The first directly considers integrating knowledge when we begin to train a model such as ERNIE [15]. Although domain-specific knowledge graphs can be injected in the pre-training phase, this training process can be expensive and time-consuming. The second considers the integration of domain-...