Statistics for Towards Effective Utilization of Pretrained Language Models — Knowledge Distillation from BERT