A KNOWLEDGE BASED APPROACH TO NATURAL LANGUAGE UNDERSTANDING (COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE)

An extremely significant feature of any Natural Language (NL) is that it is its own meta-language. One can use a NL to talk about the NL itself and to give instruction in the use and understanding of the same NL. In this thesis we present a language processing expert system that we have implemented in the role of an educable cognitive agent whose domain of expertise is language understanding and whose discourse domain includes its own language knowledge. We present a representation of language processing knowledge and a core of knowledge, including a Kernel Language, which forms the knowledge base for this AI System. Since linguistic knowledge is part of its domain of discourse, the System can be instructed in the processing and understanding of ever more sophisticated language, with instruction initially given in the predefined Kernel Language. As the System's language knowledge is expanded beyond the primitive Kernel Language, instruction of the System is expressed in an increasingly sophisticated subset of the language being taught. Thus the System's language is used as a meta-language for the self-same language.;In this thesis we discuss two experiments that we conducted. In the first experiment, our approach was to teach the System to treat linguistic knowledge in a manner that is commonly used for general knowledge (e.g. property-value pairs) and to use its acquired natural language subset as a meta-language for the same language. In the second experiment, we taught the System to process language according to a subset of a Lexical-Functional Grammar. One of our original objectives was to design a system that was as theory-independent as possible. The purpose of this second experiment was to test, at least to some extent, whether we had achieved this objective.;We also discuss the parsing and interpretation strategies of the System. Parsing is performed according to a combined bottom-up top-down strategy with a focusing context resulting from the bi-directional inference sub-system. Parsing and interpretation take place in an integrated manner in our System, governed by the language definition input to the System by a teacher-user.;Our NLU System is implemented in the form of a general purpose inference system which reasons according to the rules of its knowledge base. This knowledge base comprises the System's task domain knowledge and includes, but is not restricted to, its language processing knowledge.