Elamur Research⌕
☰
home / papers / 2013 / Dudley North visits North London: Learning When to Transliterate to Arabic

Dudley North visits North London: Learning When to Transliterate to Arabic

Paper · 2013 · Mahmoud Azab, Houda Bouamor, Behrang Mohit et al.

Summary

When translating English into Arabic, a named-entity token must either be rendered by meaning (translated) or by sound (transliterated). Ordinary words inside names cause trouble: "North" in "Dudley North" should be transliterated, but in "North London" it should be translated. The authors treat this choice as a token-level binary classification task. They first built labeled bilingual NE lexicons automatically. About 57K pairs come from a word-aligned news parallel corpus, and about 4K come from Wikipedia titles. Labels were assigned with a transliteration-similarity score, and only the uncertain cases received manual handling. An SVM classifier over 32 features reaches roughly 90% accuracy on news and up to 86% on Wikipedia, compared with a majority-class baseline near 57%. Inside a Moses English→Arabic system, the classifier gives a small but significant BLEU gain. On diverse Wikipedia text, it routes most of the would-be-mistranslated name tokens to transliteration. The work shows that a cheap, explicit decision layer can repair a systematic failure of MT on names.

Key points

  • Data construction
    • English sides of a parallel corpus were tagged with the UIUC NE tagger (PER/LOC/ORG/MISC).
    • Arabic spans were projected through word alignments and cleaned with POS checks and heuristic filters, yielding about 57K news NE pairs.
    • A separate 4K Wikipedia lexicon contains only entities absent from the parallel corpus.
  • Labeling
    • Labels come from the Freeman et al. similarity score: above 0.6 means transliteration, below 0.5 means translation.
    • The roughly 5K unique tokens in between were transliterated with Microsoft Maren and then rescored.
  • Classifier and accuracy
    • The model is LIBSVM with token-based, semantic and contextual features.
    • C_news scores 90.40 on Test_News and 84.10 on Test_Wiki.
    • C_diverse (55K news plus 4.6K Wikipedia training NEs) scores 90.42 and 86.00.
    • The baseline is about 57.
  • Feature importance: the NE category feature contributes more than the token's local context.
  • Extrinsic evaluation
    • In 10 Wikipedia articles (98,197 terms), 624 of 1031 NE tokens (60.5%) that MT would have translated wrongly were sent to transliteration.
    • On the MEDAR test set, BLEU rose from 16.63 to 16.91.
    • Classifier accuracy against gold labels on that set was 89%.

Interesting ideas

  • I1. Labels are produced automatically by a similarity score, and only the ambiguous middle band between two thresholds goes to a human; this concentrates costly annotation where it matters most.
  • I2. The same surface word ("North") is handled differently depending on the type of entity it sits in, and entity category beats local context as evidence.
  • I3. Membership of a token in WordNet as a regular noun serves as a signal against treating it as a pure name.
  • I4. The out-of-domain test set was deliberately built from entities that never occur in the training corpus, so it measures domain shift honestly instead of rewarding memorization.
  • I5. The transliteration is offered to the decoder as an extra option rather than forced, so the downstream model keeps the final say.

For Mova

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

  • I1 (learn): Silver-data agreement could use the same two-threshold scheme, auto-accepting high-agreement and auto-rejecting low-agreement parses while sending only the middle band to the LLM teacher tier.
  • I2 (world): When building entities, the decision whether a capitalized common word is part of a name or a literal descriptor (a direction, a title) could be conditioned on the entity's category rather than only on neighboring words.
  • I3 (tokens): The tagger and lemmatizer could use WordNet common-noun membership as an explainable feature when deciding between PROPN and NOUN for capitalized tokens inside multiword names.
  • I4 (learn): Evaluation of gparse or of frame extraction could include held-out splits filtered to unseen lemmas or entities, acting as a control against lexical memorization.
  • I5 (back): During translation through the meta-world, name renderings from a translate-vs-transliterate decision could be passed to generation as ranked alternatives with their justification, preserving explainability.

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 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

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

Cite

Cite as Mahmoud Azab, Houda Bouamor, Behrang Mohit and Kemal Oflazer. 2013. Dudley North visits North London: Learning When to Transliterate to Arabic. Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.

