Learning Knowledge Uncertainty from the Pretrained Language Model