On May 18, 2007, at 12:55 PM, Dan Brickley wrote:
Matt Williams wrote:
I've been lurking & reading the discussion with interest.
It might be worth pointing out that there is an ongoing attempt to
classify/ represent evidential links/ weight/ etc. started in the
legal domain by people such as Wigmore and continued by people
such as David Schum & William Twining. There's currently a
Leverhulme-sponsored research programme on "Evidence Science",
centered at UCL, London.
Such efforts don't seem to easily map to rdf (they're often based
on Bayesian models), but might provide some inspiration, although
some of the legal niceties may be unnecessary.
Interesting! There's also another related group here in W3Cland -
an incubator on Uncertainty Reasoning for the Web, see
http://www.w3.org/2005/Incubator/urw3/ ...though things are only
just starting up. The only mailing list traffic so far is about
scheduling the first telecon.
The charter has some notes on what they're doing:
http://www.w3.org/2005/Incubator/urw3/charter ...and Bayes gets a
few mentions there.
My inclination with RDF and evidence/probability, ... is that
without reinventing the RDF graph model, it is likely easier to
attach probability and other annotations to collections of
statements, rather than to individual triples. This can be done for
example by making assertions about an RDF/XML document, ... and is
somehow related to the ability in SPARQL to associate a graph with
a URI. For example see http://www.w3.org/TR/rdf-sparql-query/
#queryDataset
the idea of attaching probabilities to triples or sets of triples
bothers me. It seems like a problematic mix of monotonic logic and
probabilities.
BFO points to a cleaner solution through the use of bfo:Disposition.
We can make statements such as
HumanWithBRCA1-allele123 has_disposition some (bfo:Disposition that
towards DevelopingBreastCancer)
Which are to the best of our knowledge true in a boolean non-fuzzy-
probabilistic sense. At the same time the bayesian folks have a class
on which to hang their probabilities off of. Subclasses can be
introduced for specific environments and contributing factors.
Bayesian networks can be built on top of the (logically) reasoned
graphs. No need to worry about the probabilistic stuff feeding back
to the logical model in a non-monotonic way.
The evidence/provenance would be attached to the whole (class-level)
statement above
Of course Matt's NM approach may be useful too.. interested in
hearing more.
cheers,
Dan