On the accuracy of self-normalized log-linear models