---
title: "Automatic Expansion and Retargeting of Arabic — Mubarak 2021"
description: "The paper tackles a gap in Arabic offensive-language detection."
authors: ["Hamdy Mubarak", "Ahmed Abdelali", "Kareem Darwish", "Younes Samih"]
year: 2021
venue: "arXiv:2111.09574"
source: "https://arxiv.org/abs/2111.09574"
doi: "10.48550/arxiv.2111.09574"
arxiv: "2111.09574"
licence: "CC BY-NC-ND 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by-nc-nd/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/mubarak-2021-automatic-expansion-retargeting-arabic-offensive/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/mubarak-2021-automatic-expansion-retargeting-arabic-offensive/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2021](https://md.research.elamur.ai/free/en/papers/2021/page/5/) / Automatic Expansion and Retargeting of Arabic Offensive Language Training

# Automatic Expansion and Retargeting of Arabic Offensive Language Training

**Authors:** [Hamdy Mubarak](https://md.research.elamur.ai/free/en/authors/hamdy-mubarak/), [Ahmed Abdelali](https://md.research.elamur.ai/free/en/authors/ahmed-abdelali/), [Kareem Darwish](https://md.research.elamur.ai/free/en/authors/kareem-darwish/), [Younes Samih](https://md.research.elamur.ai/free/en/authors/younes-samih/) · **Institutions:** [Hamad bin Khalifa University](https://md.research.elamur.ai/free/en/universities/hamad-bin-khalifa-university/) · **Year:** 2021 · **Venue:** arXiv:2111.09574 · **Licence:** [CC BY-NC-ND 4.0 (arXiv)](http://creativecommons.org/licenses/by-nc-nd/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2111.09574) · [PDF](https://arxiv.org/pdf/2111.09574) · [DOI 10.48550/arxiv.2111.09574](https://doi.org/10.48550/arxiv.2111.09574) · [arXiv 2111.09574](https://arxiv.org/abs/2111.09574) · [OpenAlex](https://openalex.org/W3213412969)

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

## Summary
The paper tackles a gap in Arabic offensive-language detection. Generic classifiers often miss abuse aimed at a particular entity, because such abuse can rely on deliberately twisted spellings of the target's name or on ordinary words used with hostile intent. The authors propose growing the training data automatically, building on two observations about Twitter. First, accounts that reply often to an entity tend to oppose it. Second, a user who attacks a target once usually keeps attacking it.

The method works in three steps:
1. A baseline classifier (SVM or fastText, trained on the SemEval 2020 Arabic corpus) labels replies to 25 heavily targeted accounts.
2. The most offensive repliers are selected.
3. All of their replies are treated as offensive and added to the training set.

The procedure mirrors blind relevance feedback from information retrieval. With per-target expansion, F1 on 2,500 manually labeled target-directed replies improves by roughly 13% relative for fastText and 79% for SVM. Pooled expansion also raises SVM F1 on the general task by about 48%. The approach shows that cheap user-level signals can produce large amounts of targeted training data.

## Key points
- **Data.** The base training data is the SemEval 2020 / OSACT 2020 Arabic offensive tweet corpus: 10,000 tweets, of which 1,915 are offensive.
- **Test set.** 25 target accounts (e.g., AJArabic, BakryMP, AlSisiOfficial) were drawn from 660k tweets collected in April–May 2019. For each account, 100 replies were hand-labeled, giving 2,500 tweets. A reply counts as offensive only if it is aimed at that target.
- **Features and preprocessing.**
  - Both classifiers use character 3–5-grams.
  - Arabic is normalized: alef variants become bare alef, alef maqsoura becomes ya, and ta marbouta becomes ha.
  - The SVM is SVMLight with a linear kernel. fastText runs 50 epochs at a learning rate of 0.1.
  - In 5-fold cross-validation, the baseline SVM reaches F1 51.4 and fastText reaches 73.5.
- **User selection.** "Offensive users" are either those with at least half of their replies auto-flagged (50% users) or the top n by count of flagged replies (n = 10, 20, 50).
  - Per-target, top-50 expansion works best: fastText F1 rises from 65.6 to 74.1, and SVM from 31.1 to 55.6.
  - Expansion lowers precision but raises recall considerably.
- **General task.** Pooled expansion lifts SVM F1 from 51.4 to about 76–77, improving both precision and recall.
  - fastText F1 drops, which the authors attribute to class imbalance: the non-offensive-to-offensive ratio moves from 4:1 to 1:2.
  - One target, AlSisiOfficial, is hurt by expansion because many of its frequent repliers were supporters attacking his opponents.

