Weakly Supervised Vision-and-Language Pre-training with Relative Representations