CHINESE NATURAL LANGUAGE INTERFACE FOR EXPERT SYSTEMS

Abstract

There have been existing several types of man-machine interfaces through which people interact with expert systems. Among them, natural language interface is no doubt the most natural and easiest interacting medium. In this thesis, we will discuss Expert System shell of Chinese natural language User Interface (ESCUI). The main objective is to allow a non-professional user to access an expert system by using his most favorite language, natural language. Firstly, we combine advantages of computational grammar formalism and linguistic formalism together and propose a Chinese syntactic parser which underlies in the formalism of Unification Phrase Structure Grammar (UPSG). Meanwhile, the grammar is constrained by Government Binding (GB) Theory. So far, our parser has around 40 grammar rules, and can process simple sentences of Chinese with associative phrases, relative phrases, passive voice, ?-transformation, etc. Secondly, the semantic interpreter is designed to transfer the original sentence into its corresponding logical form. We chose the extended first-order logic formula as our representation, which provides a good, generalized formalism of representing the meaning of natural language sentences, including an adequate treatment of quantification and related constructions involving scoping. Our semantic interpreter call deal with non-standard quantifiers like ??? (a large portion), adverbs like ?? (always), and question proverbs like ?(who). The result of the semantic interpretation can serve as the input of expert systems, and the inference engine of expert systems can make reasoning based on the logical formula. The information used in reasoning processes of expert systems may be uncertain, imprecise, or even vague and incomplete. This situation seems to be more important when natural language expressions are introduced. In ESCUI, we incorporated fuzzy set theory and fuzzy logic into knowledge representation scheme and reasoning process. As a result, it allows for precise modeling of imprecise statements in natural languages, and for easy and natural specification of values for imprecise concepts. We also employ linguistic approximation to avoid the problem of unrealistically precise and consistent in the assignment of numerical rules and facts in fuzzy systems. To illustrate our work, we have built up a fairly general purposed expert system shell, and established a small financial investment advising expert system. The shell uses both forward- and backward-chaining in its fuzzy reasoning process. It also equips with the utilities such as interpretation of input rules, verification of knowledge base, and self-explanation of results. ESCUI is implemented in Sicstus Prolog under Sun Sparc workstations.