Combining phrase-based and tree-to-tree translation

Abstract

We present a novel approach to multi-engine machine translation, using a feature-based classification algorithm. Instead of just using language models, translation models, or internal confidence scores, we sought out other features that could be used to determine which of two translations to select. We combined the outputs from a phrase-based system, Moses [Koehn et al., 2007] and a tree-to-tree system [Cowan et al., 2006]. Our main result is a 0.3 to 0.4 improvement in BLEU score over the best single system used, while also improving fluency and adequacy judgments. In addition, we used the same setup to directly predict which sentences would be judged by humans to be more fluent and more adequate. In those domains, we predicted the better sentence 6% to 7% more often than a baseline of always choosing the single best system.