[
Lists Home |
Date Index |
Thread Index
]
--- Jeff Rafter <jeffrafter@defined.net> wrote:
> For example, given a sample "system" (working with
> cameras is fine) what are
> the areas of change that we are worried about-- some
> possible cases:
>
> 1) Someone removes an existing ontology (e.g. they
> decide SLRs never
> existed)
> 2) Someone adds a new ontology (e.g. a new camera
> type called an FBR is
> invented)
> 3) Someone changes an existing ontology (e.g. f-stop
> comes to mean
> Flash-Stop)
How about "digital camera technology becomes so cheap
that almost anything with a battery and a processor
can become a 'camera'." Now a "camera" ontology
would have to have links to/from cellphones, PDAs,
binoculars, email clients, etc. etc. etc.
That seems to be the essence of complex systems -- you
can't *analyze* their behavior except in a very
high-level sense, e.g. note the "attractors" around
which behavior circles. In this case, I guess that
"picture" might be the attractor around which the term
"camera" circles, but trying to do more than that
would require massive, full-time ontology maintenance
and reasoning systems that would have to chase down
all sorts of logically meaningful but empirically
useless dead ends. ("How many megapixels can that SLR
store?" "What F-stop did you use to take that picture
on your cellphone?"
I can see the utility of ontology building in domains
where things more or less sit still while we examine
them, e.g. the assumptions about human anatomy and
physiology built into SNOMED (although I suppose that
it evolves fairly quickly as disease organisms evolve
and as the etiology of known diseases is better
understood. It's just not clear to me how that is
going to help us find stuff on the Web better than we
can with heuristic / statistical approaches. For
example, Google doesn't know a stinkin' thing about
"cameras" except that the word appears on a lot of
pages with words such as "picture" in it (and its
synonyms, equivalents in other languages, etc.), so it
has no trouble with the idea that a cellphone can also
be a camera. So, we can do useful things with these
statistically useful "attractors" of one term for
another in the space of actual documents that would
utterly defeat a reasoning agent with an out-of-date
ontology that is trying to figure out why anyone would
object to people bringing cellphones into a locker
room.
Sure, the approach Google uses is beginning to fall
apart under the various strains on it, and clearly the
world needs to keep working on this problem. There
may be some way to leverage relatively static
ontologies to steer one away from "false attractors",
but the only practical way I see to keep up with
evolving language is to continuously sample real
communications.
Here's an example: What/where is "Rummy World"? The
#2 hit on Google for that term already gets it right
(Iraq). I could probably construct a tortured line of
"reasoning" to explain how this would fit into an
ontology, but AFAIK Garry Trudeau
[http://www.ucomics.com/doonesbury/2003/09/23/ ]just
created it out of thin air because the idea that Sec.
Rumsfeld and friends have created a terrorist
playground in Iraq produced a mordantly funny image in
his mind. ["They squabbled in the car all the way
here"] If you somehow told the search engine that the
phrase 'Rummy World' has nothing to do with card
games, or that "Rummy" was probably a person, the
performance could be improved by the ontological
information. ["Gin rummy" is a false attractor if you
are interested in Rumsfeld, and the Secretary of
Defense is a false attractor if you are interested in
card games.] But how on earth could you expect
ontology builders to add in these kinds of
associations in anything like real time? And more
importantly, what's the cost/benefit ratio of even
TRYING to build an ontology containing terms whose
meaning changes as fast as words do in popular
culture?
|
