This sounds like a job for the CRM114 discriminator. I'm not aware of a general purpose perl interface for it, but the SpamAssassin CRM114 plugin might give you a good start on rolling your own.

CRM114 project page on sourceforge

crm114.pm

Some training will be required. If you have a good sample of mail whose correct disposition is known, you can build corpora from them to train the discriminator. Be forewarned, though, that (as with any other solution you're likely to implement) there will be times when CRM114 gets it wrong and you'll have to reschool it.

Good luck in your quest, and have fun with it.