Elamur Research⌕
☰
home / papers / 2025 / An Exploration of Knowledge Editing for Arabic

An Exploration of Knowledge Editing for Arabic

Paper · 2025 · Basel Mousi, Nadir Durrani and Fahim Dalvi

Summary

Knowledge Editing (KE) changes specific facts inside an LLM without retraining it. Prior work has focused on English, and morphologically rich languages such as Arabic have received little attention. The authors present what they describe as the first Arabic KE study. They machine-translated the ZsRE and Counterfact benchmarks into Arabic and German with NLLB-200, keeping the KnowEdit splits. On Llama-2-7B-chat they tested four editors: ROME and MEMIT (parameter-based), ICE (in-context) and LTE (instruction tuning). Each was scored on reliability, generality, locality and portability, both within one language and across languages.

The parameter-based methods do best in English and get worse in German and worse again in Arabic. Their edits also transfer poorly from one language to another. The prompt-driven methods are much more stable across languages. The authors extended LTE with Arabic-only and joint Arabic+English training. The joint variant gave the strongest editing and the best cross-lingual transfer, at a small cost in locality. The released datasets and training data fill a resource gap for multilingual editing.

Key points

  • The four methods are ROME, MEMIT, ICE and LTE, all evaluated on Llama-2-7B-chat. Edits were tested within the same language and across Arabic, English and German.
  • The benchmarks are Arabic and German versions of ZsRE and Counterfact, produced with facebook/nllb-200-3.3B. They follow the KnowEdit splits and are released on GitHub (baselmousi/arabic-knowledge-editing).
  • Results for Arabic edits:
    • LTE-EN, trained only on English, still ranks first on all four metrics.
    • ICE comes second, but its portability collapses on Counterfact.
    • MEMIT keeps unrelated knowledge best.
    • ROME is weakest overall.
  • Cross-lingual transfer is weak in both directions (Arabic→EN/DE and EN/DE→Arabic). The weakness is most pronounced for ROME and MEMIT.
  • Three LTE variants were compared: LTE-EN, LTE-AR and LTE-AR-EN. Joint training wins on reliability, generality and cross-lingual scores, while locality drops slightly.

Interesting ideas

  • I1. Each edit is judged on four axes: whether the new fact holds, whether paraphrases also see it, whether unrelated facts stay untouched, and whether reasoning that depends on it works.
  • I2. The study separates "edit and test in the same language" from "edit in one language, test in another" and finds the two behave very differently.
  • I3. New knowledge supplied as explicit context or instructions proved sturdier than knowledge written into weights, while weight edits were the most language-fragile.
  • I4. Counterfact deliberately states plausible but false facts to check whether the model follows the new claim instead of its prior belief.
  • I5. Joint bilingual training improved generalisation but slightly hurt locality, a trade-off that only shows up when a "nothing else changed" metric is tracked.

For Mova

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

  • I1 (world): Every fact added to the knowledge graph could be checked with paraphrased queries, untouched-neighbour controls and one-hop inference probes, so that bad entries are caught early.
  • I2 (back): Translation through the meta-world could be tested this way, by asserting a fact in English text and checking that the generated target-language text carries it unchanged.
  • I3 (world): Corrections could live as an explicit, inspectable overlay on top of compiled WordNet/Wikidata knowledge rather than being merged into it, which keeps every override explainable and reversible.
  • I4 (answer): QA on fables could include probes where the tale contradicts world knowledge (for example, a talking fox), to verify that the reader answers from the text's world and not from compiled defaults.
  • I5 (learn): When retraining on silver data, negative-control sets of untouched cases should be reported next to the gains, so that broader rules that leak into unrelated cases are noticed.

