Which Words Matter in Defining Phrase Reordering Behavior in Statistical Machine Translation?

Which Words Matter in Defining Phrase Reorderings in Statistical Machine Translation? Hamidreza Ghader h.ghader@uva.nl Christof Monz c.monz@uva.nl Informatics Institute, University of Amsterdam, The Netherlands Abstract Lexicalized and hierarchical reordering models use relative frequencies of fully lexicalized phrase pairs to learn phrase reordering distributions. This results in unreliable estimation for infrequent phrase pairs which also tend to be longer phrases. There are some smoothing tech- niques used to smooth the distributions in these models. But these techniques are unable to address the similarities between phrase pairs and their reordering distributions. We propose two models to use shorter sub-phrase pairs of an original phrase pair to smooth the phrase re- ordering distributions. In the first model we follow the classic idea of backing off to shorter histories commonly used in language model smoothing. In the second model, we use syntactic dependencies to identify the most relevant words in a phrase to back off to. We show how these models can be easily applied to existing lexicalized and hierarchical reordering models. Our models achieve improvements of up to 0.40 BLEU points in Chinese-English translation compared to a baseline which uses a regular lexicalized reordering model and a hierarchical reordering model. The results show that not all the words inside a phrase pair are equally im- portant in defining phrase reordering behavior and shortening towards important words will decrease the sparsity problem for long phrase pairs. 1 Introduction The introduction of lexicalized reordering models (LRMs) (Tillmann, 2004; Koehn et al., 2005) was a significant step towards better reordering by modeling the orientation of the current phrase pair with respect to the previously translated phrase. LRMs score the order in which phrases are translated by using a distribution of distinguished orientations conditioned on phrase pairs. Typically, the set of orientations consists of: monotone (M), swap (S) and discontinuous (D). However, LRMs are limited to reorderings of neighboring phrases only. Galley and Manning (2008) proposed a hierarchical phrase reordering model (HRM) for more global reorderings. LRMs and HRMs both use relative frequencies observed in a parallel corpus to estimate the distribution of orientations conditioned on phrase pairs. As a result, they both suffer from the same problem of estimating reliable distributions for cases that occur rarely during training and therefore have to resort to smoothing methods to alleviate sparsity issues. Cherry (2013) builds on top of HRMs and proposes a sparse feature approach which uses word clusters instead of fully lexicalized forms for infrequent words to decrease the effect of sparsity on the estimated model. In this paper, we propose two types of approaches to use the most influential words from inside the original phrase pairs to estimate better orientation distributions for infrequent phrase pairs that takes phrase pair similarity more into account. In the first approach, we define a back- off model to shorten towards important words inside the original phrase pairs following the idea ������������������������������������������������������ ����������������������������������������� Dirichlet Smoothed Source Target Freq M S DL DR a 中国 政府 chinese government 2834 0.216 0.034 0.315 …

Content truncated to 500 words