Elamur Research⌕
☰
home / papers / 2017 / Learning from Relatives: Unified Dialectal Arabic Segmentation

Learning from Relatives: Unified Dialectal Arabic Segmentation

Paper · 2017 · Younes Samih, Mohamed Eldesouki, Mohammed Attia et al.

Summary

Social-media Arabic is mostly written in regional dialects. Earlier segmenters were built for one dialect at a time, so text first had to pass through dialect identification, which can be less than 90% accurate. The authors ask two questions: how well does a segmenter trained on one dialect handle the others, and can one model trained on pooled data replace per-dialect models?

They built small, manually segmented tweet corpora for Egyptian, Levantine, Gulf and Maghrebi. Annotation was done on raw spelling rather than CODA-normalized text, and the convention splits off nearly all prefixes and suffixes. Two segmenters were compared: an SVM ranker and a bi-LSTM-CRF sequence labeler. Each was tested with no lookup, with a dialectal lookup list, and with dialectal plus MSA lookup, against a Farasa MSA baseline.

The main result is that a joint bi-LSTM-CRF with DA+MSA lookup gave the best accuracy for every dialect, without any dialect identification. The cross-dialect experiments also suggest that dialect closeness follows geography. This matters for any tool that must handle a family of related, non-standardized language varieties.

Key points

  • Data collection:
    • 175 million Arabic tweets were collected in March 2014.
    • They were filtered by author location (using a gazetteer) and by 10 strongly dialectal words per dialect.
    • Native speakers kept 350 heavily dialectal tweets per dialect.
    • Corpus sizes: EGY 6,721, LEV 6,648, GLF 6,844 and MGR 5,495 tokens.
  • Evaluation setup: 5-fold cross-validation with 70/10/20 train/dev/test splits. Dropout and early stopping were used because the data is tiny.
  • Best results: the joint bi-LSTM-CRF with DA+MSA lookup reached 95.3 (EGY), 93.9 (LEV), 93.1 (GLF) and 91.4 (MGR).
  • Joint SVM: it lost 0.3–1.3 points against dialect-matched training. Restricting affixes with a dialect-ID oracle helped, but only EGY surpassed its single-dialect model.
  • Cross-dialect behaviour:
    • Without lookup, the LSTM wins in-dialect, while the SVM generalizes better across dialects.
    • BAMA analysis puts GLF closest to MSA and MGR furthest from it.
    • MGR is roughly equally distant from the other three dialects.

Interesting ideas

  • I1. Instead of routing input through a variety classifier, pool the training data of related varieties and let one model learn their shared patterns; this removes a fragile upstream step that can fail before the main model even runs.
  • I2. A full train-on-X/test-on-Y matrix serves two purposes at once: it measures how related the data sources are and shows which ones can usefully augment the others.
  • I3. Layering a lookup of known segmentations (in-domain first, then standard-language) on top of a statistical model gives steady gains and is easy to inspect.
  • I4. Annotating raw, unnormalized text, without forcing it into a spelling standard first, makes the tool robust to the variant forms it will actually meet.
  • I5. Pairing the joint model with an oracle that limits each variety's affix inventory separates how much error comes from not knowing the variety and how much comes from the model itself.

For Mova

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

  • I1 (tree): gparse could be trained on pooled English treebanks from different genres and registers, rather than detecting the genre and switching parsers, with a check that each genre's LAS does not drop.
  • I2 (learn): building such a matrix over our treebanks and silver sets would show which sources transfer well, which should be weighted down, and where domain gaps lie.
  • I3 (tokens): the en crate could let a curated lexicon of known analyses override the tagger and lemmatizer, with each override logged so the decision stays explainable.
  • I4 (tokens): the tokenizer could be trained on English with merged words, elongations and misspellings kept as they are, splitting them in the segmentation instead of normalizing them away beforehand.
  • I5 (learn): oracle-constrained runs, for example with the gold genre or a gold lexicon, could serve as upper-bound diagnostics before investing in a real classifier.

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

🏢GoogleMountain View · United States430 works · 424,363 citations🎓Hamad bin Khalifa UniversityDoha · Qatar141 works · 14,175 citations🎓Heinrich Heine University DüsseldorfDüsseldorf · Germany33 works · 1,487 citations

 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

Cite

Cite as Younes Samih, Mohamed Eldesouki, Mohammed Attia, Kareem Darwish, Ahmed Abdelali, Hamdy Mubarak and Laura Kallmeyer. 2017. Learning from Relatives: Unified Dialectal Arabic Segmentation. Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/k17-1043

@inproceedings{samih2017learning,
  title = {{Learning from Relatives: Unified Dialectal Arabic Segmentation}},
  author = {Younes Samih, Mohamed Eldesouki, Mohammed Attia, Kareem Darwish, Ahmed Abdelali, Hamdy Mubarak and Laura Kallmeyer},
  booktitle = {Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017)},
  year = {2017},
  doi = {10.18653/v1/k17-1043},
  url = {https://doi.org/10.18653/v1/k17-1043}
}

