More granular message / signature parts classification

The algo would be more accurate if the classification is more detailed. E.g. if it classifies message greeting there could be a sanity check saying that signature can't go right after a greeting.

Similarly if a closing phrase (e.g. "Kind Regards,", "Thanks,", etc) is detected in signature candidate lines we could say for sure that it should be the first signature line.

Another example - disclaimers. Some messages have them and that totally breaks signature detection logic.