Quasi-Synchronous Grammars: Alignment by Soft Projection of Syntactic Dependencies

What it is about

The paper defines quasi-synchronous grammar (QG). Given the parse tree of the source sentence, a QG builds a monolingual grammar over target trees. The target tree need not be isomorphic to the source. Each target node is aligned to at most one source node, and the allowed relations between a target parent-child pair and the aligned source nodes form a set of configurations: parent-child, child-parent, same node, siblings, grandparent-grandchild, c-command, and a catch-all "breakage". The probability is a dependency model with valence, plus a translation table, trained by EM on German-English Europarl with 1K, 10K and 100K sentence pairs. Held-out cross-entropy falls as more configurations are allowed, and the strictly synchronous model is the worst. For German-to-English word alignment, the QG has a lower alignment error rate than the IBM Model 4 channel in the same direction, at every training size, and it is within about one point of the reverse direction. Permitting breakages gives the largest drop in error.

Ideas

• I1. Count configurations on verse pairs: for each KJV or WEB verse, align each meta token to a parse head of the other side and count the configuration of each pair. Report the share in each class per author pair. This turns open problem 1 (attachment noise) into a measured table.
• I2. Soft head patches: instead of a hard head rule, let the meta-world keep several candidate heads and score them with the configuration counts, so a head patch is made only when a concept's role changes.
• I3. Use the configuration as a feature for the Hebrew reader: align Hebrew words to meta tokens with IBM Model 1 (tools/translate/src/hebrew.rs) and add a score for each configuration, as open problem 7 suggests. Measure alignment error on a hand-checked set of verses, twice, for identical results.
• I4. Learning curve as a control: train on 1K, 10K and 100K verses, and report whether the meta-world gain grows the same way. A gain that appears only at the largest size is not yet a gain.

For Mova

The code is tools/translate/src/hebrew.rs (alignment) and tools/translate/src/meta.rs (heads). The model is a generative count model over dependency configurations, which fits Mova's count-based approach. EM training is not needed for the first measurement: the configuration counts can be read off the aligned verses directly.

Topics: Machine translation

Cite

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

@inproceedings{smith2006quasi,
  title = {{Quasi-Synchronous Grammars: Alignment by Soft Projection of Syntactic Dependencies}},
  author = {Smith, David and Eisner, Jason},
  booktitle = {Proceedings on the Workshop on Statistical Machine Translation},
  year = {2006},
  url = {https://aclanthology.org/W06-3104/}
}

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. H. Alshawi, S. Bangalore, and S. Douglas. 2000. Learning dependency translation models as collections of finite state head transducers. CL, 26(1):45–60. find
2. P. F. Brown, S. A. Della Pietra, V. J. Della Pietra, and R. L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. CL, 19(2):263–311. find
3. Y. Ding and M. Palmer. 2005. Machine translation using probabilistic synchronous dependency insertion grammars. In ACL, pages 541–548. find our card
4. B. J. Dorr. 1994. Machine translation divergences: A formal description and proposed solution. Computational Linguistics, 20(4):597–633. find our card
5. J. Eisner. 2003. Learning non-isomorphic tree mappings for machine translation. In ACL Companion Vol. find our card
6. H. J. Fox. 2002. Phrasal cohesion and statistical machine translation. In EMNLP, pages 392–399. find
7. D. Gildea. 2003. Loosely tree-based alignment for machine translation. In ACL, pages 80–87. find
8. D. Gildea. 2004. Dependencies vs. constituents for tree-based alignment. In EMNLP, pages 214–221. find
9. R. Hwa, P. Resnik, A. Weinberg, and O. Kolak. 2002. Evaluating translational correspondence using annotation projection. In ACL. find
10. D. Klein and C. D. Manning. 2003. Accurate unlexicalized parsing. In ACL, pages 423–430. find our card
11. D. Klein and C. D. Manning. 2004. Corpus-based induction of syntactic structure: Models of dependency and constituency. In ACL, pages 479–486. find
12. P. Koehn. 2002. Europarl: A multilingual corpus for evaluation of machine translation. http://www.iccs.informatics.ed.ac.uk/˜pkoehn/-publications/europarl.ps. ↗
13. A. Lavie, S. Vogel, L. Levin, E. Peterson, K. Probst, A. F. Llitjós, R. Reynolds, J. Carbonell, and R. Cohen. 2003. Experiments with a Hindi-to-English transfer-based MT system under a miserly data scenario. ACM Transactions on Asian Language Information Processing, 2(2):143 – 163. find
14. I. D. Melamed, G. Satta, and B. Wellington. 2004. Generalized multitext grammars. In ACL, pages 661–668. find
15. I. D. Melamed. 2004. Statistical machine translation by parsing. In ACL, pages 653–660. find
16. F. J. Och and H. Ney. 2003. A systematic comparison of various statistical alignment models. CL, 29(1):19–51. find
17. C. Quirk, A. Menezes, and C. Cherry. 2005. Dependency treelet translation: Syntactically informed phrasal SMT. In ACL, pages 271–279. find our card
18. S. M. Shieber and Y. Schabes. 1990. Synchronous tree-adjoining grammars. In ACL, pages 253–258. find
19. D. Wu. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. CL, 23(3):377–403. find our card
20. K. Yamada and K. Knight. 2001. A syntax-based statistical translation model. In ACL. 1e+06 find our card

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/smith-2006-quasi-synchronous-grammars-alignment-soft/