basically some mutual information would give a hint about how close are the words in context! So having some mutual info in the statistics helps to enrich the feature space.