Abstract
Current state-of-the-art Statistical Machine Translation systems are based on log-linear models that combine a set of feature functions to score translation hypotheses during decoding. The models are parametrized by a vector of weights usually optimized on a set of sentences and their reference translations, called development data.
In this paper, we explore a (common and industry relevant) scenario where a system trained and tuned on general domain data needs to be adapted to a specific domain for which no or only very limited in-domain bilingual data is available. It turns out that such systems can be adapted successfully by re-tuning model parameters using surprisingly small amounts of parallel in-domain data, by cross-tuning or no tuning at all.
We show in detail how and why this is effective, compare the approaches and effort involved. We also study the effect of system hyperparameters (such as maximum phrase length and development data size) and their optimal values in this scenario.