What is NLG?

Giving an adequate general definition of the input to natural language generation (NLG), and hence to NLG itself, is a notoriously difficult problem, practically, theoretically and even methodologically. In this paper, we describe our recent experiences of implementing an NLG component of a larger question-answering system, and trying to understand and resolve some of these problems in this context. We examine the whole lifetime of an answer, from internal data structure to final expression as text, and look for characteristics of the processing which might help identify where NLG really begins. On the basis of this analysis we propose some principles to inform discussions on the scope of NLG as an individuated enterprise.