Source-linked AI summary
Mitigating Language-Dependent Ethnic Bias in BERT
Jaimeen Ahn, Alice Oh
TL;DR
Ethnic bias in language models has received less in-depth study, although it varies with cultural context. This paper quantifies it and tests two mitigation approaches, finding that the better-fitting approach depends on language-resource availability.
Problem
Ethnic bias in language models has received less in-depth investigation than gender and racial bias, despite its cultural-context dependence.
Source: §1 · p. 1Method
The paper introduces Categorical Bias score and tests mitigation with a multilingual model and contextual alignment between monolingual models.
Source: §1 · pp. 1–2Results
Takeaways & Limitations
Mitigation choice depends on language resources rather than one route being best for every language setting.
Source: §1 · p. 2The manually crafted target and attribute lists omit some ethnicities and attributes, and larger, more systematic lists remain future work.
Source: §6.2 · p. 7
Abstract
from arXiv · showhide
BERT and other large-scale language models (LMs) contain gender and racial bias. They also exhibit other dimensions of social bias, most of which have not been studied in depth, and some of which vary depending on the language. In this paper, we study ethnic bias and how it varies across languages by analyzing and mitigating ethnic bias in monolingual BERT for English, German, Spanish, Korean, Turkish, and Chinese. To observe and quantify ethnic bias, we develop a novel metric called Categorical Bias score. Then we propose two methods for mitigation; first using a multilingual model, and second using contextual word alignment of two monolingual models. We compare our proposed methods with monolingual BERT and show that these methods effectively alleviate the ethnic bias. Which of the two methods works better depends on the amount of NLP resources available for that language. We additionally experiment with Arabic and Greek to verify that our proposed methods work for a wider variety of languages.
1 Introduction
The paper addresses under-studied ethnic bias in language models and examines how it varies by language. It introduces Categorical Bias score and compares mitigation approaches whose reported effectiveness differs with language-resource availability.
1 Introduction
- Ethnic bias has received less in-depth study than gender and racial bias, and BERT predictions differ across English, German, and Korean.Source: §1 · p. 1
- Multilingual BERT mitigates bias in Chinese, English, German, and Spanish, while contextual alignment reduces bias in low-resource Korean and Turkish.The paper also presents contextual word alignment with fine-tuning as its second mitigation method.Source: §1 · p. 2
- Experiments across six languages report mitigation, an ablation study, no downstream-task performance drop, and validation in Arabic and Greek.Source: §1 · p. 2
- The paper introduces Categorical Bias (CB) score, a multi-class bias measure using log probability, and reports that ethnic bias varies across languages.Source: §1 · p. 2
2 Ethnic Bias
The section describes how ethnic bias can overlap with national bias and presents toxicity classification as an example of bias in both training data and model predictions.
2 Ethnic Bias
- The study groups ethnic and national bias and limits its analysis to coarse-grained groups because fine-grained ethnicities are poorly represented in language-model corpora.The text gives Navajo nation as an example of a fine-grained ethnicity.Source: §2 · p. 2
- Ethnic bias is prevalent in everyday-language datasets, and the authors conjecture that dataset bias can produce similar bias in models trained on those datasets.Source: §2 · p. 2
- Training data contains more toxic labels for comments mentioning Afghanistan, Iraq, or Iran, and the classifier has higher false-positive rates for those country names.Source: §2 · p. 2
3 Measuring Ethnic Bias
The section measures ethnic bias through variation in predicted ethnicity probabilities and extends a two-group bias metric to multi-class targets with the Categorical Bias score.
