The Integration of Moses into Localization Industry

The Integration of Moses into Localization Industry Tomáš Hudík Moravia Worldwide Hilleho 4, Brno, Czech Republic thudik@moraviaworldwide.com Achim Ruopp Digital Silk Road 2800 Quebec St NW #816 Washington, DC 20008 achim@digitalsilkroad.net Abstract A majority of the localization indus- try is still unable to benefit from SMT, since it does not know how to integrate SMT system into its production workflow. This paper describes a workflow on how to integrate the open source SMT sys- tem Moses into computer-aided translation (CAT) tools. It introduces a software pack- age which is able to export XLIFF data into a special InlineText format that is fur- ther processed and is sent as plain text to Moses as input. Then, the translated text is inserted back to the original file. This transformation is needed, otherwise the use of MT in the localization indus- try would hardly be advantageous, since formatting information would have to be manually added by localization profes- sionals. A few attempts to bridge the local- ization industry and Moses were already done in the past, however, the method de- scribed here represents a new approach. 1 Credits This work was done with the support of Moravia Worldwide and in cooperation with Digital Silk Road and the Let’sMT! consortium. 2 Introduction An integration of Statistical Machine Translation (SMT) into Computer Aided Tools (CAT) has been needed for a long time (Ruopp, 2006). The localization industry started to use so-called Translation Memory (TM) (Reinke, 2006) some c 2011 European Association for Machine Translation. 30 years ago. A lot of valuable information is stored in these datasets. TMs, if available, are used for each translation project, since they contain things like a translated previous version of a prod- uct, or translation of some other product from the same client. With the deployment of TMs, transla- tion has become much easier, often pre-translated text just need to be post edited. Therefore, TMs are very useful for human translators and should be heavily used also by SMT systems. Mainly universities and research centers are cre- ating various SMT systems. Each of them provides two phases – training and translation. For both phases, it is necessary to have input. Due to com- pletely different environments, the issues involved and broad variability of research questions, univer- sities do not have a uniform format similar to TM. For this reason, SMT systems usually require gen- eral plain text as an input. The localization indus- try is different, its TMs contain a lot of additional information, for example XML, RTF, HTML tags, or even binary data. Some of it is related to for- matting – e.g. end of line, italics. Other kind of the information is related to content placeholders – e.g. untranslatable company names; specific trans- lations of dates, numbers, names etc. Such addi- tional information is not a big problem in the train- ing process, since additional formatting tags can be easily removed and special tags like placehold- ers can be substituted. A much bigger problem is to use …

Content truncated to 500 words