Regularization and Search for Minimum Error Rate Training

Proceedings of the Third Workshop on Statistical Machine Translation, pages 26–34, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Regularization and Search for Minimum Error Rate Training Daniel Cer, Daniel Jurafsky, and Christopher D. Manning Stanford University Stanford, CA 94305 cerd,jurafsky,manning@stanford.edu Abstract Minimum error rate training (MERT) is a widely used learning procedure for statistical machine translation models. We contrast three search strategies for MERT: Powell’s method, the variant of coordinate descent found in the Moses MERT utility, and a novel stochastic method. It is shown that the stochastic method obtains test set gains of +0.98 BLEU on MT03 and +0.61 BLEU on MT05. We also present a method for regularizing the MERT objec- tive that achieves statistically significant gains when combined with both Powell’s method and coordinate descent. 1 Introduction Och (2003) introduced minimum error rate training (MERT) as an alternative training regime to the con- ditional likelihood objective previously used with log-linear translation models (Och & Ney, 2002). This approach attempts to improve translation qual- ity by optimizing an automatic translation evalua- tion metric, such as the BLEU score (Papineni et al., 2002). This is accomplished by either directly walk- ing the error surface provided by an evaluation met- ric w.r.t. the model weights or by using gradient- based techniques on a continuous approximation of such a surface. While the former is piecewise con- stant and thus cannot be optimized using gradient techniques, Och (2003) provides an approach that performs such training efficiently. In this paper we explore a number of variations on MERT. First, it is shown that performance gains can be had by making use of a stochastic search strategy as compare to that obtained by Powell’s method and coordinate descent. Subsequently, results are pre- sented for two regularization strategies1. Both allow coordinate descent and Powell’s method to achieve performance that is on par with stochastic search. In what follows, we briefly review minimum er- ror rate training, introduce our stochastic search and regularization strategies, and then present experi- mental results. 2 Minimum Error Rate Training Let F be a collection of foreign sentences to be translated, with individual sentences f0, f1, . . . , fn. For each fi, the surface form of an indi- vidual candidate translation is given by ei with hidden state hi associated with the derivation of ei from fi. Each ei is drawn from E, which represents all possible strings our translation sys- tem can produce. The (ei, hi, fi) triples are con- verted into vectors of m feature functions by Ψ : E × H × F → Rm whose dot product with the weight vector w assigns a score to each triple. The idealized translation process then is to find the highest scoring pair (ei, hi) for each fi, or rather (ei, hi) = argmax(e∈E,h∈H) w · Ψ(e, h, f). The aggregate argmax for the entire data set F is given by equation (1)2. This gives Ew which repre- sents the set of translations selected by …

Content truncated to 500 words