MBBQ: A Dataset for Cross-Lingual Comparison of Stereotypes in Generative LLMs

Vera Neplenbroek; Arianna Bisazza; Raquel Fernández

cs.CLJun 11, 2024

MBBQ: A Dataset for Cross-Lingual Comparison of Stereotypes in Generative LLMs

Authors: Vera Neplenbroek, Arianna Bisazza, Raquel Fernández

Organizations: Institute for Logic, Language and Computation, University of Amsterdam · Center for Language and Cognition, University of Groningen

Abstract

Generative large language models (LLMs) have been shown to exhibit harmful biases and stereotypes. While safety fine-tuning typically takes place in English, if at all, these models are being used by speakers of many different languages. There is existing evidence that the performance of these models is inconsistent across languages and that they discriminate based on demographic factors of the user. Motivated by this, we investigate whether the social stereotypes exhibited by LLMs differ as a function of the language used to prompt them, while controlling for cultural differences and task accuracy. To this end, we present MBBQ (Multilingual Bias Benchmark for Question-answering), a carefully curated version of the English BBQ dataset extended to Dutch, Spanish, and Turkish, which measures stereotypes commonly held across these languages. We further complement MBBQ with a parallel control dataset to measure task performance on the question-answering task independently of bias. Our results based on several open-source and proprietary LLMs confirm that some non-English languages suffer from bias more than English, even when controlling for cultural shifts. Moreover, we observe significant cross-lingual differences in bias behaviour for all except the most accurate models. With the release of MBBQ, we hope to encourage further research on bias in multilingual settings. The dataset and code are available at https://github.com/Veranep/MBBQ.

Figures & tables

Figure 1: Example from the Gender identity category in English and Spanish, plus the control-MBBQ counterpart of the English sample. In ambiguous contexts, the correct answer is always the “unknown” answer. In biased contexts, the correct answer to this question is the biased answer, and in counter-biased contexts this is the counter-biased answer.

Subset $\rightarrow$ Language $\downarrow$	Age	Disability status	Gender identity	Nationality	Physical appearance	Race	Religion	SES	Sexual orientation	Total
English (BBQ)	25	25	50	25	25	100	25	25	25	325
Dutch	24	25	25	-	23	-	-	24	25	146
Spanish	22	24	25	-	23	-	-	24	25	143
Turkish	23	18	24	-	16	-	-	12	6	99
MBBQ #Templates #Samples	22	18	24	-	16	-	-	12	6	98
MBBQ #Templates #Samples	3,320	1,296	528	-	1,176	-	-	3,600	152	10,072

Table 1: Number of templates about relevant stereotypes per language and bias category. The last two rows show the total number of templates and samples in MBBQ. These correspond to the intersection of those pertaining to the stereotypes relevant in the cultural contexts of English, Dutch, European Spanish, and Turkish. SES stands for Socio-economic status.

Figure 2: Percentage of templates in control-MBBQ where models perform above chance (33%) and accuracy on those templates. Models sorted by their accuracy in English.

Model	Acc D	Acc A	Overall
GPT3.5	$86.4\pm 5.83$	$83.4\pm 4.44$	$84.9\pm 5.06$
Mistral	$61.1\pm 13.2$	$78.4\pm 8.68$	$69.8\pm 13.9$
Zephyr	$61.4\pm 16.6$	$43.7\pm 9.20$	$52.6\pm 15.6$
Aya	$90.5\pm 3.20$	$18.0\pm 4.31$	$54.2\pm 38.9$
Llama	$40.9\pm 2.18$	$34.4\pm 9.07$	$37.7\pm 7.03$
Wizard	$40.6\pm 2.11$	$28.7\pm 2.29$	$34.6\pm 6.70$

Table 2: Average accuracy across languages on control-MBBQ for questions with disambiguated and ambiguous contexts. Chance accuracy is 33% in all cases.

Model	Aya				GPT 3.5				Llama
#T	94				89				67
Lang.	E	D	S	T	E	D	S	T	E	D	S	T
Acc D	94.3	91.8	91.3	85.5	87.5	85.9	82.4	73.8	36.8	39.3	43.0	38.5
Bias D	0.0050	-0.0017	0.0052	-0.0004	-0.0011	0.0002	-0.0074	0.0010	0.0119*	0.0195*	0.0294*	0.0097
Acc A	18.1	10.5	8.7	11.8	84.2	82.1	83.9	74.6	58.2	39.4	35.3	30.0
Bias A	0.0356*	0.0438*	0.1088*	0.0531*	-0.0107	0.0035	0.0080	0.0167*	0.0259*	0.0262*	0.0329*	0.0245

Table 3: Accuracy and bias scores in disambiguated and ambiguous contexts. #T refers to the number of templates and E, D, S, and T stand for English, Dutch, Spanish, and Turkish respectively. An asterisk (*) indicates that the bias score is significantly different from 0, and bold red means there is a significant difference across languages (

p<0.05

on the Kruskal-Wallis

H

-test for independent samples).

