[Paper Review] Lessons from Archives: Strategies for Collecting Sociocultural Data in Machine Learning

Eun Seo Jo, Timnit Gebru|arXiv (Cornell University)|Dec 22, 2019

Ethics and Social Impacts of AI26 references105 citations

TL;DR

The paper argues for forming an ML data-collection specialization by importing archival practices to sociocultural data, emphasizing consent, inclusivity, power, transparency, and ethics & privacy, with interventionist collection and institutional structures.

ABSTRACT

A growing body of work shows that many problems in fairness, accountability, transparency, and ethics in machine learning systems are rooted in decisions surrounding the data collection and annotation process. In spite of its fundamental nature however, data collection remains an overlooked part of the machine learning (ML) pipeline. In this paper, we argue that a new specialization should be formed within ML that is focused on methodologies for data collection and annotation: efforts that require institutional frameworks and procedures. Specifically for sociocultural data, parallels can be drawn from archives and libraries. Archives are the longest standing communal effort to gather human information and archive scholars have already developed the language and procedures to address and discuss many challenges pertaining to data collection such as consent, power, inclusivity, transparency, and ethics & privacy. We discuss these five key approaches in document collection practices in archives that can inform data collection in sociocultural ML. By showing data collection practices from another field, we encourage ML research to be more cognizant and systematic in data collection and draw from interdisciplinary expertise.

Motivation & Objective

• Motivate the need to treat data collection as a fundamental ML concern with societal impact.
• Propose drawing lessons from archival and library sciences to improve ML data collection and annotation practices.
• Identify institutional and procedural structures (mission statements, codes of ethics, documentation) to guide data collection.
• Advocate for interventionist data collection to mitigate historical and representational biases.
• Suggest concrete mechanisms (consortia, community archives, participatory approaches) to implement these practices in ML.

Proposed method

• Compare archival data collection practices with ML data collection practices to identify gaps and opportunities.
• Argument for interventionist data collection to address historical and representational biases in data.
• Map archival concepts (mission statements, documentation standards, appraisal processes) to ML data governance (Datasheets for Datasets, transparency efforts).
• Propose organizational models (data consortia, community archives, participatory archives) to democratize data collection and share resources.
• Provide guidance on implementing consent, inclusivity, power, transparency, and ethics in ML datasets and processes.

Experimental results

Research questions

• RQ1 How can archival data collection practices inform data governance in machine learning?
• RQ2 What interventionist data-collection strategies can reduce historical and representational biases in ML datasets?
• RQ3 How can ML projects implement mission statements, documentation, and ethical oversight akin to archives?
• RQ4 What organizational structures (consortia, community archives, codes of conduct) are feasible for responsible sociocultural data collection in ML?

Key findings

• Archives use mission statements to define data collection goals and promote inclusivity.
• Archives rely on multi-level supervision and documented appraisal to regulate data collection, which can inform ML transparency.
• Community/participatory archives empower underrepresented groups to define their own representations and data access protocols.
• Data consortia and shared frameworks can address cost, labor, and fairness challenges in data collection for ML.
• Ethics and privacy in archives are enforced through codes of conduct and documented procedures, offering a model for ML governance and compliance.
• Interventionist data collection helps mitigate historical and representational biases before ML modeling.