Probabilistic Assumptions Matter: Improved Models for Distantly-Supervised Document-Level Question Answering