Pith. sign in

REVIEW 3 major objections 6 minor 41 references

ASCenD-BDS: Adaptable, Stochastic and Context-aware framework for Detection of Bias, Discrimination and Stereotyping

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The ASCenD-BDS framework claims that bias, discrimination, and stereotyping in LLMs can be detected by generating test scenarios on demand from category stems, opposing XFactors, and random synonyms, rather than by relying on finite…

desk verdict A plausible product framework for Indian-context bias probing, but with no empirical validation, no released data, and a load-bearing synonym-substitution step that is unanalyzed. read the letter →

arxiv 2502.02072 v2 pith:M4YAV7MI submitted 2025-02-04 cs.CL cs.AIcs.CY

Rajiv Bahl , Venkatesan N , Parimal Aglawe , Aastha Sarasapalli , Bhavya Kancharla , Chaitanya kolukuluri , Harish Mohite , Japneet Hora
show 6 more authors
Kiran Kakollu Rahul Dhiman Shubham Kapale Sri Bhagya Kathula Vamsikrishna Motru Yogeshwar Reddy
This is my paper · ORCID
classification cs.CLcs.AIcs.CY
keywords ASCenD-BDSbiasdetectionLLMstochasticpromptingcontext-awarestereotypingIndiancontextredteaming
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that existing bias-detection resources for large language models—Civil Comments, Winogender, WinoBias, BOLD, CrowS-Pairs, and BBQ—offer only a fixed set of scenarios, so a model can be tested against those examples and no more. ASCenD-BDS is proposed as a framework that composes prompts on the fly from manually written sentence stems, opposing identity terms called XFactors, and randomly chosen synonyms, so the number of test scenarios is not bounded by a dataset. The authors claim this makes bias assessment adaptable to different domains, stochastic in the probes it generates, and context-aware, with the Indian context as the first implementation using Census 2011 categories. A sympathetic reader would care because culturally specific bias categories such as caste and language are absent from most global benchmark datasets, and an unlimited generator could in principle test an LLM more thoroughly than any finite list.

What carries the argument

The load-bearing object is the STEM-XFactor-Synonym template, a parametrized sentence scaffold with slots for two opposing identity terms. A STEM is a manually written scenario skeleton that stays fixed, XFactor1 and XFactor2 fill in the contrasting identities, and random synonym selection adds stochasticity; the assembled string is the prompt fed to the model. Around this sit a category and sub-category hierarchy, a level system for adaptive ratcheting, and a context field that selects culturally appropriate XFactors, with Indian Census 2011 supplying the categorization scheme for the Indian implementation. The machinery's job is to convert a finite set of authored scenarios into an effectively unbounded set of test prompts while keeping each prompt grammatical and culturally situated.

What would settle it

Ask independent annotators to compare original and synonym-substituted versions of a sample of STEMs and mark whether the intended identity contrast is preserved. If a large share of substitutions are judged to change the target or to be unnatural, the stochasticity feature is not generating valid bias probes. A complementary check is to run the same generated probes against a model known to be biased and one that has been debiased; the framework should separate the two.

Watch

Extended reading notes

Core claim

The central claim is that bias, discrimination, and stereotyping in LLMs can be detected by a generative prompt-construction framework rather than by static benchmark datasets. Each probe is assembled as Pre-stem + XFactor1 + Mid-stem + XFactor2 + Post-stem + Query, where the XFactors are opposing identity descriptions drawn from category and sub-category taxonomies, and synonyms are randomly substituted into the XFactors to create stochastic variation. A level field allows the framework to ratchet scenario complexity from simple to harder based on prior responses, and a context field lets the same STEMs be filtered for a nation, culture, or organization. The paper reports that 800-plus STEMs across 10 categories and 31 sub-categories were built, and that context-awareness was established for India using Census 2011 categorization as the shared vocabulary. The claim is that this design overcomes the finite-scenario limitation of prior datasets while keeping probes culturally relevant.

Load-bearing premise

The load-bearing premise is that randomly substituting synonyms into a hand-written STEM yields a prompt that still tests the same bias; if a substitute changes the meaning or the contrast, the stochastic generation produces invalid probes rather than useful ones.

