French NLP System Reduces Gender Bias with Neutral Rewriting Techniques.

Researchers developed GeNRe, the first French system for automatically neutralising gender in text. Leveraging rule-based and fine-tuned models, alongside instruction-based approaches utilising Claude 3 Opus, the system addresses gender bias inherent in the French language’s fixed-gender collective nouns. This work advances gender bias mitigation in Natural Language Processing.

Natural Language Processing (NLP) systems frequently inherit and amplify societal biases present in the textual data they process. A particular concern is the pervasive use of masculine generics – words conventionally used to represent both men and women – which can reinforce stereotypes. Researchers are now addressing this issue with automated techniques to neutralise gendered language. Enzo Doyen and Amalia Todirascu, from the Laboratoire d’Informatique et de Logique de Paris (LiLPa) at the University of Strasbourg, detail their development of GeNRe, a novel French language system designed to rewrite text using gender-neutral collective nouns. Their work, presented in the paper ‘GeNRe: A French Gender-Neutral Rewriting System Using Collective Nouns’, introduces a rule-based system and fine-tuned models, alongside an investigation into the potential of instruct-based models, to mitigate gender bias in French NLP.

GeNRe: Advancing Gender-Neutral Rewriting in French

Research details the development of GeNRe, a system designed to mitigate gender bias within the French language through automated rewriting. The core innovation lies in its focused approach to collective nouns–nouns denoting a group of individuals-and their grammatical implications for gender and number agreement.

The research team investigated multiple methodologies, balancing performance with computational demands. A rule-based system, constructed upon predefined linguistic rules, was developed alongside fine-tuned models – machine learning algorithms trained specifically for this task. Furthermore, the capabilities of the large language model (LLM) Claude 3 Opus were explored, integrated with a custom lexical resource.

Evaluations demonstrate GeNRe achieves state-of-the-art performance in French gender-neutral rewriting. Notably, the rule-based system and fine-tuned models exhibited comparable accuracy. The LLM-based approach, leveraging Claude 3 Opus and the bespoke dictionary, also yielded promising results.

However, consistent error patterns emerged during testing. Models frequently exhibited incorrect number agreement – the grammatical consistency between subjects and verbs in terms of singularity and plurality. Specifically, a tendency to convert plural verb forms to singular was observed. This suggests a difficulty in accurately interpreting the inherent plurality implied by collective nouns and maintaining grammatical coherence across extended sentence structures.

Addressing this limitation represents the primary focus for future work. Researchers propose several avenues for improvement:

• Architectural Refinement: Exploration of advanced neural network architectures, including transformers and recurrent neural networks incorporating attention mechanisms, to enhance the models’ capacity for capturing long-range dependencies.
• Enhanced Linguistic Analysis: Integration of external tools such as dependency parsers – which analyse the grammatical structure of sentences – and semantic role labelers – which identify the roles of different sentence elements – to provide more precise structural information.
• Lexical Expansion: Augmenting the custom dictionary to encompass a broader range of collective nouns and idiomatic expressions, thereby improving the system’s coverage and accuracy.
• Robustness Testing: Conducting comprehensive evaluations across diverse text genres and domains to assess the system’s generalizability and resilience to variations in language use.

This research constitutes a notable advancement in the field of gender-neutral language technologies for French. By concentrating on the specific challenges posed by collective nouns, GeNRe offers a pragmatic and effective approach to promoting more equitable and inclusive linguistic practices. The identified challenges and proposed future work delineate clear pathways for continued innovation in this increasingly important area.