Examining the Relationship between Preordering and Word Order Freedom in Machine Translation

Proceedings of the First Conference on Machine Translation, Volume 1: Research Papers, pages 118–130, Berlin, Germany, August 11-12, 2016. c⃝2016 Association for Computational Linguistics Examining the Relationship between Preordering and Word Order Freedom in Machine Translation Joachim Daiber Miloˇs Stanojevi´c Wilker Aziz Khalil Sima’an Institute for Logic, Language and Computation (ILLC) University of Amsterdam {initial.last}@uva.nl Abstract We study the relationship between word order freedom and preordering in statisti- cal machine translation. To assess word order freedom, we first introduce a novel entropy measure which quantifies how dif- ficult it is to predict word order given a source sentence and its syntactic analysis. We then address preordering for two target languages at the far ends of the word order freedom spectrum, German and Japanese, and argue that for languages with more word order freedom, attempting to predict a unique word order given source clues only is less justified. Subsequently, we ex- amine lattices of n-best word order predic- tions as a unified representation for lan- guages from across this broad spectrum and present an effective solution to a re- sulting technical issue, namely how to se- lect a suitable source word order from the lattice during training. Our experi- ments show that lattices are crucial for good empirical performance for languages with freer word order (English–German) and can provide additional improvements for fixed word order languages (English– Japanese). 1 Introduction Word order differences between a source and a tar- get language are a major challenge for machine translation systems. For phrase-based models, the number of possible phrase permutations is so large that reordering must be constrained locally to make the search space for the best hypothe- sis feasible. However, constraining the space lo- cally runs the risk that the optimal hypothesis is rendered out of reach. Preordering of the source sentence has been embraced as a way to ensure the reachability of certain target word order con- stellations for improved prediction of the target word order. Preordering aims at predicting a per- mutation of the source sentence which has min- imal word order differences with the target sen- tence; the permuted source sentence is passed on to a backend translation system trained to translate target-order source sentences into target sentences. In essence, the preordering approach makes the as- sumption that it is feasible to predict target word order given only clues from the source sentence. In the vast majority of work on preordering, a sin- gle preordered source sentence is passed on to the backend system, thereby making the stronger as- sumption that it is feasible to predict a unique pre- ferred target word order. But how reasonable are these assumptions and for which target languages? Intuitively, the assumption of a unique pre- ordering seems reasonable for translating into fixed word order languages such as Japanese, but for translation into languages with less strict word order such as German, this is unlikely to work. In such languages there are multiple comparably plausible target word orders per source sentence because the underlying …

Content truncated to 500 words