How Conservative are Language Models? Adapting to the Introduction of Gender-Neutral Pronouns

Brandl, Stephanie; Cui, Ruixiang; Søgaard, Anders

doi:10.18653/v1/2022.naacl-main.265

Cited by 8 publications

(8 citation statements)

References 25 publications

(31 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Sun et al (2021) and Vanmassenhove et al (2021) present rule-based and neural rewriting approaches to generate gender-neutral alternatives in English texts. Brandl et al (2022) find that upstream perplexity substantially increases and downstream task performance severely drops for some tasks when genderneutral language is used in English, Danish and Swedish. Amend et al (2021) show that the substitution of gendered for gender-neutral terms on image captioning models poses a viable approach for reducing gender bias.…”

Section: Bias In Vision and Languagementioning

confidence: 85%

Evaluating Bias and Fairness in Gender-Neutral Pretrained Vision-and-Language Models

Cabello,

Bugliarello,

Brandl

et al. 2023

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

View full text Add to dashboard Cite

Pretrained machine learning models are known to perpetuate and even amplify existing biases in data, which can result in unfair outcomes that ultimately impact user experience. Therefore, it is crucial to understand the mechanisms behind those prejudicial biases to ensure that model performance does not result in discriminatory behaviour toward certain groups or populations. In this work, we define gender bias as our case study. We quantify bias amplification in pretraining and after fine-tuning on three families of vision-and-language models. We investigate the connection, if any, between the two learning stages, and evaluate how bias amplification reflects on model performance. Overall, we find that bias amplification in pretraining and after fine-tuning are independent. We then examine the effect of continued pretraining on gender-neutral data, finding that this reduces group disparities, i.e., promotes fairness, on VQAv2 and retrieval tasks without significantly compromising task performance.

show abstract

Section: Bias In Vision and Languagementioning

confidence: 85%

Evaluating Bias and Fairness in Gender-Neutral Pretrained Vision-and-Language Models

Cabello,

Bugliarello,

Brandl

et al. 2023

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

View full text Add to dashboard Cite

show abstract

“…Some works have recently expressed the cruciality of dealing with non-binary identities in Natural Language Processing (NLP) (Cao and Daumé III, 2020;Dev et al, 2021), and two main approaches can be found among the works that have taken this route. Brandl, Cui and Søgaard (2022) focused on neopronouns, showing that language models have difficulties in processing them in Swedish (hen), Danish (de/høn) and English (they/xe). Others focused on standard neutral solutions, for both text classification (Attanasio et al, 2021) and natural language generation tasks, such as genderneutral rewriting (Sun et al, 2021;Vanmassenhove et al, 2021;Attanasio et al, 2021), which consists in converting gendered forms into their gender-neutral counterparts (e.g., En.…”

Section: Gender (Bias) and Machine Translationmentioning

confidence: 99%

From Inclusive Language to Gender-Neutral Machine Translation

Piergentili¹,

Fucci²,

Savoldi³

et al. 2023

Preprint

View full text Add to dashboard Cite

Gender inclusivity in language has become a central topic of debate and research. Its application in the cross-lingual contexts of human and machine translation (MT), however, remains largely unexplored. Here, we discuss Gender-Neutral Translation (GNT) as a form of gender inclusivity in translation and advocate for its adoption for MT models, which have been found to perpetuate gender bias and discrimination. To this aim, we review a selection of relevant institutional guidelines for Gender-Inclusive Language (GIL) to collect and systematize useful strategies of gender neutralization. Then, we discuss GNT and its scenarios of use, devising a list of desiderata. Finally, we identify the main technical challenges to the implementation of GNT in MT. Throughout these contributions we focus on translation from English into Italian, as representative of salient linguistic transfer problems, due to the different rules for gender marking in their grammar. 1 See https://www.cbsnews.com/amp/news/ teacher-jailed-for-contempt-of-courtrefusing-student-gender-neutralpronouns-ireland/.

show abstract

“…Existing work on noncisgender identities and machine learning is sparse (e.g., Dev et al, 2021;Cao and Daumé III, 2020;Lauscher et al, 2022). However recently, there have been a couple of works dealing with genderneutral pronouns (e.g., Brandl et al, 2022;Qian et al, 2022). As such, work by Lauscher et al (2022) explores the diversity of gender pronouns and presents five desiderata for how language models should handle (gender-neutral) pronouns.…”

Section: Related Workmentioning

confidence: 99%

“…In a similar vein, we explore potential solutions for how text-to-image models should handle non-cisgender identities. Brandl et al (2022) investigate the effect of gender-neutral pronouns on language models and demonstrate drops in performance in natural language inference. As a potential solution, Qian et al (2022) propose a perturber model for augmenting data sets which they train on texts that have been rewritten in a gender-neutral way.…”

Section: Related Workmentioning

confidence: 99%

Stereotypes and Smut: The (Mis)representation of Non-cisgender Identities by Text-to-Image Models

L.¹,

Roß²,

Lauscher³

2023

Findings of the Association for Computational Linguistics: ACL 2023

View full text Add to dashboard Cite

Warning: many of the images displayed in this paper are offensive and contain implied nudity.They are intended to illustrate potential harms.Cutting-edge image generation has been praised for producing high-quality images, suggesting a ubiquitous future in a variety of applications. However, initial studies have pointed to the potential for harm due to predictive bias, reflecting and potentially reinforcing cultural stereotypes. In this work, we are the first to investigate how multimodal models handle diverse gender identities. Concretely, we conduct a thorough analysis in which we compare the output of three image generation models for prompts containing cisgender vs. noncisgender identity terms. Our findings demonstrate that certain non-cisgender identities are consistently (mis)represented as less human, more stereotyped and more sexualised. We complement our experimental analysis with (a) a survey among non-cisgender individuals and (b) a series of interviews, to establish which harms affected individuals anticipate, and how they would like to be represented. We find respondents are particularly concerned about misrepresentation, and the potential to drive harmful behaviours and beliefs. Simple heuristics to limit offensive content are widely rejected, and instead respondents call for community involvement, curated training data and the ability to customise. These improvements could pave the way for a future where change is led by the affected community, and technology is used to positively "[portray] queerness in ways that we haven't even thought of" rather than reproducing stale, offensive stereotypes.

show abstract

How Conservative are Language Models? Adapting to the Introduction of Gender-Neutral Pronouns

Cited by 8 publications

References 25 publications

Evaluating Bias and Fairness in Gender-Neutral Pretrained Vision-and-Language Models

Evaluating Bias and Fairness in Gender-Neutral Pretrained Vision-and-Language Models

From Inclusive Language to Gender-Neutral Machine Translation

Stereotypes and Smut: The (Mis)representation of Non-cisgender Identities by Text-to-Image Models

Contact Info

Product

Resources

About