Multilingual Holistic Bias: Extending Descriptors and Patterns to Unveil Demographic Biases in Languages at Scale
Marta R. Costa-jussà, Pierre Andrews, Eric Michael Smith, Prangthip Hansanti, Christophe Ropers, Elahe Kalbassi, Cynthia Gao, Daniel Licht, Carleigh Wood
摘要
We introduce a multilingual extension of the HolisticBias dataset, the largest English template-based taxonomy of textual people references: Multilingual HolisticBias. This extension consists of 20,459 sentences in 50 languages distributed across 13 demographic axes. Source sentences are built from combinations of 118 demographic descriptors and three patterns, excluding nonsensical combinations. Multilingual translations include alternatives for gendered languages that cover gendered translations when there is ambiguity in English. Our dataset is intended to uncover demographic imbalances and be the tool to quantify mitigations towards them. Our initial findings show that translation quality for EN-to-XX translations is an average of almost 8 spBLEU better when evaluating with the masculine human reference compared to feminine. In the opposite direction, XX-to-EN, we compare the robustness of the model when the source input only differs in gender (masculine or feminine) and masculine translations are an average of almost 4 spBLEU better than feminine. When embedding sentences to a joint multilingual sentence representations space, we find that for most languages masculine translations are significantly closer to the English neutral sentences when embedded. WARNING: the current paper contains examples that may be offensive. Sentence 1: eng_Latn I'm an alcoholic. </s> </s> spa_Latn Soy un alcohólico. I ' m an alco holic . </s> <s> Masculine Ref: Yo soy un alcohólico. Feminine Ref: Yo soy una alcohólica. Sentence 2: eng_Latn I have friends who are homem ak ers . </s> <s> spa_Latn Tengo amigos que son amas de casa . Masculine Ref: Tengo amigos que son amos de casa. Feminine Ref: Tengo amigas que son amas de casa. Sentence 3: eng_Latn I ' m a homem aker . </s> <s> spa_Latn Soy una ama de casa . Masculine Ref: Soy amo de casa. Feminine Ref: Soy ama de casa. Sentence 4: spa_Latn Soy amo de casa </s> <s> eng_Latn I'm the master of the house. spa_Latn Soy ama de casa. </s> <s> eng_Latn I'm a housewife.
问问这篇 Paper
智能体会读完全文。
Lune 把这篇 Paper 索引到了最后一个公式,引用它的顶会 Paper 也一样。你提问,回答直接引用原文。
引用它的顶会 Paper5
- Metrics for What, Metrics for Whom: Assessing Actionability of Bias Evaluation Metrics in NLPPieter Delobelle, Giuseppe Attanasio, Debora Nozza, Su Lin Blodgett 等EMNLP 2024 · 被引用 4 次
- The State of Multilingual LLM Safety Research: From Measuring The Language Gap To Mitigating ItZheng Xin Yong, Beyza Ermis, Marzieh Fadaee, Stephen H. Bach 等EMNLP 2025 · 被引用 2 次
- What the Harm? Quantifying the Tangible Impact of Gender Bias in Machine Translation with a Human-centered StudyBeatrice Savoldi, Sara Papi, Matteo Negri, Ana Guerberof Arenas 等EMNLP 2024 · 被引用 1 次
- Measuring Social Bias in Vision-Language Models with Face-Only Counterfactuals from Real PhotosHaodong Chen, Qiang Huang, Jiaqi Zhao, Qiuping Jiang 等ACL 2026 · 被引用 1 次
- Social Bias in Multilingual Language Models: A SurveyLance Calvin Lim Gamboa, Yue Feng, Mark G. LeeEMNLP 2025
它引用的顶会 Paper6
- Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah 等NeurIPS 2020 · 被引用 64,255 次
- "I'm sorry to hear that": Finding New Biases in Language Models with a Holistic Descriptor DatasetEric Michael Smith, Melissa Hall, Melanie Kambadur, Eleonora Presani 等EMNLP 2022 · 被引用 56 次
- Perturbation Augmentation for Fairer NLPRebecca Qian, Candace Ross, Jude Fernandes, Eric Michael Smith 等EMNLP 2022 · 被引用 54 次
- Investigating Failures of Automatic Translationin the Case of Unambiguous GenderAdi Renduchintala, Adina WilliamsACL 2022 · 被引用 28 次
- Towards Opening the Black Box of Neural Machine Translation: Source and Target Interpretations of the TransformerJavier Ferrando, Gerard I. Gállego, Belen Alastruey, Carlos Escolano 等EMNLP 2022 · 被引用 18 次
相关 Paper
- Gender Bias in Multilingual Embeddings and Cross-Lingual TransferJieyu Zhao, Subhabrata Mukherjee, Saghar Hosseini, Kai-Wei Chang 等ACL 2020 · 被引用 59 次
- Measuring and Mitigating Name Biases in Neural Machine TranslationJun Wang, Benjamin I. P. Rubinstein, Trevor CohnACL 2022 · 被引用 31 次
- Who is better at math, Jenny or Jingzhen? Uncovering Stereotypes in Large Language ModelsZara Siddique, Liam D. Turner, Luis Espinosa AnkeEMNLP 2024 · 被引用 2 次
- A Rose by Any Other Name would not Smell as Sweet: Social Bias in Names MistranslationSandra Sandoval, Jieyu Zhao, Marine Carpuat, Hal Daumé IIIEMNLP 2023 · 被引用 2 次
- EuroGEST: Investigating gender stereotypes in multilingual language modelsJacqueline Rowe, Mateusz Klimaszewski, Liane Guillou, Shannon Vallor 等EMNLP 2025
