Explain the intuition behind distributional methods - that is, why do we believe that solving the task of predicting a word given its context yields embeddings which capture lexical semantics?