Elamur Research⌕
☰
home / papers / 2026 / Hate Speech Detection in Turkish and Arabic: A Comprehensive Study

Hate Speech Detection in Turkish and Arabic: A Comprehensive Study

Paper · 2026 · Somaiyeh Dehghan, Gokçe Uludo\ugan, Mehmet Umut Şen et al.

Summary

Hate speech detection resources for Turkish and Arabic are scarce, and most existing work treats the problem as a simple yes/no decision. This paper builds a multi-faceted benchmark from X (Twitter) posts. The Turkish portion covers five themes: refugees, the Israel–Palestine conflict (collected before October 7, 2023), anti-Greek sentiment, ethnic and religious communities, and LGBTI+. The Arabic portion focuses on refugees. Three annotators labelled each tweet with a hate category, a 0–10 intensity score, the target groups and, for span work, the specific hateful segments. The authors fine-tune BERT-based models for four tasks: category classification (using dual contrastive learning, DualCL), intensity regression, multi-label target identification, and token-level span detection. ChatGPT is used to split hashtags into words and to generate synthetic examples for rare classes. DualCL beats a cross-entropy baseline, and the gain is largest for multi-class labels and for the smaller Arabic set. Span categorization remains hard. The work matters because it gives a first joint benchmark covering these dimensions for two under-served languages.

Key points

  • Corpus size: 10,953 Turkish tweets and 2,510 Arabic tweets, annotated in Label Studio in batches of 50, with three annotators per tweet.
  • Agreement was low (average Krippendorff's Alpha of 0.234), so training relied on samples that reached agreement or a majority vote.
  • Span detection with BERTurk (dbmdz/bert-base-turkish-cased):
    • Binary span detection with BIO tags reaches about 59% F1.
    • Five-category span labelling with IO tags reaches only about 31% F1.
  • Per-category span results show explicit hate is far easier than implicit hate:
    • The category for threats of hostility, warfare, assault, killing or injury reaches F1 0.44.
    • Exclusionary/Discriminatory Discourse reaches F1 0.08.
    • Exaggeration/Generalization reaches F1 0.17.
  • Span models were evaluated with 5-fold cross-validation (mean and standard deviation) plus a held-out set of 300 tweets, scored with SeqEval.

Interesting ideas

  • I1. Categories are assigned prescriptively from guidelines, while intensity is rated descriptively as each annotator perceives it; this keeps definitional labels apart from graded subjective judgments.
  • I2. When an annotator picks several groups or categories, their single vote is divided fractionally among them instead of counting fully for each, which yields natural soft targets.
  • I3. The final test set is built only from items that annotators agreed on in the first round, while cross-validation over the rest gives a variance estimate.
  • I4. Locating the exact hateful span exposes a sharp gap between explicit and implicit categories, a gap that a single whole-text label would hide.
  • I5. An LLM splits concatenated hashtags into separate words before modelling, so content glued together in social text becomes usable.

For Mova

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

  • I1 (learn): When LLM teacher tiers produce silver labels, Mova could ask for rule-defined labels and for confidence or degree judgments as separate fields, so graded noise does not contaminate categorical decisions.
  • I2 (learn): Votes from multiple witness parsers could be split the same way when one parser hedges between attachments, giving soft silver arcs instead of forced hard choices.
  • I3 (learn): Mova could keep a small gold test set made only of sentences where all witness parsers and the teacher agree, and report fold spread alongside it to catch unstable gains.
  • I4 (frames): For speech acts, irony and implicature, Mova could record which tokens in the UD tree trigger each pragmatic label, making explanations concrete and separating easy explicit cues from inference-heavy cases.
  • I5 (tokens): A similar segmentation pass, with a teacher check, could break hashtags, run-together handles and compounds into words before tagging and lemmatization in crate en.

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

🎓Sabancı ÜniversitesiIstanbul · Türkiye5 works · 228 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSociology and Political Science · 577 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 Somaiyeh Dehghan, Gokçe Uludo\ugan, Mehmet Umut Şen, Elif Erol, Arzucan Ozgur and Berrin Yanikoglu. 2026. Hate Speech Detection in Turkish and Arabic: A Comprehensive Study. arXiv:2607.00143.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2607.00143 · arXiv 2607.00143

@misc{dehghan2026hate,
  title = {{Hate Speech Detection in Turkish and Arabic: A Comprehensive Study}},
  author = {Somaiyeh Dehghan, Gokcce Uludougan, Mehmet Umut cSen, Elif Erol, Arzucan Ozgur and Berrin Yanikoglu},
  year = {2026},
  doi = {10.48550/arxiv.2607.00143},
  eprint = {2607.00143},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2607.00143}
}

