---
title: "Word Representation Models for Arabic Dialect Identification (2022)"
description: "Systems were scored on two test sets: Test-A, with all 18 countries, and Test-B, with a number of… 8 references, cited by 3 papers here."
authors: ["Mahmoud Sobhy", "Ahmed H. Abu El-Atta", "Ahmed A. El-Sawy", "Hamada Nayel"]
year: 2022
venue: "Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP)"
source: "https://aclanthology.org/2022.wanlp-1.52/"
doi: "10.18653/v1/2022.wanlp-1.52"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/sobhy-2022-word-representation-models-arabic-dialect/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/sobhy-2022-word-representation-models-arabic-dialect/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2022](https://md.research.elamur.ai/free/en/papers/2022/page/7/) / Word Representation Models for Arabic Dialect Identification

# Word Representation Models for Arabic Dialect Identification

**Authors:** Mahmoud Sobhy, Ahmed H. Abu El-Atta, Ahmed A. El-Sawy and [Hamada Nayel](https://md.research.elamur.ai/free/en/authors/hamada-nayel/) · **Year:** 2022 · **Venue:** [Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP)](https://md.research.elamur.ai/free/en/venues/wanlp/) ([WANLP 2022](https://md.research.elamur.ai/free/en/venues/wanlp/2022/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2022.wanlp-1.52/) · [PDF](https://aclanthology.org/2022.wanlp-1.52.pdf) · [DOI 10.18653/v1/2022.wanlp-1.52](https://doi.org/10.18653/v1/2022.wanlp-1.52) · [OpenAlex](https://openalex.org/W4385573495)

**Citations:** 3 citations (OpenAlex, 2026-10-14) · [3](https://md.research.elamur.ai/free/en/papers/sobhy-2022-word-representation-models-arabic-dialect/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
This WANLP 2022 system paper from the BFCAI team (Benha University) addresses country-level Arabic dialect identification of tweets in Subtask 1 of the NADI 2022 shared task. The data covers 18 dialects and roughly 20K tweets. Systems were scored on two test sets: Test-A, with all 18 countries, and Test-B, with a number of countries the organizers did not disclose. The team built a deliberately simple pipeline. Tweets are first cleaned of non-Arabic material, letter variants are normalized, and elongated words are collapsed. Each tweet is then represented either with unigram TF/IDF or with Word2Vec (CBOW, trained in gensim on the task's own training corpus). These features feed a range of classical classifiers, from Naïve Bayes variants and SVMs to KNN, trees, Random Forest and MLPs. On the development set, CNB with TF/IDF and MLP with Word2Vec came out on top, so both were submitted. The best run reached macro-F1 of 21.25% on Test-A and 9.71% on Test-B, far behind the winning system (36.48% and 18.95%). The work serves mainly as a baseline showing how hard dialect identification is without pretrained models.

## Key points
- Task: NADI 2022 Subtask 1, country-level dialect ID over 18 Arabic dialects (~20K tweets). The official metric is macro-averaged F1, and Test-B covers an unknown number k of countries.
- Features and classifiers:
  - Features were unigram TF/IDF with smoothed IDF, or Word2Vec CBOW vectors. CBOW was chosen over Skip-gram for its speed.
  - The TF/IDF runs used MNB, CNB, SVM (linear and RBF), KNN, DT, RF and MLP with h = 10, 20, 30 and 40.
  - The Word2Vec runs used the same set minus the Naïve Bayes models.
- Development-set macro-F1:
  - CNB+TF/IDF scored 20.48 and MLP(20)+Word2Vec scored 20.19.
  - SVM (RBF)+Word2Vec had the highest accuracy (42.62) but only 12.80 macro-F1.
- Test-A macro-F1:
  - MLP(30)+WE: 21.25%
  - CNB+TF/IDF: 21.16%
  - MLP(20)+WE: 21.13%
- Test-B macro-F1:
  - MLP(30)+WE: 9.71%
  - MLP(20)+WE: 9.14%
  - CNB+TF/IDF: 7.78%
- The authors conclude that embeddings beat TF/IDF. They name BERT, AraBERT and MarBERT, together with transfer learning, as future work.

## Interesting ideas
- Collapse any character repeated more than twice down to two copies (e.g. stretched "wonderful"), which shrinks the vocabulary without wiping out the emphasis signal entirely.
- Accuracy and macro-F1 told very different stories on imbalanced labels (high-accuracy SVM, weak macro-F1), so model selection should rely on the class-balanced metric.
- Complement Naïve Bayes was picked deliberately as a cheap, transparent model that copes with skewed class distributions, and it nearly matched the MLP.
- Evaluating on a second test set with an unknown subset of labels exposed how much performance collapses when the label distribution shifts (21% to under 10%).
- Training embeddings only on the small task corpus already beat sparse TF/IDF on macro-F1, suggesting that even in-domain distributional vectors add generalization over raw counts.

