<p>NER performance of different prompt designs under zero-shot learning. The BioClinicalBERT baseline is taken from Shyr et al., who fine-tuned BioClinicalBERT on the RareDis Corpus [3,11]. The best performances among different prompt designs and BioClinicalBERT model are underscored for each task. The 95% CIs constructed by bootstrap are shown in parentheses. CIs are computed for GPT based experiments and not available for the BioClinicalBERT baseline reported in the literature.</p>

NER performance of different prompt designs under zero-shot learning. The BioClinicalBERT baseline is taken from Shyr et al., who fine-tuned BioClinicalBERT on the RareDis Corpus [3,11]. The best performances among different prompt designs and BioClinicalBERT model are underscored for each task. The 95% CIs constructed by bootstrap are shown in parentheses. CIs are computed for GPT based experiments and not available for the BioClinicalBERT baseline reported in the literature.