Quantifying Uncertainties in Natural Language Processing Tasks