Multinomial Loss on Held-out Data for the Sparse Non-negative Matrix Language Model

   Abstract