Intentional Control of Type I Error over Unconscious Data Distortion: A Neyman-Pearson Approach to Text Classification