Fig. 10.

Precision–recall curves for the relation extraction tasks. The top plots compare a majority vote of all labeling functions, Snorkel’s generative model, and Snorkel’s discriminative model. They show that the generative model improves over majority vote by providing more granular information about candidates, and that the discriminative model can generalize to candidates that no labeling functions label. The bottom plots compare the discriminative model trained on an unweighted combination of the labeling functions, hand supervision (when available), and Snorkel’s discriminative model. They show that the discriminative model benefits from the weighted labels provided by the generative model, and that Snorkel is competitive with hand supervision, particularly in the high-precision region