Table 2.

Scores for different sampling schemes on our range of text quality metrics

Method	ppv	sens	f1	CIDEr	ES
Beam (k = 3)	0.3323	0.1786	0.2118	0.2013	0.6207
Beam (k = 5)	0.3216	0.1581	0.1922	0.1865	0.5981
Beam (k = 10)	0.3190	0.1410	0.1765	0.1748	0.5733
Prob (t = 0.5)	0.3148	0.2208	0.2394	0.2186	0.6541
Prob (t = 1.0)	0.1805	0.1520	0.1492	0.1410	0.5973
Greedy	0.3608	0.2418	0.2674	0.2458	0.6688

Simplified PPV, sens, and F1-scores measure n-gram overlap between the authentic and synthetic chief complaints; and CIDEr and the ES scores measure their similarity in vector space. Top scores are shown in bold

PPV positive predictive value, sens sensitivity, ES embedding similarity