Och's (2003) minimum error rate training (MERT) procedure is the most commonly used method for training feature weights in statistical machine translation (SMT) models. The use of multiple randomized starting points in MERT is a well-established practice, although there seems to be no published systematic study of its benefits. We compare several ways of performing random restarts with MERT. We find that all of our random restart methods outperform MERT without random restarts, and we develop some refinements of random restarts that are superior to the most common approach with regard to resulting model quality and training time.
Robert C. Moore, Chris Quirk