Generalized Consultation Systems and Knowledge Acquisition 
Robert Wilensky 
Department of EECS 
University of California, Berkeley 
Berkeley, CA 94720 
Objectives 
We are developing the technology to provide helpful, natural language-capable consultation 
systems for arbitrary domains. Our approach is to develop a "'Domain Independent Retar- 
getable Consultant" (DIRC). DIRC is essentially a kit that one will be able to convert into 
an intelligent, NL-capable consultant for any domain by supplying the appropriate 
knowledge. We are also developing the knowledge acquisition technology to support DIRC. 
Previous Accomplishments 
We have previously constructed a UNIX Consultant (UC), an intelligent NL-capable "help" 
facility that allows naive users to learn about the UNIX operating system. We have also 
developed some techniques for extending the knowledge base and vocabulary of this system, 
namely, a system that allows an expert to add information in natural language, and one 
which hypothesizes new word senses by making metaphorical extrapolations. 
Plans 
We intend for every DIRC kit to come with a core vocabulary and extensive grammar, and 
build in most of the relevant pragmatics. However, supplying the vocabulary, constructions 
and world knowledge for each domain is approached as a knowledge acquisition problem. 
In particular, we plan to develop the technology for the automatic acquisition of domain 
knowledge by reading. A prototype version of such a system, called MANDI. is currently 
under construction, and is targeted to acquire knowledge about UNIX by reading the on- 
line UNIX manual. 
We are also developing techniques for the automatic acquisition of the lexicon. Our 
approach involves a theory of word sense relations that can help a system to acquire new 
word senses given old ones. The essential idea is to exploit subregularities that exist 
among, but do not successful predict, word senses. We have previously exploited one kind 
of subregularity, namely, metaphorical word sense relations. However, there appear to be 
many other useful subregularities. 
We plan to use this theory in at least two lexical acquisition methods, one which 
hypothesizes new word meanings in context, and an intelligent dictionary reader. Utilizing 
dictionary entries requires at least the language analysis required to comprehend ordinary 
text. We plan supplement such understanding with knowledge of word sense relations to 
help correctly interpret dictionary entries. 
Since inference plays an important role in the sort of text processing we propose, we have 
also been developing a general abductive inference method, based on probability theory. 
An initial implementation is being constructed to deal with the interpretation of nominal 
compounds. 
444 
