Elamur Research⌕
☰
home / papers / 2021 / Arabic Dialect Identification based on a Weighted Concatenation of TF-IDF Features

Arabic Dialect Identification based on a Weighted Concatenation of TF-IDF Features

Paper · 2021 · Mohamed Lichouri, Mourad Abbas, Khaled Lounnas et al.

Summary

This paper presents the CRSTDLA team's entry to the NADI 2021 shared task. The task is to identify which of 21 Arab countries a tweet comes from, in two subtasks: tweets in Modern Standard Arabic (subtask 1.1) and tweets in dialectal Arabic (subtask 1.2). The authors build a classical pipeline. They clean the tweets at the surface level and apply light morphology, mainly ISRI stemming. They then build TF-IDF vectors from three views of the text: word n-grams, character n-grams and boundary-aware character n-grams (char_wb). These views are joined through a feature union in which each view can carry its own weight. Eight scikit-learn classifiers are compared. The three strongest (LSVC, RidgeClassifier, SGD) are combined by majority vote, and ADASYN oversampling is tried against class imbalance. Roughly 200 n-gram configurations were explored. Final test macro-F1 is 14.87% for MSA and 21.49% for DA, a little below the task averages. The paper documents how far a carefully tuned sparse-feature baseline can go on fine-grained dialect identification.

Key points

  • Data: NADI 2021 provides an MSA subset and a DA subset, covering 21 countries and 100 provinces. Each subset is split into 21,000 train, 5,000 dev and 5,000 test tweets, and some tweets contain only a single word.
  • Preprocessing: the best cleaning setup removed emojis, diacritics and repeated words, then applied the ISRI Arabic Stemmer. NLTK lemmatization and POS tagging were also tested.
  • Unweighted feature union: SGD performed best on dev, with F1 15.28 (MSA) and 21.85 (DA).
  • Weighted union with majority vote over SVC/SGD/RDG reached 14.90 (MSA) and 20.09 (DA) on dev. Weighting alone did not clearly beat the unweighted union.
  • Final systems combined view weights such as (1.3, 1.3, 1.0), a tuned SGD (elasticnet penalty, modified_huber loss) and, for DA, ADASYN. They scored F1 14.87 (MSA test) and 21.49 (DA test), against task averages of 18.70 and 24.23.

Interesting ideas

  • I1. The same text is split into several parallel views (words, raw char n-grams, word-bounded char n-grams), and each view gets its own scalar weight before concatenation, so the views' relative influence can be tuned separately.
  • I2. Every combination of cleaning steps is tested exhaustively, and only the winning subset is kept. Steps that seem helpful, like lemmatization, did not survive this search.
  • I3. The three best different linear classifiers are combined by simple majority vote, which needs no extra training and smooths out individual model errors.
  • I4. The authors use corpus statistics, such as very short one-word tweets, to bound the n-gram range rather than guessing it.
  • I5. Dev gains from heavy tuning shrank noticeably on test (e.g., 16.30 dev vs 14.50 test for MSA), a warning about overfitting through configuration search.

For Mova

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

  • I1 (tokens): Separate weights per feature family (word form, suffix chars, shape) in the tagger would make it visible which evidence drives each tag decision.
  • I2 (learn): Ablation over normalization and preprocessing choices could become a standard evaluation step, so that every preprocessing rule must earn its place on dev data.
  • I3 (learn): This supports agreement-based silver data, where a token's label or arc is kept only when most witness parsers concur.
  • I4 (tokens): Length and character statistics from a treebank could set the suffix and prefix window sizes used in morphology features.
  • I5 (learn): After large configuration sweeps, a held-out set that is never touched would show how much of a gain is real.

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

💡 Concepts 1Key ideas and methods explained on our concept pages.TF-IDF · 26 works

Cite

Cite as Mohamed Lichouri, Mourad Abbas, Khaled Lounnas, Besma Benaziz and Aicha Zitouni. 2021. Arabic Dialect Identification based on a Weighted Concatenation of TF-IDF Features. Proceedings of the Sixth Arabic Natural Language Processing Workshop.

