Statistical Acquisition of Content Selection Rules for Natural Language Generation