What does “Weak Supervision” mean?
Using noisy, heuristic, or partial labels to train models when clean labels are scarce.
How Weak Supervision Works
Weak Supervision: Keyword rule, Existing database, Another model's guess feed into Combine noisy labeling sources, which produces Probabilistic training labels.
Practical Example
Rather than labeling 100,000 documents by hand, a team writes a few rules, such as 'contains the word invoice'. Each rule is imperfect, but combining them yields approximate labels to train a first model.
Why It Matters
It reduces the cost of labeling. The labels contain errors, so the results need checking on a small, carefully labeled set.
Common Questions About Weak Supervision
Q: Is it less accurate than manual labels?
Individual labels are noisier, but at scale the result can be useful, especially with a clean test set.
Q: What are labeling functions?
Small programs or rules that each assign a rough label to some of the data.
Q: Where is it used?
Text classification, information extraction, and other tasks where hand labeling is slow.