%A Ken Murray
%A Bruce Porter
%J Proc. 6th International Machine Learning Conference
%D 1989
%P 290-295
%I Kaufmann
%C CA
%X Adding new information to an existing knowledge base can have significant
consequences. For example, new information might contradict existing
knowledge or reveal a "gap" in the knowledge base. Most approaches
to knowledge-base refinement either ignore these consequences or compute
them exhaustively. Our approach, formalized in a task called knowledge
integration, is to partially elaborate the consequences of new 
information. A form of domain knowledge called `views' controls the
search to identify non-superficial consequences of new information.
A prototype knowledge integration program has been implemented and 
demonstrated with a complex extension to a large knowledge base.