This paper describes a method used for efficient computation of the entropy gradient in semi-supervised linear chain CRF training.
