Sparse Non-negative Matrix Language Modeling: Maximum Entropy Flexibility on the Cheap

   Abstract