Elamur Research⌕
☰
home / papers / 2013 / Reversing Morphological Tokenization in English-to-Arabic SMT

Reversing Morphological Tokenization in English-to-Arabic SMT

Summary

When a statistical machine translation (SMT) system is trained to translate from English into Arabic, the Arabic text is usually tokenized morphologically to reduce lexical sparsity. As a result, the system outputs separate morphemes, and recovering the correct word forms from them is not straightforward. Tokenization changes the spelling at morpheme boundaries, so simple concatenation is not enough. The authors treat detokenization as a character-level string transduction. They train DirecTL+, a discriminative transducer originally developed for letter-to-phoneme conversion. The training data are pairs of "tokenized form – original word". They are obtained by running a tokenizer with the PATB scheme over a large corpus of Arabic text. The model itself infers context-dependent merging rules from character n-grams within the word. It needs neither hand-written rules nor a language model. On the Arabic reference text of NIST MT05, the method slightly outperforms the strongest approach of the time, that of El Kholy and Habash (2012). In English-to-Arabic translation, it yields a small BLEU gain over rules. The method is language-independent, so it can be transferred to other morphologically rich languages.

Key points

  • The baseline methods included simple concatenation, rules, a lookup table, a table with rules as a fallback, and SRILM-disambig with a 5-gram language model. The sentence error rate (SER) was 34.3%, 14.0%, 3.2% and 2.4% respectively for the best combination. For DirecTL+ it is 2.1%, and the WER is 0.087.
  • DirecTL+ aligns characters using M2M-aligner and is trained with MIRA. The context the model takes into account extends to 11 characters. Hyperparameters were tuned by grid search on NIST MT04.
  • Translation was built with Moses and GIZA++. With DirecTL+ BLEU is 28.55, with rules 28.32, and without detokenization 26.30. Once WER drops below roughly 0.2, translation quality no longer improves.
  • DirecTL+ made errors in 21 cases. 11 of them are confusions between ta marbuta and ha. The rest relate to the choice of hamza form, case (the model does not see neighboring words), and the ambiguous combination An+y.
  • The table-based method made errors in 54 cases, and in 42 of them DirecTL+ produced the correct form. Thus, the model handles words that are absent from the table well.

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: Linguistics & languages

Topics: Machine translation

Cite

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

@inproceedings{salameh2013reversing,
  title = {{Reversing Morphological Tokenization in English-to-Arabic SMT}},
  author = {Salameh, Mohammad and Cherry, Colin and Kondrak, Grzegorz},
  booktitle = {Proceedings of the 2013 NAACL HLT Student Research Workshop},
  year = {2013},
  url = {https://aclanthology.org/N13-2007/}
}

References

11 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ibrahim Badr, Rabih Zbib, and James Glass. 2008. Segmentation for English-to-Arabic statistical machine translation. In Proceedings of ACL, pages 153–156. find
  2. Ahmed El Kholy and Nizar Habash. 2012. Orthographic and morphological processing for English-Arabic statistical machine translation. Machine Translation, 26(1-2):25–45, March. find
  3. 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. find
  4. Nizar Habash. 2007. Arabic morphological representations for machine translation. In Arabic Computational Morphology: Knowledge-based and Empirical Methods. find
  5. Nizar Habash. 2010. Introduction to Arabic Natural Language Processing. Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers. find
  6. Hieu Hoang, Alexandra Birch, Chris Callison-burch, Richard Zens, Rwth Aachen, Alexandra Constantin, Marcello Federico, Nicola Bertoldi, Chris Dyer, Brooke Cowan, Wade Shen, Christine Moran, and Ondrej Bojar. 2007. Moses: Open source toolkit for statistical machine translation. In Annual Meeting of t… find
  7. Sittichai Jiampojamarn, Grzegorz Kondrak, and Tarek Sherif. 2007. Applying many-to-many alignments and HMMs to letter-to-phoneme conversion. In Proceedings of NAACL-HLT, pages 372–379. find
  8. Sittichai Jiampojamarn, Colin Cherry, and Grzegorz Kondrak. 2008. Joint processing and discriminative training for letter-to-phoneme conversion. In Proceedings of ACL-08: HLT, pages 905–913. find
  9. Kishore Papineni, Salim Roukos, Todd Ward, and Wei jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318. find our card
  10. Fatiha Sadat and Nizar Habash. 2006. Combination of Arabic preprocessing schemes for statistical machine translation. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 1–8. find
  11. Andreas Stolcke. 2002. SRILM - an extensible language modeling toolkit. In Intl. Conf. Spoken Language Processing, pages 901–904. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/salameh-2013-reversing-morphological-tokenization-english-arabic/