Exploiting Syntactic Structure for Better Language Modeling: A Syntactic Distance Approach