Salience: The Key to the Selection Problem in Natural Language Generation

We argue that in domains where a strong notion of salience can be defined, it can be used to provide: (1) an elegant solution to the selection problem, i.e. the problem of how to decide whether a given fact should or should not be mentioned in the text; and (2) a simple and direct control framework for the entire deep generation process, coordinating proposing, planning, and realization. (Deep generation involves reasoning about conceptual and rhetorical facts, as opposed to the narrowly linguistic reasoning that takes place during realization.) We report on an empirical study of salience in pictures of natural scenes, and its use in a computer program that generates descriptive paragraphs comparable to those produced by people.