Edinburgh’s Syntax-Based Machine Translation Systems

We present the syntax-based string-totree statistical machine translation systems built for the WMT 2013 shared translation task. Systems were developed for four language pairs. We report on adapting parameters, targeted reduction of the tuning set, and post-evaluation experiments on rule binarization and preventing dropping of verbs. 1 Overview Syntax-based machine translation models hold the promise to overcome some of the fundamental problems of the currently dominating phrasebased approach, most importantly handling reordering for syntactically divergent language pairs and grammatical coherence of the output. We are especially interested in string-to-tree models that focus syntactic annotation on the target side, especially for morphologically rich target languages (Williams and Koehn, 2011). We have trained syntax-based systems for the language pairs • English-German, • German-English, • Czech-English, and • Russian-English. We have also tried building systems for FrenchEnglish and Spanish-English but the data size proved to be problematic given the time constraints. We give a brief description of the syntaxbased model and its implementation within the Moses system. Some of the available features are described as well as some of the pre-processing steps. Several experiments are described and final results are presented for each language pair.