Natural language generation utilising linguistically-motivated, general-purpose, grammars can conveniently be thought of as involving two components. The first component involves constructing some logical form that is accepted by the grammar, while the second (called tactical generation, or now more usually, realisation) involves going from the logical form to a string. My main interest is in realisation, and in particular, the implications that supporting tractable realisation has for the nature of logical form in computational grammars. I am not concerned here with denotation, but with the form of the semantic representation and with the way that composition operates. There are two main considerations. Firstly, we can make the realisation problem more or less difficult, depending on the meaning representation we choose, even though the denotation is the same. Secondly, structure in the logical form may be exploited to guide realisation, and this may be necessary for realisation to work at all. But structure that reflects the syntax of the language too closely is generally to be avoided, since it can make it impossible for the input to the realiser to be constructed without detailed knowledge of the grammar. The Minimal Recursion Semantics (MRS: Copestake et al, in press) approach was designed to be usable for generation.

Robust realisation has three requirements:

1. A realisation algorithm that is efficient with minimal guidance from the logical form syntax.
2. A formalism that allows sufficient abstraction over syntax that constructing the input to the realiser is doable.
3. An interface layer expressing those constraints imposed by the grammar that have to be known to construct the input.

Ann Copestake, Dan Flickinger, Ivan Sag and Carl Pollard Minimal Recursion Semantics: An introduction Research on Language and Computation in press