Do the Benefits of Joint Models for Relation Extraction Extend to Document-level Tasks?