Contextual Representation Learning beyond Masked Language Modeling