From: Bhushan Mandhani (bhushan_at_cs.washington.edu)
Date: Wed Oct 22 2003 - 11:23:01 PDT
Paper Title: Two Theses of Knowledge Representation
Authors: Jon Doyle and Ramesh Patil
Brief Summary: Levesque and Brachman claim in an earlier paper that
general purpose knowledge representation systems should have restricted
languages for which we have polynomial runtime for classification. Also,
the knowledge base should have separate terminological and assertional
components, with the classifier making no use of the latter. This paper
strongly counters these ideas.
Main Ideas of this Paper:
1. The restricted language thesis is false. Restricting the representation
language in order to make classification occur in polynomial time destroys
the generality of the language, and renders it uncapable of expressing
numerous common and important concepts, which are easily expressed in
first order logic. For example, the language NIKL was seen to be incapable
of expressing ternary or higher predicates. It was seen to be too
restricted to create useful knowledge bases for real applications.
2. The restricted classification thesis is false. Classification should
not be done using the terminological knowledge base alone, since the
classification of many concepts depends on contingent information in the
assertional knowledge base. Further, due to the separation between these
two, many concepts which are definable in logic have to be stored in the
terminological knowledge base as "primitives" which can't be classified.
This further reduces the utility of classification.
Flaws:
I don't feel there are any major flaws in the paper. However, I feel at
times, the authors understate the importance of classification problem.
Open Questions:
Computational efficiency was shown to be a poor measure of evaluation for
a knowledge representation system. The performance should instead be
judged according to some decision-theoretic utility function. One question
would be the design of this utilty function.
The really important problem is the design of the classifier. Given that
we are going to use a language expressive enough to create powerful
knowledge bases, the classifier can't always be complete and efficient at
the same time. Thus, provisions for incomplete deductive classification as
well as nondeductive classification are going to be needed. How these can
be done best is an important research question.
This archive was generated by hypermail 2.1.6 : Wed Oct 22 2003 - 11:23:11 PDT