This would be a pretty interesting exercise. I can imagine using path length/distance across the word2vec space or even just the inverse of an ngram's probability. E.g. "zombie" and "fungus" may be close together, so would "cicadas" and "horror" maybe, but "cicadas" and "sex" probably don't appear together very often.