Austria

We examine audience-specific how-to guides on wikiHow, in English, diachronically by comparing predictions from fine-tuned language models and human judgments. Using both early and revised versions, we quantitatively and qualitatively study how gender-specific features are identified over time. While language model performance remains relatively stable in terms of macro F$_1$-scores, we observe an increased reliance on stereotypical tokens. Notably, both models and human raters tend to overpredict women as an audience, raising questions about bias in the evaluation of educational systems and resources.

ACL 2025

A Diachronic Analysis of Human and Model Predictions on Audience Gender in How-to Guides

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

workshop paper

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Machine learning (ML) models are increasingly used to support clinical decision-making. However, real-world medical datasets are often noisy, incomplete, and imbalanced, leading to performance disparities across patient subgroups. These differences raise fairness concerns, particularly when they reinforce existing disadvantages for marginalized groups. In this work, we analyze several medical prediction tasks and demonstrate how model performance varies with patient characteristics. While ML models may demonstrate good overall performance, we argue that subgroup-level evaluation is essential before integrating them into clinical workflows. By conducting a performance analysis at the subgroup level, differences can be clearly identified—allowing, on the one hand, for performance disparities to be considered in clinical practice, and on the other hand, for these insights to inform the responsible development of more effective models. Thereby, our work contributes to a practical discussion around the subgroup-sensitive development and deployment of medical ML models and the interconnectedness of fairness and transparency.

One Size Fits None: Rethinking Fairness in Medical AI

The advancement of Large Language Models (LLMs) has transformed Natural Language Processing (NLP), enabling performance across diverse tasks with little task-specific training. However, LLMs remain susceptible to social biases, particularly reflecting harmful stereotypes from training data, which can disproportionately affect marginalised communities.
We measure gender bias in Maltese LMs, arguing that such bias is harmful as it reinforces societal stereotypes and fails to account for gender diversity, which is especially problematic in gendered, low-resource languages.
While bias evaluation and mitigation efforts have progressed for English-centric models, research on low-resourced and morphologically rich languages remains limited. This research investigates the transferability of debiasing methods to Maltese language models, focusing on BERTu and mBERTu, BERT-based monolingual and multilingual models respectively. Bias measurement and mitigation techniques from English are adapted to Maltese, using benchmarks such as CrowS-Pairs and SEAT, alongside debiasing methods Counterfactual Data Augmentation, Dropout Regularization, Auto-Debias, and GuiDebias. We also contribute to future work in the study of gender bias in Maltese by creating evaluation datasets.
Our findings highlight the challenges of applying existing bias mitigation methods to linguistically complex languages, underscoring the need for more inclusive approaches in the development of multilingual NLP.

From Measurement to Mitigation: Exploring the Transferability of Debiasing Approaches to Gender Bias in Maltese Language Models

Contributing to research on gender beyond the binary, this work introduces GENDEROUS, a dataset of gender-ambiguous sentences containing gender-marked occupations and adjectives, and sentences with the ambiguous or non-binary pronoun their. We cross-linguistically evaluate how machine translation (MT) systems and large language models (LLMs) translate these sentences from English into four grammatical gender languages: Greek, German, Spanish and Dutch. We show the systems’ continued default to male-gendered translations, with exceptions (particularly for Dutch). Prompting for alternatives, however, shows potential in attaining more diverse and neutral translations across all languages. An LLM-as-a-judge approach was implemented, where benchmarking against gold standards emphasises the continued need for human annotations.

GENDEROUS: Machine Translation and Cross-Linguistic Evaluation of a Gender-Ambiguous Dataset

Increasing attention is being dedicated by the NLP community to gender-fair practices, including emerging forms of non-binary language. Given the shift to the prompting paradigm for multiple tasks, direct comparisons between prompted and fine-tuned models in this context are lacking. We aim to fill this gap by comparing prompt engineering and fine-tuning techniques for gender-fair rewriting in Italian. We do so by framing a rewriting task where Italian gender-marked translations from English gender-ambiguous sentences are adapted into a gender-neutral alternative using direct non-binary language. We augment existing datasets with gender-neutral translations and conduct experiments to determine the best architecture and approach to complete such task, by fine-tuning and prompting seq2seq encoder-decoder and autoregressive decoder-only models. We show that smaller seq2seq models can reach good performance when fine-tuned, even with relatively little data; when it comes to prompts, including task demonstrations is crucial, and we find that chat-tuned models reach the best results in a few-shot setting. We achieve promising results, especially in contexts of limited data and resources.

