On Generative Models And Joint Architectures For Document-Level Relation Extraction