Exploiting syntactic, semantic and lexical regularities in language modeling via directed Markov random fields