Optimized Online Rank Learning for Machine Translation

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 253–262, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Optimized Online Rank Learning for Machine Translation Taro Watanabe National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0289 JAPAN {taro.watanabe}@nict.go.jp Abstract We present an online learning algorithm for statistical machine translation (SMT) based on stochastic gradient descent (SGD). Under the online setting of rank learning, a corpus-wise loss has to be approximated by a batch lo- cal loss when optimizing for evaluation mea- sures that cannot be linearly decomposed into a sentence-wise loss, such as BLEU. We pro- pose a variant of SGD with a larger batch size in which the parameter update in each iteration is further optimized by a passive-aggressive algorithm. Learning is efficiently parallelized and line search is performed in each round when merging parameters across parallel jobs. Experiments on the NIST Chinese-to-English Open MT task indicate significantly better translation results. 1 Introduction The advancement of statistical machine translation (SMT) relies on efficient tuning of several or many parameters in a model. One of the standards for such tuning is minimum error rate training (MERT) (Och, 2003), which directly minimize the loss of transla- tion evaluation measures, i.e. BLEU (Papineni et al., 2002). MERT has been successfully used in prac- tical applications, although, it is known to be un- stable (Clark et al., 2011). To overcome this insta- bility, it requires multiple runs from random start- ing points and directions (Moore and Quirk, 2008), or a computationally expensive procedure by linear programming and combinatorial optimization (Gal- ley and Quirk, 2011). Many alternative methods have been proposed based on the algorithms in machine learning, such as averaged perceptron (Liang et al., 2006), maximum entropy (Och and Ney, 2002; Blunsom et al., 2008), Margin Infused Relaxed Algorithm (MIRA) (Watan- abe et al., 2007; Chiang et al., 2008b), or pairwise rank optimization (PRO) (Hopkins and May, 2011). They primarily differ in the mode of training; on- line or MERT-like batch, and in their objectives; max-margin (Taskar et al., 2004), conditional log- likelihood (or softmax loss) (Berger et al., 1996), risk (Smith and Eisner, 2006; Li and Eisner, 2009), or ranking (Herbrich et al., 1999). We present an online learning algorithm based on stochastic gradient descent (SGD) with a larger batch size (Shalev-Shwartz et al., 2007). Like Hop- kins and May (2011), we optimize ranking in n- best lists, but learn parameters in an online fash- ion. As proposed by Haddow et al. (2011), BLEU is approximately computed in the local batch, since BLEU is not linearly decomposed into a sentence- wise score (Chiang et al., 2008a), and optimization for sentence-BLEU does not always achieve opti- mal parameters for corpus-BLEU. Setting the larger batch size implies the more accurate corpus-BLEU, but at the cost of slower convergence of SGD. There- fore, we propose an optimized update method in- spired by the passive-aggressive algorithm (Cram- mer et al., 2006), in …

Content truncated to 500 words