Re: KIF counterproposal

Ramesh S. Patil <ramesh@vaxa.isi.edu>
Message-id: <9009211719.AA12326@vaxa.isi.edu>
To: "Matthew L. Ginsberg" <ginsberg@sunburn.stanford.edu>,
        interlingua@vaxa.isi.edu
Reply-To: ramesh@venera.isi.edu
Subject: Re: KIF counterproposal 
In-reply-to: Your message of Wed, 19 Sep 90 16:46:09 -0700.
             <9009192346.AA07486@Sunburn.Stanford.EDU> 
Date: Fri, 21 Sep 90 10:19:31 PDT
From: Ramesh S. Patil <ramesh@vaxa.isi.edu>

Matt, At the global level I agree with and second your proposal.  It
clearly identifies a number of technical issues and takes a stand on
them in a progmatic fashion.  In particular, I like your position
regarding the capability of interlingua.  It shoud be clear that
different systems will be able to extract different amout of
information from the interlingua depending on the similarities between
their intenal representation stand.

On the other hand there are a number of issues at the detailed
proposal level that I have reservations about.  As you point out
however, much of the details are "trial baloons" so I will not get
into it.  I want to however comment on one of them which appears to be
fundamental.

Your Kif proposal assumes that almost all KR systems use FOL as their
basic representation.  I however think that many representation
systems use FOL as a Knowledge Level language in which to communicate
the meaning of their constructs to others (researchers).  Even so
these constructs have extra-logical meanings and implications that are
important.  One generally finds out about these when he/she goes to
use this language for serious KB building.  These are very important
to communicate if one hopes to have the knowledge (and its
organization) to be useful to  others.

In recent years it has become clear that even the same knowledge must
be organized in substantially different forms for differnt tasks.  For
example, a key finding of Hamscher's thesis was that the design
knowledge and debugging knowledge need to be organized differently, if
one wishes to make these tasks efficient (or even computationally
reasonable).  Thus one of my main concern with MLG's KIF proposal was
that it did not provide any means for describing the organization of
the knowledge.  I feel that your proposal also suffers from similar
problem. 

To make this point clear, allow me to overstate it:  It appears that
going to interlingua is more like taking your KR and putting through a
meat grinder.  The nutritional content is still there, but the gross
structure is lost.  Yes I can translate  my KB to its meaning
postulates, but given the meaning postulates it is not possible to
re-assemble the original organization back.
 
One reason I have been pushing for some form of annotations is to take
account of this problem.  I would like "nonsemantic" information
regarding my knowledge organization to be available to others who can
then use it to re-constitute the meaning sentences to their favorate
language.  For example, if I could describe my frame as a set of WFF's
(a la Hayes) with annotation stating that they are chunked into a
frame, which of the relations are slots which are constraints etc.  It
will simplify other frame oriented systems to extract the information
and re-use it.  If they have some differences then they can choose to
re-organize this information in a manner that best maintains the
spirit of the KB.  No guarentee that the effect will be sound but will
be the best knowledge transfer we can achieve between agents with
different architectures.

In general, I would like a set of annotations developed to address the
issues of chunking, indexing and accessing of knowledge.  Furthermore,
if we expect that the two KR sysetms will be used for similar tasks
(e.g. diagnosis) then it is also reasonable to specify what kind of
inferences and queries are more likely than others.  In general some
specification of average case distribution of the use of the knowledge
should also be communicated.  This afterall is the rationale behind
the organization of knowledge.  Notice that both these concerns are
essentially "nonlogical" in your terminology.

I believe that such an extension is possible to either of the two
interlingua proposals and I would like to see both of these respond to
my concerns.

I appreciate your efforts towards developing an alternative proposal.
I believe it will lead to substantial positive contribution and will,
by virtue of providing an alternative, help sharpen the focus of
debate.

Cheers, Ramesh