Active learning improves rare class detection despite initial bias
Active Learning for Finding a Needle in a Haystack
Computation and Language
Summary
Finding rare examples, like hate speech, in a large collection is hard because there are very few of them to learn from. The researchers studied special methods called active learning to see if they can better find these rare cases even if the starting examples are biased or incomplete. They found some methods recover missing groups well on news data, but performance varies on hate speech data. Their work shows that different active learning methods handle initial bias differently and that overall scores can hide these important differences.
What this means in practice
- •For content moderation teams: Improve detection of rare harmful content by choosing training examples that better represent all hate speech types despite biased initial data.
- •For document classification teams: Develop classifiers for rare document categories more efficiently by using active learning strategies that correct initial sampling biases.
Authors
Emily Silcock, Simon Löwe
Abstract
Training a classifier to find a rare class of documents, such as hate speech directed at a particular religious group, requires labelled examples that are expensive to obtain when the class makes up only a small fraction of a corpus. Active learning reduces this cost, but it must start from an initial labelled set that contains examples of the target class. Random sampling cannot supply these at reasonable cost, so they must be selected in some other way, such as keyword search, and any such selection biases the initial set towards part of the target class. We evaluate whether active learning can correct this bias. On AG News and Measuring Hate Speech, at a prevalence of around 1%, we compare random sampling with eight active learning strategies, starting from initial sets that cover or omit known subgroups of the target class, and measure recall on the omitted subgroups as well as F1. On AG News, almost every strategy recovers the omitted subgroup. On Measuring Hate Speech, uncertainty-based strategies recover less of the omitted subgroups than random sampling, whereas discriminative active learning recovers 55% of them without a significant loss of F1. Aggregate metrics conceal these differences.