Xintong Wang, Seid Muhie Yimam, Daniil Moskovskiy, Elisei Stakovskii, Eran Kaufman, Ashraf Elnagar, Animesh Mukherjee, and Alexander Panchenko. 2025. Multilingual and explainable text detoxification with parallel corpora. In Proceedings of the 31st International Conference on Computational Linguistics, pages 7998–8025, Abu Dhabi, UAE. Association for Computational Linguistics. ↗
  11. Mahmoud El-Haj, Paul Rayson, Steve Young, Andrew Moore, Martin Walker, Thomas Schleicher, and Vasiliki Athanasakou. 2016. Learning tone and attribution for financial text mining. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 1820–1825. find
  12. Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. 2023. Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30):e2305016120. ↗
  13. Hatem Haddad, Hala Mulki, and Asma Oueslati. 2019. T-hsab: A tunisian hate speech and abusive dataset. In International conference on Arabic language processing, pages 251–263. Springer. find
  14. Myra S Hunter, Mahmoud El-Haj, Eleanor Thorne, Amanda Griffiths, and Claire Hardy. 2023. # menopause: Examining the frequency of communications about menopause on twitter between 2014 and 2022. Maturitas, 177:107806. ↗
  15. Hung Nguyen Huy, Mo El-Haj, Dawn Knight, and Paul Rayson. 2026. Freetxt-vi: A benchmarked vietnamese-english toolkit for segmentation, sentiment, and summarisation. arXiv preprint arXiv:2603.05690. ↗
  16. 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, Kyiv, Ukraine (Online). Association for Computational Linguistics.
  17. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: Transformer-based model for Arabic language understanding. 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. ↗
  18. Varvara Logacheva, Daryna Dementieva, Sergey Ustyantsev, Daniil Moskovskiy, David Dale, Irina Krotova, Nikita Semenov, and Alexander Panchenko. 2022. ParaDetox: Detoxification with parallel data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6804–6818, Dublin, Ireland. Association for Computational Linguistics.
  19. Daryna Dementieva, Nikolay Babakov, and Alexander Panchenko. 2024. MultiParaDetox: Extending text detoxification with parallel data to new languages. In Proceedings of the 2024 Conference of the North Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. 2024. On llms-driven synthetic data generation, curation, and evaluation: A survey. In Findings of the Association for Computational Linguistics: ACL 2024, pages 11065–11082. ↗
  20. Hamdy Mubarak, Hend Al-Khalifa, and Abdulmohsen Al-Thubaity. 2022. Overview of OSACT5 shared task on Arabic offensive language and hate speech detection. In Proceedinsg of the 5th Workshop on Open-Source Arabic Corpora and Processing Tools with Shared Tasks on Qur’an QA and Fine-Grained Hate Speech Detection, pages 162–166, Marseille, France. European Language Resources Association.
  21. Hamdy Mubarak, Kareem Darwish, Walid Magdy, Tamer Elsayed, and Hend Al-Khalifa. 2020. Overview of osact4 arabic offensive language detection shared task. In Proceedings of the 4th Workshop on open-source arabic corpora and processing tools, with a shared task on offensive language detection, pages 48–52.
  22. Omar F. Zaidan and Chris Callison-Burch. 2014. Arabic dialect identification. Computational Linguistics, 40(1):171–202. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/el-haj-2026-aradetox-multi-dialect-arabic-detoxification/