Probabilistic ontology inference

1
Probabilistic ontology inference

• Octavian Udrea
• Yu Deng
• Cihan Tas
• V.S. Subrahmanian

2
Constrained probabilistic ontologies

• Constrained probabilistic ontology IS-A graph
  with
• Multiple disjoint decompositions
• Conditional probability labels on edges
• (A ?p B) Given an individual in class B, the
  probability of it belonging to class A is p.
• Described in detail in a previous JIKD talk

3
Goals and problems

• Goal improve the quality of answers to queries
  by associating CPOs with entities in the system.
• The problem addressed here
• There are numerous ontologies publicly available
• But none with probabilities
• How can we infer a good probability
  distribution for the classes in a CPO?

4
The solution in a nutshell

• Assume the existence of a dataset D where each
  item has already been classified the ground
  truth set.
• By annotations, NLP methods, entity resolution,
  etc.
• Let Q be a set of random queries.
• Let C be the CPO with random initial
  probabilities.

5
The solution (continued)

• For item j and query q
• With the CPO, item j appears in the answer to q
  with probability p
• The ground truth dataset D either contains or
  does not contain j in the answer to q
• This leads to a set of constraints for the
  probabilities gt an LP problem