The notion of a reversible MT system was first expressed by Landsbergen . Such a system will in principle produce a set of possible translations, by employing linguistic knowledge only. Choosing the best translation from the set of linguistically possible translations will usually require other sources of knowledge, either incorporated in the system or provided (interactively) by the user. The relation `possible translation' is symmetric whereas the relation `best translation' is not. Thus an MT system may consist of a reversible core, implementing the symmetric relation `possible translation', and additional components (not necessarily reversible) to select the best translation.
Not only is it possible to build reversible (modules of) MT systems; it has also been claimed that reversible systems are preferable. For example Isabelle  claims that reversible MT systems are to be preferred to others because in reversible MT systems a better understanding of the translation relation is achieved; such systems will eventually exhibit better practical performance. Moreover, the arguments in favour of using bidirectional grammars in NLP, such as those given in  carry over to translation as well.
Because of the declarative nature of unification- and logic grammar formalisms grammars written in these formalisms are increasingly used in a bidirectional way, thus the same grammar is used for both parsing and generation. Some recent developments are reported in .
In this paper I will show how to use such bidirectional unification grammars to build a completely reversible, multilingual, MT system. For each language there is a unification grammar that defines a reversible relation between strings and language dependent meaning representations (logical forms). Moreover, for each language pair (or set of languages) there is a unification grammar that defines a reversible relation between such language dependent logical forms. Translation is thus defined by a series of three unification grammars.
A specific version of the system that is described here is implemented as the core of the experimental MiMo2 translation system . This system aims at translating international news items on Teletext. Apart from unification grammars the system uses a bidirectional two-level orthography component. Language dependent meanings are represented as simple predicate argument structures with some extra labels indicating `universal' meaning such as tense and aspect. The current system (November 1989) includes grammars for Dutch, Spanish and English.
The paper is set up as follows. In section 2, I will give some examples that show how bidirectional unification grammars can be used to define relations between logical forms of different languages. In section 3, reversibility is defined in terms of symmetry and computability. Possible approaches to obtain reversibility are discussed. In section 4, I will compare the current approach with some other approaches in the unification based translation paradigm and discuss some problems and future directions.