The abstract describes the challenge of training classifiers to find rare document types, such as hate speech directed at a particular religious group. Because these documents make up only a small fraction of a larger corpus, obtaining enough labeled examples through random sampling is costly and impractical.
The proposed approach uses active learning, a technique that lets the model choose which unlabeled examples to query for human annotation. By focusing labeling effort on the most informative candidates, the method seeks to reduce the number of labeled examples needed while still achieving strong classification performance.
As described in the abstract, the work addresses a practical imbalance problem common in content moderation and targeted abuse detection. The full paper likely details the specific query strategies and experimental results, but the core idea is to make rare-class classification more efficient.