Agree completely -- which is why I sent a base file which had the
conditional probabilities, the mapping, and the values to be able to
compute marginals.
About the URIs, I should have added in my email that because freebase
types are not URIs, and have types such as /people/person, we added a
base URI: http://freebase.com to the types. Sorry I missed
mentioning that...
Kavitha
On Aug 10, 2009, at 4:42 PM, Tim Finin wrote:
Kavitha Srinivas wrote:
I understand what you are saying -- but some of this reflects the
way types are associated with freebase instances. The types are
more like 'tags' in the sense that there is no hierarchy, but each
instance is annotated with multiple types. So an artist would in
fact be annotated with person reliably (and probably less
consistently with /music/artist). Similar issues with Uyhurs,
murdered children etc. The issue is differences in modeling
granularity as well. Perhaps a better thing to look at are types
where the YAGO types map to Wordnet (this is usually at a coarser
level of granularity).
One way to approach this problem is to use a framework to mix logical
constraints with probabilistic ones. My colleague Yun Peng has been
exploring integrating data backed by OWL ontologies with Bayesian
information,
with applications for ontology mapping. See [1] for recent papers
on this
as well as a recent PhD thesis [2] that I think also may be relevant.
[1] http://ebiquity.umbc.edu/papers/select/search/html/
613a353a7b693a303b643a37383b693a313b643a303b693a323b733a303a22223b693a
333b733a303a22223b693a343b643a303b7d/
[2] http://ebiquity.umbc.edu/paper/html/id/427/Constraint-
Generation-and-Reasoning-in-OWL
