This got me thinking on how to train such models. At first, it could be that DailyBot allows any recipient to flag a message they have received, then they provide feedback whether it was aggressive, etc... DailyBot can use that to train the model, and also to provide some neutral feedback to the sender.

The good thing about flagging is that DailyBot won't need to process all the messages, only the ones that a person wants to flag.