## Interesting ideas
- **Labels from actor persistence.** Spread a classifier's partial hits to everything the same actor wrote about the same target, on the assumption that stances are stable, much like pseudo-relevance feedback in retrieval. — For Mova (learn): Silver data could be expanded by propagating trusted labels across all sentences tied to the same source or construction pattern, with a held-out gold set to check that the propagation helps.
- **A precise seed makes a safer expander.** The higher-precision SVM gave more cautious expansion and gained on the general task, while the high-recall fastText flooded the training data and lost. — For Mova (learn): When witness parsers or LLM teacher tiers supply silver trees, favour the most precise witness as the expansion seed and watch how the class or label distribution shifts after each round.
- **Offensiveness depends on the target.** A tweet was judged offensive only if aimed at the specific entity. The Sisi failure shows that a hostile reply can attack a third party rather than the addressee. — For Mova (frames): Speech-act and irony frames should record who is being attacked, separately from who is being addressed, so that hostility is attached to the right participant.
- **Deliberate name distortions.** Abusers bend target names into insulting near-rhymes (Aljazeera → Alkhanzeera), which character n-grams plus orthographic normalization can still catch. — For Mova (tokens): The tokenizer and lemmatizer could flag out-of-vocabulary forms that are close in spelling to a known entity name and pass this to the next stage as a possible pejorative variant rather than an unknown word.
- **Report per-target outcomes, not just averages.** Checking results target by target revealed the one case where expansion backfired and explained why. — For Mova (learn): Evaluation of self-induced repair rules should break results down by construction or domain and treat regressions in any slice as negative controls.

*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: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

Topics: [Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/)

## Cite

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

```bibtex
@misc{mubarak2021automatic,
  title = {{Automatic Expansion and Retargeting of Arabic Offensive Language Training}},
  author = {Mubarak, Hamdy and Abdelali, Ahmed and Darwish, Kareem and Samih, Younes},
  year = {2021},
  doi = {10.48550/arxiv.2111.09574},
  eprint = {2111.09574},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2111.09574}
}
```

- [← previous A Study of Acoustic Features in Arabic Speaker Identification under Noisy Environmental Conditions](https://md.research.elamur.ai/free/en/papers/benhafid-2021-study-acoustic-features-arabic-speaker/)
- [next → BERT-based Multi-Task Model for Country and Province Level Modern Standard Arabic and Dialectal Arabic Identification](https://md.research.elamur.ai/free/en/papers/mekki-2021-bert-based-multi-task-model-country/)
- [↑ back to papers of 2021](https://md.research.elamur.ai/free/en/papers/2021/page/5/)