Editorial extensions

If this is right

  • If ASCenD-BDS works as claimed, an organization can generate fresh bias probes for every evaluation run instead of reusing the same benchmark items, making test-set overfitting harder.
  • The same STEMs can be repopulated with country- or organization-specific XFactors and contexts, extending bias checks to socio-cultural categories that global benchmarks omit, such as caste and regional language.
  • The level field implies a grading of bias severity: models that pass simple contrasts can be pushed to harder ones automatically, yielding a profile of where a model starts to show biased behavior.
  • If the framework is adopted as a red-teaming tool, the stochasticity means repeated runs sample different parts of the scenario space, so a single clean run would not certify a model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper describes a product-level test but publishes no benchmark numbers, so a direct next step is to measure whether models scored as biased by fixed datasets receive higher or more varied scores on generated probes; that comparison would tell whether the generative space actually adds signal.
  • The same STEM-plus-XFactor decomposition could be reused for tasks beyond text, such as measuring whether an image-generation model reproduces the same category contrasts when the XFactors are rendered as visual descriptions.
  • A less obvious consequence of stochasticity is that a single clean evaluation run becomes hard to interpret; reliable use of the framework would require aggregating many generated probes and reporting a distribution of outcomes rather than one pass or fail result.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes ASCenD-BDS, a framework for detecting bias, discrimination, and stereotyping in LLMs by dynamically generating evaluation prompts from hand-built templates (STEMs), opposing attribute groups (XFactor1/XFactor2), and random synonym substitution. The authors argue that static bias datasets (Civil Comments, Winogender, WinoBias, BOLD, CrowS-Pairs, BBQ) provide only finite 'point solutions,' whereas their framework adds Adaptability, Stochasticity, and Context Awareness, with context awareness instantiated for India using Indian Census 2011 categories. The paper reports the development of 800+ STEMs across 10 categories and 31 sub-categories by consultants at Saint Fox Consultancy Private Ltd., states that the concept was tested in SFCLabs as part of product development, and compares the framework with six existing datasets in descriptive terms. No experimental results, evaluation metrics, data release, or code artifacts are presented.

Significance. If the framework worked as claimed, it would address a genuine and underserved need: culturally specific, open-ended bias evaluation that is not limited to a fixed set of benchmark items, with categories such as caste, language, and regional identity that are largely absent from Western-centric benchmarks. The paper deserves credit for naming this gap and for proposing a concrete generative mechanism (STEM/XFactor/synonym decomposition) rather than only arguing abstractly. However, the significance is entirely conditional: the central claim that ASCenD-BDS detects bias is unsupported by any empirical evidence in the manuscript, the load-bearing stochasticity mechanism is unvalidated, and neither the taxonomy nor the test results are released, so the claims are not independently checkable. As submitted, the paper is a framework proposal without the validation that a detection claim requires.

