This chapter presents the architecture of a reversible Machine Translation (MT) system. This architecture is at the core of the MiMo2 prototype which was developed by the author and colleagues at the University of Utrecht [103,99,104]. The chapter is organized as follows.
In the first section I discuss the notion `linguistically possible translation', and ways in which we could go about defining linguistically possible translations. A problem, called the subset problem, will be encountered which motivates us to use a specific architecture for MT. In this architecture, which will be presented in some detail in section 5.3, translation is defined by a series of three reversible constraint-based grammars.
In section 5.4 I present some exemplification of the use of constraint-based grammars to define transfer relations between pairs of languages.
In section 5.5 I discuss how to ensure that such transfer components are reversible. It turns out that by requiring that translation is compositional in a sense to be made precise, transfer grammars can be shown to be reversible.
Section 5.6 goes to show that using reversible constraint-based grammars for transfer is expressive enough to define some so-called non-compositional translations, which were problematic for less powerful formalisms. Our conclusion then will be that reversible transfer grammars may be a good compromise between expressive power and computational feasibility.