How Far Can It Go?: On Intrinsic Gender Bias Mitigation for Text Classification