major comments (3)
  1. [Section 3, 'Stochasticity'] The framework's ability to overcome the finite-scenario limitation rests on the statement that 'For each XFactor synonyms are selected randomly to provide Stochasticity feature,' but the manuscript provides no evidence that random synonym substitution preserves the semantic contrast and bias-probing intent of the STEM. The illustrative prompt itself illustrates the risk: synonyms for XFactor1 'person with hearing impairment' (e.g., 'deaf person,' 'hearing-impaired individual,' 'person with a hearing disability') differ in register, naturalness, and the specific stereotype they activate, so generated prompts may differ in ways other than the protected attribute. As written, the mechanism invites lexical and format confounds that would be conflated with measured bias. To support the claim, the authors should report, for a sample of STEMs, within-XFactor variance versus between-XFactor variance of model completions across the full synonym set, and human annotations (with inter-annotator agreement) confirming that each generated prompt is a natural, valid probe that isolates the protected attribute.
  2. [Abstract and Section 5] The central claim that ASCenD-BDS detects bias, discrimination, and stereotyping is not supported by any reported evaluation. The statement that 'The concept has been tested out in SFCLabs as part of product development' appears without results, metrics, model names, sample sizes, or error analysis, and Section 4's 'comparison' with Civil Comments, Winogender, WinoBias, BOLD, CrowS-Pairs, and BBQ is entirely descriptive, containing no experiments on any of these benchmarks. For the paper's claims to be assessable, the authors need to (a) run the framework on one or more publicly available LLMs and report detection results per category, (b) quantify stability of results across synonym draws, (c) compare against at least one existing benchmark on shared categories, and (d) release the STEM, XFactor, and synonym tables so that the work is reproducible.
  3. [Section 3, categories and STEMs] The representativeness and validity of the hand-built content are asserted rather than demonstrated. The paper gives no information about how the 800+ STEMs were developed, by how many annotators, with what instructions, or with what agreement, and the claim that Indian Census 2011 provides a 'commonality of categorization' is not justified. This matters because if the STEMs and XFactors encode the consultants' own priors, the framework risks measuring its own assumptions rather than model bias. A concrete way to address this is to document the annotation protocol, report inter-annotator agreement on a sample, compare the category taxonomy with established Indian social-science classifications, and publish the full artifact for independent audit.
minor comments (6)
  1. [Section 2 and References] Several in-text citations do not match the reference list: Section 1 cites 'Blodgett et al. [18]' but reference [18] is Kheya et al. (2024); Section 2 attributes 'linguistic hegemony' to 'Ramakrishnan et al. [15]' and a statistic to 'Pandey and Srinivasan [16],' but references [15] and [16] are Bolukbasi et al. (2016) and Gehman et al. (2020), respectively; and reference [19] duplicates [12] (both Crenshaw 1989).
  2. [Section 3, category count] The paper states that the STEMs span '10 different categories' but then enumerates eleven: Age, Disability status, Regional Identity, Language, Socio-economic status, Urban/Rural Background, Religion, Gender, Linguistic Identity, Appearance, and Caste; 'Language' and 'Linguistic Identity' also appear to overlap, so the taxonomy needs to be reconciled and consistently reported.
  3. [Figures 1-6] The comparison figures are labeled as 'distribution' comparisons but contain no quantitative data, axis labels, or legend definitions, so they do not support the comparative claims made in the surrounding text.
  4. [Reproducibility] No link, appendix, or supplementary material provides the 800+ STEMs, the XFactor and synonym tables, or the Census 2011 mapping, which makes the framework non-reproducible as submitted.
  5. [Abstract, Introduction, Conclusion] The rationale ('point solutions' of finite datasets) and the three feature bullets (Adaptability, Stochasticity, Context Awareness) are repeated nearly verbatim in the Abstract, Section 1, Section 5, and the Introduction; this redundancy should be condensed.
  6. [Language and terminology] The manuscript would benefit from professional language editing; there are numerous grammatical and typographical errors (e.g., 'The framework presents approach to detecting' in the Abstract, 'A/An person' in the example prompt, and inconsistent capitalization of 'Sub-Categories' vs 'Subcategories'), and the acronym STEM collides with the established meaning (Science, Technology, Engineering, Mathematics), so renaming the template component (e.g., 'PromptTemplate') would avoid ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No structural circularity: ASCenD-BDS is a hand-built template framework with no fitted parameters, no derived predictions, and no load-bearing self-citations.

full rationale

The paper does not contain a derivation chain in the mathematical or statistical sense. ASCenD-BDS is described as a manually constructed framework of Categories, Sub-Categories, STEMs, XFactors, and Synonyms, with 800+ STEMs developed by the authors' consultancy. There are no fitted parameters, no equations, no quantitative evaluation, and no claim that a numerical prediction is computed from inputs. The central claim—that dynamic, stochastic, context-aware scenario generation overcomes the finite-scenario limitation of static datasets—is an architectural assertion rather than a result derived from the framework's own definitions. The stochastic synonym substitution step ('For each XFactor synonyms are selected randomly to provide Stochasticity feature', Section 3) is an unvalidated design assumption: it presumes random substitutions preserve the semantic contrast between XFactor1 and XFactor2 and isolate the protected attribute. If this assumption fails, outputs would be confounded, but that is a validity threat, not a circularity. The statement that 'The concept has been tested out in SFCLabs as part of product development' (Abstract and Section 1) is an in-house validation claim with no reported results; it is self-referential in provenance but it does not function as a premise from which the framework's outputs are derived, and no circular step can be exhibited from it. No self-citations are load-bearing, no uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed as a prediction. Therefore the honest finding is no significant circularity, with the caveat that the framework's effectiveness remains unvalidated rather than circularly established.

Assumptions & free parameters 1 free parameters · 3 assumptions · 3 invented entities

The framework introduces conceptual entities (STEM, X-Factor, levels) and relies on hand-curated content and domain assumptions. There are no numeric parameters fitted to data, but the hand-chosen level count and the unvalidated assumptions about synonym substitution and content validity are significant burdens.

