Data Statements: Professional Practices for NLP Bias Mitigation and Scientific Rigor

In the field of Natural Language Processing (NLP), Emily M. Bender and Batya Friedman proposed the concept of "Data Statements" in 2018. The paper points out that as technology becomes widely integrated into social applications, researchers need to critically consider how design decisions impact people. The field currently faces two core problems: insufficient generalizability of scientific results and potential real-world harm from an ethical standpoint. Many NLP systems fail or reinforce existing biases against uncovered groups because their training data comes from specific populations.

A Data Statement is defined as a description of dataset characteristics, intended to provide developers and users with context to understand the scope of experimental result generalization, the appropriateness of software deployment, and any biases that may be reflected in the system. Drawing on practices for standardizing research populations in psychology and medicine, the authors advocate for incorporating Data Statements into NLP papers, experiment reports, and system documentation. This increased transparency helps address ethical issues such as exclusion, overgeneralization, and underexposure.

The paper details key roles involved in dataset creation, including speakers, annotators, curators, and stakeholders. For example, indirect stakeholders do not directly use the system but are affected by it, such as when search engine algorithms determine whether their web content is visible. By clarifying these roles and dataset characteristics, Data Statements help the field better understand and describe results, enabling more rigorous scientific and engineering practices.

Consider a hiring screening system trained on historical resume data. If the data shows a low proportion of female candidates for computer programming positions and lower ratings for them, the system may learn and amplify this bias, ranking female candidates lower. Data Statements require disclosure of the source population of training data and potential biases, allowing developers to identify such risks and make adjustments before deployment to avoid unfair treatment of specific groups by the system.