Domain specific MT in use

Domain specific MT in use Lene Offersgaard1, Claus Povlsen1, Lisbeth Almsten2, and Bente Maegaard1 1Center for Sprogteknologi, Københavns Universitet, Njalsgade 140, 2300 København S, Denmark 2Inter-set Translation,Vestergade 13A, 6500 Vojens, Denmark {leneo, cpovlsen, bmaegaard}@hum.ku.dk, LKJ@intersetgroup.dk Abstract. This paper focuses on domain specific use of MT with a special focus on SMT in the workflow of a Language Service Provider (LSP). We report on the feedback of post-editors using fluency/adequacy evaluation and the evaluation metric ’Usability’, understood in this context as where users on a three point scale evaluate the sentence from the point of view of the post-editor. The post-editor profile defined by the LSP is based on the experiences of introducing MT in the LSP workflow. The relation between the Translation Edit Rate (TER) scores and ‘Usability’ scores is tested. We find TER a candidate for an automatic metric simulating the post-editors’ usability judgements. LSP tests show 67% saved time in post-editing for the tested domain. Finally, the use of weighted sub-domain phrase tables in a SMT system is shown to improve translation quality. Introduction As part of a general strategy to strengthen cooperation between the research community and small and medium-sized enterprises, the Danish Council for Strategic Research has decided to co-finance two projects involving machine translation. The aim of both projects has been to explore the possibilities of using statistical machine translation (SMT) approaches in small and medium sized companies (SMEs). The project goals were to find out not only whether it would possible for translation companies to integrate SMT systems in their daily translation flow, but also to assess whether it would be financially beneficial. The primary tasks of the involved translation companies have been to provide bilingual corpora consisting of sentence aligned documents and then subsequently to test and evaluate the translation results as a first step to uncover the commercial potential of using SMT in their translation process. This paper builds on the results of the second project, mainly carried out in 2007, involving at the research side Copenhagen Business School and the University of Copenhagen and on the business side Inter-Set, a medium-sized language service provider (LSP) with subsidiaries in other countries. The Open Source Moses MT system [1] was used both for training the SMT system and for translation. MOSES is currently mainly supported under the EuroMatrix project, funded by the European Commission. The language model was 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 150 trained using the language modelling toolkit IRSTLM [2]. The language models were trained with order 5. The maximum length of phrases in the phrase tables was set to 5. Domain Issues in SMT The assumption that there would be a productivity gain using SMT was the prime motivation factor for the LSP to investigate the potential of SMT systems. One of the issues to be considered in this context was the handling of subject domains. In an ideal world, all users involved with translation of technical documents would apply the same large-scale general subject classification system …

Content truncated to 500 words