A labeling function is a cheap, imperfect heuristic that labels otherwise unlabeled examples. Collections of these rules produce noisy training data for multisource weak supervision. section 13.4
Use labeling functions when:
- A model issue needs a prompt workaround while retraining or a full release would take too long. section 13.4
- Existing overrides can be collected into a noisy dataset rather than remaining untracked technical debt. section 13.4
From overrides to training data
Rule-based overrides can correct outputs for particular scenarios and later become labeling functions. Their combined, imperfect labels provide a foundation for model training, although the rules must be tracked and maintained. section 13.4
