An Unofficial Implementation for Distilling Task-Specific Knowledge from BERT into Simple Neural Networks