Minimum Error Rate Training by Sampling the Translation Lattice

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 606–615, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Minimum Error Rate Training by Sampling the Translation Lattice Samidh Chatterjee∗ Department of Computer Science Florida State University, USA chatterj@cs.fsu.edu Nicola Cancedda Xerox Research Centre Europe 6 Chemin de Maupertuis, 38240 Meylan, France nicola.cancedda@xrce.xerox.com Abstract Minimum Error Rate Training is the algo- rithm for log-linear model parameter train- ing most used in state-of-the-art Statistical Machine Translation systems. In its original formulation, the algorithm uses N-best lists output by the decoder to grow the Transla- tion Pool that shapes the surface on which the actual optimization is performed. Recent work has been done to extend the algorithm to use the entire translation lattice built by the decoder, instead of N-best lists. We pro- pose here a third, intermediate way, consist- ing in growing the translation pool using sam- ples randomly drawn from the translation lat- tice. We empirically measure a systematic im- provement in the BLEU scores compared to training using N-best lists, without suffering the increase in computational complexity as- sociated with operating with the whole lattice. 1 Introduction Most state-of-the-art Statistical Machine Translation (SMT) systems are based on a log-linear model of the conditional probability of generating a certain translation given a specific source sentence. More specifically, the conditional probability of a transla- tion e and a word alignment a given a source sen- tence f is modeled as: ∗The work behind this paper was done during an intern- ship at the Xerox Research Centre Europe. The author was par- tially supported by NSF through Grant CCF-0643593 and the AFOSR Young Investigator Research Program. P(e, a|f) ∝ exp � K � k=1 λkhk (e,a,f) � (1) where the hk(e,a,f) are feature functions provid- ing complementary sources of information on the quality of the produced translation (and alignment). Once such a model is known: the decoder (i.e. the actual translation program), which builds a transla- tion by searching in the space of all possible transla- tions the one that maximizes the conditional proba- bility: (e∗, a∗) = arg max e,a K � k=1 λkhK(e,a,f) (2) where we have taken into account that the exponen- tial is monotonic. The parameters λk determine the relative impor- tance of the different feature functions in the global score. Best results are typically obtained by search- ing in the space of all possible parameter vectors ¯λ for the one that minimizes the error on a held-out development dataset for which one or more refer- ence human translations are available, as measured by some automatic measure. This procedure is re- ferred to as Minimum Error Rate Training (MERT). 1.1 Minimum Error Rate Training on N-best Lists The most widespread MERT algorithm is the one described in (Och, 2003). This algorithm starts by initializing the parameter vector ¯λ. For each source sentence in the development set, the decoder 606 is used to initialize a translation pool with a list of N- best …

Content truncated to 500 words