Re: the mythical RDF inference engine was: Re: What is truth anyways?

patrick hayes wrote:

> ... If there's a CWM agent out there  disguised as an RDF
> agent, and my RDF agent reads its output, then all of *its*
> inferences are likely to be corrupted by invalid RDF conclusions that
> the CWM agent has drawn.  I just want each spec to wear its meaning
> on its sleeve, as it were, *as part of that spec*. And CWM and N3 are
> NOT parts of the RDF spec; and in fact, it is easy to prove that the
> claimed meanings for much of N3 are not writable in anything that
> could possibly conform to the RDF spec.

Right. The problem is that RDF is not (currently) making itself a good
platform for such _other agents_ particularly WebOnt/OWL to as you say "wear
its meaning on its sleeve"

Since we agree that useful inferencing engines are going to extend RDF in
some way, rather than direct the work of RDF _inward_ toward describing how
RDF inferencing ought work (do I care ? :-) perhaps a bit more effort ought
be directed at figuring out how other systems might use RDF triples _as
data_.
>
> >Indeed since base RDF itself is just a bunch of assertions, I'm not sure
> >what inferencing you can do _with RDF alone_ rather it seems just the
> >mechanism for carrying 'facts'. I worry that all this emphasis on
> >inferencine _within_ RDF is over constraining the ability to do
inferencing
> >_on_ RDF.
>
> The problem is that last phrase is meaningless. Writing all the
> triples backwards is inferencing on RDF. Erasing the RDF and
> replacing it with a quote from the Koran is inferencing on RDF. If
> you disagree, tell us what YOU mean by inferencing on RDF.

Shrug... this last statement is no more or less meaningless than anything
else I choose to say in English.

But OK, I'll bite. OWL might choose to represent individuals and facts about
individuals _in RDF_. OWL might have a model theory that provides for
certain inferences to be drawn given these individuals and a bunch of
classes. That might be useful. I might describe these inferences as being
drawn _on_ the particular facts. Nonetheless we consider such facts as data
from which we draw conclusions. What is the big deal?

>
> >So why not just say that RDF is a bunch of assertions
>
> Becasue that is an extremely dangerous thing to say: it sounds
> meaningful but in fact is meaningless. A bunch of assertions *in what
> language*?? In RDF? If so, then it is INCORRECT to use non-rdf-valid
> processes on them.

Huh? INCORRECT according to whom? Certainly this is (actually) begging the
question if you define CORRECT as according to your model theory, eh :-)

I am just trying to be practical. A bunch of RDF triples is like a
database -- well really a single table with three columns. Why not define
"rdf-valid" loosely -- let whomever is using the triples decide what is
valid. Isn't that how most all software works: there are data structures and
algorithms which operate on the data. A triple is a data structure, right?

>
> >and forget about this
> >pure RDF inferencing engine that doesn't seem to exist and let the folks
who
> >care about inferencing decide what inferences might be drawn from a
certain
> >bunch of facts and be done with it.
>
> The whole point of having a precise account of meaning in the
> language spec is to sanction some inferences and not others. If we
> can allow anyone to draw any conclusions from an RDF graph, then RDF
> doesn't differ from XML.

Well N-triples looks alot different to me than XML. And a triple store fits
into a single relational table with three columns, so that is the
different -- perhaps not earth shattering, but what magic are you looking
for?

> In fact, it doesn't differ from ASCII or
> graffiti: its just character strings that you can do whatever the
> hell you like with and call it inference. So why are we even
> bothering with it?
>

ASCII, or to be modern, Unicode, will get you pretty far. But yeah, that's
about right, we have Unicode, XML and RDF. And on top of that we develop
OWL, DAML-S etc. Why not just consider RDF as a standard way to transmit
graphs, or collections of triples. Isn't that useful enough?

Jonathan

Received on Wednesday, 12 June 2002 16:55:39 UTC