Large-scale Expected BLEU Training of Phrase-based Reordering Models

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1250–1260, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Large-scale Expected BLEU Training of Phrase-based Reordering Models Michael Auli, Michel Galley, Jianfeng Gao Microsoft Research Redmond, WA, USA {michael.auli,mgalley,jfgao}@microsoft.com Abstract Recent work by Cherry (2013) has shown that directly optimizing phrase-based re- ordering models towards BLEU can lead to significant gains. Their approach is lim- ited to small training sets of a few thou- sand sentences and a similar number of sparse features. We show how the ex- pected BLEU objective allows us to train a simple linear discriminative reordering model with millions of sparse features on hundreds of thousands of sentences re- sulting in significant improvements. comparison to likelihood training demon-A strates that expected BLEU is vastly more effective. Our best results improve a hi- erarchical lexicalized reordering baseline by up to 2.0 BLEU in a single-reference setting on a French-English WMT 2012 setup. 1 Introduction Modeling reordering for phrase-based machine translation has been a long standing problem. Contrary to synchronous context free grammar- based translation models (Wu, 1997; Galley et al., 2004; Galley et al., 2006; Chiang, 2007), phrase- based models (Koehn et al., 2003; Och and Ney, 2004) have no in-built notion of reordering beyond what is captured in a single phrase pair, and the first phrase-based decoders simply scored inter- phrase reorderings using a restricted linear dis- tortion feature, which scores a phrase reordering proportionally to the length of its displacement. While phrase-based models allow in theory com- pletely unrestricted reordering patterns, move- ments are generally limited to a finite distance for complexity reasons. To address this limitation, extensive prior work focused on richer feature sets, in particular on lexicalized reordering mod- els trained with maximum likelihood-based ap- proaches (Tillmann, 2003; Xiong et al., 2006; Gal- ley and Manning, 2008; Nguyen et al.,2009;§2). More recently, Cherry (2013) proposed a very effective sparse ordering model relying on a set of only a few thousand indicator features which are trained towards a task-specific metric such as BLEU (Papineni et al., 2002). These features are simply added to the log-linear framework of translation that is trained with the Margin Infused Relaxed Algorithm (MIRA; Chiang et al., 2009) on a small development set of a few thousand sentences. While simple, the approach outper- forms the state-of-the-art hierarchical reordering model of Galley and Manning (2008), a maximum likelihood-based model trained on millions of sen- tences to fit millions of parameters. Ideally, we would like to scale sparse reorder- ing models to similar dimensions but recent at- tempts to increase the amount of training data for MIRA was met with little success (Eidelman et al., 2013). In this paper we propose much larger sparse ordering models that combine the scalabil- ity of likelihood-based approaches with the higher accuracy of maximum BLEU training (§3). We train on the output of a hierarchical reordering model-based system and scale to millions of fea- tures learned on hundreds of thousands …

Content truncated to 500 words