Revisiting Sample Size Determination in Natural Language Understanding