References

32 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. 2013. Speech recognition with deep recurrent neural networks. In Acoustics, speech and signal processing (icassp), 2013 ieee international conference on. IEEE, pages 6645–6649. ↗
  2. Ahmed Abdelali, Kareem Darwish, Nadir Durrani, and Hamdy Mubarak. 2016. Farasa: A fast and furious segmenter for arabic. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations. Association for Computational Linguistics, San Diego, California, pages 11–16.
  3. Alex Graves and Jürgen Schmidhuber. 2005. Frame-wise phoneme classification with bidirectional lstm and other neural network architectures. Neural Networks 18(5):602–610. find
  4. Nizar Habash, Mona T Diab, and Owen Rambow. 2012. Conventional orthography for dialectal arabic. In LREC. pages 711–718.
  5. Mohammed Attia, Pavel Pecina, Antonio Toral, Lamia Tounsi, and Josef van Genabith. 2011. An open-source finite state morphological transducer for modern standard arabic. In Proceedings of the 9th International Workshop on Finite State Methods and Natural Language Processing. Association for Computational Linguistics, pages 125–133.
  6. Nizar Habash, Ryan Roth, Owen Rambow, Ramy Eskander, and Nadi Tomeh. 2013. Morphological analysis and disambiguation for dialectal arabic. In Hlt-Naacl. pages 426–432.
  7. Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. 2012. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580 . ↗
  8. Yoshua Bengio, Patrice Simard, and Paolo Frasconi. 1994. Learning long-term dependencies with gradient descent is difficult. IEEE transactions on neural networks 5(2):157–166. ↗
  9. Fadi Biadsy, Julia Hirschberg, and Nizar Habash. 2009. Spoken arabic dialect identification using phonotactic modeling. In Proceedings of the EACL 2009 Workshop on Computational Approaches to Semitic Languages. Association for Computational Linguistics, Stroudsburg, PA, USA, Semitic ’09, pages 53– 61.
  10. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation 9(8):1735–1780. ↗
  11. Houda Bouamor, Nizar Habash, and Kemal Oflazer. 2014. A multidialectal parallel corpus of arabic. In Nicoletta Calzolari (Conference Chair), Khalid Choukri, Thierry Declerck, Hrafn Loftsson, Bente Maegaard, Joseph Mariani, Asuncion Moreno, Jan Odijk, and Stelios Piperidis, editors, Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14). European Language Resources Association (ELRA), Reykjavik, Iceland.
  12. Najib Ismail Jarad. 2014. The grammaticalization of the motion verb ”ra” as a prospective aspect marker in syrian arabic. Al-’Arabiyya 47:101–118. find
  13. Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional LSTM-CRF models for sequence tagging. CoRR abs/1508.01991. ↗
  14. Thorsten Joachims. 2006. Training linear svms in linear time. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, pages 217–226. find
  15. John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proc. ICML. find
  16. Tim Buckwalter. 2004. Buckwalter arabic morphological analyzer version 2.0 . ↗
  17. Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer. 2016. Neural architectures for named entity recognition. arXiv preprint arXiv:1603.01360 . ↗
  18. Rich Caruana, Steve Lawrence, and Lee Giles. 2000. Overfitting in neural nets: Backpropagation, conjugate gradient, and early stopping. In NIPS. pages 402–408. find
  19. Kareem Darwish, Hassan Sajjad, and Hamdy Mubarak. 2014. Verifiably effective arabic dialect identification. In EMNLP. pages 1465–1468.
  20. Zachary C Lipton, David C Kale, Charles Elkan, and Randall Wetzell. 2015. A critical review of recurrent neural networks for sequence learning. CoRR abs/1506.00019. ↗
  21. Mohamed Eldesouki, Fahim Dalvi, Hassan Sajjad, and Kareem Darwish. 2016. Qcri@ dsl 2016: Spoken arabic dialect identification using textual features. VarDial 3 page 221.
  22. Xuezhe Ma and Eduard Hovy. 2016. End-to-end sequence labeling via bi-directional lstm-cnns-crf. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Berlin, Germany, pages 1064–1074. http://www.aclweb.org/anthology/P16-1101. ↗
  23. Mohamed Maamouri, Ann Bies, Seth Kulick, Michael Ciul, Nizar Habash, and Ramy Eskander. 2014. Developing an egyptian arabic treebank: Impact of dialectal morphology on annotation and tool development. In LREC. pages 2348–2354. find
  24. Emad Mohamed, Behrang Mohit, and Kemal Oflazer. 2012. Annotating and learning morphological segmentation of egyptian colloquial arabic. In LREC. pages 873–877.
  25. Will Monroe, Spence Green, and Christopher D Manning. 2014. Word segmentation of informal arabic with domain adaptation. In ACL (2). pages 206–211.
  26. Hamdy Mubarak and Kareem Darwish. 2014. Using twitter to collect a multi-dialectal corpus of arabic. In Proceedings of the EMNLP 2014 Workshop on Arabic Natural Language Processing (ANLP). pages 1–7.
  27. Arfath Pasha, Mohamed Al-Badrashiny, Mona Diab, Ahmed El Kholy, Ramy Eskander, Nizar Habash, Manoj Pooleery, Owen Rambow, and Ryan M Roth. 2014. Madamira: A fast, comprehensive tool for morphological analysis and disambiguation of Arabic. Proc. LREC .
  28. Maria Persson. 2008. The role of the b-prefix in gulf arabic dialects as a marker of future, intent and/or irrealis 8:26–52. find
  29. Younes Samih, Mohammed Attia, Mohamed Eldesouki, Hamdy Mubarak, Ahmed Abdelali, Laura Kallmeyer, and Kareem Darwish. 2017. A neural architecture for dialectal arabic segmentation. WANLP 2017 (co-located with EACL 2017) page 46.
  30. Mike Schuster and Kuldip K Paliwal. 1997. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11):2673–2681. ↗
  31. Omar F Zaidan and Chris Callison-Burch. 2014. Arabic dialect identification. Computational Linguistics 40(1):171–202. ↗
  32. Rabih Zbib, Erika Malchiodi, Jacob Devlin, David Stallard, Spyros Matsoukas, Richard Schwartz, John Makhoul, Omar F. Zaidan, and Chris Callison-Burch. 2012. Machine translation of arabic dialects. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, Stroudsburg, PA, USA, NAACL HLT ’12, pages 49–59.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/samih-2017-learning-relatives-unified-dialectal-arabic/