It seems the way this works is that you get a higher score if the 2-gram frequencies for all pairs in your set are low. But this can easily be gamed just by choosing very specific words, since they will naturally occur rarely with other niche words, regardless of whether or not their meanings are "related" in any sense.