The control of inferencing in natural language understanding

The understanding of a natural language text requires that a reader (human or computer program) be able to resolve ambiguities at the syntactic and lexical levels; it also requires that a reader be able to recover that part of the meaning of a text which is over and above the collection of meanings of its individual sentences taken in isolation. The satisfaction of this requirement involves complex inferencing from a large database of world-knowledge. While human readers seem able to perform this task easily, the designer of computer programs for natural language understanding faces the serious difficulty of algorithmically defining precisely the items of world-knowledge required at any point in the processing, i.e. the problem of controlling inferencing. This paper discusses the problems involved in such control of inferencing; an approach to their solution is presented, based on the notion of determining where each successive sentence "fits" into the text as a whole.