mlsanityData to 13 Oct 2026

Unmasking the Mask -- Evaluating Social Biases in Masked Language Models

arXiv 2104.07496 ↗First posted 15 Apr 2021cs.CLcited by 60

Not on the map. Why?

Abstract (arXiv)

Masked Language Models (MLMs) have shown superior performances in numerous downstream NLP tasks when used as text encoders. Unfortunately, MLMs also demonstrate significantly worrying levels of social biases. We show that the previously proposed evaluation metrics for quantifying the social biases in MLMs are problematic due to following reasons: (1) prediction accuracy of the masked tokens itself tend to be low in some MLMs, which raises questions regarding the reliability of the evaluation metrics that use the (pseudo) likelihood of the predicted tokens, and (2) the correlation between the prediction accuracy of the mask and the performance in downstream NLP tasks is not taken into consideration, and (3) high frequency words in the training data are masked more often, introducing noise due to this selection bias in the test cases. To overcome the above-mentioned disfluencies, we propose All Unmasked Likelihood (AUL), a bias evaluation measure that predicts all tokens in a test case given the MLM embedding of the unmasked input. We find that AUL accurately detects different types of biases in MLMs. We also propose AUL with attention weights (AULA) to evaluate tokens based on their importance in a sentence. However, unlike AUL and AULA, previously proposed bias evaluation measures for MLMs systematically overestimate the measured biases, and are heavily influenced by the unmasked tokens in the context.

Key facts. Cited by 60 papers in the map (data through 13 Oct 2026). Builds on 32 of its 50 references that are in the map. First posted on arXiv 15 Apr 2021.

Builds on

32 of 50 references.

References · 32

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,304
  2. 2020Language Models are Few-Shot Learnerscited by 32,320
  3. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,322
  4. 2013Efficient Estimation of Word Representations in Vector Spacecited by 4,360
  5. 2015Neural Machine Translation of Rare Words with Subword Unitscited by 2,910
  6. 2019ALBERT: A Lite BERT for Self-supervised Learning of Language Representationscited by 2,779
  7. 2018Deep contextualized word representationscited by 2,286
  8. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,309
  9. 2015Aligning Books and Movies: Towards Story-like Visual Explanations by Watching Movies and Reading Bookscited by 1,042
  10. 2020StereoSet: Measuring stereotypical bias in pretrained language modelscited by 918
  11. 2016Semantics derived automatically from language corpora contain human-like biasescited by 895
  12. 2020CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Modelscited by 705
20 more
  1. 2018Gender Bias in Coreference Resolution: Evaluation and Debiasing Methodscited by 658
  2. 2019Attention is not not Explanationcited by 633
  3. 2019On Measuring Social Biases in Sentence Encoderscited by 397
  4. 2021Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLPcited by 337
  5. 2019Measuring Bias in Contextualized Word Representationscited by 275
  6. 2019Gender Bias in Contextualized Word Embeddingscited by 203
  7. 2019BERT has a Mouth, and It Must Speak: BERT as a Markov Random Field Language Modelcited by 185
  8. 2018Learning Gender-Neutral Word Embeddingscited by 169
  9. 2019Black is to Criminal as Caucasian is to Police: Detecting and Removing Multiclass Bias in Word Embeddingscited by 146
  10. 2019Assessing Social and Intersectional Biases in Contextualized Word Representationscited by 135
  11. 2019On Measuring and Mitigating Biased Inferences of Word Embeddingscited by 112
  12. 2021Debiasing Pre-trained Contextualised Embeddingscited by 107
  13. 2019Attention Interpretability Across NLP Taskscited by 71
  14. 2019Gender-preserving Debiasing for Pre-trained Word Embeddingscited by 61
  15. 2019Understanding Undesirable Word Embedding Associationscited by 59
  16. 2019Rare Words: A Major Problem for Contextualized Embeddings And How to Fix it by Attentive Mimickingcited by 40
  17. 2021Dictionary-based Debiasing of Pre-trained Word Embeddingscited by 24
  18. 2021Attention Can Reflect Syntactic Structure (If You Let It)cited by 22
  19. 2019Conceptor Debiasing of Word Representations Evaluated on WEATcited by 18
  20. 2020Autoencoding Improves Pre-trained Word Embeddingscited by 11

Built on by

Cited by year

Citing papers per year: 2021: 1; 2022: 10; 2023: 19; 2024: 17; 2025: 13.

60 papers.

