Elamur Research⌕
☰
home / papers / 2015 / Exploiting Out-of-Domain Data Sources for Dialectal Arabic Statistical Machine Translation

Exploiting Out-of-Domain Data Sources for Dialectal Arabic Statistical Machine Translation

Paper · 2015 · Katrin Kirchhoff, Bing Zhao and Wen Wang

Summary

Translating dialectal Arabic is hard because dialects are rarely written. Building parallel data means recording speech, transcribing it, and translating it, which is slow and costly. The authors target Iraqi Arabic (IA), where the in-domain resource is the Transtac corpus. They ask whether useful IA–English training pairs can be found inside large out-of-domain parallel collections: 22 LDC corpora of MSA, Egyptian and Levantine Arabic, roughly 5.7M source words. Instead of running a dialect identifier, they score candidate sentences by how well they match n-gram features from an IA sample. They compare two selectors: the Moore–Lewis cross-entropy ranking and a submodular objective built on tf-idf-weighted n-gram coverage, optimized greedily. A very small selected set of about 7k sentences gives small but consistent BLEU gains, and the submodular method is slightly better than cross-entropy. A side experiment uses machine-translated IA telephone speech as extra parallel data and finds no benefit. The work shows that a tiny, targeted slice of mismatched data can help a low-resource variety.

Key points

  • In-domain data is the Iraqi Arabic Transtac corpus: about 7.6M training words, with small tune, dev and three test sets of 8k–64k words.
  • The out-of-domain pool is 22 LDC corpora, including GALE, TIDES, BOLT, NIST OpenMT, the GALE Levantine treebank, Egyptian (ARZ) word alignments and crowd-translated dialect web text, totalling about 5.69M words.
  • The submodular function sums a concave transform of tf-idf-weighted n-gram counts, so repeated features add less and less value. Greedy selection is guaranteed to reach at least about 63% (1−1/e) of the optimal value.
  • Adding 7k submodular-selected sentences raised EN→IA BLEU by up to 0.5 absolute (e.g., test2 rose from 32.2 to 32.7). Larger selections of up to 20k sentences did not help.
  • The selected data came mostly from GALE broadcast conversations (47%), the dialectal web corpus (35.7%) and BOLT training data (9.9%). Auto-translated Iraqi CTS speech (LDC2006T16) made up 80% of a re-run selection but gave no real gain.

Interesting ideas

  • I1. The authors skip dialect identification and instead pull material whose n-grams resemble a small sample of the target variety, which also tends to match topic and register.
  • I2. Their submodular objective rewards covering many relevant features and penalizes redundancy, unlike per-sentence ranking that can keep picking near-duplicates.
  • I3. A corpus labelled as the standard language can still contain dialect segments through code-switching, especially in conversational broadcast material.
  • I4. Self-translated data matched the target well on the source side but added noise, and the authors suggest keeping only high-confidence outputs.
  • I5. A small, focused addition beat larger additions (7k sentences helped, 20k did not).

For Mova

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

  • I1 (learn): Silver training sentences for a new genre (fables, word problems) could be chosen from large corpora by n-gram similarity to a small seed set, rather than by a genre classifier.
  • I2 (learn): A coverage-based greedy picker could choose which sentences to send to the LLM teacher or witness parsers, so the labelling budget reaches varied constructions rather than repeats.
  • I3 (tree): When harvesting informal or spoken-style English for treebank or UD dialect conversion work, transcribed conversation inside "formal" corpora could be mined at the sentence level.
  • I4 (learn): This is a useful negative control: self-generated silver data should be filtered by confidence or witness agreement, because a high share of selected data alone does not show it helps.
  • I5 (learn): Evaluation could sweep the size of added silver data instead of assuming more is better, and pick the point where held-out accuracy peaks.

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 2Research subjects and languages the work is about.Machine translation · 1,087 worksSemitic languages · 2,908 works

Cite

Cite as Katrin Kirchhoff, Bing Zhao and Wen Wang. 2015. Exploiting Out-of-Domain Data Sources for Dialectal Arabic Statistical Machine Translation. arXiv:1509.01938.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1509.01938 · arXiv 1509.01938

@misc{kirchhoff2015exploiting,
  title = {{Exploiting Out-of-Domain Data Sources for Dialectal Arabic Statistical Machine Translation}},
  author = {Katrin Kirchhoff, Bing Zhao and Wen Wang},
  year = {2015},
  doi = {10.48550/arxiv.1509.01938},
  eprint = {1509.01938},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1509.01938}
}

