Monday 9:00-10:30 Back on Track
OK - that other place was a project startup meeting - looked like a load of cool guys, but felt like an interloper. Now found the probabilities semantics strand.
Had a talk on bayes networks OWL - they said the semantics of the two are not 1:1 - no shit sherlock. Now somebody is describing RDF/RDF-S in painfull detail. Extending RDF with a 'value' attached to each triple to annotate it e.g. with probabilities.
So far, not got wrireless working - I smell proxy hell.
This next talk started with a statement of what the problem is they are attacking - an improvement on those so far. Jives with what we've been doing. First big example from him is Star Trek, how to work out what type of starships are on the scanner, with or without cloaking devices. I think he's telling us why bayesian networks are inherently limited.
Semantic discovery of services is up next. Not quite sure where they are going with this. The lady acting as chair is fierce.
Wireless works now!
Apparently, keywords aren't enough. These guys are trying to extract ontologies from natural text. With a genomics focus, suprise, suprise. He's done a good explanation of how unlikely it is that ontologies reaped from plain text will be consistent. I like this one. He even has graphs about how as you vary stringency, you vary inconsistencies. I think the probabilities attached to axioms in his system are statements about the probability that the axiom should be present/absent, so it does make sence in this concept to prune low probability edges. It's not a statement that the axiom holds at a given rate. I guess this is different from the bayesian approach where low evidence axioms would be pruned and some low probability arks could be very high evidence. Perhaps part of the problem is this philosophical one about ontologies being 'right descriptions' of reality rather than 'good models'. Possibly, ascribing source to axioms would result in much better multi-world properties, but there's nothing so far to make me think that the people here conceive of the multi-world, multiple-viewpoint thing as being worthy of modelling.