# Data Governance in the Age of Large-Scale Data-Driven Language Technology

**Authors:** Yacine Jernite, Huu Nguyen, Stella Biderman, Anna Rogers, Maraim Masoud, Valentin Danchev, Samson Tan, Alexandra Sasha Luccioni, Nishant Subramani, Gérard Dupont, Jesse Dodge, Kyle Lo, Zeerak Talat, Isaac Johnson, Dragomir Radev, Somaieh Nikpoor, Jörg Frohberg, Aaron Gokaslan, Peter Henderson, Rishi Bommasani  
**arXiv:** [2206.03216](https://arxiv.org/abs/2206.03216)  
**Date:** 2022-05-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.03216v1/

## Key points

- Machine Learning technology, particularly Large Language Models, has brought attention to the need for systematic and transparent management of language data.
- The authors propose a global language data governance approach that organizes data management among stakeholders, values, and rights.
- The framework presented is a multi-party international governance structure focused on language data, incorporating technical and organizational tools.
- There is tension between the goals of reproducible research and the need to update datasets for personal information removal.
- The example of distributed data governance in the Wikimedia Project offers valuable insights into collaborative and self-regulated data curation.
- Core stakeholders in Wikimedia projects align with those in the proposed governance structure.
- Challenges faced by Wikimedia projects parallel those that would arise in global digital language data governance.
- Content licenses play a role in governing Wikimedia data but may conflict with cultural values or concerns about exploitation.
- Editors enforce policies that evolve through contestation to ensure adherence to chosen licenses and regulations.
- A new model called Data Stewardship Organization (DSO) is proposed to empower data subjects and involve them in decisions regarding their data use.
- The DSO facilitates collaboration among multiple stakeholders to build and manage language resources responsibly.

## Summary

The recent emergence and adoption of Machine Learning technology, particularly Large Language Models, has highlighted the importance of systematic and transparent management of language data. In response to this need, the authors propose a global language data governance approach that aims to organize data management among stakeholders, values, and rights. This proposal is informed by prior work on distributed governance and is supported by an international research collaboration involving researchers and practitioners from 60 countries. The framework presented in this work is a multi-party international governance structure focused specifically on language data. It incorporates both technical and organizational tools necessary to support its work. The authors emphasize the tension between the goals of reproducible research which require public recording of datasets and the need to update datasets to accommodate requests for personal information removal. They also highlight the potential circulation of unredacted copies of datasets for extended periods. To provide further context, the authors discuss the example of distributed data governance in the Wikimedia Project. This project offers valuable insights into highly collaborative and self-regulated data curation, similar to the goals proposed in their governance structure. The core stakeholders in Wikimedia projects align with those in Figure 1 of their proposed governance structure: contributors (data rights holders), editors (data custodians), Wikimedia Foundation (data stewards and helpers), researchers, digital platforms, and end-users (data modelers). The challenges faced by Wikimedia projects such as diverse editor needs and goals navigating local laws and addressing power imbalances parallel those that would arise in global digital language data governance. The authors highlight how content licenses play a role in governing Wikimedia data but may conflict with cultural values or concerns about exploitation. Editors enforce policies that constantly evolve through contestation to ensure adherence to chosen licenses and regulations. Similar to their proposed framework success within the Wikimedia editor community relies on a wide range of tools facilitating data governance at scale. In light of identified needs for a comprehensive governance structure outlined throughout previous sections, the authors propose a new model called Data Stewardship Organization (DSO). This organization aims to empower data subjects and rights holders by involving them in decisions regarding the use of their data. The DSO facilitates collaboration among multiple stakeholders to build and manage language resources responsibly supporting the development of data-driven language technology. The authors acknowledge that coordination across stakeholders remains a significant challenge.