## References

20 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ehab Abozinadah. 2017. Detecting Abusive Arabic Language Twitter Accounts Using a Multidimensional Analysis Model. Ph.D. thesis, George Mason University. [find](https://scholar.google.com/scholar?q=Ehab+Abozinadah%2E+2017%2E+Detecting+Abusive+Arabic+Language+Twitter+Accounts+Using+a+Multidimensional+Analysis+Model%2E+Ph%2ED%2E+thesis%2C+George+Mason+University%2E)
2. Sweta Agrawal and Amit Awekar. 2018. Deep learning for detecting cyberbullying across multiple social media platforms. In European Conference on Information Retrieval, pages 141–153. Springer. [find](https://scholar.google.com/scholar?q=Sweta+Agrawal+and+Amit+Awekar%2E+2018%2E+Deep+learning+for+detecting+cyberbullying+across+multiple+social+media+platforms%2E+In+European+Conference+on+Information+Retrieval%2C+pages+141%E2%80%93153%2E+Springer%2E)
3. Azalden Alakrot, Liam Murray, and Nikola S Nikolov. 2018. Towards accurate detection of offensive language in online communication in arabic. Procedia computer science, 142:315–320. [find](https://scholar.google.com/scholar?q=Azalden+Alakrot%2C+Liam+Murray%2C+and+Nikola+S+Nikolov%2E+2018%2E+Towards+accurate+detection+of+offensive+language+in+online+communication+in+arabic%2E+Procedia+computer+science%2C+142%3A315%E2%80%93320%2E)
4. Nuha Albadi, Maram Kurdi, and Shivakant Mishra. 2018. Are they our brothers? analysis and detection of religious hate speech in the arabic twittersphere. In 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 69–76. IEEE. [find](https://scholar.google.com/scholar?q=Nuha+Albadi%2C+Maram+Kurdi%2C+and+Shivakant+Mishra%2E+2018%2E+Are+they+our+brothers%3F+analysis+and+detection+of+religious+hate+speech+in+the+arabic+twittersphere%2E+In+2018+IEEE%2FACM+International+Conference+on+A)
5. Pinkesh Badjatiya, Shashank Gupta, Manish Gupta, and Vasudeva Varma. 2017. Deep learning for hate speech detection in tweets. In Proceedings of the 26th International Conference on World Wide Web Companion, pages 759–760. International World Wide Web Conferences Steering Committee. [find](https://scholar.google.com/scholar?q=Pinkesh+Badjatiya%2C+Shashank+Gupta%2C+Manish+Gupta%2C+and+Vasudeva+Varma%2E+2017%2E+Deep+learning+for+hate+speech+detection+in+tweets%2E+In+Proceedings+of+the+26th+International+Conference+on+World+Wide+Web+Comp)
6. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2016. Enriching word vectors with subword information. arXiv preprint arXiv:1607.04606. [↗](https://arxiv.org/abs/1607.04606)
7. Javier Borge-Holthoefer, Walid Magdy, Kareem Darwish, and Ingmar Weber. 2015. Content and network dynamics behind egyptian political polarization on twitter. In Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing, pages 700–711. ACM. [find](https://scholar.google.com/scholar?q=Javier+Borge%2DHolthoefer%2C+Walid+Magdy%2C+Kareem+Darwish%2C+and+Ingmar+Weber%2E+2015%2E+Content+and+network+dynamics+behind+egyptian+political+polarization+on+twitter%2E+In+Proceedings+of+the+18th+ACM+Conference+)
8. Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. In Eleventh International Conference on Web and Social Media (ICWSM), pages 512–515. [find](https://scholar.google.com/scholar?q=Thomas+Davidson%2C+Dana+Warmsley%2C+Michael+Macy%2C+and+Ingmar+Weber%2E+2017%2E+Automated+hate+speech+detection+and+the+problem+of+offensive+language%2E+In+Eleventh+International+Conference+on+Web+and+Social+Medi) [our card](https://md.research.elamur.ai/free/en/papers/davidson-2017-automated-hate-speech-detection-problem/)
9. Nemanja Djuric, Jing Zhou, Robin Morris, Mihajlo Grbovic, Vladan Radosavljevic, and Narayan Bhamidipati. 2015. Hate speech detection with comment embeddings. In Proceedings of the 24th international conference on world wide web, pages 29–30. ACM. [find](https://scholar.google.com/scholar?q=Nemanja+Djuric%2C+Jing+Zhou%2C+Robin+Morris%2C+Mihajlo+Grbovic%2C+Vladan+Radosavljevic%2C+and+Narayan+Bhamidipati%2E+2015%2E+Hate+speech+detection+with+comment+embeddings%2E+In+Proceedings+of+the+24th+international+c)
