Characterizing a subset of natural language

Methods are discussed for choosing a dictionary and grammar for the subset of natural language used in referring to data base specific information. Two possibilities are explored for bettering the usual ad hoc generation of these definitions. First, methods for extracting the grammar from instances of the language are reviewed and determined to be attractive but impractical. Second, given a starter system, methods for monitoring its usage and performance are discussed. Techniques are being investigated using a developing natural language system which is a front end to medical information stored in an INGRES data base--a relational data base running under UNIX on a PDP II.