We recently did some analysis on our bayesian spam filter corpus (spam assassin token database), and came up with a list of words with a high spam/ham ratio.

By using the spam/ham ratio, and not the spam count, we came up with a better list of words to avoid.