multilingual pre-trained language models have shown impressive performance on
cross-lingual tasks. It greatly facilitates the applications of natural
language processing on low-resource languages. However, there are still some
languages that the current multilingual models do not perfo