3.1 Normalized probability
- Figure 2 uses the difference in normalized probabilities for two target groups and their variance for more than two groups.Both metrics use normalized probabilities of target terms replacing the mask token.Source: Figure 2 · p. 3
- Normalized probability compares target-word probabilities with prior probabilities obtained when the attribute is masked, measuring association rather than word-occurrence likelihood.The score uses the change in probability given the presence or absence of an attribute word.Source: §3.1 · p. 3
3.2 Categorical Bias Score
- The Categorical Bias score is the variance of log normalized probabilities, generalizing the metric to multi-class ethnicity targets.With two templates, it is equivalent to the bias metric in Kurita et al. (2019).Source: §3.2 · p. 3
- For words split into multiple WordPiece tokens, the method multiplies the token probabilities to obtain each word’s probability.It adds as many mask tokens as the word has WordPiece tokens.Source: §3.2 · p. 3
- The CB score assumes no ethnicity word has a remarkably different normalized probability: uniform probabilities yield zero, while a markedly higher probability for one group yields a high score.Source: §3.2 · p. 3
4 Mitigation
The paper proposes multilingual BERT and contextual word alignment as mitigation approaches for language-dependent ethnic bias. Alignment maps a source representation into a lower-bias target space and requires an additional fine-tuning stage.
4 Mitigation
- The proposed mitigation approaches are multilingual BERT and contextual word alignment of two monolingual models.The authors seek techniques applicable across languages because ethnic bias varies by language.Source: §4 · pp. 3–4
4.1 Method 1: Multilingual BERT
- Multilingual BERT is intended to counterbalance monolingual ethnic biases by training multiple languages in one embedding space.The authors note a concern that M-BERT performance can degrade for relatively low-resource languages such as Korean and Turkish.Source: §4.1 · p. 4
4.2 Method 2: Contextual Word Alignment
- Contextual word alignment maps a source-language representation into the space of a language with a low CB score, using parallel-corpus anchors and a Procrustes approach.The alignment matrix is computed from contextual representations of the anchor tokens.Source: §4.2 · p. 4
- Unlike prior contextual alignment approaches, this method keeps the source-language MLM head and fine-tunes it on an additional source-language corpus while freezing BERT and the alignment matrix.The source embedding is moved into the target embedding space, so the MLM layer must be adapted to that space.Source: §4.2 · p. 4
5 Experiments
The experiments assess ethnic bias with translated sentence templates and compare BERT-based models and mitigation approaches across languages. They also describe downstream NER evaluation and a limitation on comparisons with embedding-space mitigation methods.
5 Experiments
- The template-based evaluation uses ten semantically equivalent singular and plural templates, thirty ethnicities, and seventy social-position attributes, translated and professionally revised for five languages.Redundant singular/plural templates may be excluded for languages such as Chinese.Source: §5 · p. 4
- The experiments use six monolingual base-uncased BERT baselines across English, German, Spanish, Korean, Turkish, and Chinese, with XNLI or KorNLI as alignment anchors.The paper reports BERT-based results in the main paper and places XLM results in Appendix C.Source: §5 · p. 4
- Counterfactual data augmentation is a baseline that replaces ethnicity terms in training data to balance their counts during fine-tuning.The described approach is one-sided CDA.Source: §5 · pp. 4–5
- The authors cannot compare their method with embedding-space bias mitigation approaches because their measurement is probability-based.Source: §5 · p. 5
- The proposed method is also evaluated on named-entity recognition for each language.Source: §5 · p. 5
6 Results & Discussion
Ethnic bias in monolingual BERT varies across languages, and the paper evaluates mitigation methods whose effectiveness depends on language resource availability. English alignment reduces bias across the studied monolingual models, while multilingual BERT with fine-tuning performs best for high-resource languages.
