You reap what you sow: On the Challenges of Bias Evaluation Under Multilingual Settings

By Zeerak Talat1, Aurélie Névéol, Stella Biderman, Miruna Clinciu, Manan Dey, Shayne Longpre, Alexandra Sasha Luccioni, Maraim Masoud, Margaret Mitchell, Dragomir R. Radev and 6 more. 2022

Evaluating bias, fairness, and social impact in monolingual language models is a difficult task. This challenge is further compounded when language modeling occurs in a multilingual context. Considering the implication of evaluation biases for large multilingual language models, we situate the discussion of bias evaluation within a wider context of social scientific research with computational work.We highlight three dimensions of developing multilingual bias evaluation frameworks: (1) increasing transparency through documentation, (2) expanding targets of bias beyond gender, and (3) addressing cultural differences that exist between languages.We further discuss the power dynamics and consequences of training large language models and recommend that researchers remain cognizant of the ramifications of developing such technologies.

Subfields: Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

BLOOM: A 176B-Parameter Open-Access Multilingual Language Model (2022, background): “Evaluation and interpretability lists individuals who helped evaluate the BLOOMmodel as well as authors of Talat et al. (2022).”
The Gradient of Generative AI Release: Methods and Considerations (2023, background): “These figures are based on tracking and evaluation initiatives [50, 49, 94], are not exhaustive, and intend to show release trends over time.”
The BigScience ROOTS Corpus: A 1.6TB Composite Multilingual Dataset (2023, methodology): “Throughout the process of building this corpus we engaged in simultaneous investigation of ethical (Talat et al., 2022), sociopolitical (McMillan-Major et al.”
Undesirable biases in NLP: Averting a crisis of measurement (2022, methodology): “…law, sociology and psychology — expertise that NLP (and other AI) practitioners, who are themselves not specialized on these topics, should make use of (Blodgett et al., 2020; Kiritchenko et al., 2021; Balayn & Gürses, 2021; Cheng et al., 2021; Talat et al., 2022; Weinberg, 2022, i.a.).”
On Measures of Biases and Harms in NLP (2021, background): “Bias in NLP Bias in language models is commonly defined as “skew that produces a type of harm” (Crawford, 2017) towards different social groups, though it is a complex notion that is often not well-defined in existing literature (Blodgett et al., 2020; Delobelle et al., 2022; Talat et al., 2022).”
Exploring AI Ethics of ChatGPT: A Diagnostic Analysis (2023, background): “Multilingualism: The monolingual bias in multilingualism is a type of bias that can occur in language models [39].”
Gender bias and stereotypes in Large Language Models (2023, background): “However, as is well known, language models perpetuate and occasionally amplify biases, stereotypes, and negative perceptions of minoritized groups in society [10, 13, 14, 66, 69, 84, 85, 90].”