chinese word segmentation (CWS) models have achieved very high performance when the training data is sufficient and in-domain. However, the performance drops drastically when shifting to cross-domain and low-resource sc