GeNRe: A French Gender-Neutral Rewriting System Using Collective Nouns
Enzo Doyen, Amalia Todirascu · 2025
Read GeNRe: A French Gender-Neutral Rewriting System Using Collective Nouns by Enzo Doyen, Amalia Todirascu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-23630.txt.
Introduction
Enzo Doyen and Amalia Todirascu present GeNRe, a system that rewrites French text in gender-neutral form. Much natural language processing data carries gender bias, especially through masculine generics, masculine words meant to cover mixed groups of men and women. Such forms can perpetuate and amplify stereotypes.
Gender rewriting automatically detects gendered forms and replaces them with neutral or opposite forms. Neutralization techniques, as opposed to inclusive or gender-switching ones, had previously been studied only for English. GeNRe is the first French neutralization system built on collective nouns, which have fixed gender in French. The authors build a rule-based system tailored to French, plus two fine-tuned language models trained on data the rule-based system generated. They also test instruction-based models, finding one large model paired with their dictionary comes close to the rule-based system. The work extends bias mitigation techniques for French language technology.
Catalog record
| CREATOR | Enzo Doyen, Amalia Todirascu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 60 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-23630
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GeNRe: A French Gender-Neutral Rewriting System Using Collective Nouns Enzo Doyen, Amalia Todirascu Abstract A significant portion of the textual data used in the field of Natural Language Processing (NLP) exhibits gender biases, particularly due to the use of masculine generics (masculine words that are supposed to refer to mixed groups of men and women), which can perpetuate and amplify stereotypes. Gender rewriting, an NLP task that involves automatically detecting and replacing gendered forms with neutral or opposite forms (e.g., from masculine to feminine), can be employed to mitigate these biases. While such systems have been developed in a number of languages (English, Arabic, Portuguese, German, French), automatic use of gender neutralization techniques (as opposed to inclusive or gender-switching techniques) has only been studied for English. This paper presents GeNRe, the very first French gender-neutral rewriting system using collective nouns, which are gender-fixed in French. We introduce a rule-based system (RBS) tailored for the French language alongside two fine-tuned language models trained on data generated by our RBS. We also explore the use of instruct-based models to enhance the performance of our other systems and find that Claude 3 Opus combined with our dictionary achieves results close to our RBS. Through this contribution, we hope to promote the advancement of gender bias mitigation techniques in NLP for French. \noautomath 1 Introduction Since the 1970s, several psycholinguistic studies have shown how language influences thoughts Berlin and Kay (1969) ; Kay and McDaniel (1978) . Further studies examining gender in language proved that it could lead to cognitive biases Jacobson and Insko (1985) ; Sczesny et al. (2016) , particularly when it comes to the use of masculine generics (MG), that is masculine words that are supposed to refer to mixed groups of men and women or to someone who gender is unknown Braun et al. (2005) ; Gygax et al. (2008) ; Gygax et al. (2019) ; Richy and Burnett (2021) . For example, Stahlberg et al. (2001) showed that when asked to name a celebrity in a certain field in German, respondents were more likely to give the name of a man when a MG was used in the question. Gender bias in natural language processing (NLP) models is a critical issue that can lead to biased predictions and the amplification of bias in the training data Lu et al. (2020) ; Stanczak and Augenstein (2021) ; Kotek et al. (2023) ; Ducel et al. (2024) . This problem is particularly relevant for machine translation systems, which are highly susceptible to gender bias when translating between languages with different grammatical gender systems Savoldi et al. (2021) ; Wisniewski et al. (2021) ; Vanmassenhove (2024) . Data augmentation, which involves balancing the amount of data for all genders in a specific language, has been suggested as a potential solution to debias NLP systems Zhao et al. (2018) . This led to an NLP task known as “gender rewriting,” whose goal is to automatically propose alternatives to gendered sentences. As of yet, automatic gender neutralization techniques have only been developed in English Vanmassenhove et al. (2021) ; Sun et al. (2021) . In French, the only gender rewriting system created is, to our knowledge, that of Lerner and Grouin (2024) , which converts MG to gender-fair forms highlighting the feminine suffix 1 1 1 For example, “danseurs” (male dancers or MG) and “danseuses” (female dancers) can be merged into “danseur·euses” with an interpunct (·) to refer to both men and women. . Thus, we develop a French gender-neutral rewriting system using human collective nouns (CN), defined by Lecolle (2019) as “nouns referring to entities comprised of groups of individuals.” 