ReGAL: Rule-Generative Active Learning for Model-in-the-Loop Weak Supervision

Anonymous

Keywords: active learning, data programming, rule generation, text classification, human-in-the-loop

TL;DR: We create an active learning framework where humans provide feedback to multiple datapoints simultaneously via automatically-generated labeling functions.

Abstract: One of the main bottlenecks to extending deep learning systems to new domains is the prohibitive cost of acquiring sufficient training labels. While many previous works have sought to alleviate this problem with weak supervision and data programming, rule and label noise prevent them from approaching fully-supervised performance. This work-in-progress provides a principled, AI-guided approach to improve rule-based and weakly supervised text classification by performing active learning not on individual data instances, but on entire labeling functions. We argue that such a framework can guide users and subject matter experts to select labeling rules that expand label function coverage without sacrificing clarity. Our experiments show that our framework, ReGAL, is able to generate coherent labeling rules while simultaneously obtaining state-of-the-art performance in weakly supervised text classification.