{gwb24ad465wjb31}@cam.ac.uk重点.PDFVIP

Fluency Constraints for Minimum Bayes-Risk Decoding of Statistical Machine Translation Lattices ` Graeme Blackwood and Adria de Gispert and William Byrne Cambridge University Engineering Department, Trumpington Street, CB2 1PZ, U.K. {gwb24|ad465|wjb31}@cam.ac.uk Abstract 2009) show broad correlation with human rank- ings of MT quality, but are not capable of making A novel and robust approach to improv- fine distinctions between fluency and adequacy. ing statistical machine translation flu- There is concern that the fluency of current ency is developed within a minimum SMT is inadequate (Knight, 2007b). SMT is ro- Bayes-risk decoding framework. By seg- bust, in that a translation is nearly always pro- menting translation lattices according to duced. But unlike translators who should be confidence measures over the maximum skilled in at least one of the languages, SMT sys- likelihood translation hypothesis we are tems are limited in both source and target lan- able to focus on regions with potential guage competence. Fluency and accuracy there- translation errors. Hypothesis space con- fore tend to suffer together as translation quality straints based on monolingual coverage degrades. This should not be the case. Ideally, are applied to the low confidence regions an SMT system should never be any less fluent to improve overall translation fluency. than the best stochastic text generation system 1 Introduction and Motivation