Fine-Tuning vs Prompting Techniques for Gender-Fair Rewriting of Machine Translations

This paper critiques common assumptions about gender bias in NLP, focusing primarily on word vector-based methods for detecting and mitigating bias. It argues that these methods assume a kind of "binary thinking" that goes beyond the gender binary toward a conceptual model that structures and limits the effectiveness of these techniques. Drawing its critique from the Humanities field of Queer Studies, this paper demonstrates that binary thinking drives two "myths" in gender bias research: first, that bias is categorical, measuring bias in terms of presence/absence, and second, that it is zero-sum, where the relations between genders are idealized as symmetrical. Due to their use of binary thinking, each of these myths flattens bias into a measure that cannot distinguish between the types of bias and their effects in language. The paper concludes by briefly pointing to methods that resist binary thinking, such as those that diversify and amplify gender expressions.

Some Myths About Bias: A Queer Studies Reading Of Gender Bias In NLP

Gendered language is the use of words that indicate an individual's gender. Though useful in certain context, it can reinforce gender stereotypes and introduce bias, particularly in machine learning models used for tasks like occupation classification. When textual content such as biographies contains gender cues, it can influence model predictions, leading to unfair outcomes such as reduced hiring opportunities for women. To address this issue, we propose GenWriter, an approach that integrates Case-Based Reasoning (CBR) with Large Language Models (LLMs) to rewrite biographies in a way that obfuscates gender while preserving semantic content. We evaluate GenWriter by measuring gender bias in occupation classification before and after rewriting the biographies used for training the occupation classification model. Our results show that GenWriter significantly reduces gender bias by 89% in nurse biographies and 62% in surgeon biographies, while maintaining classification accuracy. In comparison, an LLM-only rewriting approach achieves smaller bias reductions (by 44% and 12% in nurse and surgeon biographies, respectively) and leads to some classification performance degradation.

GenWriter: Reducing Gender Cues in Biographies through Text Rewriting

We investigate biases related to ageism, ableism, nationality, and gender in four Norwegian and two multilingual language models. Our methodology involves using a set of templates. constructed around stimuli and attributes relevant to these categories. We use statistical and predictive evaluation methods, including Kendall’s Tau correlation and dependent variable prediction rates, to assess model behaviour and output bias. Our findings indicate that models frequently associate older individuals, people with disabilities, and poorer countries with negative attributes, potentially reinforcing harmful stereotypes. However, most tested models appear to handle gender-related biases more effectively. Our findings indicate a correlation between the sentiment of the input and that of the output.

Ableism, Ageism, Gender, and Nationality bias in Norwegian and Multilingual Language Models

Natural language processing (NLP) systems often inadvertently encode and amplify social biases through entangled representations of demographic attributes and task-related attributes. To mitigate this, we propose a novel framework that combines causal analysis with practical intervention strategies. The method leverages attribute-specific prompting to isolate sensitive attributes while applying information-theoretic constraints to minimize spurious correlations. Experiments across six language models and two classification tasks demonstrate its effectiveness. We hope this work will provide the NLP community with a causal disentanglement perspective for achieving fairness in NLP systems.

Disentangling Biased Representations: A Causal Intervention Framework for Fairer NLP Models

This paper investigates human gender bias and its relation to bias in machine translation (MT), focussing on the role of context in gender interpretation. We measured human implicit gender bias and conducted an annotation study in context, followed by a linguistic and computational analysis to compare human gender perceptions among themselves and with machine translations.

Gender Bias and the Role of Context in Human Perception and Machine Translation

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.

Content not yet available

Downloads

Next from ACL 2025

One Size Fits None: Rethinking Fairness in Medical AI

Similar lecture

On Shortcuts and Biases: How Finetuned Language Models Distinguish Audience-Specific Instructions in Italian and English

Content not yet available

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2025

One Size Fits None: Rethinking Fairness in Medical AI

Similar lecture

On Shortcuts and Biases: How Finetuned Language Models Distinguish Audience-Specific Instructions in Italian and English

Downloads