On the Accuracy of Self-Normalized Log-Linear Models