Learning Language Representations with Logical Inductive Bias