Generalizing Local and Non-Local Word-Reordering Patterns for Syntax-Based Machine Translation

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 572–581, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Generalizing Local and Non-Local Word-Reordering Patterns for Syntax-Based Machine Translation Bing Zhao IBM T.J. Watson Research Yorktown Heights, NY-10598 zhaob@us.ibm.com Yaser Al-onaizan IBM T.J. Watson Research Yorktown Heights, NY-10598 onaizan@us.ibm.com Abstract Syntactic word reordering is essential for translations across different grammar struc- tures between syntactically distant language- pairs. In this paper, we propose to em- bed local and non-local word reordering de- cisions in a synchronous context free gram- mar, and leverages the grammar in a chart- based decoder. Local word-reordering is ef- fectively encoded in Hiero-like rules; whereas non-local word-reordering, which allows for long-range movements of syntactic chunks, is represented in tree-based reordering rules, which contain variables correspond to source- side syntactic constituents. We demonstrate how these rules are learned from parallel cor- pora. Our proposed shallow Tree-to-String rules show significant improvements in trans- lation quality across different test sets. 1 Introduction One of the main issues that a translator (human or machine) must address during the translation pro- cess is how to match the different word orders be- tween the source language and the target language. Different language-pairs require different levels of word reordering. For example, when we translate between English and Spanish (or other Romance languages), most of the word reordering needed is local because of the shared syntactical features (e.g., Spanish noun modifier constructs are written in English as modifier noun). However, for syn- tactically distant language-pairs such as Chinese- English, long-range reordering is required where whole phrases are moved across the sentence. The idea of “syntactic cohesion” (Fox, 2002) is characterized by its simplicity, which has attracted researchers for years. Previous works include sev- eral approaches of incorporating syntactic informa- tion to preprocess the source sentences to make them more like the target language in structure. Xia and McCord (2004) (Niessen and Ney, 2004; Collins et al., 2005) described approaches applied to language- pairs such as French-English and German-English. Later, Wang et al. (2007) presented specific rules to pre-order long-range movements of words, and improved the translations for Chinese-to-English. Overall, these works are similar, in that they design a few language-specific and linguistically motivated reordering rules, which are generally simple. The eleven rules described in Wang et al. (2007) are ap- pealing, as they have rather simple structure, mod- eling only NP, VP and LCP via one-level sub-tree structure with two children, in the source parse-tree (a special case of ITG (Wu, 1997)). It effectively en- hances the quality of the phrase-based translation of Chinese-to-English. One major weakness is that the reordering decisions were done in the preprocessing step, therefore rendering the decoding process un- able to recover the reordering errors from the rules if incorrectly applied to. Also the reordering decisions are made without the benefits of additional models (e.g., the language models) that are typically used during decoding. Another method to address the re-ordering prob- lem in translation is the Hiero …

Content truncated to 500 words