Enhancing Text Representation By Minimizing Bias For Downstream Natural Language Processing Tasks