10. Timothy Jay and Kristin Janschewitz. 2008. The pragmatics of swearing. Journal of Politeness Research. Language, Behaviour, Culture, 4(2):267–288. [find](https://scholar.google.com/scholar?q=Timothy+Jay+and+Kristin+Janschewitz%2E+2008%2E+The+pragmatics+of+swearing%2E+Journal+of+Politeness+Research%2E+Language%2C+Behaviour%2C+Culture%2C+4%282%29%3A267%E2%80%93288%2E)
11. T. Joachims. 1999. Making large-scale SVM learning practical. In B. Schölkopf, C. Burges, and A. Smola, editors, Advances in Kernel Methods - Support Vector Learning, chapter 11, pages 169–184. MIT Press, Cambridge, MA. [find](https://scholar.google.com/scholar?q=T%2E+Joachims%2E+1999%2E+Making+large%2Dscale+SVM+learning+practical%2E+In+B%2E+Sch%C3%B6lkopf%2C+C%2E+Burges%2C+and+A%2E+Smola%2C+editors%2C+Advances+in+Kernel+Methods+%2D+Support+Vector+Learning%2C+chapter+11%2C+pages+169%E2%80%93184%2E+MIT+Pr)
12. Irene Kwok and Yuzhou Wang. 2013. Locate the hate: Detecting tweets against blacks. In Twenty-seventh AAAI conference on artificial intelligence. [find](https://scholar.google.com/scholar?q=Irene+Kwok+and+Yuzhou+Wang%2E+2013%2E+Locate+the+hate%3A+Detecting+tweets+against+blacks%2E+In+Twenty%2Dseventh+AAAI+conference+on+artificial+intelligence%2E)
13. Shervin Malmasi and Marcos Zampieri. 2017. Detecting hate speech in social media. arXiv preprint arXiv:1712.06427. [↗](https://arxiv.org/abs/1712.06427)
14. Hamdy Mubarak and Kareem Darwish. 2019. Arabic offensive language classification on twitter. In International Conference on Social Informatics, pages 269–276. Springer. [find](https://scholar.google.com/scholar?q=Hamdy+Mubarak+and+Kareem+Darwish%2E+2019%2E+Arabic+offensive+language+classification+on+twitter%2E+In+International+Conference+on+Social+Informatics%2C+pages+269%E2%80%93276%2E+Springer%2E)
15. Hamdy Mubarak, Kareem Darwish, Walid Magdy, Tamer Elsayed, and Hend Al-Khalifa. 2020a. 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. [find](https://scholar.google.com/scholar?q=Hamdy+Mubarak%2C+Kareem+Darwish%2C+Walid+Magdy%2C+Tamer+Elsayed%2C+and+Hend+Al%2DKhalifa%2E+2020a%2E+Overview+of+osact4+arabic+offensive+language+detection+shared+task%2E+In+Proceedings+of+the+4th+Workshop+on+Open%2DSo)
16. Hamdy Mubarak, Ammar Rashed, Kareem Darwish, Younes Samih, and Ahmed Abdelali. 2020b. Arabic offensive language on twitter: Analysis and experiments. arXiv preprint arXiv:2004.02192. [↗](https://arxiv.org/abs/2004.02192) [our card](https://md.research.elamur.ai/free/en/papers/mubarak-2021-arabic-offensive-language-twitter-analysis/)
17. Chikashi Nobata, Joel Tetreault, Achint Thomas, Yashar Mehdad, and Yi Chang. 2016. Abusive language detection in online user content. In Proceedings of the 25th international conference on world wide web, pages 145–153. International World Wide Web Conferences Steering Committee. [find](https://scholar.google.com/scholar?q=Chikashi+Nobata%2C+Joel+Tetreault%2C+Achint+Thomas%2C+Yashar+Mehdad%2C+and+Yi+Chang%2E+2016%2E+Abusive+language+detection+in+online+user+content%2E+In+Proceedings+of+the+25th+international+conference+on+world+wide+)
18. Gerard Salton and Chris Buckley. 1990. Improving retrieval performance by relevance feedback. Journal of the American society for information science, 41(4):288–297. [find](https://scholar.google.com/scholar?q=Gerard+Salton+and+Chris+Buckley%2E+1990%2E+Improving+retrieval+performance+by+relevance+feedback%2E+Journal+of+the+American+society+for+information+science%2C+41%284%29%3A288%E2%80%93297%2E)
19. Dawei Yin, Zhenzhen Xue, Liangjie Hong, Brian D Davison, April Kontostathis, and Lynne Edwards. 2009. Detection of harassment on web 2.0. Proceedings of the Content Analysis in the WEB, 2:1–7. [find](https://scholar.google.com/scholar?q=Dawei+Yin%2C+Zhenzhen+Xue%2C+Liangjie+Hong%2C+Brian+D+Davison%2C+April+Kontostathis%2C+and+Lynne+Edwards%2E+2009%2E+Detection+of+harassment+on+web+2%2E0%2E+Proceedings+of+the+Content+Analysis+in+the+WEB%2C+2%3A1%E2%80%937%2E)
20. Marcos Zampieri, Preslav Nakov, Sara Rosenthal, Pepa Atanasova, Georgi Karadzhov, Hamdy Mubarak, Leon Derczynski, Zeses Pitenis, and Çağrı Çöltekin. 2020. Semeval-2020 task 12: Multilingual offensive language identification in social media (offenseval 2020). arXiv preprint arXiv:2006.07235. [↗](https://arxiv.org/abs/2006.07235)