Interest in statistical natural language generation is rapidly increasing. This work sheds important light from theoretical linguistics on the type of information crucial to statistical NLG algorithms.