Towards Human-Like Machine Comprehension: Few-Shot Relational Learning in Visually-Rich Documents