[Top] [All Lists]

Re: [ontolog-forum] ISO merged ontology effort "MCO"

To: "[ontolog-forum]" <ontolog-forum@xxxxxxxxxxxxxxxx>
From: "A. J. Vizedom" <ajvizedom@xxxxxxxxxxxxxx>
Date: Sat, 11 Apr 2009 17:34:32 -0400
Message-id: <73158bfd0904111434u5bc80749g86b3d040bf23752d@xxxxxxxxxxxxxx>
John, et al.,

I've largely stayed out of this last round of upper ontology comparison, partly due to lack of time and partly due to the deep sense that the framing assumptions of the discussion are fundamentally off-track,

Your remarks about the *methodology* at the heart of useful upper ontology get at one aspect of this problematic framing. Thanks for articulating it just so.  On the heals of some good conversations at last week's Ontology Summit, I'd like to add a few comments of my own.

On Methodology and Ontology -- Yes. Yes, indeed. This is a little bit different from, but related to, a concern I've raised previously.  It's a soap-box issue for me as an epistemologist, so let me minimize the rant by stating the critical observations as a list without full explanation. Further comment on any particular points is available on request, of course (see: soap-box). ;-)

Too many proponents of this or that Upper Ontology appear to be some combination of the following:
epistemologically naive; confusing intuitive clarity with truth; unconcerned with application (or at least, with the possibility of clearly determining how the stuff folks need to reason about relates to their theoretical model);
confusing elegance with truth; confusing simplicity with truth (that is, rejecting complex, context-incorporating models as "relativistic," without being willing or able to consider that possibility that reality requires complex, context-senstitive models; assuming that a single path into modeling decisions (be it logical, introspective, or a particular instance or type of application) will provide sufficient information on which to ground a broadly applicable model; or making some other, less uniquely ontological error conflating the outcome of personal debates with evidentially significant trials.

I would urge a different way of understanding upper ontology as practice, and upper ontologies as artifacts.  I thnk we've often got things the wrong way-round. And I don't mean a simplistic version of Bottom-up vs Top-down.  I am talking about emergence, but it seems to me essential that models be in development whereever people are driven to develop them.  The good decision points seem to be where models meet, top and bottom, bottom and top.  There is some very good reason not to let the top drive the bottom, in its initial development.* There is also reason to develop partial upper models independently. But the adequacy and correctness, not to mention usefulness (and its cousin, confirmability), of Upper Ontologies, and the formation of a whole, coherent one, if it will happen at all, must be a process of emergence from those meeting points.

I offer some points in support of this understanding:

The task or process of modeling all of the things, types, and relations in some chunk of world is undoable. Applied ontology projects that begin from this conception of their goal fail. They produce less structured, complete, and coherent models than those that begin in more focused way, if they produce anything at all.  In as many cases, they never get going or exhaust themselves in a futile quest for completion.  After all, however small a chunk of world you choose, you can always find more to model.  Ontology is fractal.

The ordinary process of modeling the things, types, and relations in some chunk of world is given focus by a need. It could be some reasoning to support, some analysis to provide, some information to translate, some service to expose semantically.  If done well, the ontology developers are able to get decent abstraction and reusability, not by going up to an upper ontology, but by bringing in multiple angles on the focus. For example, if the driving task deals with data, bringing in Subject Matter Experts (SMEs) in the business processes or other activities that use or produce the data, focusing on what they do with it (decisions, belief formation)  or to get it (factors, conditions, limitations), can provide the balancing factor that moves the ontology toward reusability, and brings more of the implicit assumptions to light. That's generally good enough locally, and where the local context does not require metaphysical specificity and future extension and interoperability needs needs are unknown, it is better to make as few metaphysical commitments as needed. Whatever are made beyond that likely will not be grounded in what is known about the world being modeled.

Within a small-scale, well-defined domain, it's often the case that little or no upper ontology is needed. The need comes when we want to cross contexts, however defined. There are many of us working projects with semantic interoperability at their core, now.  This may be one reason a broader range of people are caring about upper ontologies recently, and why some existing ones are being tried out in new contexts.

