Robust Sentence Analysis and Habitability

Abstract

Systems for using subsets of English with computers have progressed much in the area of linguistic coverage of well-formed sentences for a specific task. Some methods have also been devised for the treatment of input that is almost well-formed. Nevertheless, it is still quite easy to stray over the bounds imposed by current natural language systems. Without proper diagnosis, this leads to interactive systems that are not habitable, i.e. systems that are not able to perform up to the user's expectations. This thesis presents an overall system for the treatment of several areas normally outside the limits of natural language systems, and for the diagnosis of any input. The system, Robust Sentence Analysis, includes procedures for handling ambiguous input, resolving input with anaphors (e.g. pronouns), making several kinds of major and minor corrections to input, and the interactions of all these areas. The system does not treat every aspect of these methods of human interaction, but does provide for the more prevalent forms as found in simulations of user interaction in several modes: face-to-face, terminal-to-terminal, and human-to-computer (using a previously implemented natural language system). Thus the system incorporates the most likely forms found in human performance. Diagnostics are designed to lead the user back into the boundaries of the system. The Robust Sentence Analysis system is implemented as a part of the ASK system, A Simple Knowledgeable System.