Abstract
System combination has proved to be a successful technique in the
pattern recognition field. However, several difficulties arise when combining the
outputs of tasks, e.g. machine translation, that generate structured patterns. So
far, machine translation system combination approaches either implement sophisticated classifiers to select one of the provided translations, or generate new sentences by combining the "best" subsequences of the provided translations. We present minimum Bayes' risk system combination (MBRSC), a system combination method for machine translation that gathers together the advantages of sentence-selection and subsequence-combination methods. MBRSC is able to detect and utilize the "best" subsequences of the provided translations to generate
the optimal consensus translation with respect to a particular performance met-
ric. Experiments show that MBRSC yields significant improvements in translation
quality.