Wednesday, November 09, 2005

Wednesday afternoon

Ontology alignment part I

Given two ontologies O1 and O2, align one to the othher means for each entityh in 01, find a concept in O2 with the same intended meaning - ignoring structure issues. End up with equivalence links between the two ontologies. Compare a concept by text-comparison of the label (wordnet, etc.) and by properties it has (including name, cardinalty, ...) and then the classes on the end of the properties. Calculate some weights. Take average (or otherwise agregate). Threshold. Call those similar. So - how to choose the weights? They end up with a decission tree that does the job better than people. It doesn't seem to use the taxonomy - I guess looking at the decission tree rules would tell you at what level people are capturing the 'core' relationships in their domain.

The next talk makes the good point that equivalence between terms in different ontologies is less likely than subsumption between them. Example of mapping eu agric food to us grocery food. So far, these approaches have only worked because of injecting extra info e.g. from wordnet - this kind of indicates to me that the way ontologies are being used is impoverished relative to dictionaries. Users or representation?

The next one is a string-similarity-metric based approach. His laptop isn't talking to the projector :-( Let's hope there's no associated powercut like yesterday. Humph - not sure what this guy was doing other than getting fancy string matching.

0 Comments:

Post a Comment

<< Home