This page contains some code that
either my coauthors or I wrote.
Pytorch code for the ZEN method
This is an unsupervised pretraining method for Chinese that can take advantage of n-gram (multi-character) information, achieving better results than BERT.