Black is to Criminal as Caucasian is to Police: Detecting and Removing\n Multiclass Bias in Word Embeddings

2019/04/03 by Thomas Manzini, Yao Chong Lim, Manzini, Thomas +5 · 6 citations
Computer Science · #Text Readability and Simplification #Hate Speech and Cyberbullying Detection #Natural Language Processing Techniques

paper · pdf · doi:10.48550/arxiv.1904.04047

Abstract

Online texts -- across genres, registers, domains, and styles -- are riddled\nwith human stereotypes, expressed in overt or subtle ways. Word embeddings,\ntrained on these texts, perpetuate and amplify these stereotypes, and propagate\nbiases to machine learning models that use word embeddings as features. In this\nwork, we propose a method to debias word embeddings in multiclass settings such\nas race and religion, extending the work of (Bolukbasi et al., 2016) from the\nbinary setting, such as binary gender. Next, we propose a novel methodology for\nthe evaluation of multiclass debiasing. We demonstrate that our multiclass\ndebiasing is robust and maintains the efficacy in standard NLP tasks.\n

Cited by

• Eliminating Inductive Bias in Reward Models with Information-Theoretic Guidance
• Textual Data Bias Detection and Mitigation -- An Extensible Pipeline with Experimental Evaluation
• Erasing 'Ugly' from the Internet: Propagation of the Beauty Myth in Text-Image Models
• Fair-GPTQ: Bias-Aware Quantization for Large Language Models
• From Charts to Fair Narratives: Uncovering and Mitigating Geo-Economic Biases in Chart-to-Text
• Exploring Gender Bias in Large Language Models: An In-depth Dive into the German Language

Related

• Joint Multiclass Debiasing of Word Embeddings
• From Prejudice to Parity: A New Approach to Debiasing Large Language Model Word Embeddings
• "Thy algorithm shalt not bear false witness": An Evaluation of\n Multiclass Debiasing Methods on Word Embeddings
• DebIE: A Platform for Implicit and Explicit Debiasing of Word Embedding\n Spaces
• What are the biases in my word embedding?