Sort:

By area:

  1. 2023Bias and Fairness in Large Language Models: A Surveycited by 636
  2. 2023Having Beer after Prayer? Measuring Cultural Bias in Large Language Modelscited by 153
  3. 2022Probing Pre-Trained Language Models for Cross-Cultural Differences in Valuescited by 147
  4. 2023A Survey on Fairness in Large Language Modelscited by 93
  5. 2022Gender Bias in Masked Language Models for Multiple Languagescited by 60
  6. 2022Debiasing isn't enough! -- On the Effectiveness of Debiasing MLMs and their Social Biases in Downstream Taskscited by 48
  7. 2022BERTScore is Unfair: On Social Bias in Language Model-Based Metrics for Text Generationcited by 48
  8. 2023Social Bias Probing: Fairness Benchmarking for Language Modelscited by 29
  9. 2024Deconstructing The Ethics of Large Language Models from Long-standing Issues to New-emerging Dilemmas: A Surveycited by 23
  10. 2022Debiasing Methods for Fairer Neural Models in Vision and Language Research: A Surveycited by 22
  11. 2024Advancing Social Intelligence in AI Agents: Technical Challenges and Open Questionscited by 22
  12. 2021Measuring Fairness with Biased Rulers: A Survey on Quantifying Biases in Pretrained Language Modelscited by 20
38 more
  1. 2023An Efficient Multilingual Language Model Compression through Vocabulary Trimmingcited by 19
  2. 2023Understanding the Effect of Model Compression on Social Bias in Large Language Modelscited by 19
  3. 2022Sense Embeddings are also Biased--Evaluating Social Biases in Static and Contextualised Sense Embeddingscited by 18
  4. 2024CEB: Compositional Evaluation Benchmark for Fairness in Large Language Modelscited by 18
  5. 2023Indian-BhED: A Dataset for Measuring India-Centric Biases in Large Language Modelscited by 17
  6. 2023Fairness in Language Models Beyond English: Gaps and Challengescited by 16
  7. 2023Evaluating Gender Bias of Pre-trained Language Models in Natural Language Inference by Considering All Labelscited by 16
  8. 2024Cross-Care: Assessing the Healthcare Implications of Pre-training Data on Language Model Biascited by 15
  9. 2023In-Contextual Gender Bias Suppression for Large Language Modelscited by 12
  10. 2023Tackling Bias in Pre-trained Language Models: Current Trends and Under-represented Societiescited by 11
  11. 2024Fairness Definitions in Language Models Explainedcited by 10
  12. 2024Understanding Intrinsic Socioeconomic Biases in Large Language Modelscited by 9
  13. 2022Towards an Enhanced Understanding of Bias in Pre-trained Neural Language Models: A Survey with Special Emphasis on Affective Biascited by 8
  14. 2022HERB: Measuring Hierarchical Regional Bias in Pre-trained Language Modelscited by 8
  15. 2023Comparing Intrinsic Gender Bias Evaluation Measures without using Human Annotated Examplescited by 8
  16. 2023The Impact of Debiasing on the Performance of Language Models in Downstream Tasks is Underestimatedcited by 7
  17. 2025Evaluating the Effect of Retrieval Augmentation on Social Biasescited by 7
  18. 2022Gender Bias in Meta-Embeddingscited by 6
  19. 2025Evaluate Bias without Manual Test Sets: A Concept Representation Perspective for LLMscited by 6
  20. 2023Solving Cosine Similarity Underestimation between High Frequency Words by L2 Norm Discountingcited by 5
  21. 2023Solving NLP Problems through Human-System Collaboration: A Discussion-based Approachcited by 5
  22. 2024Robust Evaluation Measures for Evaluating Social Biases in Masked Language Modelscited by 5
  23. 2024Large Language Models as Mirrors of Societal Moral Standardscited by 5
  24. 2023A Predictive Factor Analysis of Social Biases and Task-Performance in Pretrained Masked Language Modelscited by 4
  25. 2024Quantifying Stereotypes in Languagecited by 4
  26. 2024Eagle: Ethical Dataset Given from Real Interactionscited by 4
  27. 2024A Study of Nationality Bias in Names and Perplexity using Off-the-Shelf Affect-related Tweet Classifierscited by 4
  28. 2025Dutch CrowS-Pairs: Adapting a Challenge Dataset for Measuring Social Biases in Language Models for Dutchcited by 4
  29. 2025LLM Bias Detection and Mitigation through the Lens of Desired Distributionscited by 4
  30. 2023Cross-Modal Attribute Insertions for Assessing the Robustness of Vision-and-Language Learningcited by 3
  31. 2024Evaluating Short-Term Temporal Fluctuations of Social Biases in Social Media Data and Masked Language Modelscited by 3
  32. 2025McBE: A Multi-task Chinese Bias Evaluation Benchmark for Large Language Modelscited by 3
  33. 2025Mechanistic Interpretability with SAEs: Probing Religion, Violence, and Geography in Large Language Modelscited by 3
  34. 2022Learning Dynamic Contextualised Word Embeddings via Template-based Temporal Adaptationcited by 2
  35. 2024Measuring Social Biases in Masked Language Models by Proxy of Prediction Qualitycited by 2
  36. 2024What is Your Favorite Gender, MLM? Gender Bias Evaluation in Multilingual Masked Language Modelscited by 2
  37. 2024REFINE-LM: Mitigating Language Model Stereotypes via Reinforcement Learningcited by 2
  38. 2024A Note on Bias to Completecited by 1

50 most cited of 60 All 60 (JSON)