6.1 Language Dependency
- Ethnic-bias distributions differ across languages; English-German divergence is lower than English-Korean, though language is not a clean proxy for culture.Middle Eastern countries, especially Iraq, rank high in English, German, and Spanish, while Korean, Turkish, and Chinese distributions differ; English has the lowest CB score. English-language models may reflect multiple cultures, complicating culture-specific interpretation.Source: §6.1 · pp. 5–6, §6.2 · p. 8
6.2 Mitigation Result
- English alignment lowers CB scores across all five aligned languages; fine-tuned M-BERT generally performs best for high-resource languages, including Chinese.Alignment also beats the CDA baseline for Korean and Turkish, where M-BERT mitigation is not very effective.Source: Table 2 · p. 6, §6.2 · p. 7
- Both fine-tuning and proper English alignment contribute to mitigation; properly aligned models are mostly better than randomly initialized or unaligned models.Together, the ablation results support contextual word alignment as a mitigation approach across the monolingual models.Source: §6.2 · p. 7
- Aligning English to a higher-bias target language increases CB scores; among the tested variants, M-BERT trained with additional corpus performs best.Table 5 reports the reverse alignment direction.Source: §6.2 · p. 7, Table 5 · p. 7
- Aligned models have downstream-task performance comparable to original BERT under the same conditions, but remain below the best reported performance.With BERT frozen, the aligned-versus-unaligned performance difference is insignificant across all five languages.Source: §6.2 · p. 7, Table 6 · p. 7
- Adding five targets and five attributes changes overall CB scores slightly, while the alignment-related decrease remains; the original manually crafted lists omit some groups and attributes.The authors identify larger, more systematically constructed lists as future work.Source: §6.2 · p. 7, Table 7 · p. 7
- After contextual alignment, top-ranked ethnicity associations become more uniform in four of five languages, but the highest-probability ethnicity can change.Chinese is the exception to the reported distributional reduction; the paper gives Japan-to-China as an example of a rank change.Source: §6.2 · pp. 7–8, Figure 4 · p. 8
6.3 Additional Languages
- In Arabic and Greek, contextual word alignment outperforms the multilingual model in CB score, using templates and target/attribute lists translated without human revision.These additional-language experiments extend the mitigation evaluation beyond the six main languages.Source: §6.3 · p. 8, Table 8 · p. 8
7 Related Work
Prior bias measurement focused on word embeddings and later adapted association tests to language-model representations. This paper extends masked-token measurement to multi-class ethnic bias and proposes mitigation through transfer and model alignment.
How to measure semantic bias in NLP
- Earlier bias measures centered on word embeddings, including WEAT, before later work adapted association tests to sentence representations.WEAT was inspired by the Implicit Association Test.Source: §7 · p. 8
- The paper extends masked-token prediction measurement to multi-class bias, generalizing the Log Probability Bias Score.Kurita et al. (2019) introduced masked-token prediction to estimate bias in neural language models.Source: §7 · p. 9
How to mitigate semantic bias in NLP
- Prior mitigation approaches include bias-subspace subtraction, data augmentation, constrained retraining, and transfer learning.The section notes that constrained retraining comes with the difficulty of retraining.Source: §7 · p. 9
- The paper contributes a multi-class bias measure and mitigation methods based on multilingual fine-tuning or aligning monolingual models.The proposed approaches are presented as useful across multiple languages, with multilingual fine-tuning suited to high-resource languages and alignment to relatively low-resource ones.Source: §7 · pp. 8–9
8 Conclusion & Future Work
The paper measures language-dependent ethnic bias in BERT and proposes multilingual-model and contextual-alignment strategies, with the better fit depending on language resources. Its study remains limited in language and ethnicity coverage and in the granularity available from public monolingual models.
8 Conclusion & Future Work
- For resource-rich languages, multilingual models mitigate bias, while alignment with English is a better solution for low-resource languages.The conclusion also reports that alignment reduces bias for all languages.Source: §8 · p. 9
- The study addresses a gap in bias research by examining multiple languages, including relatively low-resource languages.Its comparisons use the same templates and attributes translated across languages.Source: §8 · p. 9
- The study omits some languages and detailed ethnicities because it relies on publicly available monolingual models and cannot use fine-grained ethnicity that may be under-represented.Future work is proposed using templates, attributes, and ethnic groups better suited to each language.Source: §8 · p. 9
9 Ethical Considerations
The authors report that their methods mitigate some ethnic bias and might help alleviate it in real-world applications, while emphasizing that models and measurements have important representational and scope limits. They also warn that reducing the overall bias score can shift which ethnicity has the highest probability.
