DebIE: A Platform for Implicit and Explicit Debiasing of Word Embedding Spaces

TL;DR

DebIE is an integrated platform that measures and mitigates biases in word embeddings, providing accessible tools for researchers and developers to evaluate and reduce stereotypes in NLP models.

Contribution

This work introduces DebIE, the first unified platform for bias measurement and debiasing in word embeddings, supporting multiple interfaces and customizable bias specifications.

Findings

DebIE effectively measures implicit and explicit biases in various embedding spaces.

DebIE successfully reduces biases using its debiasing models across different datasets.

The platform is accessible via web, desktop, API, and command-line interfaces.

Abstract

Recent research efforts in NLP have demonstrated that distributional word vector spaces often encode stereotypical human biases, such as racism and sexism. With word representations ubiquitously used in NLP models and pipelines, this raises ethical issues and jeopardizes the fairness of language technologies. While there exists a large body of work on bias measures and debiasing methods, to date, there is no platform that would unify these research efforts and make bias measuring and debiasing of representation spaces widely accessible. In this work, we present DebIE, the first integrated platform for (1) measuring and (2) mitigating bias in word embeddings. Given an (i) embedding space (users can choose between the predefined spaces or upload their own) and (ii) a bias specification (users can choose between existing bias specifications or create their own), DebIE can (1) compute several measures of implicit and explicit bias and modify the embedding space by executing two (mutually composable) debiasing models. DebIE's functionality can be accessed through four different interfaces: (a) a web application, (b) a desktop application, (c) a REST-ful API, and (d) as a command-line application. DebIE is available at: debie.informatik.uni-mannheim.de.