Figure 3: Bias scores in ambiguous contexts per subset. Bias scores that are significantly different from 0 ( $p<0.05$ ) are marked with a star ( $\star$ ).

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Dutch		English		Spanish		Turkish
Male	Female	Male	Female	Male	Female	Male	Female
Noah	Emma	Liam	Olivia	Martín	Lucía	Alparslan	Zeynep
Liam	Julia	Noah	Emma	Mateo	Sofía	Yusuf	Asel
Luca	Mila	Oliver	Charlotte	Hugo	Martina	Miraç	Defne
Lucas	Sophie	James	Amelia	Leo	Valeria	Göktuğ	Zümra
Mees	Olivia	Elijah	Sophia	Lucas	María	Ömer	Elif

Appendix

Table 4: Names used in the control data set.

Model	Language	Acc D	Acc A
	English	94.0	24.4
	Dutch	90.9	16.3
	Spanish	90.7	15.4
Aya	Turkish	86.2	15.9
	English	38.6	28.5
	Dutch	17.8	13.6

Appendix

Table 5: Accuracy on control-MBBQ in disambiguated and ambiguous contexts, red means there is a significant difference across languages (

p<0.05

on the Kruskal-Wallis H-test for independent samples).

Figure 4: Bias scores in disambiguated contexts per subset. Bias scores that are significantly different from 0 ( $p<0.05$ ) are marked with a star (*).

Social Bias in Language Models

Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.

See all 192 papers

StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs
May 11, 2026Pierre Le Jeune, Étienne Duchesne, Weixuan Xiao +4Open-Ended Generation Multilingual Language Model Evaluation
Giskard AI · École Normale Supérieure, INSERM, Paris
Scalable and Culturally Specific Stereotype Dataset Construction via Human-LLM Collaboration
Jul 8, 2026Weicheng Ma, John Guerrerio, Soroush VosoughiMultilingual Language Model Evaluation Social Bias in Language Models
College of Computing, Georgia Institute of Technology · Computer Science Department, Dartmouth College
Anchoring LLM Gender Bias to Human Baselines: A Cross-Lingual Audit
May 29, 2026Jiwoo Choi, Seonwoo Ahn, Tongxin Zhang +1Gender Bias in Language Models Multilingual Language Model Evaluation
School of Digital Humanities and Computational Social Sciences, KAIST, South Korea

May 11, 2026J/K move · Enter open · S save

StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs

Pierre Le Jeune, Étienne Duchesne, Weixuan Xiao, Stefano Palminteri+3

Giskard AI · École Normale Supérieure, INSERM, Paris

Open-Ended Generation Multilingual Language Model Evaluation Language Model Generation Evaluation Social Bias in Language Models Cultural Bias in Language Models Language Model Bias Evaluation

Open paper AskPDF arXiv

Multilingual studies of social bias in open-ended LLM generation remain limited: most existing benchmarks are English-centric, template-based, or restricted to recognizing pre-specified stereotypes. We introduce StereoTales, a multilingual dataset and evaluation pipeline for systematically studying the emergence of social bias in open-ended LLM generation. The dataset covers 10 languages and 79 socio-demographic attributes, and comprises over 650k stories generated by 23 recent LLMs, each annotated with the socio-demographic profile of the protagonist across 19 dimensions. From these, we apply statistical tests to identify more than 1{,}500 over-represented associations, which we then rate for harmfulness through both a panel of humans (N = 247) and the same LLMs. We report three main findings. (i) Every model we evaluate emits consequential harmful stereotypes in open-ended generation, regardless of size or capabilities, and these associations are largely shared across providers rather than isolated misbehaviors. (ii) Prompt language strongly shapes which stereotypes appear: rather than transferring as a shared set of biases, harmful associations adapt culturally to the prompt language and amplify bias against locally salient protected groups. (iii) Human and LLM harmfulness judgments are broadly aligned (Spearman

ρ=0.62

), with disagreements concentrating on specific attribute classes rather than specific providers. To support further analyses, we release the evaluation code and the dataset, including model generations, attribute annotations, and harmfulness ratings.

Abstract

Social Bias in Language Models

Explore similar work

StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs

Scalable and Culturally Specific Stereotype Dataset Construction via Human-LLM Collaboration

Anchoring LLM Gender Bias to Human Baselines: A Cross-Lingual Audit