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

@inproceedings{lichouri2021arabic,
  title = {{Arabic Dialect Identification based on a Weighted Concatenation of TF-IDF Features}},
  author = {Mohamed Lichouri, Mourad Abbas, Khaled Lounnas, Besma Benaziz and Aicha Zitouni},
  booktitle = {Proceedings of the Sixth Arabic Natural Language Processing Workshop},
  year = {2021},
  url = {https://aclanthology.org/2021.wanlp-1.33/}
}

References

17 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mourad Abbas, Mohamed Lichouri, and Abed Alhakim Freihat. 2019. St madar 2019 shared task: Arabic fine-grained dialect identification. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 269–273. ↗
  2. Muhammad Abdul-Mageed, Chiyu Zhang, Abdel-Rahim Elmadany, Houda Bouamor, and Nizar Habash. 2021. NADI 2021: The Second Nuanced Arabic Dialect Identification Shared Task. In Proceedings of the Sixth Arabic Natural Language Processing Workshop (WANLP 2021). ↗
  3. Kathrein Abu Kwaik and Motaz K Saad. 2019. Arbdi-alectid at madar shared task 1: Language modelling and ensemble learning for fine grained arabic dialect identification. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, Proceedings of the Fourth Arabic Natural Language Processing Workshop. Association for Computational Linguistics.
  4. Ahmad Beltagy, Abdelrahman Wael, and Omar ElSherief. 2020. Arabic dialect identification using bert-based domain adaptation. arXiv preprint arXiv:2011.06977. ↗
  5. Houda Bouamor, Sabit Hassan, and Nizar Habash. 2019. The madar shared task on arabic fine-grained dialect identification. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 199–207.
  6. Abdellah El Mekki, Ahmed Alami, Hamza Alami, Ahmed Khoumsi, and Ismail Berrada. 2020. Weighted combination of bert and n-gram features for nuanced arabic dialect identification. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 268–274.
  7. Kamel Gaanoun and Imade Benelallam. 2020. Arabic dialect identification: An arabic-bert model with data augmentation and ensembling strategy. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 275–281. ↗
  8. Guillaume Lemaı̂tre, Fernando Nogueira, and Christos K. Aridas. 2017. Imbalanced-learn: A python toolbox to tackle the curse of imbalanced datasets in machine learning. Journal of Machine Learning Research, 18(17):1–5. find
  9. Mohamed Lichouri and Mourad Abbas. 2020a. Simple vs oversampling-based classification methods for fine grained arabic dialect identification in twitter. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 250–256.
  10. Mohamed Lichouri and Mourad Abbas. 2020b. Speechtrans@ smm4h’20: Impact of preprocessing and n-grams on automatic classification of tweets that mention medications. In Proceedings of the Fifth Social Media Mining for Health Applications Workshop & Shared Task, pages 118–120. find
  11. Mohamed Lichouri, Mourad Abbas, and Besma Benaziz. 2020. Profiling fake news spreaders on twitter based on tfidf features and morphological process. find
  12. Mohamed Lichouri, Mourad Abbas, Abed Alhakim Freihat, and Dhiya El Hak Megtouf. 2018. Word-level vs sentence-level language identification: Application to algerian and arabic dialects. Procedia Computer Science, 142:246–253. find
  13. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830. ↗
  14. Pavel Přibáň and Stephen Taylor. 2019. Zcu-nlp at madar 2019: Recognizing arabic dialects. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 208–213.
  15. Ahmad Ragab, Haitham Seelawi, Mostafa Samir, Abdelrahman Mattar, Hesham Al-Bataineh, Mohammad Zaghloul, Ahmad Mustafa, Bashar Talafha, Abed Alhakim Freihat, and Hussein Al-Natsheh. 2019. Mawdoo3 ai at madar shared task: Arabic fine-grained dialect identification with ensemble learning. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 244–248.
  16. Bashar Talafha, Mohammad Ali, Muhy Eddin Za’ter, Haitham Seelawi, Ibraheem Tuffaha, Mostafa Samir, Wael Farhan, and Hussein T Al-Natsheh. 2020. Multi-dialect arabic bert for country-level dialect identification. arXiv preprint arXiv:2007.05612. ↗
  17. Samia Touileb. 2020. Ltg-st at nadi shared task 1: Arabic dialect identification using a stacking classifier. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 313–319.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lichouri-2021-arabic-dialect-identification-based-weighted/