Position: Measure Dataset Diversity, Don't Just Claim It
Dora Zhao, Jerone T. A. Andrews, Orestis Papakyriakopoulos, Alice Xiang
TL;DR
This paper argues that ML datasets are value-laden and not inherently neutral, and thus require rigorous measurement-theory-based practices to define, operationalize, and validate diversity. Through a systematic review of 135 image and text datasets, the authors develop a taxonomy of how diversity is defined, identify pervasive documentation gaps, and propose concrete guidelines for defining variables, indicators, and reliability/validity assessments. A SA-1B case study demonstrates how explicit conceptualization and transparent operationalization can improve the assessment of geographic and object-variation diversity, while revealing limitations like inference errors and collection opacity. Overall, the work contributes to improved transparency, reliability, and reproducibility in ML data curation by providing a structured framework to quantify and validate value-laden properties in datasets.
Abstract
Machine learning (ML) datasets, often perceived as neutral, inherently encapsulate abstract and disputed social constructs. Dataset curators frequently employ value-laden terms such as diversity, bias, and quality to characterize datasets. Despite their prevalence, these terms lack clear definitions and validation. Our research explores the implications of this issue by analyzing "diversity" across 135 image and text datasets. Drawing from social sciences, we apply principles from measurement theory to identify considerations and offer recommendations for conceptualizing, operationalizing, and evaluating diversity in datasets. Our findings have broader implications for ML research, advocating for a more nuanced and precise approach to handling value-laden properties in dataset construction.