[
Lists Home |
Date Index |
Thread Index
]
Through every change of season, throughout its hundred year
history of ring by ring growth, even in its fiery death, a tree is still
a tree, is still a redwood, is still the species that it is. All
perception of stable objects requires some degree of abstraction from
the buzz of subatomic particles. An ontology that is constantly revised
is not sufficiently abstract (how many shades of green leaf do you need
to distinguish?). An ontology that never requires change is perhaps too
abstract, and makes too few distinctions to be of practical use (is it
animal, vegetable, or mineral?).
Implementing a change to an ontology, provoked by change in its
domain, has a cost that ought to be exceeded by the cost of not making
the change. When the number of patents in one of the ~140,000 subclass
of the US Patent Classification becomes too large (that is, there are so
many patents in it that it takes too much time for an examiner to find
the ones relevant to his search), he asks a classifier to reclassify
that subclass. That means analyzing the documents in the subclass and
any related subclasses, creating a new subclass structure (more
subclasses, fewer documents per subclass), and then resorting all the
affected documents (from the 6.5 million) into the new subclasses. The
result is that examiners can process more applications in less time, and
they are more likely to find the documents they need. As you can
imagine, this process requires a significant institutional commitment.
The consequences of not making that commitment are significant as well,
as you all know from recent stories about patents for what many agree to
have been technologies already well known. (Don't read too much into
that comment about why those patents were granted -- software is a
particularly difficult field of search, no matter how good the
classification.)
I suspect that the more serious problems for a web ontology than
modeling change are finding the appropriate level of abstraction such
that its utility is maximized, and then attracting sufficient
institutional commitment to maintain it. But those aren't the only
problems. In the PTO, only those who have been trained as classifiers
are permitted to apply classifications to patent documents. In the
wilds of the web, no such control is likely to be possible. Achieving
consistency in the use of an ontology will be a challenge that will have
a direct affect on quality and utility, and that will directly affect
the likelihood of gaining the support required for maintenance.
Of course, the Library of Congress classification is widely
used, not just because it is a great classification, but largely because
it has enormous institutional support. Libraries that use it do not
necessarily implement every change from LC, nor do they do it in
synchrony, but that's OK because most library collections (until
recently) were isolated groves visited by local users, with little or no
conflict resulting from the inconsistency. I think this is not true for
objects living on the Web.
I don't mean to distract from the discussion of maintaining
ontologies. Just wanted to make some points about what provokes the
desire for change in a real system.
Bruce B. Cox
SA4XMLT
USPTO/OCIO/AETS
703-306-2606
-----Original Message-----
From: Simon St.Laurent [mailto:simonstl@simonstl.com]
Sent: Wednesday, October 01, 2003 10:00 AM
To: xml-dev@lists.xml.org
Subject: Re: [xml-dev] Beyond Ontologies
jonathan@openhealth.org (Jonathan Borden) writes:
>1) There is nothing that prevents ontologies from changing fairly
>regularly. Ontologies themselves are just :-) documents,
>representations of resources :-) that can change as often as the author
>desires. Ontologies can be generated from databases etc. The only
>tricky point of this is designing systems that can deal with changing
>ontologies, but that is a software issue, not necessarily an ontology
>issue. In any case the WebOnt WG, which has a good number of people who
>have lots of experience with ontologies has considered these issues,
>indeed from the "use cases and requirements" document:
>http://www.w3.org/TR/webont-req/#goal-evolution . Note *requirements*
>R3 "explicit ontology extension" and R6 "versioning information" which
>are relevent to this issue.
Technically, ontologies can certainly change. Culturally, however, the
paeans to ontologies that I see appear driven by a deep need for a sense
of stability at least, a sense of monolithic truth at worst.
Economically, the whole thing is sold as "once you do this, you'll be
able to make all your information meaningful at low low cost", and I'm
not sure that's compatible with regular change.
>2) although the world is constantly changing, relationships need not so
>constantly change. For example, *you* and *your father* -- I don't need
>to know any details about you, nor about your father, and indeed both
>you and your father are constantly changing -- even for deceased
>people, for example, the "time since birth" is a property whose value
>is in a constant state of change. "physical location" is another
>property which might be in a constant state of change etc. etc.
>Nonetheless, the *relationship* <#Simon> :sonOf _:1 between the two of
>you need not change (I haven't even assigned a URI to your father!)
If all relationships changed all the time, we'd all be pretty confused.
The questions seem to arise around cases where changes in classification
possible, necessary, and often unpredictable.
>The point about this is that the constraints imposed by any ontology do
>not (typically) result in any single state of affairs, rather a *range
>of states of affairs*. A good ontology might capture a wide range of
>states while at the same time imposing the proper constraints on these
>states.
>
>Admittedly time dependent changes in the state of the world remains an
>area of current research for ontologies, OWL/RDF in specific, but while
>OWL may not have detailed specific mechanisms for dealing with time
>dependent state changes, the fact that such state changes might be
>important has been factored into and considered in the design of OWL
>itself i.e. future extensions to OWL might indeed directly capture time
>dependencies.
Great to hear, but I await experience on the matter. Versioning seems
to be a consistently hard problem - not because new versions are hard to
create, even annotate that they're new - but because change in one part
of a model generally has cascading effects through other parts of the
model and implementations based on the model.
>There is no reason that one should conclude that one has to throw out
>OWL, or move beyond OWL in order to model a changing world. Language
>itself is in a constant state of flux, yet Wordnet.
There are plenty of other reasons for throwing out such projects. None
of them seem particularly likely to win those suffused with ontological
fervor, though.
|