References

23 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. [19] Bäumler J, Blöcher L, Frey L-J, Chen X, Bayer M, and Reuter C. A survey of machine learning models and datasets for the multi-label classification of textual hate speech in english. ACM Trans. Knowl. Discov. Data, June 2026. Just Accepted.
  2. [20] Santos G L, Santos V G d, and Kearns C. All together against hate: ensemble based llms for multi class hate speech classification in the football context, 2026. find
  3. [21] Ayele A A, Jalew E A, Ali A C, Yimam S M, and Biemann C. Exploring boundaries and intensities in offensive and hate speech: Unveiling the complex spectrum of social media discourse, 2024. Preprint at https://doi.org/10.48550/arXiv.2404.12042. ↗
  4. [22] Geleta R R, Eckelt K, Parada-Cabaleiro E, and Schedl M. Exploring intensities of hate speech on social media: A case study on explaining multilingual models with XAI. In Proceedings of the 4th Conference on Language, Data and Knowledge, pages 532–537, Vienna, Austria, September 2023. NOVA CLUNL, Portugal. find
  5. [23] Riyadi S, Masyhur A M, and Andriyani A D. Deep learning for hate speech intensity analysis: Distilbert classification algorithm. In 2024 IEEE Symposium on Industrial Electronics and Applications (ISIEA), pages 1–6, 2024. find
  6. [24] Büyükdemirci K, Kucukkaya I E, Ölmez E, and Toraman C. JL-hate: An annotated dataset for joint learning of hate speech and target detection. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 9543–9553, Torino, Italia, may 2024. ELRA and ICCL. find
  7. [25] G Uludoğan, Yüksel A E, Tunçer U, Işık B, Korkmaz Y, Akar D, and Özgür A. Detecting hate speech in Turkish print media: A corpus and a hybrid approach with target-oriented linguistic knowledge. In Proceedings of the 7th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE 2024), pages 205–214, St. find
  8. Julians, Malta, March 2024. Association for Computational Linguistics. find
  9. [26] Hoang P G, Luu C D, Tran K Q, Nguyen K V, and Nguyen N L-T. ViHOS: Hate speech spans detection for Vietnamese. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 652–669, Dubrovnik, Croatia, May 2023. Association for Computational Linguistics. find
  10. [27] Calabrese A, Neves L, Shah N, Bos M, Ross B, Lapata M, and Barbieri F. Explainability and hate speech: Structured explanations make social media moderators faster. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 398–408, Bangkok, Thailand, August 2024. Association for Computational Linguistics. find
  11. [28] Arın I, Işık Z, Kutal S, Dehghan S, Özgür A, and Yanikoglu B. Siu2023-nst - hate speech detection contest. In 2023 31st Signal Processing and Communications Applications Conference (SIU), pages 1–4, 2023. find
  12. [29] Dehghan S, Sen M U, and Yanikoglu B. Dealing with annotator disagreement in hate speech classification, 2025. Preprint at https://doi.org/10.48550/arXiv.2502.08266. ↗
  13. [30] Uludoğan G, Dehghan S, Arin I, Erol E, Yanikoglu B, and Özgür A. Overview of the hate speech detection in Turkish and Arabic tweets (HSD-2Lang) shared task at CASE 2024. In Proceedings of the 7th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE 2024), pages 229–233, St.
  14. Julians, Malta, March 2024. Association for Computational Linguistics. find
  15. [31] Dehghan S and Yanikoglu B. Evaluating ChatGPT’s ability to detect hate speech in Turkish tweets. In Proceedings of the 7th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE 2024), pages 54–59, St. find
  16. Julians, Malta, March 2024. Association for Computational Linguistics. find
  17. [32] Şeker H K, Uludoğan G, Önal P, and Özgür A. HATECAT-TR: A hate speech span detection and categorization dataset for Turkish. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors, Findings of the Association for Computational Linguistics: EMNLP 2025, pages 25568–25579, Suzhou, China, November 2025. Association for Computational Linguistics. find
  18. [33] Devlin J, Chang M W, Lee K, and Toutanova K. 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, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics.
  19. [34] Khosla P, Teterwak P, Wang C, Sarna A, Tian Y, Isola P, Maschinot A, Liu C, and Krishnan D. Supervised contrastive learning. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red Hook, NY, USA, 2020. Curran Associates Inc. find
  20. [35] Chen T, Kornblith S, Norouzi M, and Hinton G. A simple framework for contrastive learning of visual representations. In Proceedings of the 37 th International Conference on Machine Learning, Vienna, Austria, 2020.
  21. [36] Rodrigues R C, Inuzuka M A, Gomes J R S, Rocha A S, Calixto I, and Nascimento H A D. Zero-shot hashtag segmentation for multilingual sentiment analysis, 2021. Preprint at https://doi.org/10.48550/arXiv.2112.03213. ↗
  22. [37] Li Z, Zhu H, Lu Z, and Yin M. Synthetic data generation with large language models for text classification: Potential and limitations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10443–10461, Singapore, December 2023. Association for Computational Linguistics.
  23. [38] Guo X and Chen Y. Generative ai for synthetic data generation: Methods, challenges and the future, 2024. Preprint at https://doi.org/10.48550/arXiv.2403.04190. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/dehghan-2026-hate-speech-detection-turkish-arabic/