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

@inproceedings{azab2013dudley,
  title = {{Dudley North visits North London: Learning When to Transliterate to Arabic}},
  author = {Mahmoud Azab, Houda Bouamor, Behrang Mohit and Kemal Oflazer},
  booktitle = {Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2013},
  url = {https://aclanthology.org/N13-1046/}
}

References

20 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yaser Al-Onaizan and Kevin Knight. 2002a. Named-Entity translation. In Proceedings of HLT, San Francisco, USA. find
  2. Yaser Al-Onaizan and Kevin Knight. 2002b. Translating Named Entities Using Monolingual and Bilingual Resources. In Proceedings of ACL, Philadelphia, USA.
  3. Chih-Chung Chang and Chih-Jen Lin. 2011. LIB-SVM: A Library for Support Vector Machines. ACM Transactions on Intelligent Systems and Technology, 2:27:1–27:27. Software available at http://www. csie.ntu.edu.tw/˜cjlin/libsvm. ↗
  4. Silviu Cucerzan. 2007. Large-Scale Named-Entity Disambiguation Based on Wikipedia Data. In Proceedings of EMNLP-CoNLL, Prague, Czech Republic. find
  5. Ahmed El Kholy and Nizar Habash. 2010. Techniques for Arabic Morphological Detokenization and Orthographic Denormalization. In Proceedings of LREC, Valletta, Malta. find
  6. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. The MIT Press. ↗
  7. Andrew Freeman, Sherri Condon, and Christopher Ackerman. 2006. Cross Linguistic Name Matching in English and Arabic. In Proceedings of NAACL, New York City, USA.
  8. Bonnie Glover-Stalls and Kevin Knight. 1998. Translating Named and Technical Terms in Arabic Text. In Proceeding of the COLING/ACL Workshop on Computational Approaches to Semitic Languages, Montreal, Canada. find
  9. Nizar Habash, Owen Rambow, and Ryan Roth. 2009. Mada+Tokan: A Toolkit for Arabic Tokenization, Diacritization, Morphological Disambiguation, POS Tagging, Stemming and Lemmatization. In Proceedings of the Second International Conference on Arabic Language Resources and Tools (MEDAR), Cairo, Egypt. find
  10. Ahmed Hassan, Haytham Fahmy, and Hany Hassan. 2007. Improving Named Entity Translation by Exploiting Comparable and Parallel Corpora. In Proceedings of RANLP, Borovets, Bulgaria. find
  11. Ulf Hermjakob, Kevin Knight, and Hal Daumé III. 2008. Name Translation in Statistical Machine Translation - Learning When to Transliterate. In Proceedings of ACL-HLT, Columbus, Ohio. find
  12. Jun’ichi Kazama and Kentaro Torisawa. 2007. Exploiting Wikipedia as External Knowledge for Named-Entity Recognition. In Proceedings of EMNLP-CoNLL, Prague, Czech Republic.
  13. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In Proceedings of ACL: Demo session, Prague, Czech Republic. find
  14. Wang Ling, Pavel Calado, Bruno Martins, Isabel Trancoso, and Alan Black. 2011. Named-Entity Translation using Anchor Texts. In Proceedings of IWSLT, San Francisco, USA. find
  15. Bente Maegaard, Mohamed Attia, Khalid Choukri, Olivier Hamon, Steven Krauwer, and Mustafa Yaseen. 2010. Cooperation for Arabic Language Resources and Tools–The MEDAR Project. In Proceedings of LREC, Valetta, Malta.
  16. Behrang Mohit, Nathan Schneider, Rishav Bhowmick, Kemal Oflazer, and Noah A. Smith. 2012. Recall-Oriented Learning of Named Entities in Arabic Wikipedia. In Proceedings of EACL, Avignon, France.
  17. Joel Nothman, Nicky Ringland, Will Radford, Tara Murphy, and James R. Curran. 2013. Learning Multilingual Named Entity Recognition from Wikipedia. Artificial Intelligence, 194(0):151 – 175. ↗
  18. Franz Josef Och. 2003. Minimum Error Rate Training in Statistical Machine Translation. In Proceedings of ACL, Sapporo, Japan. find
  19. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a Method for Automatic Evaluation of Machine Translation. In Proceedings of ACL, Philadelphia, USA.
  20. Lev Ratinov and Dan Roth. 2009. Design Challenges and Misconceptions in Named Entity Recognition. In Proceedings of CONLL, Boulder, USA.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/azab-2013-dudley-north-visits-north-london/