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

🎓Hamad bin Khalifa UniversityDoha · Qatar141 works · 14,175 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Basel Mousi, Nadir Durrani and Fahim Dalvi. 2025. An Exploration of Knowledge Editing for Arabic. Proceedings of The Third Arabic Natural Language Processing Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.arabicnlp-main.34

@inproceedings{mousi2025exploration,
  title = {{An Exploration of Knowledge Editing for Arabic}},
  author = {Basel Mousi, Nadir Durrani and Fahim Dalvi},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-main.34},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-main.34}
}

References

28 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Muhammad Asif Ali, Nawal Daftardar, Mutayyba Waheed, Jianbin Qin, and Di Wang. 2025. MQA-KEAL: Multi-hop question answering under knowledge editing for Arabic language. In Proceedings of the 31st International Conference on Computational Linguistics, pages 5629–5644, Abu Dhabi, UAE. Association for Computational Linguistics. ↗
  2. Himanshu Beniwal, Kowsik D, and Mayank Singh. 2024. Cross-lingual editing in multilingual language models. In Findings of the Association for Computational Linguistics: EACL 2024, pages 2078–2128, St. Julian’s, Malta. Association for Computational Linguistics. find
  3. Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491– 6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. find
  4. Nadir Durrani, Basel Mousi, and Fahim Dalvi. 2025. Editing across languages: A survey of multilingual knowledge editing. In In The Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (To Appear), Suzhou, China. Association for Computational Linguistics. find
  5. Imane Guellil, Houda Saâdane, Faical Azouaou, Billel Gueni, and Damien Nouvel. 2021. Arabic natural language processing: An overview. Journal of King Saud University - Computer and Information Sciences, 33(5):497–507.
  6. Nizar Habash, Houda Bouamor, Ramy Eskander, Nadi Tomeh, Ibrahim Abu Farha, Ahmed Abdelali, Samia Touileb, Injy Hamed, Yaser Onaizan, Bashar Alhafni, Wissam Antoun, Salam Khalifa, Hatem Haddad, Imed Zitouni, Badr AlKhamissi, Rawan Almatham, and Khalil Mrini, editors. 2024. Proceedings of the Second Arabic Natural Language Processing Conference. Association for Computational Linguistics, Bangkok, Thailand. find
  7. Yuxin Jiang, Yufei Wang, Chuhan Wu, Wanjun Zhong, Xingshan Zeng, Jiahui Gao, Liangyou Li, Xin Jiang, Lifeng Shang, Ruiming Tang, Qun Liu, and Wei Wang. 2024. Learning to edit: Aligning LLMs with knowledge editing. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4689– 4705, Bangkok, Thailand. Association for Computational Linguistics. ↗
  8. Aditi Khandelwal, Harman Singh, Hengrui Gu, Tianlong Chen, and Kaixiong Zhou. 2024. Cross-lingual multi-hop knowledge editing. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11995–12015, Miami, Florida, USA. Association for Computational Linguistics. find
  9. Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. 2017. Zero-shot relation extraction via reading comprehension. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 333–342, Vancouver, Canada. Association for Computational Linguistics.
  10. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022a. Locating and editing factual associations in GPT. In Advances in Neural Information Processing Systems, volume 35 of NeurIPS, New Orleans, LA. ↗
  11. Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2022b. Mass editing memory in a transformer. arXiv preprint arXiv:2210.07229. ↗
  12. Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D. Manning, and Chelsea Finn. 2022. Memory-based model editing at scale. Preprint, arXiv:2206.06520. ↗
  13. Basel Mousi, Nadir Durrani, Fahim Dalvi, Majd Hawasly, and Ahmed Abdelali. 2024. Exploring alignment in shared cross-lingual spaces. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6326–6348, Bangkok, Thailand. Association for Computational Linguistics. find
