Regularized Minimum Error Rate Training

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1948–1959, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Regularized Minimum Error Rate Training Michel Galley Microsoft Research mgalley@microsoft.com Chris Quirk Microsoft Research chrisq@microsoft.com Colin Cherry National Research Council colin.cherry@nrc-cnrc.gc.ca Kristina Toutanova Microsoft Research kristout@microsoft.com Abstract Minimum Error Rate Training (MERT) re- mains one of the preferred methods for tun- ing linear parameters in machine translation systems, yet it faces significant issues. First, MERT is an unregularized learner and is there- fore prone to overfitting. Second, it is com- monly used on a noisy, non-convex loss func- tion that becomes more difficult to optimize as the number of parameters increases. To ad- dress these issues, we study the addition of a regularization term to the MERT objective function. Since standard regularizers such as ℓ2 are inapplicable to MERT due to the scale invariance of its objective function, we turn to two regularizers—ℓ0 and a modification of ℓ2— and present methods for efficiently integrating them during search. To improve search in large parameter spaces, we also present a new direc- tion finding algorithm that uses the gradient of expected BLEU to orient MERT’s exact line searches. Experiments with up to 3600 features show that these extensions of MERT yield re- sults comparable to PRO, a learner often used with large feature sets. 1 Introduction Minimum Error Rate Training emerged a decade ago (Och, 2003) as a superior training method for small numbers of linear model parameters of machine translation systems, improving over prior work using maximum likelihood criteria (Och and Ney, 2002). This technique quickly rose to prominence, becom- ing standard in many research and commercial MT systems. Variants operating over lattices (Macherey et al., 2008) or hypergraphs (Kumar et al., 2009) were subsequently developed, with the benefit of reducing the approximation error from n-best lists. The primary advantages of MERT are twofold. It directly optimizes the evaluation metric under consid- eration (e.g., BLEU) instead of some surrogate loss. Secondly, it offers a globally optimal line search. Un- fortunately, there are several potential difficulties in scaling MERT to larger numbers of features, due to its non-convex loss function and its lack of reg- ularization. These challenges have prompted some researchers to move away from MERT, in favor of lin- early decomposable approximations of the evaluation metric (Chiang et al., 2009; Hopkins and May, 2011; Cherry and Foster, 2012), which correspond to easier optimization problems and which naturally incorpo- rate regularization. In particular, recent work (Chiang et al., 2009) has shown that adding thousands or tens of thousands of features can improve MT quality when weights are optimized using a margin-based approximation. On simulated datasets, Hopkins and May (2011) found that conventional MERT strug- gles to find reasonable parameter vectors, where a smooth loss function based on Pairwise Ranking Op- timization (PRO) performs much better; on real data, this PRO method appears at least as good as MERT on small feature sets, and also scales better …

Content truncated to 500 words