In this talk, I will describe an approach to sentence generation based on
probabilistic tree adjoining grammar and AI planning. The approach treats
sentence generation as an integrated task comprising of sentence planning 
and surface realization using declarative descriptions of the syntactic and 
semantic contribution of each word and contextual constraints on its use. 
Our system uses AI planning technology to find a TAG derivation that fulfills 
a given communicative goal while satisfying these constraints.
I show that our probabilistic extension makes it possible to use such an
approach with large, treebank induced grammars for the first time.
Furthermore, while most previous statistical NLG systems use discriminative 
models to select an output sentence, our system employs a generative model 
for TAG derivations. I show how AI planning systems can exploit this model 
for improved search heuristics.