Appendix of " Decoupling Knowledge from Memorization: Retrieval-augmented Prompt Learning "

In this section, we introduce the datasets as shown in Table 1 and list the templates we use in experiments as follows. T (x) = [CLS]x It was [MASK]. We set Verbalizer: (great/terrible) (positive/negative) for SST-2 MR and CR. For the Yahoo dataset, we assign the Verbalizer following the original labels. FewNERD, SemEval and TACRED are datasets for information extraction, which require inserting the entity into the template.