Towards More Accurate Uncertainty Estimation In Text Classification