## For Mova

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

- **I1** (tokens): Add a reversible elongation-normalization step for English social text ("soooo good"), keeping the original surface form plus a flag the tagger and lemmatizer can explain.
- **I2** (learn): Report macro-averaged scores per label class (rare dependency relations, rare frame roles) alongside accuracy, so majority-class gains are not mistaken for real progress.
- **I3** (learn): Keep a simple, explainable count-based baseline next to the perceptron and gparse models as a sanity check; if it nearly ties, the bigger model has not yet learned anything structural.
- **I4** (answer): Build held-out test variants with shifted or restricted label inventories, such as fables from unseen genres or question types, to measure robustness rather than in-distribution fit.
- **I5** (tree): In-domain distributional vectors, trained cheaply on the target genre, could serve as backoff features for rare words in the perceptron parser when converting or parsing new UD dialects.

*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

- [Benha University](https://md.research.elamur.ai/free/en/universities/benha-university/) — University, Banhā, Egypt · 14 works · 89 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works

## Cite

Cite as: Mahmoud Sobhy, Ahmed H. Abu El-Atta, Ahmed A. El-Sawy and Hamada Nayel. 2022. Word Representation Models for Arabic Dialect Identification. Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP).

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

```bibtex
@inproceedings{sobhy2022word,
  title = {{Word Representation Models for Arabic Dialect Identification}},
  author = {Mahmoud Sobhy, Ahmed H. Abu El-Atta, Ahmed A. El-Sawy and Hamada Nayel},
  booktitle = {Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP)},
  year = {2022},
  doi = {10.18653/v1/2022.wanlp-1.52},
  url = {https://doi.org/10.18653/v1/2022.wanlp-1.52}
}
```

- [← previous Universal Dependencies for Punjabi](https://md.research.elamur.ai/free/en/papers/arora-2022-universal-dependencies-punjabi/)
- [next → Affrication as the cause of /s/-retraction: Evidence from Manchester English](https://md.research.elamur.ai/free/en/papers/bailey-2022-affrication-as-cause-s-retraction/)
- [↑ back to papers of 2022](https://md.research.elamur.ai/free/en/papers/2022/page/7/)

## References

8 references, 7 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. 2021a. [ARBERT & MARBERT: Deep bidirectional transformers for Arabic](https://md.research.elamur.ai/free/en/papers/abdul-mageed-2021-arbert-marbert-deep-bidirectional-transformers/). 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. [↗](https://doi.org/10.48550/arxiv.2101.01785)
2. Muhammad Abdul-Mageed, Chiyu Zhang, Houda Bouamor, and Nizar Habash. 2020. [NADI 2020: The first nuanced Arabic dialect identification shared task](https://md.research.elamur.ai/free/en/papers/mageed-2020-nadi-2020-first-nuanced-arabic/). In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 97–110, Barcelona, Spain (Online). Association for Computational Linguistics. [↗](https://doi.org/10.48550/arxiv.2010.11334)
3. Muhammad Abdul-Mageed, Chiyu Zhang, Abdel-Rahim Elmadany, Houda Bouamor, and Nizar Habash. 2021b. [NADI 2021: The second nuanced Arabic dialect identification shared task](https://md.research.elamur.ai/free/en/papers/mageed-2021-nadi-2021-second-nuanced-arabic/). In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 244–259, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. [↗](https://doi.org/10.48550/arxiv.2103.08466)
4. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. [BERT: Pre-training of deep bidirectional transformers for language understanding](https://md.research.elamur.ai/free/en/papers/devlin-2019-bert-pre-training-deep-bidirectional/). 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. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/n19-1423)
5. Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. [Efficient estimation of word representations in vector space](https://md.research.elamur.ai/free/en/papers/mikolov-2013-efficient-estimation-word-representations-vector/). In 1st International Conference on Learning Representations, ICLR 2013, Scottsdale, Arizona, USA, May 2-4, 2013, Workshop Track Proceedings. [↗](https://doi.org/10.48550/arxiv.1301.3781)
6. Hamada Nayel. 2020. [NAYEL at SemEval-2020 task 12: TF/IDF-based approach for automatic offensive language detection in Arabic tweets](https://md.research.elamur.ai/free/en/papers/nayel-2020-nayel-semeval-2020-task-12/). In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 2086–2089, Barcelona (online). International Committee for Computational Linguistics.
7. Muhammad Abdul-Mageed, Chiyu Zhang, Abdel-Rahim Elmadany, Houda Bouamor, and Nizar Habash. 2022. [NADI 2022: The Third Nuanced Arabic Dialect Identification Shared Task](https://md.research.elamur.ai/free/en/papers/mageed-2022-nadi-2022-third-nuanced-arabic/). In Proceedings of the Seven Arabic Natural Language Processing Workshop (WANLP 2022).
8. Ahmed Aliwy, Hawraa Taher, and Zena AboAltaheen. 2020. Arabic dialects identification for all Arabic [find](https://scholar.google.com/scholar?q=Ahmed+Aliwy%2C+Hawraa+Taher%2C+and+Zena+AboAltaheen%2E+2020%2E+Arabic+dialects+identification+for+all+Arabic)

*Updated 1 November 2026*