French CrowS-Pairs: Extending a challenge dataset for measuring social bias in masked language models to a language other than English

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8521 - 8531 May 22-27, 2022 c⃝2022 Association for Computational Linguistics French CrowS-Pairs: Extending a challenge dataset for measuring social bias in masked language models to a language other than English Aurélie Névéol Université Paris-Saclay, CNRS, LISN 91400, Orsay, France neveol@lisn.fr Yoann Dupont1, Julien Bezançon2 1ObTIC, 1,2Sorbonne Université, 28 rue Serpente, 75006 Paris, France {first}.{last}@2[.etu] .sorbonne-universite.fr Karën Fort Université de Lorraine, CNRS, Inria, LORIA, F-54000 Nancy, France Sorbonne Université, F-75006 Paris, France karen.fort@loria.fr Abstract Warning: This paper contains explicit state- ments of offensive stereotypes which may be upsetting Much work on biases in natural language pro- cessing has addressed biases linked to the so- cial and cultural experience of English speak- ing individuals in the United States. We seek to widen the scope of bias studies by cre- ating material to measure social bias in lan- guage models (LMs) against specific demo- graphic groups in France. We build on the US-centered CrowS-pairs dataset to create a multilingual stereotypes dataset that allows for comparability across languages while also characterizing biases that are specific to each country and language. We introduce 1,677 sentence pairs in French that cover stereo- types in ten types of bias like gender and age. 1,467 sentence pairs are translated from CrowS-pairs and 210 are newly crowd- sourced and translated back into English. The sentence pairs contrast stereotypes concern- ing underadvantaged groups with the same sentence concerning advantaged groups. We find that four widely used language models (three French, one multilingual) favor sen- tences that express stereotypes in most bias cat- egories. We report on the translation process, which led to a characterization of stereotypes in CrowS-pairs including the identification of US-centric cultural traits. We offer guide- lines to further extend the dataset to other lan- guages and cultural environments. 1 Introduction Human language technologies can have a direct impact on people’s everyday life. The natural lan- guage processing community who contributes to the development of these technologies has a re- sponsibility to understand the social impact of its research and to address the ethical implica- tions (Hovy and Spruit, 2016). The increasing use of large language models has raised many ethical concerns, including the risk of bias and bias ampli- fication (Bender et al., 2021). Biases in NLP have received a lot of attention in recent years (Blodgett et al., 2020). However, the bulk of the work has addressed biases linked to the social and cultural experience of English speaking individuals in the United States. In this work, we seek to widen the scope of bias studies by creating material to mea- sure social bias in multiple languages and social contexts. As a case study, we chose to address bi- ases against specific demographic groups in France. The CrowS-pairs dataset (Nangia et al., 2020) was recently developed to address nine types of bias. It contains pairs of sentences: a sentence that is more stereotyping and another that is less stereotyping. …

Content truncated to 500 words