Dependency Tree Abstraction for Long-Distance Reordering in Statistical Machine Translation

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 424–433, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Dependency Tree Abstraction for Long-Distance Reordering in Statistical Machine Translation Chenchen Ding Department of Computer Science University of Tsukuba 1-1-1 Tennodai, Tsukuba, Ibaraki , Japan tei@mibel.cs.tsukuba.ac.jp Yuki Arase Microsoft Research No. 5 Danling St., Haidian Dist. Beijing, P.R. China yukiar@microsoft.com Abstract Word reordering is a crucial technique in statistical machine translation in which syntactic information plays an important role. Synchronous context-free gram- mar has typically been used for this pur- pose with various modifications for adding flexibilities to its synchronized tree gen- eration. We permit further flexibilities in the synchronous context-free grammar in order to translate between languages with drastically different word order. Our method pre-processes a parallel corpus by abstracting source-side dependency trees, and performs long-distance reordering on top of an off-the-shelf phrase-based sys- tem. Experimental results show that our method significantly outperforms previous phrase-based and syntax-based models for translation between English and Japanese. 1 Introduction Since the inception of statistical machine trans- lation (SMT), long-distance word reordering has been a notable challenge, particularly when trans- lating between languages with drastically different word orders, such as subject-verb-object (SVO) and subject-object-verb (SOV) languages like En- glish and Japanese, respectively. Phrase-based models (Koehn et al., 2003; Och and Ney, 2004; Xiong et al., 2006) have been strong in local translation and reordering. However, phrase-based models cannot effectively conduct long-distance reordering because they are based purely on statis- tics of syntax-independent phrases. As a comple- mentary approach to phrase-based models, some researchers have incorporated syntactic informa- tion into an SMT framework (Wu, 1997; Yamada and Knight, 2001; Liu et al., 2006) using syn- chronous context-free grammar (SCFG) (Aho and when the fluid pressure cylinder 31 is used, fluid is gradually applied. , is gradually applied . fluid when is used [N] the fluid pressure cylinder 31 [N] [X] [X] (root) Figure 1: English abstraction tree example Ullman, 1972). The original SCFG assumes that the syntactic trees of the source and target lan- guages can be derived synchronously. However, this assumption is too strict for handling paral- lel sentences that are often comparable rather than parallel. For alleviating this assumption, some re- searchers have added flexibilities in synchronized tree generation (Wu, 1997; Burkett et al., 2010). In addition, in the SMT framework, there is an approach that alleviates the assumption by only generating the source-side syntactic tree and pro- jecting it to the target-side sentence (Yamada and Knight, 2001; Liu et al., 2006). In practice, these existing methods are not flex- ible enough to handle parallel sentence pairs, es- pecially those of SVO and SOV languages. There- fore, we permit further flexibility in SCFG aiming to effectively conduct long-distance reordering. We design our method as a pre-processing proce- dure so that we can use a well-developed phrase- based system without adding heavy computational complexity to the system. Specifically, we propose an abstraction tree that is a …

Content truncated to 500 words