Despite concerns that Large Language Models (LLMs) are vectors for reproducing and amplifying social biases such as sexism, transphobia, islamophobia, and racism, there is a lack of work qualitatively analyzing how such patterns of bias are generated by LLMs. We use mixed-methods approaches and apply a feminist, intersectional lens to the problem across two language domains, Swedish and English, by generating narrative texts using LLMs. We find that hegemonic norms are consistently reproduced; dominant identities are often treated as ‘default’; and discussion of identity itself may be considered ‘inappropriate’ by the safety features applied to some LLMs. Due to the differing behaviors of models, depending both on their design and the language they are trained on, we observe that strategies of identifying “bias” must be adapted to individual models and their socio-cultural contexts._Content warning: This research concerns the identification of harms, including stereotyping, denigration, and erasure of minoritized groups. Examples, including transphobic and racist content, are included and discussed._

Part of Speech (POS) taggers for Swedish routinely fail for the third person gender-neutral pronoun “hen”, despite the fact that it has been a well-established part of the Swedish language since at least 2014. In addition to simply being a form of gender bias, this failure can have negative effects on other tasks relying on POS information. We demonstrate the usefulness of semi-synthetic augmented datasets in a case study, retraining a POS tagger to correctly recognize “hen” as a personal pronoun. We evaluate our retrained models for both tag accuracy and on a downstream task (dependency parsing) in a classicial NLP pipeline. Our results show that adding such data works to correct for the disparity in performance. The accuracy rate for identifying “hen” as a pronoun can be brought up to acceptable levels with only minor adjustments to the tagger’s vocabulary files. Performance parity to gendered pronouns can be reached after retraining with only a few hundred examples. This increase in POS tag accuracy also results in improvements for dependency parsing sentences containing hen.

Gender bias has been identified in many models for Natural Language Processing, stemming from implicit biases in the text corpora used to train the models. Such corpora are too large to closely analyze for biased or stereotypical content. Thus, we argue for a combination of quantitative and qualitative methods, where the quantitative part produces a view of the data of a size suitable for qualitative analysis. We investigate the usefulness of semi-supervised topic modeling for the detection and analysis of gender bias in three corpora (mainstream news articles in English and Swedish, and LGBTQ+ web content in English). We compare differences in topic models for three gender categories (masculine, feminine, and nonbinary or neutral) in each corpus. We find that in all corpora, genders are treated differently and that these differences tend to correspond to hegemonic ideas of gender.