Abstract
Factored models have been successfully used in many language pairs to improve translation quality in various aspects. In this work, we analyze this paradigm in an attempt at automating the search for well-performing machine translation systems.
We examine the space of possible factored systems, concluding that a fully automatic search for good configurations is not feasible. We demonstrate that even if results of automatic evaluation are available, guiding the search is difficult due to small differences between systems, which are further blurred by randomness in tuning.
We describe a heuristic for estimating the complexity of factored models. Finally, we discuss the possibilities of a "semi-automatic" exploration of the space in several directions and evaluate the obtained systems.