Learning to Transform and Select Elementary Trees for Improved Syntax-based Machine Translations

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 846–855, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Learning to Transform and Select Elementary Trees for Improved Syntax-based Machine Translations Bing Zhao†, and Young-Suk Lee†, and Xiaoqiang Luo†, and Liu Li‡ IBM T.J. Watson Research† and Carnegie Mellon University‡ {zhaob, ysuklee, xiaoluo}@us.ibm.com and liul@andrew.cmu.edu Abstract We propose a novel technique of learning how to transform the source parse trees to improve the trans- lation qualities of syntax-based translation mod- els using synchronous context-free grammars. We transform the source tree phrasal structure into a set of simpler structures, expose such decisions to the decoding process, and find the least expensive transformation operation to better model word re- ordering. In particular, we integrate synchronous bi- narizations, verb regrouping, removal of redundant parse nodes, and incorporate a few important fea- tures such as translation boundaries. We learn the structural preferences from the data in a generative framework. The syntax-based translation system in- tegrating the proposed techniques outperforms the best Arabic-English unconstrained system in NIST- 08 evaluations by 1.3 absolute BLEU, which is sta- tistically significant. 1 Introduction Most syntax-based machine translation models with syn- chronous context free grammar (SCFG) have been re- lying on the off-the-shelf monolingual parse structures to learn the translation equivalences for string-to-tree, tree-to-string or tree-to-tree grammars. However, state- of-the-art monolingual parsers are not necessarily well suited for machine translation in terms of both labels and chunks/brackets. For instance, in Arabic-to-English translation, we find only 45.5% of Arabic NP-SBJ struc- tures are mapped to the English NP-SBJ with machine alignment and parse trees, and only 60.1% of NP-SBJs are mapped with human alignment and parse trees as in § 2. The chunking is of more concern; at best only 57.4% source chunking decisions are translated contiguously on the target side. To translate the rest of the chunks one has to frequently break the original structures. The main issue lies in the strong assumption behind SCFG-style nonterminals – each nonterminal (or variable) assumes a source chunk should be rewritten into a contiguous chunk in the target. Without integrating techniques to mod- ify the parse structures, the SCFGs are not to be effec- tive even for translating NP-SBJ in linguistically distant language-pairs such as Arabic-English. Such problems have been noted in previous literature. Zollmann and Venugopal (2006) and Marcu et al. (2006) used broken syntactic fragments to augment their gram- mars to increase the rule coverage; while we learn opti- mal tree fragments transformed from the original ones via a generative framework, they enumerate the fragments available from the original trees without learning pro- cess. Mi and Huang (2008) introduced parse forests to blur the chunking decisions to a certain degree, to ex- pand search space and reduce parsing errors from 1-best trees (Mi et al., 2008); others tried to use the parse trees as soft constraints on top of unlabeled grammar such as Hiero (Marton and Resnik, 2008; Chiang, 2010; Huang et al., 2010; Shen …

Content truncated to 500 words