Moses SMT

SAP has been heavily involved in the implementation and deployment of machine translation (MT) within the company since the early 1990s. In 2013, SAP initiated an extensive proof of concept project, based on the statistical MT system Moses (Koehn, et al., 2007), in collaboration with the external implementation partner CrossLang. The project focused on the use of Moses SMT as an aid to translators in the production process. This paper describes the outcome of the productivity evaluation for technical documents pertaining to SAP’s Rapid Deployment Solutions (RDS), which was performed as part of this proof of concept project. 1 Background and Project Description The use of machine translation at SAP dates back to the early 1990s. Originally the rule-based approach was deployed mainly for the translation of technical troubleshooting documents (SAP Notes), test cases, documentation, training materials, and as a gist translation tool for customer messages. MT systems used were METAL (German-English/English-German) and Logos (English-French mainly), followed by the next generation system Lucy LT (for these same languages). In 2012, SAP started experimenting with statistical machine translation (SMT). A prototype system was built at SAP Language Services (SLS) for the Chinese-to-English and English-to-Chinese language pairs. This prototype was based on the Moses SMT technology. In 2013, SLS initiated a more extensive proof of concept project, again based on Moses, in collaboration with the external implementation part© 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC-BY-ND. ner CrossLang. The project focused on the use of Moses SMT as an aid to translators in the production process. In that context CrossLang developed a plugin for SDL Trados Studio, thus enabling a seamless integration of Moses SMT into the SDL Trados Studio environment. MT suggestions were provided to translators during the proof of concept projects in addition to translation memory (TM) segments, which translators were free to accept, edit or discard just as they would TM matches. The overall timeline for the project was rather ambitious as all project phases (MT engine development, piloting, evaluation and engine improvement) had to be run between July and December 2013. In 2014, the SLS MT team will take additional steps to align machine translation landscapes and further extend the MT offering to various usage scenarios and more content types. The proof of concept projects were carried out for two different content types: sap.com and RDS (Rapid Deployment Solutions) texts. While sap.com materials are typically texts used for SAP’s official website, RDS texts are technical documents related to SAP’s RDS product offering. Consequently the former content type can be classified as being of a more creative nature and thus more marketing-like than the latter, which is more technical by nature and hence more similar to documentation. The present paper will focus on the RDS content type. The language scope of the proof of concept phase comprised the eight target languages Chinese, French, German, Italian, Japanese, Portuguese (Brazil), Russian and Spanish with source language English as well as the respective reverse language directions. However, the evaluations subject to this paper were carried out only for the target languages Chinese, French, German, and Russian. For each language pair and content type Moses engines were built in three iterations:  Iteration 1: Engines built with content type-specific data only (in-domain engines)