Snuba: Automating Weak Supervision to Label Training Data Paroma Varma Christopher Re´ Stanford University Stanford University
[email protected] [email protected] ABSTRACT benign, malignant if area > 210.8: As deep learning models are applied to increasingly diverse return False Label if area < 150: Aggregator problems, a key bottleneck is gathering enough high-quality return Abstain training labels tailored to each task. Users therefore turn Labeled Data 25% benign, 75%benign, ... ?, ?, ?, ? if perim > 120: to weak supervision, relying on imperfect sources of labels return True like pattern matching and user-defined heuristics. Unfor- if perim > 80 Terminate? return Abstain tunately, users have to design these sources for each task. Heuristic Generation Training Labels This process can be time consuming and expensive: domain Unlabeled Data Snuba experts often perform repetitive steps like guessing optimal numerical thresholds and developing informative text pat- Figure 1: Snuba uses a small labeled and a large unlabeled terns. To address these challenges, we present Snuba,a dataset to iteratively generate heuristics. It uses existing la- system to automatically generate heuristics using a small bel aggregators to assign training labels to the large dataset. labeled dataset to assign training labels to a large, unla- beled dataset in the weak supervision setting. Snuba gen- pervision, or methods that can assign noisy training labels erates heuristics that each labels the subset of the data it to unlabeled data, like crowdsourcing [9, 22, 60], distant su- is accurate for, and iteratively repeats this process until the pervision [8,32], and user-defined heuristics [38,39,50]. Over heuristics together label a large portion of the unlabeled the past few years, we have been part of the broader effort data.