2 2 2 In French: “nom désignant une entité composée d’un ensemble d’individus humains.” CNs have been widely discussed in the literature, especially when it comes to French Flaux (1999) ; Lammert (2010) ; Lammert and Lecolle (2014) ; Lecolle (2019) . Since, in French, this type of noun has a gender which does not depend upon the referent’s, 3 3 3 For instance, “la police” ( “police” ) refers to both policemen and policewomen. it is an effective way of achieving gender neutralization. By focusing on gender neutralization, our work targets a writing technique that, compared to visibilization (used by Lerner and Grouin (2024) ), tends to be less contentious among native French speakers, as it does not alter the spelling of existing words nor does it introduce non-standard or new punctuation marks to separate the feminine suffix from the base word form Burnett and Pozniak (2021) . Finally, gender neutralization challenges the binary male/female gender dichotomy and is better adapted for people whose gender falls outside of the traditional categories. Our gender-neutral rewriting system, GeNRe ( Ge nder- N eutral Re writing System Using French Collective Nouns), is the very first gender-neutral rewriting system for French 4 4 4 Code and data are made publicly available on GitHub, under license CC BY-SA 4.0 https://github.com/spidersouris/GeNRe . We present three different versions of our system: one relying on a rule-based system (RBS), one relying on fine-tuned language models and one relying on an instruct-based model. 2 Gender in French In French, nouns (N) are classified as either masculine or feminine, and the gender of a noun influences the form of determiners (D), adjectives (A) and past participle verbs (V) that are syntactically related. Similarly, coreferent pronouns (P), that is pronouns referring to a previously mentioned entity, also feature the same gender. Examples 2 (masculine) and 2 (feminine) highlight the syntactic differences that arise when using either an inanimate masculine noun ( “courrier” , mail ) and an inanimate feminine noun ( “lettre” , letter ). {exe} \ex \overset{\mathbf{D}}{\text{{Le}}} \overset{\mathbf{N}}{\text{courrier}} \overset{\mathbf{A}}{\text{{recommandé}}} a été \overset{\mathbf{V}}{\text{{écrit}}} récemment. \overset{\mathbf{P}}{\text{{Il}}} est \overset{\mathbf{V}}{\text{{adressé}}} à son mari. (The registered [m.] mail [m.] was written [m.] recently. It [m.] is addressed [m.] to her husband.) \ex \overset{\mathbf{D}}{\text{{La}}} \overset{\mathbf{N}}{\text{lettre}} \overset{\mathbf{A}}{\text{{recommandée}}} a été \overset{\mathbf{V}}{\text{{écrite}}} récemment. \overset{\mathbf{P}}{\text{{Elle}}} est \overset{\mathbf{V}}{\text{{adressée}}} à son mari. (The registered [f.] letter [f.] was written [f.] recently. It [f.] is addressed [f.] to her husband.) The gender of human role nouns reflects the sociological gender of the referent and is motivated (for instance, “danseuse” refers to a female dancer), while gender of nouns referring to unanimated beings is arbitrary Watbled (2012) . The masculine gender for human nouns is considered to be the “default” gender in French, and can be used in a non-specific context (in the singular form, as in Example 2 5 5 5 In this example, “professeur” is considered as a masculine generics insofar as it does not refer to one specific male individual, but to any individual serving as “professor” . ) or to refer to groups of people composed of both men and women (in the plural form, as in Example 2 ). {exe} \ex Un professeur doit savoir faire preuve d’autorité. (A professor [m.] has to know how to show authority.) \ex Les lecteurs assidus financent le journal chaque mois. (Avid readers [m.] provide financial support to the newspaper every month.) However, the use of masculine as the default gender (masculine generics, abbreviated MG) in gender-marked languages promotes androcentric mental representations and participates in invisibilizing women Jacobson and Insko (1985) ; Stahlberg et al.