Simulating Discriminative Training for Linear Mixture Adaptation in Statistical Machine Translation.
George Foster, Boxing Chen, Roland Kühn · NPARC · 2013
Linear mixture models are a simple and effective technique for performing domain adaptation of translation models in statis-tical MT. In this paper, we identify and correct two weaknesses of this method. First, we show that standard maximum-likelihood weights are biased toward large corpora, and that a straightforward pre-processing step that down-samples phrase tables can be used to counter this bias. Second, we show that features inspired by prototypical linear mixtures can be used to loosely simulate discriminative training for mixture models, with results that are almost certainly superior to true discrimi-native training. Taken together, these en-hancements yield BLEU gains of approx-imately 1.5 over existing linear mixture techniques for translation model adapta-tion. 1