I’ve come up with the idea, and I’m wondering if someone has already tried to implement this.
As I`ve seen from various algorithms when an NN tries to answer a question, it actually uses a text passage, a question and all the contextual knowledge it got from the big corpus of text it has been trained on. Therefore, it cannot answer a difficult question that a human can, because it has no reasoning ability, it tries to draw conclusions only from the context.
Why if we add, in addition to a big corpus of text, a bunch of predicates, know facts, like “water boils at 100 C”, and then while labeling if a human answers a question he will specify, that, in order to answer a specific question he also used some specific set of facts.
This will add some specific form of attention to some known predicates.