On Subjective Uncertainty Quantification and Calibration in Natural Language Generation