Using Syntactic Coupling Features for Discriminating Phrase-Based Translations (WMT-08 Shared Translation Task)

Proceedings of the Third Workshop on Statistical Machine Translation, pages 159–162, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Using syntactic coupling features for discriminating phrase-based translations (WMT-08 Shared Translation Task) Vassilina Nikoulina and Marc Dymetman Xerox Research Centre Europe Grenoble, France {nikoulina,dymetman}@xrce.xerox.com Abstract Our participation in the shared translation task at WMT-08 focusses on news translation from English to French. Our main goal is to con- trast a baseline version of the phrase-based MATRAX system, with a version that incor- porates syntactic “coupling” features in order to discriminate translations produced by the baseline system. We report results comparing different feature combinations. 1 Introduction Our goal is to try to improve the fluency and ad- equacy of a baseline phrase-based SMT system by using a variety of “syntactic coupling features”, ex- tracted from parses for the source and target strings. These features are used for reranking the n-best can- didates of the baseline system. The phrase-based SMT system MATRAX, devel- oped at XRCE, is used as the baseline in the experi- ments. MATRAX is based on a fairly standard log- linear model, but one original aspect of the system is the use of non-contiguous bi-phrases such as ne ... plus / not ... anymore, where words in the source and target phrases may be separated by gaps, to be filled at translation time by lexical material provided by some other such pairs (Simard et al., 2005). For parsing, we use the Xerox Incremental Parser XIP (A¨ıt-Mokhtar et al., 2002), which is a robust dependency parser developed at the Xerox Research Centre Europe. XIP is fast (around 2000 words per second for English) and is well adapted to a situ- ation, like the one we have here, were we need to parse on the order of a few hundred target candi- dates on the fly. Also of interest to us is the fact that XIP produces labelled dependencies, a feature that we use in some of our experiments. 1.1 Decoding and Training We resort to a standard reranking approach in which we produce an n-best list of MATRAX candidate translations (with n = 100 in our experiments), and then rerank this list with a linear combination of our parse-dependent features. In order to train the fea- ture weights, we use an averaged structured percep- tron approach (Roark et al., 2004), where we try to learn weights such that the first candidate to emerge is equal to the “oracle” candidate, that is, the candi- date that is closest to the reference in terms of NIST score. 1.2 Coupling Features Our general approach to computing coupling fea- tures between the dependency structure of the source and that of a candidate translation produced by MA- TRAX is the following: we start by aligning the words between the source and the candidate trans- lation, we parse both sides, and we count (possi- bly according to a weighting scheme) the number of configurations (“rectangles”) that are of the follow- ing type: ((s1, s12, s2), …

Content truncated to 500 words