Source-linked AI summary
Datasheets for Datasets
Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé, Kate Crawford
TL;DR
Machine learning lacks a standardized process for documenting datasets, despite their importance to model behavior and risks in high-stakes domains. This paper proposes datasheets documenting datasets’ motivation, composition, collection, and recommended uses, and reports adoption by researchers and industry teams.
Problem
Machine learning lacks a standardized process for documenting datasets, despite datasets’ fundamental influence on model behavior and their importance in high-stakes domains.
Method
The authors developed questions and a workflow through iterative feedback and by testing example datasheets for two widely used datasets.
Results
Researchers adopted datasheets, while Microsoft, Google, and IBM began piloting them internally within product teams.
Takeaways & Limitations
Datasheets can facilitate communication between dataset creators and consumers and provide an opportunity to prioritize transparency and accountability.
Takeaways & Limitations
Wider adoption may require adapting the questions and workflow to organizational infrastructure, and the approach may pose problems for dynamic datasets.
Abstract
from arXiv · showhide
The machine learning community currently has no standardized process for documenting datasets, which can lead to severe consequences in high-stakes domains. To address this gap, we propose datasheets for datasets. In the electronics industry, every component, no matter how simple or complex, is accompanied with a datasheet that describes its operating characteristics, test results, recommended uses, and other information. By analogy, we propose that every dataset be accompanied with a datasheet that documents its motivation, composition, collection process, recommended uses, and so on. Datasheets for datasets will facilitate better communication between dataset creators and dataset consumers, and encourage the machine learning community to prioritize transparency and accountability.
1 Introduction
Datasets fundamentally influence machine-learning behavior, yet the field lacks a standardized process for documenting them. The paper proposes datasheets that record datasets’ motivation, composition, collection process, and recommended uses to improve transparency, accountability, and informed dataset selection.
- Dataset characteristics fundamentally influence model behavior, and mismatches between datasets and deployment contexts or societal biases can have severe consequences.
- The machine-learning community has no standardized process for documenting datasets, despite extensive study of data provenance in databases.
- Datasheets are proposed as dataset documentation covering motivation, composition, collection process, recommended uses, operating characteristics, and test results by analogy to electronics datasheets.
- Datasheets may increase transparency, accountability, reproducibility, and appropriate dataset selection, while their questions should vary by domain and organizational workflow.They are intended to facilitate reproducibility even when researchers and practitioners lack direct access to a dataset.
- Datasheets aim to prompt creators to reflect on assumptions, risks, harms, and use implications while giving consumers information for informed dataset-use decisions.
2 Development Process
The datasheet questions and workflow were developed over roughly two years through researcher experience, example datasheets, and repeated feedback. This process refined question content and organization to better align with dataset lifecycle stages and elicit factual compliance information without requiring legal judgments.
- The questions and workflow were refined over roughly two years through many rounds of feedback.
- Researchers drew on diverse experiences and dataset-related concerns, then tested initial questions by creating datasheets for two widely used datasets.
- Feedback led to added, removed, refined, and reordered questions that better matched key stages of the dataset lifecycle.
- Product teams were more likely to address legal and ethical considerations when those questions were integrated into relevant lifecycle sections rather than grouped together.
- Lawyer feedback removed explicit regulatory-compliance questions and introduced factual questions that elicited compliance information without requiring dataset creators to make legal judgments.
3 Questions and Workflow
This section presents a datasheet workflow organized around dataset-lifecycle stages, with questions that help creators document datasets and consumers assess their suitability. It emphasizes answering these questions before and after collection-related activities, including motivation, composition, collection, preprocessing, and intended uses.
- Workflow: The workflow groups datasheet questions by dataset-lifecycle stages: motivation, composition, collection, preprocessing/cleaning/labeling, uses, distribution, and maintenance.This grouping is intended to encourage dataset creators to reflect on how the dataset was created.
- Motivation: Motivation questions ask creators to explain the dataset’s purpose, identify its creators and affiliated entities, and disclose funding interests.The questions are intended to promote transparency, especially for research datasets.
- Composition: Composition questions document instance types and counts, sampling, errors, external dependencies, subpopulations, identifiability, and sensitive data.These questions help consumers evaluate dataset contents and potential privacy concerns.
- Collection process: Collection questions cover how data were acquired and validated, collection mechanisms, notice and consent, consent revocation, and impact analysis.Creators should review these questions before collection to flag potential issues and answer them after collection is complete.
- Preprocessing and uses: Preprocessing and intended-use questions help consumers determine whether transformations preserve suitability for chosen tasks and which uses may create risks or harms.For example, converting text to a bag-of-words removes information needed for tasks involving word order.
4 Impact and Challenges
Datasheets for datasets have gained adoption across academia, industry, and AI governance efforts, while implementation requires workflow changes and does not eliminate all risks. Despite added overhead and contextual challenges, the authors argue that datasheets’ communication, transparency, and accountability benefits outweigh their costs.
- Adoption: Datasheets have been adopted by academic researchers, piloted by Microsoft, Google, and IBM, and incorporated into Partnership on AI documentation guidance.Google researchers have also published follow-up work on the proposal.
- Implementation challenges: Dataset creators must adapt the proposed questions and workflow to existing organizational infrastructure, and dynamic datasets may require updated datasheets as data changes.The paper recommends accompanying infrequently changing datasets’ updated versions with updated data documentation.
- Risks and expertise: Datasheets cannot fully mitigate societal bias or potential harms because creators cannot anticipate every use and may lack demographic labels constrained by privacy and data protection.Identifying unwanted societal biases often requires additional demographic information about individuals.
- Risks and expertise: Datasheets for datasets may require collaboration with anthropology, sociology, and science and technology studies experts to collect data from people respectfully.Appropriate collection depends on complex, contextual social, historical, and geographical factors.
- Costs and benefits: Creating datasheets imposes time and organizational overhead, although it may reduce one-off questions and requires modified workflows and incentives.The authors nevertheless conclude that datasheets’ benefits for communication, transparency, and accountability outweigh their costs.
A Appendix
The appendix illustrates a datasheet for Pang and Lee’s movie-review polarity dataset, documenting its sentiment-classification purpose, composition, preprocessing, and limitations. It cautions that models trained on this movie-review data may not generalize to other sentiment tasks or support consequential decisions without verification.
- Purpose: The dataset was created to support research on predicting positive or negative sentiment polarity in English movie reviews.It focuses intentionally on movie reviews, where affect and sentiment are frequently expressed.
- Composition: Instances are movie reviews extracted from newsgroup postings, with 1,400 instances in original v1.x versions and 2,000 in v2.0.No tests were run to determine representativeness, and no more than 40 posts by a single author were included.
- Preprocessing: Preprocessing discarded reviews without explicit ratings, retained only strongly positive or strongly negative ratings, and removed non-English reviews in v1.1.The text was also down-cased, stripped of HTML and boilerplate newsgroup text, and subjected to additional automatic filtering.
- Other uses: The dataset could also support modeling or understanding movie reviews, including inducing sentiment lexicons and automatically generating movie reviews.Its instances include a sentiment polarity label derived from star ratings.
- Limitations and use: Because the dataset covers only the movie-review domain, systems trained on it may not generalize to other sentiment-prediction tasks.The datasheet advises against using such systems for consequential decisions about people without additional verification.