Implicit Representations of Meaning in Neural Language Models