When a well-defined, context-specific ontology project occurs with cross-domain interoperability in mind, part of the task becomes the inclusion and/or connection with of defining concepts and relations that possible non-domain users would recognize.  Middle-to-upper ontology questions come in quickly here, and naturally.  SMEs understand and can supply clarifying relations to such cross-domain concepts as  Event, Process, Place, Time, Individual, Organization, Information, Physical thing.  Supporting ontologists can ask the questions to identify what well-defined concept a SME is invoking.  Ontologists are better able to ask the right questions and specifiy the right concepts because an experienced applied ontologist is a sme about these mid-to-upper concepts. The ontologist is therefore able to *ask the right questions*: the questions that distinguish the upper categories from each other, define the upper relations.

If a single Upper Ontology has been pre-ordained, there is likely to be difficulty here. In my experience, an upper ontology developed in one context (application type or lack thereof, domain, people, etc), never fits a new context nearly as well as one would have expected if it were truly the sort of grounding, universal model its authors claim. From an epistemological perspective, I find this immensely unsurprising.  As a dilletante fan of the cognitive sciences, I find it nearly impossible to imagine how it could be other wise.  Based on unscientific surveys of other experienced, working, applied ontologists, I suspect that I am in good company. From this perspective, it is a matter of head-shaking and amazement that certain schools of ontologists continue to think that they can generate something that fits naturally over ever good middle-to-lower ontology by attempting the best abstraction, etc., they can muster, with various rules and practices to guide them. Nevertheless, all the normal rules of modeling apply. Things may be missed or emphasized, overlooked or overfit, elegant/seductive but unused/extra baggage.  Ontological models are really not so different from scientific models, in this way. Yes, you are trying to model something real, and accurately. You never try to model all of it, nor could you. You model the parts that matter for your activities. You model the characteristics that might explain the observables of your field, and so on.

*This is not at all a disaster.* Nor does it spell the end of abstraction, proper generality, and conceptual analysis across domains. No more than location-based variations in perspective spell the end of physics. It does mean a certain tolerance of complexity, of pockets of unknownness. It does mean that we may ourselves not be able to see how all the different perspectives fit together; that's a kind of unknownness that some people find intolerable or equate with relativism (in the sense that is opposed to realism).  Nevertheless, acknowledging the reality and complexity of context, and the fact that we have little basis on which to declare its irrelevance or to predict the patterns that might characterize it -- well, these all seem both unavoidable and something we must grapple with. They are also a far, far cry from the "anything goes" characterization that some send up in alarm. 

The best way to develop an upper ontology may be to let it emerge.  Pursue applied ontology projects as described above, and bring in the upper (and upper-middle) as they are needed. Bring in higher level concepts as you need them to connect concepts accross domains, to model the logical behavior and other characteristics of the domain-specific concepts, to avoid redundant and oddly-placed assertions on domain concepts that can accurately be placed on shared concepts shared across the domains.  Now, every ontology project cannot and should not develop its own theory of whatever middle and upper levels it needs.  A better approach is to import only what is needed, and to select from multiple models based on evaluation against the specific need.

If existing models are inadequate, the provision of this information may inform efforts to extend, combine, or replace upper ontologies or portions thereof. Patterns in which projects tend to make use of which upper ontologies may also reveal errors (that matter only to certain types of project), assumptions, dependencies and/or missing compentents that can be modeled. Who knows, maybe such a process would emerge, in the end, in a clear winner: a single Upper Ontology that is really the best one. I'm happy to be neutral on that question.  What seems clear to me is that such clear winner -- a Best Upper Ontology -- sure as hell *isn't* going to come about from Theorize and Impose from Above approach, or be developed from a single domain or application context, short of somebody showing up with an Improbability Drive.

To enable this emergence, to let the meeting points be tested, it's essential to keep the channells of visibility, criticism, shared use, and implementable importation open. That way folks working as described in #1 can find and select candidates to fill the higher-needs. Supporting these channels is also essential to enabling semantic technologies without forcing a single, inadequate model on everyone.  What most kills us is when the multiple models aren't visible, to potential users or to each other, and time and energy are wasted building things we don't need. It also kills us when people spend hundreds of person-years trying to argue idealogically for one upper ontology over another, instead of putting energy into supporting attempts to hook real world middle and lower ontologies up to it, or (Imagine!) listening, watching, and gaining understanding of what makes and upper ontology useable or not usable, trying to improve it.  Both of these -- emergence of better middle-upper and potentially upper ontology (via testing, improving, and so on across many use contexts) and avoidance of wasted time and energy -- are among the reasons that Ontolog and more domain-specific but still vital knowledge and ontology sharing organizations, and their ontology repositories, are so vital.

