Scalable Training of Markov Logic Networks Using Approximate Counting