Gender-Inclusive Language in Bias Measurement and Mitigation for Large Language Models

Research on gender bias in Large Language Models (LLMs) echoes the aims of feminist language reform to mitigate sexism and gender stereotyping in newly generated language.
Feminist linguistics has long highlighted how linguistic structures in English support a male-centred and heteronormative worldview. To address this, gender-inclusive language strategies such as the replacement of masculine generic nouns and pronouns with gender-neutral alternatives were developed. However, research in Natural Language Processing (NLP) has rarely incorporated methodologies from feminist linguistics. The resulting research gaps include a limited scope of resources capturing gender(-stereotyping) linguistically, unexplored potential of transferring equality-promoting effects of gender-inclusive language to LLMs, and a lack of understanding of LLMs' processing of gender-inclusive text. This thesis integrates the two fields of feminist linguistics and NLP. We investigate the potential of training and fine-tuning models with gender-inclusive language for reducing bias in language technology. An affix-based strategy was developed to identify and neutralise gender-specific wording in English text. Psycholinguistic research was adapted for LLMs to assess the processing of gender-neutral language in a coreference context. This method was also adapted to evaluate German gender-inclusive language strategies. The findings suggest that using gender-neutral text in training/fine-tuning data and prompts presents a promising approach for, respectively, mitigating gender-stereotyping and reducing masculine biases in language models.