A Comparison of Update Strategies for Large-Scale Maximum Expected BLEU Training

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1516–1526, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics A Comparison of Update Strategies for Large-Scale Maximum Expected BLEU Training Joern Wuebker, Sebastian Muehr, Patrick Lehnen, Stephan Peitz and Hermann Ney Human Language Technology and Pattern Recognition Group RWTH Aachen University Aachen, Germany {surname}@cs.rwth-aachen.de Abstract This work presents a flexible and efficient discriminative training approach for statisti- cal machine translation. We propose to use the RPROP algorithm for optimizing a max- imum expected BLEU objective and experi- mentally compare it to several other updat- ing schemes. It proves to be more effi- cient and effective than the previously pro- posed growth transformation technique and also yields better results than stochastic gra- dient descent and AdaGrad. We also report strong empirical results on two large scale tasks, namely BOLT Chinese→English and WMT German→English, where our final sys- tems outperform results reported by Setiawan and Zhou (2013) and on matrix.statmt.org. On the WMT task, discriminative training is per- formed on the full training data of 4M sen- tence pairs, which is unsurpassed in the litera- ture. 1 Introduction The main advantage of learning parameters in a dis- criminative fashion is the possibility to directly opti- mize towards a quality or error measure on the task that is being performed. This stands in contrast to the generative approach, where parameters are cho- sen to maximize likelihood under a generative story, which often bears little correspondence with the ac- tual application of the model. In statistical machine translation (SMT), ex- tending the generative noisy-channel formulation (Brown et al., 1993) as a discriminative, log-linear combination of multiple models (Och, 2003) has be- come the state of the art. However, most of the component models are still estimated by heuristics or generative training. In this paper, a flexible, effi- cient and easy to implement discriminative training scheme for SMT is presented. It can be applied to any kind and any number of features. We use the RPROP algorithm to optimize a maximum expected BLEU objective. n-best lists approximate the infea- sibly large space of translation hypotheses. They are generated with the application of leave-one-out to make them more representative with respect to un- seen data. We make the following main contributions: 1. We propose to apply the RPROP algorithm for maximum expected BLEU training and per- form an experimental comparison with growth transformation (GT) (He and Deng, 2012; Setiawan and Zhou, 2013), stochastic gradi- ent descent (Auli et al., 2014) and AdaGrad (Green et al., 2013). RPROP yields supe- rior performance, reaching a total improve- ment of 1.2 BLEU points over our IWSLT German→English baseline using 5.22M fea- tures. 2. In terms of time and memory efficiency, RPROP clearly outperforms GT. The latter needs to update a much larger number of fea- tures due to its renormalization component. On the IWSLT data, RPROP is 6.4 times faster than GT and requires a third of the …

Content truncated to 500 words