If we are keeping those channels open, the cost of allowing multiple representations (and letting better and/or more comprehensive ontologies emerge) is not so high. It's certainly lower than the cost of attempting to enforce any existing Upper Ontology on everyone.  The modularity of ontology, and the support for modularity in the existing representations and technologies, already go a great distance toward enabling successful implentations and interoperability without a single, unifying theory at the top. There is certainly more to do, more to solve.  I'd love to see us, the broader ontology and semantic technologies communities, putting more energy into such areas as comparison, testing, evaluation, and inter-ontology connections beyond full mapping and translation.  Maybe some of that energy we could get by putting less into attempting to agree on, or get people to adopt, any particular Upper Ontology. 


* I mentioned above that there is good reason not to let the top drive the bottom. I don't want to leave that totally unexplained, so let me just throw in one more point that I see as badly under-appreciated:  An preponderance of evidence - research and experience - suggest that middle-to-lower ontologies must be developed with as few mediators between the SMEs and the semantically-explicit model as possible. In practice, every additional layer of assistance/intervention adds a heap of lost context (that is, meaning and dependence that is then never captured -- a loss that reduces the reusability of the model), not to mention misunderstandings, oversights, and the familiar bottleneck factor. It's also multiply evident that SME's knowledge is often only elicted accurately when they are able to exercise the kind of thinking (pattern recognition over scenarios, for example) that they use when applying the knowledge normally. A consequence of these facts (beyond the prioritization of more and better tools to support SME knowledge entry in such contextual, domain-appropriate, ontology-under-the-hood, modes) is that when accuracy and efficiency really matter, as much as possible, the development of mid-to-lower ontologies should be by SMEs themselves, or by SMEs with ontologist assistance that does not derail their expert thinking.

On Sat, Apr 11, 2009 at 11:11 AM, John F. Sowa <sowa@xxxxxxxxxxx> wrote:
Pat, Ed, Azamat, Dick, and Mike,

PC> After all these years of talking about standard ontologies,
 > I have arrived at the feeling that an ontology is more complex
 > and of a different character than most standards that have been
 > formalized. The detailed means of applying ontologies to practical
 > applications appear to me to be a lot less easy to envision than
 > for most standards.  For that reason I think that no ontology
 > should be a formal 'standard' until it has had a lot of public
 > vetting in more than a few applications of the kind it is intended
 > to support.

I strongly agree.

EB> ... the purpose of an ISO "study period" is to determine whether
 > there is a specification or parts of a specification that has
 > sufficient consensus for standardization and meets some perceived
 > communal need (either in industry, or in the making of other
 > standards).

That is a laudable aim.  But those terms 'specification' and 'parts
of a specification' are very unclear by themselves.  When it comes
to ontologies, people tend to focus on the names of the categories
and their placement in some hierarchy.  But pioneers in the field,
such as Aristotle, Leibniz, and Kant, emphasized the *methodology*
for deriving the categories.

Aristotle's methodology was to associate a question with each of
his ten categories.  Each subtype of substance is the answer to
"What is it?"  For Quality, "What kind?"  For Quantity, "How much?"

Leibniz's methodology was to define his categories as conjunctions
of primitive features.  Then he assigned prime numbers to each of
his features.  The universal category, Entity, had no distinguishing
features, and its number was 1.  All other categories were assigned
the product of the primes for each of their features.

Leibniz's goal was to replace philosophical disputes with calculation.
If the number for category A divides the number for B, then A is
a supertype of B.  The result of Leibniz's methodology is a lattice
with all possible combinations of the features.

Kant's methodology borrowed some aspects from both Aristotle and
Leibniz.  I won't go into the details, but he formed a table of
twelve types of judgments, each of which corresponded to one of
his twelve top-level categories.

EB> The idea that some amalgam of BFO, DOLCE and SUO would have
 > such consensus would be supported by findings from the study
 > period that some useful common set of concepts is essentially
 > identical across these ontologies, differing only in terminology.

