Fluency Constraints for Minimum Bayes-Risk Decoding of Statistical Machine Translation Lattices

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 71–79, Beijing, August 2010 Fluency Constraints for Minimum Bayes-Risk Decoding of Statistical Machine Translation Lattices Graeme Blackwood and Adri`a de Gispert and William Byrne Machine Intelligence Laboratory, Department of Engineering, Cambridge University {gwb24|ad465|wjb31}@cam.ac.uk Abstract A novel and robust approach to improv- ing statistical machine translation fluency is developed within a minimum Bayes- risk decoding framework. By segment- ing translation lattices according to con- fidence measures over the maximum like- lihood translation hypothesis we are able to focus on regions with potential transla- tion errors. Hypothesis space constraints based on monolingual coverage are ap- plied to the low confidence regions to im- prove overall translation fluency. 1 Introduction and Motivation Translation quality is often described in terms of fluency and adequacy. Fluency reflects the ‘na- tiveness’ of the translation while adequacy indi- cates how well a translation captures the meaning of the original text (Ma and Cieri, 2006). From a purely utilitarian view, adequacy should be more important than fluency. But fluency and adequacy are subjective and not easy to tease apart (Callison-Burch et al., 2009; Vilar et al., 2007). There is a human tendency to rate less fluent trans- lations as less adequate. One explanation is that errors in grammar cause readers to be more crit- ical. A related phenomenon is that the nature of translation errors changes as fluency improves so that any errors in fluent translations must be rel- atively subtle. It is therefore not enough to fo- cus solely on adequacy. SMT systems must also be fluent if they are to be accepted and trusted. It is possible that the reliance on automatic met- rics may have led SMT researchers to pay insuffi- cient attention to fluency: BLEU (Papineni et al., 2002), TER (Snover et al., 2006), and METEOR (Lavie and Denkowski, 2009) show broad corre- lation with human rankings of MT quality, but are incapable of fine distinctions between fluency and adequacy. There is concern that the fluency of current SMT is inadequate (Knight, 2007b). SMT is ro- bust, in that a translation is nearly always pro- duced. But unlike translators who should be skilled in at least one of the languages, SMT sys- tems are limited in both source and target lan- guage competence. Fluency and accuracy there- fore tend to suffer together as translation quality degrades. This should not be the case. Ideally, an SMT system should never be any less fluent than the best stochastic text generation system avail- able in the target language (Oberlander and Brew, 2000). What is needed is a good way to enhance the fluency of SMT hypotheses. The maximum likelihood (ML) formulation (Brown et al., 1990) of translation of source lan- guage sentence F to target language sentence ˆE ˆE = argmax E P(F|E)P(E) (1) makes it clear why improving SMT fluency is a difficult modelling problem. The language model P(E), the closest thing to a ‘fluency component’ in the original formulation, only affects candidates likely …

Content truncated to 500 words