Peking University, University of California Berkeley and others jointly | domain adaptive text classification with structured knowledge from unlabeled data (Domain Adaptive Text Classification Based o

author ：Tian Li,Xiang Chen,Zhen Dong etc.

brief introduction ： Domain adaptive text classification is a challenging task in large-scale pre training problem Language model , Because they often require expensive additional tag data to adapt to new areas . Existing works usually fail to take advantage of the implicit relationships between cross domain words . In this paper , A new method is proposed , It is called structured knowledge domain adaptation (DASK), Domain adaptation is enhanced by utilizing word level semantic relationships .DASK First, build a knowledge map to capture the main words in the target domain （ Domain independent words ） And non main words . Then during training ,DASK Inject the knowledge map information related to the pivot into the source domain text . For downstream tasks , These knowledge injection texts are input to the... That can process the knowledge injection text data BERT In variant . Thanks for knowledge injection , The author's model is an invariant feature of non pivot learning domain according to the relationship with pivot .DASK The polarity score of the candidate pivot is dynamically inferred during training using the pseudo label , Make sure the pivot has domain invariant behavior . The author has verified the results on a wide range of cross domain emotion classification tasks DASK, And observed ​​20 The baseline absolute performance improvement of different domain pairs is up to 2.9%. The code will be https://github.com/hikaru-nara/DASK The provision of . 

 

Paper download ：https://arxiv.org/pdf/2206.09591.pdf