Minimum error rate training is a crucial component to many state-of-the-art NLP applications, such as machine translation and speech recognition. However, common evaluation functions such as BLEU or word error rate are generally highly non-convex and thus prone to search errors. In this paper, we present LP-MERT, an exact search algorithm for minimum error rate training that reaches the global optimum using a series of reductions to linear programming. Given a set of N-best lists produced from S input sentences, this algorithm finds a linear model that is globally optimal with respect to this set. We find that this algorithm is polynomial in N and in the size of the model, but exponential in S. We present extensions of this work that let us scale to reasonably large tuning sets (e.g., one thousand sentences), by either searching only promising regions of the parameter space, or by using a variant of LP-MERT that relies on a beam-search approximation. Experimental results show improvements over the standard Och algorithm.

}, author = {Michel Galley and Chris Quirk}, booktitle = {Proc. of Empirical Methods in Natural Language Processing}, month = {July}, title = {Optimal Search for Minimum Error Rate Training}, url = {http://research.microsoft.com/apps/pubs/default.aspx?id=150580}, year = {2011}, }