Ethical Considerations
- The authors report that the proposed methods mitigate some ethnic bias and might help alleviate it in real-world applications.They caution that the methods cannot detect and remove all bias.Source: §9 · pp. 9–10
- A monolingual model does not represent all speakers or ethnic groups, and internet-based training data may reflect a skewed user population.The authors caution that model stereotypes do not establish that the language or its users are themselves biased.Source: §9 · p. 9
- Nation-level ethnic categories may be too broad to capture distinct cultural backgrounds or represent minorities within diverse nations.The authors used this broad scope to define groups that could occur across languages.Source: §9 · pp. 9–10
- Reducing the overall Categorical Bias score can change which ethnicity has the highest probability, such as Japan changing to Germany in Korean.The authors describe this as a possible side effect of the proposed method.Source: §9 · p. 10
- The study covers only a handful of languages, ethnicities, and attributes, so its measurement and mitigation cannot detect or remove all bias.The authors cite time and resource constraints and urge care in real-world language-model deployment.Source: §9 · p. 10
A Task Description
The task uses translated sentence templates and target and attribute terms to examine ethnic bias across languages. The appendix describes language-specific template adjustments and the term sets used in the experiments.
A.1 Templates
- Templates adapted from prior work were translated across languages, with Korean josa and Chinese singular-plural differences prompting adjustments.The authors chose templates over crowdsourcing because it was difficult to find experts for six languages, while noting crowdsourcing may offer better ecological validity.Source: §A.1 · p. 14
A.2 Targets and Attributes
- Experiments used 30 target terms and 70 attribute terms, translated into each language, with masculine forms used in gender-rich languages.The appendix also describes five additional English-selected targets and attributes translated using Google Translator.Source: §A.2 · pp. 14–15
B Detailed Setting
The section details the models and language-modeling and downstream-task settings, then reports that XLM is more biased than the proposed methods in five languages and that contextual alignment reduces JSD across five languages.
B Detailed Setting
- The study uses monolingual BERT models for English, German, Spanish, Korean, Chinese, Turkish, Arabic, and Greek plus a multilingual model, trains language models on five corpora, and evaluates NER on language-specific datasets.Language-modeling data are split train:validation:test at 0.64:0.16:0.2. German and Spanish unfrozen models train for 3 epochs, versus 100 and 300 epochs, respectively, when frozen.Source: B Detailed Setting · p. 15
C Another model type: XLM
- XLM’s Categorical Bias scores exceed those of the proposed mitigation methods in every language except Chinese.The authors report XLM scores of 8.95 (English), 12.72 (German), 9.97 (Spanish), 30.25 (Korean), 42.11 (Turkish), and 12.40 (Chinese).Source: C Another model type: XLM · p. 16
D Efficacy in terms of distance
- Alignment reduces the Jensen-Shannon Divergence (JSD) between each of five language models and English, while moving English farther from its original distribution.The authors conclude that contextual alignment changes each language’s distribution toward the target embedding space.Source: D Efficacy in terms of distance · p. 16
E Another Case Study
The case study compares ethnic-group predictions for equivalent pirate-related prompts across languages before and after mitigation. Somalia is often ranked first before mitigation, while mitigation generally flattens the distributions except in Chinese.
E Another Case Study
- Somalia ranks first for the pirate prompt in four of six languages and exceeds 40% in Korean and Spanish.The examples use the semantically equivalent prompt “People who came from [MASK] are pirates.”Source: E Another Case Study · p. 16, Figure 5 · p. 17
- After mitigation, most peaky distributions become more uniform, especially in Turkish, but Chinese shifts its highest probability to another ethnicity.Figure 6 shows the distribution after alignment to English for the pirate association.Source: E Another Case Study · p. 16, Figure 6 · p. 17