Source-Side Classifier Preordering for Machine Translation

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 513–523, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Source-Side Classifier Preordering for Machine Translation Uri Lerner Google Inc. Mountain View, CA, USA uri@google.com Slav Petrov Google Inc. New York, NY, USA slav@google.com Abstract We present a simple and novel classifier-based preordering approach. Unlike existing pre- ordering models, we train feature-rich dis- criminative classifiers that directly predict the target-side word order. Our approach com- bines the strengths of lexical reordering and syntactic preordering models by performing long-distance reorderings using the structure of the parse tree, while utilizing a discrimina- tive model with a rich set of features, includ- ing lexical features. We present extensive ex- periments on 22 language pairs, including pre- ordering into English from 7 other languages. We obtain improvements of up to 1.4 BLEU on language pairs in the WMT 2010 shared task. For languages from different families the improvements often exceed 2 BLEU. Many of these gains are also significant in human eval- uations. 1 Introduction Generating the appropriate word order for the tar- get language has been one of the fundamental prob- lems in machine translation since the ground setting work of Brown et al. (1990). Lexical reordering ap- proaches (Tillmann, 2004; Zens and Ney, 2006) add a reordering component to standard phrase-based translation systems (Och and Ney, 2004). Because the reordering model is trained discriminatively, it can use a rich set of lexical features. However, it only has access to the local context which often times is insufficient to make the long-distance re- ordering decisions that are necessary for language pairs with significantly different word order. Preordering (sometimes called pre-reordering or simply reordering) approaches (Xia and McCord, 2004; Collins et al., 2005) preprocess the input in such a way that the words on the source side appear closer to their final positions on the target side. Be- cause preordering is performed prior to word align- ment, it can improve the alignment process and can then be combined with any subsequent translation model. Most preordering models use a source-side syntactic parser and perform a series of tree trans- formations. Approaches that do not use a parser ex- ist as well and typically induce a hierarchical rep- resentation that also allows them to perform long- distance changes (Tromble and Eisner, 2009; DeN- ero and Uszkoreit, 2011; Neubig et al., 2012). Models that use a source-side parser differ on two main dimensions: the way tree transformations are expressed, and whether they are built manually or learned from data. One common type of tree trans- formation are rewrite rules. These typically involve some condition under which the transformation can be applied (e.g., a noun and an adjective found in the same clause) and the transformation itself (e.g., move the adjective after the noun). These rules can be designed manually (Collins et al., 2005; Wang et al., 2007) or learned from data (Xia and McCord, 2004; Habash, 2007; Genzel, 2010; Wu et al., …

Content truncated to 500 words