Table 1.

Accuracy of differently trained models on the test set

Model	Full test set: 13 545 QAs		Test HasAns: 12 376 QAs		Test NoAns: 1169 QAs
	Exact	Partial	Exact	Partial	Exact	Partial
BERTbase + emrQAwhy	0.633	0.688	0.599	0.659	0.995	0.995
BERTbase + SQuADwhy + emrQAwhy	0.660	0.728	0.628	0.703	0.994	0.994
BERTbase + i2b2Notespre + emrQAwhy	0.663	0.718	0.631	0.692	0.997	0.997
BERTbase + ClinBERTpre + emrQAwhy	0.695	0.744	0.666	0.720	0.994	0.994
BERTbase + ClinBERTpre + SQuADwhy + emrQAwhy	0.700	0.757	0.672	0.735	0.995	0.995
BERTbase + ClinBERTpre + SQuADwhy + emrQAwhy (10 epochs)	0.707	0.760	0.679	0.737	0.999	0.999

Abbreviations: HasAns: answerable according to the gold; NoAns: unanswerable according to the gold.