Two points in that comment seem arbitrary:  the phrase 'some amalgam'
and the choice of BFO, DOLCE, and SUO.  As I said before, Cyc is the
largest formal ontology on planet earth, and it has undergone almost
a quarter century of continuous development with contributions from
many highly respected logicians, linguists, philosophers, and experts
in artificial intelligence.

Almost everybody (including me) has found many aspects of Cyc that
can be criticized.  BFO, DOLCE, and SUO can also be criticized for
many reasons, and if there are fewer criticisms, part of the reason
may be that they aren't as well known and haven't been as widely
and thoroughly tested.  I suspect that excluding Cyc from the
process would be a recipe for producing another half vast product
that would create more problems than solutions.

The word 'amalgam' makes me think of a dentist mixing mercury with
powdered silver.  Far more critical to the development of an ontology
is a well thought-out methodology that is able to accommodate an
open-ended range of categories.  Ideally, it should be able, in
principle, to classify *every* category that anyone has ever
conceived -- including everything in BFO, DOLCE, SUO, *and* Cyc
*and* anything that anyone might ever come up with in the future.

If you have such a methodology and suitable software to support it,
you can take all the categories of all the ontologies, dump them
in the hopper, turn the crank, and see how they are related to
one another in the universal scheme.

 From that approach, you can pick a useful subset with confidence
that more categories can be added at any time without disrupting
applications that use the subset.

AA> The ontologies proposed are of restricted importance, and
 > hardly will make any useful standards, see the synopsis on
 > STANDARD ONTOLOGY: the Standard Model of Reality, Representation
 > and Reasoning....

Since I believe that all serious contenders should be considered,
I would recommend that your categories be added to the mix as
well as Cyc's.

RHM> I offer as a reasonable standard the tabula rasa ontology

Starting with a blank slate is useful, but any concept that
anyone has ever found useful must be accommodated.  Leibniz
had an even more fundamental starting point:  the number 1,
which represented the universal Entity.

RHM> It has a small number of fundamental concepts which are
 > likely to be used in all applications.

Those are included in the starting ontologies recommended above.

RHM> By contrast, most other ontologies are similar to Sowa's
 > ontology... which "mixes" three different views of existents,
 > thereby "cluttering" the ontology with concepts which may not
 > be used in a particular application.

When people buy a dictionary, they get thousands of words they
don't know -- that's the point.  For children and students who
are learning a language, there are smaller learner's dictionaries.
There are also specialized dictionaries for special applications,
but it's important to ensure that all the special cases are
compatible with the general framework.

For beginners who don't want to see the clutter, I suggest that
somebody should write an uncluttered primer with pointers to
the "Big Book" for anybody who wants more.

MB> The three sets of top level classes in the KR ontology are
 > fundamental, technology neutral and easy to explain to non
 > technical business subject matter experts. Any class of "Thing"
 > in a mid-level ontology of common things (for example contracts,
 > money, processes, events, parameters) can be identified as being
 > a child or descendant of one thing in each of those layers (e.g.
 > an independent, concrete, continuant thing). To this I would add
 > whole/part, set, time and so on.
 > I have done this, and it works, and people like it, understand it
 > and use it to provide practical business review and validation
 > of semantics models.

Thanks for the note of support.  But I'd like to emphasize that
I proposed the KR ontology as an *example* of how an ontology can
be derived by a systematic methodology:

 1. Start with fundamental distinctions, not categories.

 2. Each distinction generates a pair (for a binary distinction)
    or a triplet (for a ternary distinction) of basic features.

 3. Leibniz's methodology derives all possible combinations of
    those features, which include many unnamed categories. Other
    methodologies (such as FCA) create lattices that minimize
    the number of unnamed categories.

 4. If more categories and distinctions are added to the list,
    a new lattice can be derived that subsumes the earlier
    lattice as a proper subset.

As an example of such a methodology with open-source software,
I suggest Formal Concept Analysis (FCA).  See the FCA home page:



Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/  
Config Subscr: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/  
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/ 
To join: http://ontolog.cim3.net/cgi-bin/wiki.pl?WikiHomePage#nid1J
To Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx    (01)

<Prev in Thread] Current Thread [Next in Thread>