free parameters (1)
  • Adaptability levels 1 to 3 = 1-3
    The number of difficulty levels is chosen by hand with no calibration or evidence that three levels capture the range of bias severity.
assumptions (3)
  • domain assumption Indian Census 2011 provides a valid and sufficient categorization of social groups for bias testing.
    Used as the basis for categories and sub-categories in Section 1 and 3.
  • ad hoc to paper Random synonym substitution preserves the semantic contrast and bias-probing intent of the STEM.
    Stated in Section 3 as the Stochasticity mechanism, with no validation.
  • domain assumption The manually curated STEMs and XFactors are unbiased and representative of real-world bias scenarios.
    The framework's validity rests on the quality of these unlisted artifacts, but no inter-rater agreement or content validation is reported.
invented entities (3)
  • STEM
    purpose: A sentence template with placeholders for generating bias test prompts.
    Introduced as a new component; no independent evidence of its efficacy.
  • X-Factor
    purpose: Opposing demographic or social descriptors inserted into STEM templates.
    No evidence that the chosen XFactor pairs cover the relevant bias space.
  • Adaptability level
    purpose: A 1-3 ratcheting mechanism to increase complexity based on prior responses.
    No defined failure criterion or validation of the progression.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ASCenD-BDS: Adaptable, Stochastic and Context-aware framework for Detection of Bias, Discrimination and Stereotyping." pith.science (2026). https://pith.science/paper/M4YAV7MI

@misc{pith2026250202072,
  author       = {Pith},
  title        = {Pith review of: ASCenD-BDS: Adaptable, Stochastic and Context-aware framework for Detection of Bias, Discrimination and Stereotyping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M4YAV7MI}},
  note         = {Machine review of arXiv:2502.02072}
}
read the original abstract

The rapid evolution of Large Language Models (LLMs) has transformed natural language processing but raises critical concerns about biases inherent in their deployment and use across diverse linguistic and sociocultural contexts. This paper presents a framework named ASCenD BDS (Adaptable, Stochastic and Context-aware framework for Detection of Bias, Discrimination and Stereotyping). The framework presents approach to detecting bias, discrimination, stereotyping across various categories such as gender, caste, age, disability, socioeconomic status, linguistic variations, etc., using an approach which is Adaptive, Stochastic and Context-Aware. The existing frameworks rely heavily on usage of datasets to generate scenarios for detection of Bias, Discrimination and Stereotyping. Examples include datasets such as Civil Comments, Wino Gender, WinoBias, BOLD, CrowS Pairs and BBQ. However, such an approach provides point solutions. As a result, these datasets provide a finite number of scenarios for assessment. The current framework overcomes this limitation by having features which enable Adaptability, Stochasticity, Context Awareness. Context awareness can be customized for any nation or culture or sub-culture (for example an organization's unique culture). In this paper, context awareness in the Indian context has been established. Content has been leveraged from Indian Census 2011 to have a commonality of categorization. A framework has been developed using Category, Sub-Category, STEM, X-Factor, Synonym to enable the features for Adaptability, Stochasticity and Context awareness. The framework has been described in detail in Section 3. Overall 800 plus STEMs, 10 Categories, 31 unique SubCategories were developed by a team of consultants at Saint Fox Consultancy Private Ltd. The concept has been tested out in SFCLabs as part of product development.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages

  1. [1]

    Introduction ................................ ................................ ................................ ............... 3

    work page

  2. [2]

    Literature Review ................................ ................................ ................................ ....... 4

    work page

  3. [3]

    ASCenD-BDS Framework ................................ ................................ ............................ 6

    work page

  4. [4]

    Comparison of ASCenD-BDS with existing frameworks ................................ ................. 8 a) Civil Comments ................................ ................................ ................................ ...... 9 b) Winogender................................ ................................ ................................ ............ 9 c...

    work page

  5. [5]

    Conclusion and Future work ................................ ................................ ..................... 13

    work page

  6. [6]

    References ................................ ................................ ................................ .............. 15 3

    work page

  7. [7]

    Examples include datasets such as Civil Comments [1], Wino Gender [2], WinoBias [ 3], BOLD [ 4], CrowS -Pairs [ 5] and BBQ [ 6]

    Introduction The existing frameworks rely heavily on usage of datasets to generate scenarios for detection of Bias, Discrimination and Stereotyping. Examples include datasets such as Civil Comments [1], Wino Gender [2], WinoBias [ 3], BOLD [ 4], CrowS -Pairs [ 5] and BBQ [ 6]. As a result, these datasets provide a finite number of scenarios. The current f...

    work page 2011

  8. [8]

    It works well for tasks such as identifying organizational bias, evaluating societal stereotypes, or aiding in inclusive policymaking

    Adaptability: The framework can be adjusted for use across different domains. It works well for tasks such as identifying organizational bias, evaluating societal stereotypes, or aiding in inclusive policymaking

    work page

Show all 41 references
  1. [9]

    Stochasticity: By creating scenarios dynamically, the framework avoids the repetitive, limited nature of predefined datasets and enables more in-depth evaluations

  2. [10]

    Context Awareness: The framework’s ability to adapt to cultural nuances ensures that it identifies and addresses biases in a way that is both meaningful and appropriate to the context of the organisation. The rapid advancement of Large Language Models (LLMs) has revolutionized...

