The Mikrokosmos project utilizes an efficient, constraint-directed control architecture named Hunter-Gatherer (HG). [Beale et al. 1996] overviews how it enables semantic analysis to be performed in near linear-time. Its use in generation is quite similar. [Beale1997] describes the architecture in detail.
HG uses knowledge of constraints
to partition the
input problem into smaller, relatively independent subproblems which
can be processed in isolation. Solution synthesis techniques are
then used to combine sets of optimized solutions to subproblems into
global solutions. HG is able to guarantee optimal solutions (optimal
as far as the internal scoring mechanisms are concerned) and can do so
in near linear time for many types of problems, including semantic
analysis and, we expect, text generation.
The following sums up the advantages HG provides for text generation: