SampleRank Training for Phrase-Based Machine Translation

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 261–271, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics SampleRank Training for Phrase-Based Machine Translation Barry Haddow School of Informatics University of Edinburgh bhaddow@inf.ed.ac.uk Abhishek Arun Microsoft UK abarun@microsoft.com Philipp Koehn School of Informatics University of Edinburgh pkoehn@inf.ed.ac.uk Abstract Statistical machine translation systems are normally optimised for a chosen gain func- tion (metric) by using MERT to find the best model weights. This algorithm suffers from stability problems and cannot scale beyond 20-30 features. We present an alternative al- gorithm for discriminative training of phrase- based MT systems, SampleRank, which scales to hundreds of features, equals or beats MERT on both small and medium sized systems, and permits the use of sentence or document level features. SampleRank proceeds by repeatedly updating the model weights to ensure that the ranking of output sentences induced by the model is the same as that induced by the gain function. 1 Introduction In phrase-based machine translation (PBMT), the standard approach is to express the probability dis- tribution p(a, e|f) (where f is the source sentence and (a, e) is the aligned target sentence) in terms of a linear model based on a small set of feature func- tions p(a, e|f) ∝ exp � n � i=1 wihi(a, e, f) � (1) The feature functions {hi} typically include log probabilities of generative models such as trans- lation, language and reordering, as well as non- probabilistic features such as word, phrase and dis- tortion penalties. The feature weights w = {wi} are normally trained using MERT (minimum error rate training) (Och, 2003), to maximise performance as measured by an automated metric such as BLEU (Papineni et al., 2002). MERT training uses a par- allel data set (known as the tuning set) consisting of about 1000-2000 sentences, distinct from the data set used to build the generative models. Optimis- ing the weights in Equation (1) is often referred to as tuning the MT system, to differentiate it from the process of training the generative models. MERT’s inability to scale beyond 20-30 features, as well as its instability (Foster and Kuhn, 2009) have led to investigation into alternative ways of tuning MT systems. The development of tuning methods is complicated, however by, the use of BLEU as an objective function. This objective in its usual form is not differentiable, and has a highly non-convex error surface (Och, 2003). Furthermore BLEU is evaluated at the corpus level rather than at the sentence level, so tuning methods either have to consider the entire corpus, or resort to a sentence- level approximation of BLEU. It is unlikely, how- ever, that the difficulties in discriminative MT tun- ing are due solely to the use of BLEU as a metric – because evaluation of translation is so difficult, any reasonable gain function is likely to have a complex relationship with the model parameters. Gradient-based tuning methods, such as mini- mum risk training, have been investigated as pos- sible alternatives to …

Content truncated to 500 words