Elamur Research⌕
☰
home / papers / 2026 / HACS-TL: Cross-Script Transfer Learning for Hausa Ajami Hate Speech Detection Using Transformer-Based Architecture

HACS-TL: Cross-Script Transfer Learning for Hausa Ajami Hate Speech Detection Using Transformer-Based Architecture

Paper · 2026 · Abdulkadir Shehu Bichi, Muqaddar Ali, Prashant Sharma et al.

Summary

The paper tackles hate speech detection in Hausa written in Ajami, the Arabic-derived script, which NLP tools barely cover. No native Ajami hate speech corpus exists, so the authors start from the Latin-script (Boko) Hausa Hate Speech Dataset of 2,000 labelled comments. They convert it into Ajami with a hand-built rule system. The converter splits words into (C)V(C) syllables, maps them to Arabic letters by longest match, with special cases for implosives, ejectives, long vowels and digraphs, and then adds sukun to consonant clusters.

On the converted text they train HACS-TL. It is an mBERT backbone extended with:

  • a cross-script attention module,
  • a small orthographic transformer encoder,
  • dialect-aware attention pooling,
  • multi-scale CNN features.

Training uses focal loss and augmentation of the offensive class. Under 2-fold cross-validation, HACS-TL reaches 75.33% macro F1 against nine baselines; mBERT is the strongest at 73.70%. AraBERT does worst, which suggests that sharing a script does not by itself make a model transfer well. The work offers a first baseline for both Ajami conversion and Ajami classification, and the authors release the rules, models and data.

Key points

  • Data: 2,000 Boko samples, about 66% non-offensive. The offensive class is enlarged by synonym replacement, random swap and random deletion. The paper reports two different post-augmentation sizes, 1,741 and 2,712.
  • Converter quality: checked on 100 samples, with 63.57% mean character similarity. Only 51 samples pass the 70% similarity threshold; the failures are mostly code-mixing, dialect forms and informal spelling.
  • Main results (macro F1):
    • HACS-TL 75.33
    • mBERT 73.70
    • CharCNN 70.01
    • AfroXLMR 69.67
    • XLM-R 67.20
    • LoRA-mBERT 66.03
    • LoRA-XLMR 63.71
    • ByteLSTM 63.24
    • AraBERT 56.96
  • ROC AUC: mBERT (0.820) and HACS-TL (0.818) are essentially tied, while AraBERT scores 0.590.
  • Ablation:
    • Removing cross-script attention costs the most (75.33 → 73.45).
    • Removing the orthographic encoder drops F1 to 74.21.
    • Removing dialectal pooling drops F1 to 74.87.
    • The diacritic stress test shows only 32% prediction consistency, which the authors list as a weakness.

Interesting ideas

  • I1. A deterministic, phonology-aware transliterator, checked by round-trip conversion and a character-similarity pass threshold, can turn labelled data from one script into training data for another.
  • I2. Comparing baselines from several families (multilingual, script-matched, char/byte-level, PEFT) shows that an "obviously related" model like AraBERT can be the worst one.
  • I3. An orthographic stress test perturbs only spelling marks, such as diacritics, and measures how often the prediction stays the same.
  • I4. The error analysis compares the model's confidence on correct predictions against its confidence on false positives and false negatives.
  • I5. Hand-written rules with explicit classes for sounds the target script lacks (implosives, long vowels, digraphs) keep the conversion step transparent and fixable.

For Mova

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

  • I1 (back): Mova could score its text → graph → text compressor the same way, reporting round-trip similarity and the share of outputs above a pass threshold rather than only an average.
  • I2 (learn): Mova's evaluation could routinely include a superficially similar but mismatched witness as a negative control, so that agreement with it is not mistaken for real competence.
  • I3 (tokens): Mova's tagger and lemmatizer could be probed with spelling, casing and punctuation variants, flagging tokens whose tags flip when the meaning does not change.
  • I4 (learn): when silver data is filtered by parser confidence, Mova could check whether the wrong parses carry nearly as much confidence as the right ones, a sign that the threshold alone is unsafe.
  • I5 (tokens): Mova's normalization of dialectal or variant English spellings could use a similar ordered longest-match rule table, keeping each rewrite 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

🎓Vivekananda Global UniversityJaipur · India2 works

Also on the byline: Federal University Gusau.

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 3Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 worksBias, fairness & safety · 517 works

Cite

Cite as Abdulkadir Shehu Bichi, Muqaddar Ali, Prashant Sharma and Ismail Dauda Abubakar. 2026. HACS-TL: Cross-Script Transfer Learning for Hausa Ajami Hate Speech Detection Using Transformer-Based Architecture. 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. DOI 10.18653/v1/2026.abjadnlp-1.36

@inproceedings{bichi2026hacs,
  title = {{HACS-TL: Cross-Script Transfer Learning for Hausa Ajami Hate Speech Detection Using Transformer-Based Architecture}},
  author = {Abdulkadir Shehu Bichi, Muqaddar Ali, Prashant Sharma and Ismail Dauda Abubakar},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.36},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.36}
}

