[topicmaps-comment] OpenCyc and Topic Maps: Progress Report

["W. Eliot Kimber" <eliot@isogen.com>]

I spent the weekend coming up to speed--I think I now have a basic idea
of how Cyc represents knowledge--I can at least create simple constants,
associations, and rules and then ask simple questions about them and get
the expected answer. 

Unless I've missed something, there is clearly a direct mapping between
Cyc's knowledge representation and topics, in particular:

- Cyc "constants" correspond to topics, that is, reified things about
which assertions can be made. Topic characteristics become other
constants to which the base constant is associated. 

- Cyc "assertions" correspond to topic associations, to the degree that
they relate topics together.  Any topic map association can be
completely modeled as one or more Cyc assertions, including roles and
constraints on members.

- Cyc's "isa" and "genls" (generalizes) assertions directly correspond
to the subclass and instance-of associations in topic maps (no
surprise).

- Cyc's concept of "microtheory" (bounded context) seems to correspond
to scope (although it's not yet 100% clear to me that scope exactly
matches this concept or is the best way to represent it). In particular,
microtheories scope assertions, which can only be represented indirectly
in topic maps by setting the scope on association role members. It
probably comes to the same thing but the mapping might not always be
obvious to the casual observer.

From this it seems clear that there should be a direct mapping from
topics and associations to constants and assertions within Cyc, although
I'm sure there will be some subtle problems of exactly how to map
certain topic map constructs. This would then enable Cyc to reason about
the topic map and, possibly, further analyse occurrences of topics in
the input topic map in order to learn more (but I have no idea how much
of the Cyc natural language features are exposed in the OpenCyc
project). 

In addition, Cyc constants would make very good public subjects. In any
case, fully useful import of topics into Cyc would require mapping the
subjects in the topic map to the corresponding subjects in Cyc. Also,
Cyc appears to have a pretty robust mechanism for talking about the
temporal nature of things, so it might offer a ready solution to the
problem of temporal scoping.

It's also clear that even the subset of the Cyc KB released in OpenCyc
reflects a huge body of very deep thought about how to classify and
relate things--the range of subtle distinctions of association and
typing within Cyc is quite astounding. This thinking could certainly be
of great value to anyone developing a complex topic map that models more
than some very narrow subject domain.

Likewise, it should be possible to export or expose some or all of a Cyc
database as a topic map, although I'm not entirely sure what value that
would have in practice (it might have value for interchange of knowledge
bases between disparate expert systems, but I can't really see using a
generic topic map engine in any useful way on such a topic map when the
Cyc engine is already optimized for that sort of processing). As the Cyc
ontology defines everything within the KB all the way up to "thing", it
must be possible to generate a topic map that captures all the
information in the KB without loss. But I find that prospect much less
interesting than the prospect of importing topic maps into Cyc in order
to then use Cyc to reason about the topic map.

My next task is to experiment with importing topic maps into OpenCyc.
Given that OpenCyc provides a Java API it should be relatively easy to
integrate any Java-based topic map engine with it (e.g., TM4J). The real
challenge will be mapping the input topic map's public subjects to
existing Cyc constants and figuring out what the mapping conventions
should be.

Cheers,

E.
-- 
W. Eliot Kimber, eliot@isogen.com
Consultant, ISOGEN International

1016 La Posada Dr., Suite 240
Austin, TX  78752 Phone: 512.656.4139