Research map / Large language models / AI and society
Position: Measure Dataset Diversity, Don't Just Claim It
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper that applies measurement theory (conceptualization, operationalization, reliability, validity) to critique how 135 image and text ML datasets define and validate "diversity", with a case study on SA-1B.
Problem
Dataset curators describe datasets as "diverse" using value-laden terms that lack clear definitions and validation, so claimed diversity cannot be checked or compared across datasets.
Approach
Treats dataset collection as a measurement process and borrows social-science measurement theory (conceptualization, operationalization, reliability, construct validity) to give concrete recommendations for defining and validating diversity; grounds this in a systematic review of 135 datasets and a case study of SA-1B.
Key results
- Of 135 surveyed image and text datasets, only 52.9% explicitly justify the need for diverse data.
- Methodological caveats/limitations are discussed in only 74.8% of cases, and only 40 of 135 papers are standalone dataset papers.
- Only 56.3% of datasets describe quality control and 36.8% include annotation quality details; 48.5% rely on downstream model performance for evaluation.
Surprising findings
- Diversity claims are often validated only through downstream model performance, which may measure the wrong construct.
- Quantitative inter-annotator agreement metrics are reported only in text datasets, not image datasets.
Limitations
- Corpus limited to image and text datasets from selected venues up to September 2022.
- Recommendations are illustrated qualitatively rather than validated empirically with experiments.
- Case study on SA-1B relies on the authors' datasheet and paper without access to collection details.
Experimental setup
Datasets: SA-1B COCO OpenImages
Metrics: percentage of datasets reporting justification, trade-offs,
Builds on: measurement theory datasheets for datasets
Closest papers by meaning
- A Taxonomy of Challenges to Curating Fair Datasets (NeurIPS 2024 · similarity 0.85)
- The State of Data Curation at NeurIPS: An Assessment of Dataset Development Practices in the Datasets and Benchmarks Track (NeurIPS 2024 · similarity 0.85)
- "What is Different Between These Datasets?" A Framework for Explaining Data Distribution Shifts (NeurIPS 2026 · similarity 0.83)
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.83)
- A Decade's Battle on Dataset Bias: Are We There Yet? (ICLR 2025 · similarity 0.83)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.