Bridging SMT and TM with Translation Recommendation

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 622–630, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Bridging SMT and TM with Translation Recommendation Yifan He Yanjun Ma Josef van Genabith Centre for Next Generation LocalisationAndy Way School of Computing Dublin City University {yhe,yma,josef,away}@computing.dcu.ie Abstract We propose a translation recommendation framework to integrate Statistical Machine Translation (SMT) output with Transla- tion Memory (TM) systems. The frame- work recommends SMT outputs to a TM user when it predicts that SMT outputs are more suitable for post-editing than the hits provided by the TM. We describe an im- plementation of this framework using an SVM binary classifier. We exploit meth- ods to fine-tune the classifier and inves- tigate a variety of features of different types. We rely on automatic MT evalua- tion metrics to approximate human judge- ments in our experiments. Experimental results show that our system can achieve 0.85 precision at 0.89 recall, excluding ex- act matches. Furthermore, it is possible for the end-user to achieve a desired balance between precision and recall by adjusting confidence levels. 1 Introduction Recent years have witnessed rapid developments in statistical machine translation (SMT), with con- siderable improvements in translation quality. For certain language pairs and applications, automated translations are now beginning to be considered acceptable, especially in domains where abundant parallel corpora exist. However, these advances are being adopted only slowly and somewhat reluctantly in profes- sional localization and post-editing environments. Post-editors have long relied on translation memo- ries (TMs) as the main technology assisting trans- lation, and are understandably reluctant to give them up. There are several simple reasons for this: 1) TMs are useful; 2) TMs represent con- siderable effort and investment by a company or (even more so) an individual translator; 3) the fuzzy match score used in TMs offers a good ap- proximation of post-editing effort, which is useful both for translators and translation cost estimation and, 4) current SMT translation confidence esti- mation measures are not as robust as TM fuzzy match scores and professional translators are thus not ready to replace fuzzy match scores with SMT internal quality measures. There has been some research to address this is- sue, see e.g. (Specia et al., 2009a) and (Specia et al., 2009b). However, to date most of the research has focused on better confidence measures for MT, e.g. based on training regression models to per- form confidence estimation on scores assigned by post-editors (cf. Section 2). In this paper, we try to address the problem from a different perspective. Given that most post- editing work is (still) based on TM output, we pro- pose to recommend MT outputs which are better than TM hits to post-editors. In this framework, post-editors still work with the TM while benefit- ing from (better) SMT outputs; the assets in TMs are not wasted and TM fuzzy match scores can still be used to estimate (the upper bound of) post- editing labor. There are three specific goals we …

Content truncated to 500 words