References

23 references, 13 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 7085–7105, Online. Association for Computational Linguistics. ↗
  2. This work introduces HACS-TL, a transformer- David Ifeoluwa Adelani, Jade Abbott, Graham Neubig, based architecture for hate speech detection in and 1 others. 2021. Masakhaner: Named entity recognition for african languages. Transactions of the Hausa Ajami that addresses critical gaps in NLP Association for Computational Linguistics, 9:1116– for low-resource, non-Latin script languages. Our 1131. approach achieved 75.33% macro F1-score, outperforming multilingual baselines (mBERT: 73.70%) David Ifeoluwa Adelani, Graham Neubig, Sebastian Ruder, and 1 others. 2022. Masakhaner 2.0: Africa-through three key innovations: a linguistically val-centric transfer learning for named entity recognition. idated Latin-to-Hausa Ajami converter (63.57% In Proceedings of the 2022 Conference on Empiri-character similarity), cross-script attention mecha-cal Methods in Natural Language Processing, pages 4488–4508. nisms, and dialect-aware pooling. The findings demonstrate that script similar- Jesujoba O Alabi, David Ifeoluwa Adelani, Marius Mos-ity alone does not ensure effective transfer learn-bach, and Dietrich Klakow. 2022. Adapting pre-ing—AraBERT’s poor performance (56.96% F1) trained language models to african languages via 293 multilingual adaptive fine-tuning. In Proceedings of the 29th International Conference on Computational Linguistics, pages 4336–4349.
  3. Shamsuddeen Hassan Muhammad, Idris Abdulmumin, Yusuf Idris, and 1 others. 2022. Naijasenti: A nigerian twitter sentiment corpus for multilingual sentiment analysis. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 590–602.
  4. Sai Saket Aluru, Binny Mathew, Punyajoy Saha, and Animesh Mukherjee. 2020. Deep learning models for multilingual hate speech detection. arXiv preprint arXiv:2004.06465. ↗
  5. 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, pages 9–15. ↗
  6. Tommaso Caselli, Valerio Basile, Jelena Mitrović, and Michael Granitzer. 2021. Hatebert: Retraining bert for abusive language detection in english. In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 17–25.
  7. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, and 1 others. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440–8451. find
  8. Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. In Proceedings of the International AAAI Conference on Web and Social Media, volume 11, pages 512– 515. ↗
  9. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171– 4186. ↗
  10. Paula Fortuna and Sérgio Nunes. 2018. A survey on automatic detection of hate speech in text. ACM Computing Surveys (CSUR), 51(4):1–30. ↗
  11. Philip J Jaggar. 2001. Hausa. John Benjamins Publishing Company, Amsterdam. find
  12. Fajri Koto, Afshin Rahimi, Jey Han Lau, and Timothy Baldwin. 2020. Indolem and indobert: A benchmark dataset and pre-trained language model for indonesian nlp. In Proceedings of the 28th International Conference on Computational Linguistics, pages 757– 770. find
  13. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980–2988. ↗
  14. Marzieh Mozafari, Reza Farahbakhsh, and Noel Crespi. 2020. A bert-based transfer learning approach for hate speech detection in online social media. In International Conference on Complex Networks and Their Applications, pages 928–940. Springer. find
  15. Shamsuddeen Hassan Muhammad, David Ifeoluwa Adelani, Sebastian Ruder, and 1 others. 2023. Afrisenti: A twitter sentiment analysis benchmark for african languages. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13968–13981.
  16. Meikal Mumin and Kees Versteegh, editors. 2014. The Arabic Script in Africa: Studies in the Use of a Writing System. Brill. find
  17. Paul Newman. 2000. The Hausa Language: An Encyclopedic Reference Grammar. Yale University Press. find
  18. Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, and Dit-Yan Yeung. 2019. Multilingual and multi-aspect hate speech analysis. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4675–4684.
  19. Prachi Patil, Sakshi Raul, Dhanisha Raut, and Tat-wadashi Nagarhalli. 2023. Hate speech detection using deep learning and text analysis. In 2023 7th International Conference on Intelligent Computing and Control Systems (ICICCS), pages 322–330. find
  20. Edoardo Maria Ponti, Goran Glavaš, Olga Majewska, and 1 others. 2020. Xcopa: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362– 2376.
  21. Phillip Rust, Jonas Pfeiffer, Ivan Vulić, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. 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 3118–3135.
  22. Francielle Vargas, Samuel Guimarães, Shamsuddeen Hassan Muhammad, Diego Alves, Ibrahim Said Ahmad, Idris Abdulmumin, Diallo Mohamed, Thiago Pardo, and Fabrício Benevenuto. 2024. HausaHate: An expert annotated corpus for Hausa hate speech detection. In Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024), pages 52–58, Mexico City, Mexico. Association for Computational Linguistics. find
  23. Shijie Wu and Mark Dredze. 2020. Are all languages created equal in multilingual bert? In Proceedings of the 5th Workshop on Representation Learning for NLP, pages 120–130.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bichi-2026-hacs-tl-cross-script-transfer/