References

22 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Al-Sabbagh, R. and R. Girju. 2012. A supervised POS tagger for written [Habash et al.2012] Habash, N., R. Eskander, and Arabic social networking corpora. In Proceedings A. Hawwari. 2012. A morphological analyzer for of KONVENS. Egyptian Arabic. In Proceedings of the Twelfth Meeting of the Special Interest Group on Com- find
  2. Alorifi, F.S. 2008. Automatic Identifica-putational Morphology and Phonology (SIGMOR-tion of Arabic Dialects USING Hidden Markov Mod-PHON2012), page 19. els. Ph.D. thesis, Swanson School of Engineering, University of Pittsburgh. find
  3. Aminian, M., M. Ghoneim, and M. Diab. 2014. Handling OOV words in dialectal Arabic to English machine translation. In EMNLP Workshop on Language Technology for Closely Related Languages and Language Variants (LT4CloseLang), page 99108. find
  4. Axelrod, A., X. He, and J. Gao. 2011. Domain adaptation via pseudo in-domain data selection. In Proceedings of EMNLP, pages 355– 362, Edinburgh, Scotland. translation using targeted clarification questions. In Proceedings of ICASSP. neural language models: Experiments in machine translation. In Proceedings of ACL.
  5. Feige, U. 1998. A threshold of ln n for approximating set cover. Journal of the ACM (JACM), 45(4):634–652. matics, volume 58. Elsevier Science. ↗
  6. Habash, N., O. Rambow, and G. Kiraz. 2005. Morphological analysis and generation for Arabic dialects. In Proceedings of the ACL Workshop on Computational Approaches to Semitic Languages.
  7. Habash, N., R. Roth, O. Rambow, [Sadat et al.2014] Sadat, F., F. Kazemi, and A. Farzindar. 2014. Automatic identification of Arabic language R. Eskander, and N. Tomeh. 2013. Morphological varieties and dialects in social media. In Proceedings analysis and disambiguation for dialectal Arabic. In of the Second Workshop on Natural Language Pro-Proceedings of NAACL, pages 426–432. cessing for Social Media (SocialNLP), page 2227. find
  8. Habash, N. 2010. Introduction to Arabic natural language processing. Synthesis Lectures on [Salloum and Habash2011] Salloum, W. and N. Habash. 2011. Dialectal to standard Arabic paraphrasing to Human Language Technologies, 3:1–187. improve Arabic-English statistical machine transla- ↗
  9. Jegelka, S. and J.A. Bilmes. tion. In Proceedings of the First Workshop on Algo-2011. Submodularity beyond submodular energies: rithms and Resources for Modelling of Dialects and coupling edges in graph cuts. In Computer Vision Language Varieties, pages 10–21. and Pattern Recognition (CVPR), Colorado Springs, find
  10. Salloum, W. and N. Habash. CO, June. 2013. Dialectal Arabic to English machine translation: Pivoting through modern standard Arabic. In
  11. Kirchhoff, K. and Proceedings of NAACL. J. Bilmes. 2014. Submodularity for data selection in statistical machine translation. In Proceedings of find
  12. Wei, K., Y. Liu, K. Kirchhoff, and EMNLP, pages 131–141. J. Bilmes. 2013. Using document summarization techniques for speech data subset selection. In Pro- find
  13. Kirchhoff, K., Y.C. Tam ceedings of NAACL. dn C. Richey, and W. Wang. 2015. Morphological modeling for machine translation of English-iraqi find
  14. Zaidan, O. and Arabic spoken dialogs. In Proceedings of NAACL. C. Callison-Burch. 2014. Arabic dialect identification. Computational Linguistics, 40:171–202.
  15. Krause, A. and C. Guestrin. 2011. Submodularity and its applications in optimized information gathering. ACM Transactions on [Zbib et al.2012] Zbib, Rabih, Erika Malchiodi, Jacob Devlin, David Stallard, Spyros Matsoukas, Richard Intelligent Systems and Technology, 2(4). Schwartz, John Makhoul, Omar F. Zaidan, and Chris Callison-Burch. 2012. Machine translation of Ara- ↗
  16. Krause, A., H.B. McMahan, bic dialects. In Proceedings of the 2012 Conference C. Guestrin, and A. Gupta. 2008. Robust sub-of the North American Chapter of the Association for modular observation selection. Journal of Machine Computational Linguistics: Human Language Tech-Learning Research, 9:2761–2801. nologies, page 4959. find
  17. Lin, H. and J. Bilmes. 2012. Learning mixtures of submodular shells with appli- [Zhao et al.2014] Zhao, B., Y.-C. Tam, and J. Zheng. 2014. An autoencoder with bilingual sparse features cation to document summarization. In Uncertainty for improved statistical machine translation. In Pro-in Artifical Intelligence (UAI), Catalina Island, USA, ceedings of ICASSP. July. AUAI. find
  18. Maamouri, M., A. Bies, T. Buckwalter, M. Diab, N. Habash, O. Rambow, and D. Tabessi. 2006. Developing and using a pilot dialectal Arabic treebank. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC).
  19. Mediani, M., J. Winebarger, and A. Waibel. 2014. Improving in-domain data selection for small in-domain sets. In Proceedings of IWSLT. ↗
  20. Moore, R. and W. Lewis. 2010. Intelligent selection of language model training data. In Proceedings EMNLP.
  21. Narasimhan, M. and J. Bilmes. 2007. Local search for balanced submodular clustering. In Proceedings of IJCAI. find
  22. Nemhauser, G.L., L.A. Wolsey, and M.L. Fisher. 1978. An analysis of approximations for maximizing submodular set functions i. Mathematical Programming, 14:265–294. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/kirchhoff-2015-exploiting-out-domain-data-sources/