Visual Exploration of Indirect Bias in Language Models

Information

• Publication Type: Conference Paper
• Workgroup(s)/Project(s):
  • Joint Human-Machine Data Exploration
• Date: June 2023
• ISBN: 978-3-03868-219-6
• Publisher: The Eurographics Association
• Open Access: yes
• Location: Leipzig, Germany
• Lecturer: Manuela Waldner
• Event: 25th EG Conference on Visualization (EuroVis 2023)
• DOI: 10.2312/evs.20231034
• Booktitle: EuroVis 2023 - Short Papers
• Pages: 5
• Conference date: 12. June 2023 – 16. June 2023
• Keywords: visual analytics, language models, bias

Abstract

Language models are trained on large text corpora that often include stereotypes. This can lead to direct or indirect bias in downstream applications. In this work, we present a method for interactive visual exploration of indirect multiclass bias learned by contextual word embeddings. We introduce a new indirect bias quantification score and present two interactive visualizations to explore interactions between multiple non-sensitive concepts (such as sports, occupations, and beverages) and sensitive attributes (such as gender or year of birth) based on this score.

Additional Files and Images

Weblinks

• Eurographics Digital Library
• online demo
• Entry in reposiTUm (TU Wien Publication Database)
• DOI: 10.2312/evs.20231034

BibTeX

@inproceedings{indirectBiasLanguageModels-2023,
  title =      "Visual Exploration of Indirect Bias in Language Models",
  author =     "Judith Louis-Alexandre Dit Petit-Frere and Manuela Waldner",
  year =       "2023",
  abstract =   "Language models are trained on large text corpora that often
               include stereotypes. This can lead to direct or indirect
               bias in downstream applications. In this work, we present a
               method for interactive visual exploration of indirect
               multiclass bias learned by contextual word embeddings. We
               introduce a new indirect bias quantification score and
               present two interactive visualizations to explore
               interactions between multiple non-sensitive concepts (such
               as sports, occupations, and beverages) and sensitive
               attributes (such as gender or year of birth) based on this
               score.",
  month =      jun,
  isbn =       "978-3-03868-219-6",
  publisher =  "The Eurographics Association",
  location =   "Leipzig, Germany",
  event =      "25th EG Conference on Visualization (EuroVis 2023)",
  doi =        "10.2312/evs.20231034",
  booktitle =  "EuroVis 2023 - Short Papers",
  pages =      "5",
  keywords =   "visual analytics, language models, bias",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2023/indirectBiasLanguageModels-2023/",
}