Surface generation for a variety of applications

If a knowledge-based computer program is to produce English texts that are fluent and appropriate to the context of their audience and situation, the program's model of the conceptual information to be communicated must be augmented by a model of the English language, including linguistic information of the kind one finds in grammars and books of rhetoric, such as its constraints, alternative formulations, styles, carrying capacity, etc. Since this information is for the most part uniform across subject matters and tasks, it is included as part of any task-specific generation system, thereby relieving other designers of the need to be concerned with its unavoidable complexities. This paper describes the design issues that arise in developing such a linguistic component, using illustrations from a system that has been successfully interfaced to eight text-planning programs, and a variety of knowledge representation formalisms.