The Cunei Machine Translation Platform for WMT '10

This paper describes the Cunei Machine Translation Platform and how it was used in the WMT ’10 German to English and Czech to English translation tasks. 1 The Cunei Machine Translation Platform The Cunei Machine Translation Platform (Phillips and Brown, 2009) is open-source software and freely available at http://www.cunei. org/. Like Moses (Koehn et al., 2007) and Joshua (Li et al., 2009), Cunei provides a statistical decoder that combines partial translations (either phase pairs or grammar rules) in order to compose a coherent sentence in the target language. What makes Cunei unique is that it models the translation task with a non-parametric model that assesses the relevance of each translation instance. The process begins by encoding in a lattice all possible contiguous phrases from the input.1 For each source phrase in the lattice, Cunei locates instances of it in the corpus and then identifies the aligned target phrase(s). This much is standard to most data-driven MT systems. The typical step at this stage is to model a phrase pair by computing relative frequencies over the collection of translation instances. This model for the phrase pair will never change and knowledge of the translation instances can subsequently be discarded. In contrast to using a phrase pair as the basic unit of modeling, Cunei models each translation instance. A distance function, represented by a log-linear model, scores the relevance of each translation instance. Our model then sums the scores of translation instances that predict the same target hypothesis. The advantage of this approach is that it provides a flexible framework for novel sources of Cunei offers limited support for non-contiguous phrases, similar in concept to grammar rules, but this setting was disabled in our experiments. information. The non-parametric model still uses information gleaned over all translation instances, but it permits us to define a distance function that operates over one translation instance at a time. This enables us to score a wide-variety of information represented by the translation instance with respect to the input and the target hypothesis under consideration. For example, we could compute how similar one translation instance’s parse tree or morpho-syntactic information is to the input. Furthermore, this information will vary throughout the corpus with some translation instances exhibiting higher similarity to the input. Our approach captures that these instances are more relevant and they will have a larger effect on the model. For the WMT ’10 task, we exploited instance-specific context and alignment features which will be discussed in more detail below.