I don't know if there's ever going to be a foolproof way to do what you ask, since in order to match the level of discernment we humans are capable of you will have, in the end, ended up with a module capable of true language comprehension.

Probably a compromise between a full-blown human brain and some dirty matching would be some method (or several, cross-correlated) of fingerprinting the linguistic patterns. Lingua::EN::Fathom might be a good place to start, along with some Bayesian filtering scheme such as Mail::SpamTest::Bayesian. Toss a well-designed neural net in there and you might have something.

I've considered playing with this sort of thing myself -- let me know if you run with it or find something relevant that I have neglected to mention.

Matt