The A+ Alliance funded paper, prototype, and current pilot, E.D.I.A. has been selected as one of five funded projects in Mozilla’s Data Futures Lab 2024 Infrastructure Fund Award which celebrates groundbreaking research and supports projects committed to open source and the creation of trustworthy data economies.
E.D.I.A. is an acronym In Spanish for “Stereotypes and Discrimination in Artificial Intelligence”.
The tool works on Democratising Technical Barriers for Bias Assessment in NLP in order to address discrimination in Large Language Models (LLMs) and word embeddings (WE).
E.D.I.A. led by computer scientists Laura Alonso Alemany and Luciana Benotti and developed by the Buenos Aires based feminist civil society organization Fundación Vía Libre led by Beatriz Busaniche, was developed and funded from a concept to a paper in the first A+ Alliance cohort, then to a prototype in the second A+ Alliance cohort, and now as a pilot in the final and fourth cohort of the 2021-2024 A+ Alliance + f<A+I>r Feminist AI Research network Incubating Feminist AI: Paper to Prototype to Pilot project funded by IDRC, Canada’s International Development Research Centre.
In this new cycle, E.D.I.A. will take the tool built from the A+ Alliance prototype and provide a methodology for social scientists and domain experts in Latin America to explore these biases and discriminatory stereotypes.
Fundación Vía Libre has a longstanding commitment to using community-centered methods in their work, and are now building an ecosystem to gather community-built datasets that represent stereotypes in Argentina. Such datasets are the keystone to audit language technologies, to detect and characterize discriminatory behaviors and hate speech.
E.D.I.A. allows users to define the type of bias they wish to explore. Additionally, E.D.I.A. supports an intersectional analysis by considering two binary dimensions, such as female-male intersected with fat-skinny.
With E.D.I.A., Via Libre aims to
publish programming libraries to integrate the dataset in audit processes for public and private institutions who use language models;publish structured content and teaching materials so that others can replicate their methods for other languages and contexts,inspect core components of automatic language processing technologies to detect and characterize discriminatory behaviors.