Re: SUO: RE: About SUMO's Mereotopology
Jean-Luc,
Thank you for citing my message from August 31, 2001.
Since I wrote it, there have been new developments that
bring the technology even closer. I'm enclosing a copy
of that note below, but I'd like to add some further
information about new developments.
In that note, I mentioned "Ripple-Down Rules" (RDR), which are
very efficient rules that can be compiled into multibranching
if-then-else form. Paul Compton started a company that was based
on developing such tools. Their execution style is very simple,
but their power comes from the ability of any user to add a new
condition whenever an error occurs.
RDR rules are widely used in bioinformatics because pathologists
who have no knowledge of programming, KIF, or other formalisms
can use them to develop expert systems with a very low error rate.
That low error rate comes from the technology: whenever an error
occurs, the user simply states a new condition that corrects the
error, and the system automatically creates a new if-then-else
rule and inserts it into the appropriate place in the RDR.
To quote myself,
JS> I'm not saying that it would be done overnight, but I believe that
> some outstanding tools could be developed by one person in about
> a few month's time.
A few months have passed, and Gerard Ellis did that. He took the
original RDR approach, which was based on a simple vector of features,
and replaced the simply binary tests at each stage with an arbitary
conceptual graph. As a result, it is now possible for a user to
state any arbitrary condition in a stylized or controlled English
sentence, which can then be mapped to a CG that does the test.
The control structure is still RDR in style.
That is just one of the "outstanding tools" that have been developed,
and there are others that cover other important areas of knowledge
acquisition, reasoning, language processing, and machine learning.
As another "outstanding tool", I'd like to mention Bob Levinson's
learning algorithms, which are based on CGs. He developed a system
called Morph, which learned to play chess starting from no knowledge
other than how to make legal moves. It didn't even know the rules
for determining a win or a loss. Today, its rating is 1395, which
is almost a B level -- and it can reach that level with only a few
hundred games of learning.
To demonstrate that his algorithms are not specific to chess, he
also applied them to the stockmarket in his system called Stocko.
Last year, when everybody was losing money, Bob used Stocko to
manage his retirement plan -- and he achieved a 12% growth rate
during a recession.
There are other tools that I could mention, but I'll say more
about them when they are ready for prime time.
John Sowa
____________________________________________________________________
* From: "John F. Sowa" <sowa@bestweb.net>
* Date: Fri, 31 Aug 2001 22:52:23 -0400
Adam,
The theorem prover is the centerpiece. All the rest is ASMOP
(A Simple Matter of Programming). And it's not very difficult
programming either.
Following is Seth's wish list, broken down with a sketch of how
each item could be provided:
> I believe that the most
> useful SUMO would have a front end tool where the ~user~ answers a set of
> questions and clicks on a set of contexts .. the tool would then generate
> the needed ontology.
You could handle this with ripple-down programming tools, which are
essentially multi-branching if-then-else rules generated from
questionnaires. For more info, see Paul Compton's home page:
http://www.cse.unsw.edu.au/~compton/
Following is a brief excerpt that suggests how easy it is for
non-programmers to generate such things:
This approach is now used in commercial systems in medicine
for pathology reporting developed by Pacific Knowledge Systems.
Pathologists can add rules to these systems at the rate of about
one per minute. By correcting a few errors per day they end up
with very sophisticated systems of thousands of rules, with very
little effort. These systems are now in routine use in a number
of laboratories. Other tools based on RDR have been developed by
HNK and Sricom. but I have little information on these developments.
Basically, you could use the RDR tools to ask the questions and to
select the modules needed for any customized ontology. RDR rules
have been used for database customization, and they could be adapted
to selecting the necessary modules for ontologies as well. Note that
the RDR tools are easy enough to use that pathologists create their
own rules -- by a process a lot like answering questions and making
corrections when it gives the wrong answers.
I'm not saying that it would be done overnight, but I believe that
some outstanding tools could be developed by one person in about
a few month's time.
> Then where certain axioms were still
> unacceptable, one
> could replace them individually ... the tool would then thread
through the
> selected knowledge and delete any contradictions and suggest where new
> axioms were needed.
This is a task for which a theorem prover is needed. It could check
for compatibility of different modules and suggest where there might
be contradictions. It could be combined with RDR rules.
John Sowa