  14. Ercong Nie, Bo Shao, Zifeng Ding, Mingyang Wang, Helmut Schmid, and Hinrich Schütze. 2025. Bmike-53: Investigating cross-lingual knowledge editing with in-context learning. Preprint, arXiv:2406.17764. ↗
  15. Hassan Sawaf, Samhaa El-Beltagy, Wajdi Zaghouani, Walid Magdy, Ahmed Abdelali, Nadi Tomeh, Ibrahim Abu Farha, Nizar Habash, Salam Khalifa, Amr Keleg, Hatem Haddad, Imed Zitouni, Khalil Mrini, and Rawan Almatham, editors. 2023. Proceedings of ArabicNLP 2023. Association for Computational Linguistics, Singapore (Hybrid). find
  16. Nianwen Si, Hao Zhang, and Weiqiang Zhang. 2024. Mpn: Leveraging multilingual patch neuron for cross-lingual model editing. Preprint, arXiv:2401.03190. ↗
  17. Daniel Tamayo, Aitor Gonzalez-Agirre, Javier Hernando, and Marta Villegas. 2024. Mass-editing memory with attention in transformers: A cross-lingual exploration of knowledge. In Findings of the Association for Computational Linguistics: ACL 2024, pages 5831–5847, Bangkok, Thailand. Association for Computational Linguistics. ↗
  18. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2022. No language left behind: Scaling human-centered machine translation. Preprint, arXiv:2207.04672. ↗
  19. Jiaan Wang, Yunlong Liang, Zengkui Sun, Yuxuan Cao, Jiarong Xu, and Fandong Meng. 2024a. Cross-lingual knowledge editing in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11676–11686, Bangkok, Thailand. Association for Computational Linguistics. ↗
  20. Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. 2024b. Knowledge editing for large language models: A survey. Preprint, arXiv:2310.16218. ↗
  21. Weixuan Wang, Barry Haddow, and Alexandra Birch. 2024c. Retrieval-augmented multilingual knowledge editing. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 335–354, Bangkok, Thailand. Association for Computational Linguistics. 2025. Multilingual knowledge editing with language-agnostic factual neurons. In Proceedings of the 31st International Conference on Computational Linguistics, pages 5775–5788, Abu Dhabi, UAE. Association for Computational Linguistics.
  22. Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023. Can we edit factual knowledge by in-context learning? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4862–4876, Singapore. Association for Computational Linguistics. find
  23. Zihao Wei, Jingcheng Deng, Liang Pang, Hanxing Ding, Huawei Shen, and Xueqi Cheng. 2025. MLaKE: Multilingual knowledge editing benchmark for large language models. In Proceedings of the 31st International Conference on Computational Linguistics, pages 4457–4473, Abu Dhabi, UAE. Association for Computational Linguistics. ↗
  24. Suhang Wu, Minlong Peng, Yue Chen, Jinsong Su, and Mingming Sun. 2023. Eva-kellm: A new benchmark for evaluating knowledge editing of llms. Preprint, arXiv:2308.09954. ↗
  25. Yuchen Wu, Liang Ding, Li Shen, and Dacheng Tao. 2025. Edit once, update everywhere: A simple framework for cross-lingual knowledge synchronization in llms. Preprint, arXiv:2502.14645. ↗
  26. Yang Xu, Yutai Hou, Wanxiang Che, and Min Zhang. 2023. Language anisotropic cross-lingual model editing. In Findings of the Association for Computational Linguistics: ACL 2023, pages 5554–5569, Toronto, Canada. Association for Computational Linguistics. find
  27. Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. Editing large language models: Problems, methods, and opportunities. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10222–10240, Singapore. Association for Computational Linguistics. find
  28. Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, Siyuan Cheng, Ziwen Xu, Xin Xu, Jia-Chen Gu, Yong Jiang, Pengjun Xie, Fei Huang, Lei Liang, Zhiqiang Zhang, and 3 others. 2024. A comprehensive study of knowledge editing for large language models. Preprint, arXiv:2401.01286. Xue Zhang, Yunlong Liang, Fandong Meng, Songming Zhang, Yufeng Chen, Jinan Xu, and Jie Zhou. *Figure 5: Comparison of ROME, MEMIT, and ICE on LLaMA2-7B-Chat across four metrics: reliability, generality, locality, and Portability on the counterfact dataset* ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mousi-2025-exploration-knowledge-editing-arabic/