Exploiting Causal Independence in Markov Logic Networks: Combining Undirected and Directed Models