Table 1.

Modeling advantage $A_w$ attained using a generative model for several applications in Snorkel (Sect. 4.1), the upper bound ${\tilde{A}}^{\ast }$ used by our optimizer, the modeling strategy selected by the optimizer—either majority vote (MV) or generative model (GM)—and the empirical label density $d_{\mathrm{\Lambda }}$

Dataset	$A_w$ (%)	${\tilde{A}}^{\ast }$ (%)	Modeling strategy	$d_{\mathrm{\Lambda }}$
Radiology	7.0	12.4	GM	2.3
CDR	4.9	7.9	GM	1.8
Spouses	4.4	4.6	GM	1.4
Chem	0.1	0.3	MV	1.2
EHR	2.8	4.8	GM	1.2