Combining Neural and Symbolic Learning to Revise Probabilistic Rule Bases