Incremental learning of transfer rules for customized machine translation

Abstract

In this paper we present a machine translation system, which translates Japanese into German. We have developed a transfer-based architecture in which the transfer rules are learnt incrementally from translation examples provided by a user. This means that there are no handcrafted rules, but, on the contrary, the user can customize the system according to his own preferences. The translation system has been implemented by using Amzi! Prolog. This programming environment had the big advantage of offering sufficient scalability even for large lexicons and rule bases, powerful unification operations for the application of transfer rules, and full Unicode support for Japanese characters. Finally, the application programming interface to Visual Basic made it possible to design an embedded translation environment so that the user can use Microsoft Word to work with the Japanese text and invoke the translation features directly from within the text editor. We have integrated the machine translation system into a language learning environment for German-speaking language students to create a Personal Embedded Translation and Reading Assistant (PETRA). © Springer- Verlag Berlin Heidelberg 2005.