LOGON — A Norwegian MT Effort Jan Tore Lønning ♣ , Stephan Oepen ♣♠ , Dorothee Beermann ♥ , Lars Hellan ♥ , John Carroll △ , Helge Dyvik ♦ , Dan Flickinger ♠ , Janne Bondi Johannessen ♣ , Paul Meurer ♦ , Torbjørn Nordg˚ ard ♥ , Victoria Ros´ en ♦ , Erik Velldal ♣ ♣ Universitetet i Oslo, Boks 1102 Blindern; 0317 Oslo (Norway) ♦ Universitetet i Bergen, Sydnesplassen 7, 5007 Bergen (Norway) ♥ Norges Teknisk-Naturvitenskapelige Universitet, 7491 Trondheim (Norway) ♠ Center for the Study of Language and Information, Stanford, CA 94305 (USA) △ University of Sussex, Falmer, Brighton BN1 9QH (UK) all@emmtee.net Abstract We present an overview over the Nor- wegian LOGON initiative, an ongo- ing project to produce high-quality MT from Norwegian to English based on grammar-based analysis, semantic transfer, and full generation in the target language. 1 Introduction The LOGON project, focusing on machine trans- lation from Norwegian to English, involves three universities (Oslo, Bergen and NTNU Trondheim) and has a public funding from the Norwegian Re- search Council’s program for language technology (KUNSTI, 2002 – 2006) on the order of 20 mil- lions NOK. KUNSTI emerged as a reaction to the global growth in language technology at the turn of the millennium with the following two con- cerns. Firstly, language technology applications should not only be available for English and other large languages but also for Norwegian. Secondly, Norway as a country should not fall behind in this new and growing industry. As a response to the first concern, KUNSTI asked for projects with a focus on Norwegian language, over time giving rise to reusable language technology resources. As a reaction to the second concern, it was important to raise the national competence, in particular ed- ucate more PhDs within the field. It was also a goal to initiate large projects resulting in working demonstrators that involve several sub-areas of the field and a diversification of methods. Machine translation turned out to be a suitable task. 2 The Approach We have chosen a traditional semantic transfer- based approach as our starting point. In spite of strong winds in the direction of statistical meth- ods during the last decennium, in language tech- nology in general, and machine translation in par- ticular, we are still firm believers in symbolic and ‘deep’ linguistic methods. Although statistical ap- proaches can deliver good initial results to MT, they seem to sooner or later suffer from ‘ceiling’ effects in performance and ask to be augmented with more linguistic structure. Our approach is to start with a firm and the- oretically sound symbolic backbone, while aug- menting this with probabilistic methods to direct the choices where the symbolic methods fan out. The transfer-based approach falls into three steps. (i) An in-depth grammatical and semantic anal- ysis of Norwegian resulting in language-specific logical semantic representations. (ii) A transfer of these representations into language-specific En- glish representations. (iii) And, finally, generation from the semantic representations to English sen- tences. A central locus in the project is the format of the semantic representations, where our start- ing point is Minimal Recursion Semantics (MRS;