Exploring the Influence of Label Aggregation on Minority Voices: Implications for Dataset Bias and Model Training
Mugdha Pandya, Nafise Sadat Moosavi, Diana Maynard
TL;DR
This paper examines how standard label aggregation methods in sexism-detection datasets influence minority opinion representation and downstream model behavior. It compares majority, expert, and minority aggregation across two datasets (GE and EDOS), analyzing data distributions, gold-label alignment, and model predictions, along with a qualitative study of annotator disagreements. The findings show that minority aggregation preserves nuanced and potentially more harmful forms of sexism but can lead to broader, sometimes excessive, classifications, while majority/expert approaches emphasize easily identifiable labels and can bias models toward those perspectives. The work highlights the need for task-dependent aggregation strategies and distribution-aware evaluation to mitigate bias amplification in downstream models and to better represent minority viewpoints in sensitive domains.
Abstract
Resolving disagreement in manual annotation typically consists of removing unreliable annotators and using a label aggregation strategy such as majority vote or expert opinion to resolve disagreement. These may have the side-effect of silencing or under-representing minority but equally valid opinions. In this paper, we study the impact of standard label aggregation strategies on minority opinion representation in sexism detection. We investigate the quality and value of minority annotations, and then examine their effect on the class distributions in gold labels, as well as how this affects the behaviour of models trained on the resulting datasets. Finally, we discuss the potential biases introduced by each method and how they can be amplified by the models.