Machine Learning and statistical corpus techniques are often very useful in Natural Language Generation, but they are not the best way to solve all NLG problems.