On Uncertainty And Robustness In Deep Learning For Natural Language Processing