Table 3:
An ablation study of Snorkel DryBell using only labeling functions that depend on servable features (“Servable LFs”) compared with all labeling functions, including non-servable resources. All scores are normalized to the precision, recall, and F1 of the logistic regression classifier trained directly on the development set. Lift is reported relative to Servable LFs.
| Relative: | P | R | F1 | Lift |
|---|---|---|---|---|
| Topic Classification | ||||
| Servable LFs | 50.9% | 159.2% | 86.1% | |
| + Non-Servable LFs | 100.6% | 132.1% | 117.5% | +36.4% |
| Product Classification | ||||
| Servable LFs | 38.0% | 119.2% | 62.5% | |
| + Non-Servable LFs | 99.2% | 110.1% | 105.2% | +68.2% |