Phrase-Based and Deep Syntactic English-to-Czech Statistical Machine Translation

What it is about

The paper reports two English-to-Czech systems for WMT 2008. The first is a factored phrase-based system in Moses. Its target side is generated in two steps: word forms first, then full morphological tags, with six language models, including 7-gram models over tags. It is the stronger system, at 16.4 BLEU on the WMT08 news-commentary test. The second is a probabilistic tree transfer at the deep syntactic (t-) layer, which uses synchronous treelet grammars and a binode model of governor and child. Its node labels have more than 20 attributes, and the transfer maps them as factors. It reaches only about 4.9 BLEU on the same test. The authors list the causes: errors add up over analysis steps, parses of the two languages do not match and lose data at alignment, the target attributes multiply the search space, and the deterministic generation has no n-gram model to rescore with.

Ideas

• I1. Fill the fine attributes last: the explosion comes from producing all attributes during search. In the translator, let the encoder fix structure and lexical choice first, and write the function-word and feature values (open problems 4 and 5) only once those are set, as the paper recommends.
• I2. Two-step target generation: first the concept and its chunk, then the grammatical features from UD FEATS, as the two-step scheme of the paper. Measure the residual bits of each step on the Bible pairs in tools/translate/src/profile.rs.
• I3. A count-based binode score: estimate p(child | governor) on the target trees, and use it as one more ranking term in the encoder, with the counts in BTreeMap. The paper's gain is in the same direction as its own explanation, so the control is the encoder without the term.
• I4. Measure the parse-loss cost: the paper says parsing both sides on its own loses many pairs. Count, for KJV and WEB verses, the share of meta tokens that have no partner after the reader, and report it per author pair (open problem 1).

For Mova

The code is tools/translate/src/profile.rs and tools/translate/src/meta.rs. The paper's search over treelets is not for Mova, but its lesson about attribute explosion is. The table-based two-step generation fits the count-based encoder. Report exact match as well as BLEU, since the paper's BLEU gap is large and its causes are structural.

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.3115/1626394.1626413

@inproceedings{bojar2008phrase,
  title = {{Phrase-Based and Deep Syntactic English-to-Czech Statistical Machine Translation}},
  author = {Bojar, Ondřej and Hajič, Jan},
  booktitle = {Proceedings of the Third Workshop on Statistical Machine Translation},
  year = {2008},
  doi = {10.3115/1626394.1626413},
  url = {https://doi.org/10.3115/1626394.1626413}
}

References

25 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ondřej Bojar and Martin Čmejrek. 2007. Mathematical Model of Tree Transformations. Project EuroMatrix - Deliverable 3.2, ÚFAL, Charles University, Prague. find
2. Ondřej Bojar, Zdeněk Žabokrtský, Pavel Češka, Peter Beňa, and Miroslav Janı́ček. 2008. CzEng 0.7: Parallel Corpus with Community-Supplied Translations. In Proc. of LREC 2008. ELRA. find
3. Ondřej Bojar. 2007. English-to-Czech Factored Machine Translation. In Proc. of ACL Workshop on Statistical Machine Translation, pages 232–239, Prague. find
4. Thorsten Brants. 2000. TnT - A Statistical Part-of-Speech Tagger . In Proc. of ANLP-NAACL. find our card
5. Peter F. Brown et al. 1992. Class-based n-gram models of natural language. Computational Linguistics, 18(4):467–479. find
6. Pavel Češka. 2006. Segmentace textu. Bachelor’s Thesis, MFF, Charles University in Prague. find
7. Silvie Cinková et al. 2004. Annotation of English on the tectogrammatical level. Technical Report TR-2006-35, ÚFAL/CKL, Prague, Czech Republic. find
8. Michael Collins. 1996. A New Statistical Parser Based on Bigram Lexical Dependencies. In Proc. of ACL. find
9. Yuan Ding and Martha Palmer. 2005. Machine Translation Using Probabilistic Synchronous Dependency Insertion Grammars. In Proc. of ACL. find our card
10. Jan Hajič. 2004. Disambiguation of Rich Inflection (Computational Morphology of Czech). Nakladatel-stvı́ Karolinum, Prague. find
11. Jan Hajič et al. 2006. Prague Dependency Treebank 2.0. LDC2006T01, ISBN: 1-58563-370-4. find
12. Liang Huang, Kevin Knight, and Aravind Joshi. 2006. Statistical Syntax-Directed Translation with Extended Domain of Locality. In Proc. of AMTA, Boston, MA. find our card
13. Václav Klimeš. 2006. Analytical and Tectogrammatical Analysis of a Natural Language. Ph.D. thesis, ÚFAL, MFF UK, Prague, Czech Republic. find
14. Jan Kocek, Marie Kopřivová, and Karel Kučera, editors. 2000. Český národnı́ korpus - úvod a přı́ručka uživatele. FF UK - ÚČNK, Praha. find
15. Philipp Koehn and Hieu Hoang. 2007. Factored Translation Models. In Proc. of EMNLP. find
16. Philipp Koehn, Hieu Hoang, et al. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In Proc. of ACL Demo and Poster Sessions. find
17. Ryan McDonald et al. 2005. Non-Projective Dependency Parsing using Spanning Tree Algorithms. In Proc. of HLT/EMNLP 2005. find our card
18. Guido Minnen, John Carroll, and Darren Pearce. 2001. Applied morphological processing of English. Natural Language Engineering, 7(3):207–223. find
19. Franz Josef Och and Hermann Ney. 2000. A Comparison of Alignment Models for Statistical Machine Translation. In Proc. of COLING, pages 1086–1090. find
20. Franz Josef Och. 2003. Minimum Error Rate Training in Statistical Machine Translation. In Proc. of ACL. find
21. Jan Ptáček and Zdeněk Žabokrtský. 2006. Synthesis of Czech Sentences from Tectogrammatical Trees. In Proc. of TSD, pages 221–228. find
22. Chris Quirk, Arul Menezes, and Colin Cherry. 2005. Dependency Treelet Translation: Syntactically Informed Phrasal SMT. In Proc. of ACL, pages 271–279. find our card
23. Stefan Riezler and John T. Maxwell III. 2006. Grammatical Machine Translation. In Proc. of HLT/NAACL. find
24. Petr Sgall, Eva Hajičová, and Jarmila Panevová. 1986. The Meaning of the Sentence and Its Semantic and Pragmatic Aspects. Academia, Prague. find
25. Zdeněk Žabokrtský. 2008. Tecto MT. Technical report, ÚFAL/CKL, Prague, Czech Republic. In prep. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/bojar-2008-phrase-based-deep-syntactic-english/