Enhancing function word translation with syntax-based statistical post-editing
What it is about
A post-editor for the top-1 output of a dependency tree-to-tree system (Japanese to English, ASPEC scientific abstracts, 3 million training sentences). A function-word model P(f | t, t') is counted over (token, head) pairs of target trees, with tags inside the tokens and separate counts for left and right children. Three greedy operations, insert, replace and delete (delete is replace with a null token), apply when a function word's probability beats the alternative. Replacement and deletion are allowed only for function words not aligned to a source content word, because the filter stopped losses such as a lost preposition. Human judges compared 20 edited sentences per setting as better, worse or neutral. The final test gave 23 wins, 10 losses and 7 neutral, a change rate of 3.9%; Fleiss kappa was 0.663 on wins and losses and 0.285 with neutral. BLEU was judged too insensitive for this change rate.
Ideas
- I1. A post-editor on the encoder output. After the standard stack writes the target tree, count P(f | head, token) over the training trees and insert or replace a function word when the best candidate beats the null token, as the paper's rule does. Measure the function-word residual before and after on dev, run twice.
- I2. A source filter from the meta-world. Meta tokens keep the source words they were read from, so allow a deletion or replacement only when the token has no source word or a function-word source. Count the edits that would remove a word with a source counterpart.
- I3. A small blind human check, as in the paper: 20 verses, better, worse or neutral against the plain encoder, with Fleiss kappa reported next to the exact-mode bits.
For Mova
The post-editor is a count model over trees, deterministic and fitting the translator's tables. Its human check is small, so it does not replace Mova's exact-mode residual, but it is a second measure of the function-word class, which open problem 4 needs. The paper is in NC licence: summary only.
Cite
@inproceedings{richardson2015enhancing,
title = {{Enhancing function word translation with syntax-based statistical post-editing}},
author = {Richardson, John and Nakazawa, Toshiaki and Kurohashi, Sadao},
booktitle = {Proceedings of the 6th Workshop on Patent and Scientific Literature Translation},
year = {2015},
url = {https://aclanthology.org/2015.mtsummit-wpslt.7/}
}References
- Arnold, D. and Sadler, L. (1991). EuroTra: An assessment of the current state of the ECs MT Programme. In Working Papers in Language Processing. find
- Charniak, E. and Johnson, M. (2005). Coarse-to-fine n-best parsing and maxent discriminative reranking. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, ACL ’05, pages 173–180. Association for Computational Linguistics. find our card
- Elming, J. (2006). Transformation-based corrections of rule-based MT. In EAMT 11th Annual Conference. find
- Fleiss, L., Levin, B., and Paik, M. C. (1981). The measurement of interrater agreement. In Statistical methods for rates and proportions (2nd ed), pages 212–236. Wiley. find
- Heafield, K. (2011). KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation. find
- Huang, A., Kuo, T., Lai, Y., and Lin, S. (2010). Discovering correction rules for auto editing. International Journal of Computational Linguistics and Chinese Language Processing, 15(3-4). find
- Kawahara, D. and Kurohashi, S. (2006). A fully-lexicalized probabilistic model for Japanese syntactic and case structure analysis. In Proceedings of the Main Conference on Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, HLT-NAACL ’06, pages 176–183. Association for Computational Linguistics. find
- Menezes, A. and Quirk, C. (2008). Syntactic models for structural word insertion and deletion during translation. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 735–744, Honolulu, Hawaii. Association for Computational Linguistics. find our card
- Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J. (2002). BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, pages 311–318. Association for Computational Linguistics. find our card
- Richardson, J., Cromières, F., Nakazawa, T., and Kurohashi, S. (2014). KyotoEBMT: An example-based dependency-to-dependency translation framework. In Proceedings of 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 79–84, Baltimore, Maryland. Association for Computational Linguistics. find
- Riesa, J., Irvine, A., and Marcu, D. (2011). Feature-rich language-independent syntax-based alignment for statistical machine translation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pages 497–507. Association for Computational Linguistics. find
- Rosa, R., Mareček, D., and Tamchyna, A. (2013). Deepfix: Statistical post-editing of statistical machine translation using deep syntactic analysis. In Proceedings of the Student Research Workshop at the 51st Annual Meeting of the Association for Computational Linguistics. find
- Simard, M., Goutte, C., and Isabelle, P. (2007). Statistical phrase-based post-editing. In NAACL. find
- Wu, X., Matsuzaki, T., and Tsujii, J. (2011). Effective use of function words for rule generalization in forest-based translation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 22–31, Portland, Oregon, USA. Association for Computational Linguistics. find our card
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/richardson-2015-enhancing-function-word-translation-syntax/