Thorny Roses: Investigating the Dual Use Dilemma in Natural Language Processing