SUO: Re: CG: Architectures for Intelligent Systems
Philippe,
Sorry for this late response, but I have been quite busy.
Quoting you from: http://mars.virtual-earth.de/pipermail/cg/2002q2/004222.html
Philippe Martin wrote:
> Yes, I have read your original proposition http://suo.ieee.org/email/msg08132.html
> (where, by the way, the word "incrementality" does not appear) and some of the answers.
Yes, my mistake, I sould have told about "incrementality" explicitly
and even put a LOT of emphasis on it, because, obviously, some
misunderstandings appearing in further argumentations stem from
the fact that people did not care to get a true understanding of
my proposal and of what incrementality implies.
This is not so much incrementality per se which have the most import
but all what ensue when you have successfully implemented it.
NUMBER ONE: You have a *consistent* ontology with *fragments* of
any number of other ontologies seamlessly included.
You don't care anymore about "domains", all you need to know about
part of a "domain" from another party will be downloaded on demand.
NUMBER TWO: You don't have to crave after "updating" your ontology with
respect to the ontologies of other parties you are interacting with.
It is done "on the fly" as you keep interacting, and both ways, they
will be updated if you upgrade your concepts.
> On a more technical side, in your proposition, the "definitions" of the categories
> are simply attributes (and your classification technique is based on FCA). That is
> very restricting for a representation formalism (the fact that FCA can be used to
> represent CGs is unrelated). Your proposition could be extended to any formalism
> and any classification technique (e.g. subsumption between simple CGs) simply by
> replacing the words "attribute" by (formal) "definition".
Exactly. The word "attribute" is indeed misleading, it is taken from FCA but,
actually, the way I define it is more akin to a CG relation with the incoming
arrow (to the relation) coming from the concept, and the outgoing arrow (out
of the relation) going to the "part" or "quality" tied to the concept.
In fact, the logic is just the CG logic with the following differences:
- There is no emphasis on the visuo-spatial representation, the graph is
(seemingly) split into a bunch of nodes and arrows.
- There are no gimmicks or shorthands like referents which are collections
and cardinality constraints on those collections.
- Referents of the PROP type (embedded sub-graphs) are allowed to have,
in some cases, a "high order" meaning.
Any CG can easily be translated to this format. I do not quite catch the point
you make above: "the fact that FCA can be used to represent CGs is unrelated".
> One problem I have with your approach, or any other automatic ontology merging
> approach, is that if the ontologies are not already "tightly interconnected"
> (especially via subsumption or identity links between the categories) classification
> cannot be done because the categories (including the categories for attributes, in
> your case) are incomparable. Even if some lexical matching is done (which is of course
> very dangerous and render the ontology merging quite dubious), unless the ontologies
> are big and/or about the same domain, it is likely that many categories of one
> ontology are more or less unconnected to the categories of the other ontology,
> and hence graph matching or other classification techniques are not usable.
The solution that, to me, seems feasible is to look for a match *only* on
the "leaves" of the concepts/attributes tree, that is, upon concepts and
attributes that CANNOT be described further by references to other concepts.
Like 'red' has no "components" or "qualities" as such even if you can tie
a wavelength or a "warm feeling" to it.
While a 'car' is certainly NOT a "leaf" concept.
I do acknowledge that a 'car' can be a leaf concept for some inventory
purposes, all that implies is, that just as an human being will have to
convey the "meaning" of the words he use to communicate, one way or another
(and this is mostly done implicitly thru shared vocabulary), the owner of
such an ontology will have to care to "beef up" it's definition of a 'car'
if he wants to communicate with other ontologies.
I also acknowledge that the "most prominent" attributes or qualities of
a same and given concept appearing in ontologies with an interest in
totally different fields may have *nothing* in common.
This is just an extension of the case above, one or both of the ontology
owners will have to add some *commonly recognised* attribute to his
concept definition in order to communicate.
And *one* such attribute, properly chosen to be specific enough, will be
sufficient to ensure matching of the concepts.
This is the "pars pro toto" human capability that Hubert Dreyfus used
to deride AI, but in this case it can be used by computers too.
> Manual ontology merging is not far better: although humans have the advantage of
> having a huge background knowledge and the ability to understand informal
> descriptions, they too make dubious connections or are not able to make
> connections if they do not have enough information to exploit.
THANK YOU VERY MUCH for this remark!
I have been bitterly arguing with naysayers on the SUO list about feasibility
of my approach, but it is NO MORE INFEASIBLE than a hand-crafted approach.
It will just go at COMPUTER SPEED instead of COMMITTEE SPEED!
> Another problem in your case is that, if I remember well, you ask that each
> category has a definition (a set of attributes) that distinguish it from
> other categories. That is *very* unlikely to be the case in most ontologies
> because writing such definitions is difficult and most ontology authors don't
> bother or don't know how to do it. Try finding distinguishing attributes (or more
> complete definitions) to common categories such as those referred by the words
> justice, eating, animal, cat, vehicle, car, Toyota_car, Toyota_Corrola_car, etc.
As explained above, a single one, well choosen, attribute will do and there
is NO NEED to try to plan for which one in advance, trial and test by ontology
builders will quickly converge to an adequate consensus for every concept of
every field. Unsuccessfull attempts will just result in "redundant" or "useless"
attributes being attached to some concepts, given the current cost of RAM and
disk space, WHO CARES?
> A large set of distributed (loosely connected small) KBs is not exploitable because
> - accessing (a selected subset of) them takes time,
> - automatic ontology merging takes time and leads to dubious results.
I hope you will change your mind if you seriously investigate my proposal.
> A large set of distributed (loosely connected small) KBs is undesirable because
> - connecting the knowledge of one KB to all the other KB is impossible
No, it is not.
> - it is easier for a user to enter knowledge into a large KB
Don't see why.
> - an inference engine has much more information to exploit for checking newly
> entered knowledge, helping the user to correct and refine its categories
> and knowledge, and connect them to other users'categories and knowledge.
> Here, you can have the possibility and advanges of "incrementality": knowledge
> may only be compared and connected to other users' knowledge *when* it is added
> into a shared KB (and with the help of the user).
That sounds quite close to what I suggest. I have explained elsewhere how
consistency should be ensured from the start by "definitional extension"
and how the consistency can be maintained across interoperating ontologies
by carrying consistency proofs along with each transmitted assertion.
See: http://suo.ieee.org/email/msg08252.html
Also: http://suo.ieee.org/email/msg08307.html
> - indexing the content of the KBs is difficult (the perfect index for arbitrary
> queries would be the whole KB itself); this shortcoming does not exist with
> large KB servers that try to piggy back from each other whenever they can.
I am not too sure I understand which kind of "indexing" you are talking about
and which purposes this indexing is meant to serve.
Cheers.
-- Jean-Luc Delatre
--------------------------------------------------------------
"You can always count on Americans to do the right thing
- after they've tried everything else." - Winston Churchill.
--------------------------------------------------------------
http://perso.club-internet.fr/jld/ -- GSM: +33 6 11 24 06 29