So, there's a lot of literature on natural language... Assuming that you want a bit quicker ramp-up than a few months of reading, I'd do a search on citeseer for data on this, since someone has probably written a paper.

Failing that, you could scan corpora for sentences that contain the words on the magnets, or you could generate a model from corpora (such as a bayesian one), then constrain your model to only sentences containing those words, and choose sentences with the highest probability.