Random Restarts in Minimum Error Rate Training for Statistical Machine Translation

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 585–592 Manchester, August 2008 Random Restarts in Minimum Error Rate Training for Statistical Machine Translation Robert C. Moore and Chris Quirk Microsoft Research Redmond, WA 98052, USA bobmoore@microsoft.com, chrisq@microsoft.com Abstract Och’s (2003) minimum error rate training (MERT) procedure is the most commonly used method for training feature weights in statistical machine translation (SMT) mod- els. The use of multiple randomized start- ing points in MERT is a well-established practice, although there seems to be no published systematic study of its bene- fits. We compare several ways of perform- ing random restarts with MERT. We find that all of our random restart methods out- perform MERT without random restarts, and we develop some refinements of ran- dom restarts that are superior to the most common approach with regard to resulting model quality and training time. 1 Introduction Och (2003) introduced minimum error rate train- ing (MERT) for optimizing feature weights in sta- tistical machine translation (SMT) models, and demonstrated that it produced higher translation quality scores than maximizing the conditional likelihood of a maximum entropy model using the same features. Och’s method performs a series of one-dimensional optimizations of the feature weight vector, using an innovative line search that takes advantage of special properties of the map- ping from sets of feature weights to the resulting translation quality measurement. Och’s line search is guaranteed to find a global optimum, whereas more general line search methods are guaranteed only to find a local optimum. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. Global optimization along one dimension at a time, however, does not insure global optimization in all dimensions. Hence, as Och briefly men- tions, “to avoid finding a poor local optimum,” MERT can be augmented by trying multiple ran- dom starting points for each optimization search within the overall algorithm. However, we are not aware of any published study of the effects of ran- dom restarts in the MERT optimization search. We first compare a variant of Och’s method with and without multiple starting points for the op- timization search, selecting initial starting points randomly according to a uniform distribution. We find that using multiple random restarts can sub- stantially improve the resulting model in terms of translation quality as measured by the BLEU met- ric, but that training time also increases substan- tially. We next try selecting starting points by a random walk from the last local optimum reached, rather than by sampling from a uniform distrib- ution. We find this provides a slight additional improvement in BLEU score, and is significantly faster, although still slower than training without random restarts. Finally we look at two methods for speeding up training by pruning the set of hypotheses consid- ered. We find that, under some circumstances, this can speed up training so that it takes very little additional time compared to the original method without restarts, with no significant reduction …

Content truncated to 500 words