Rethinking Uncertainty Estimation in Natural Language Generation