A Phrase Orientation Model for Hierarchical Machine Translation

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 452–463, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics A Phrase Orientation Model for Hierarchical Machine Translation Matthias Huck and Joern Wuebker and Felix Rietig and Hermann Ney Human Language Technology and Pattern Recognition Group Computer Science Department RWTH Aachen University D-52056 Aachen, Germany {huck,wuebker,rietig,ney}@i6.informatik.rwth-aachen.de Abstract We introduce a lexicalized reordering model for hierarchical phrase-based ma- chine translation. The model scores mono- tone, swap, and discontinuous phrase ori- entations in the manner of the one pre- sented by Tillmann (2004). While this type of lexicalized reordering model is a valuable and widely-used component of standard phrase-based statistical machine translation systems (Koehn et al., 2007), it is however commonly not employed in hi- erarchical decoders. We describe how phrase orientation prob- abilities can be extracted from word- aligned training data for use with hierar- chical phrase inventories, and show how orientations can be scored in hierarchi- cal decoding. The model is empirically evaluated on the NIST Chinese→English translation task. We achieve a signifi- cant improvement of +1.2 %BLEU over a typical hierarchical baseline setup and an improvement of +0.7 %BLEU over a syntax-augmented hierarchical setup. On a French→German translation task, we obtain a gain of up to +0.4 %BLEU. 1 Introduction In hierarchical phrase-based translation (Chiang, 2005), a probabilistic synchronous context-free grammar (SCFG) is induced from bilingual train- ing corpora. In addition to continuous lexical phrases as in standard phrase-based translation, hierarchical phrases with usually up to two non- terminals are extracted from the word-aligned par- allel training data. Hierarchical decoding is typically carried out with a parsing-based procedure. The parsing al- gorithm is extended to handle translation candi- dates and to incorporate language model scores via cube pruning (Chiang, 2007). During decod- ing, a hierarchical translation rule implicitly spec- ifies the placement of the target part of a sub- derivation which is substituting one of its non- terminals in a partial hypothesis. The hierarchical phrase-based model thus provides an integrated re- ordering mechanism. The reorderings which are being conducted by the hierarchical decoder are a result of the application of SCFG rules, which generally means that there must have been some evidence in the training data for each reordering operation. At first glance one might be tempted to believe that any additional designated phrase ori- entation modeling would be futile in hierarchical translation as a consequence of this. We argue that such a conclusion is false, and we will pro- vide empirical evidence in this work that lexical- ized phrase orientation scoring can be highly ben- eficial not only in standard phrase-based systems, but also in hierarchical ones. The purpose of a phrase orientation model is to assess the adequacy of phrase reordering dur- ing search. In standard phrase-based translation with continuous phrases only and left-to-right hy- pothesis generation (Koehn et al., 2003; Zens and Ney, 2008), phrase reordering is implemented by jumps within the input sentence. The choice of the best order for the target sequence is …

Content truncated to 500 words