Ouch, feeding probabilistic models training data scored with a gradient of truthfulness tags generated by humans and all their biases... surely this won't end horribly and simply serve as a method to algorithmically institute the tyranny of the majority.

If you really want to do this (You really don't, I assure you - you'll hate the end result), you've got to reach back through the AI winter and drag the granddaddy of NLP, propositional logic, into modern AI development. We'll see this employed by lawyers long before journalists.

https://en.wikipedia.org/wiki/Attempto_Controlled_English