  3. [11]

    Language (technology) is power: A critical survey of

    Blodgett, S. L., Barocas, S., Daumé III, H., & Wallach, H. (2020). “Language (technology) is power: A critical survey of" bias” in nlp. arXiv preprint arXiv:2005.14050

    2020 arXiv

  4. [12]

    linguistic hegemony in machine learning

    Literature Review As language models advance in capabilities, their development increasingly aligns with the availability of digital content, often at the expense of balanced representation. The fundamental architecture of most LLMs tends to prioritize dominant languages an d ...

  5. [13]

    religion, disability, caste, etc

    ASCenD-BDS Framework A lexical hierarchy has been created for categories e.g. religion, disability, caste, etc. and in some case sub -categories have been identified e.g. disability ( visually challenged, physically challenged, etc). The categories and sub -categories have bee...

  6. [14]

    Comparison of ASCenD-BDS with existing frameworks Comparisons with existing datasets/frameworks is given below: Civil Comments Dataset This dataset includes over 2 million comments from the Civil Comments platform [1], annotated for toxicity and identity-based attributes. It e...

  7. [15]

    Examples include datasets such as Civil Comments, Wino Gender, WinoBias, BOLD, CrowS -Pairs and BBQ

    Conclusion and Future work Why ASCenD-BDS? The existing frameworks rely heavily on usage of datasets to generate scenarios for detection of Bias, Discrimination and Stereotyping. Examples include datasets such as Civil Comments, Wino Gender, WinoBias, BOLD, CrowS -Pairs and BB...

    2011

  8. [16]

    A benchmark for toxic comment classification on civil comments dataset

    Duchene, C., Jamet, H., Guillaume, P ., & Dehak, R. (2023). “A benchmark for toxic comment classification on civil comments dataset”. arXiv preprint arXiv:2301.11125

    2023 arXiv

  9. [17]

    Revisiting English Winogender Schemas for Consistency, Coverage, and Grammatical Case

    Vagrant Gautam, Julius Steuer, Eileen Bingert, Ray Johns, Anne Lauscher, Dietrich Klakow,(2024). “Revisiting English Winogender Schemas for Consistency, Coverage, and Grammatical Case”

    2024

  10. [18]

    have criticized the Anglo-centric NLP paradigm, advocating for more contextually relevant approaches. In India, where social categorizations often extend beyond binary classifications, addressing these complexities demands tailored, inclusive methodologi es that reflect the re...

  11. [19]

    Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, Kai-Wei Chang. “Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods” NAACL-2018

    2018

  12. [20]

    W., & Gupta, R

    Dhamala, J., Sun, T., Kumar, V ., Krishna, S., Pruksachatkun, Y ., Chang, K. W., & Gupta, R. (2021, March). Bold: Dataset and metrics for measuring biases in open-ended language generation. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency...

    2021

  13. [21]

    CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models

    Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. “CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models ”. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1953–196...

    2020

  14. [22]

    BBQ: A hand -built bias benchmark for question answering

    Parrish, Alicia, et al. "BBQ: A hand -built bias benchmark for question answering. " arXiv preprint arXiv:2110.08193 (2021)

    2021 arXiv

  15. [23]

    Language Models are Few-Shot Learners

    Brown, T. B., et al. (2020). “Language Models are Few-Shot Learners”. NeurIPS

    2020

  16. [24]

    Language Models are Unsupervised Multitask Learners

    Radford, A., et al. (2019). “Language Models are Unsupervised Multitask Learners ”. OpenAI

    2019

  17. [25]

    On the Dangers of Stochastic Parrots

    Bender, E. M., et al. (2021). “On the Dangers of Stochastic Parrots”. FAccT Conference

    2021

  18. [26]

    On the Opportunities and Risks of Foundation Models

    Bommasani, R., et al. (2021). “On the Opportunities and Risks of Foundation Models ”. https://doi.org/10.48550/arXiv.2108.07258

    2021 doi arXiv

  19. [27]

    Demarginalizing the Intersection of Race and Sex

    Crenshaw, K. (1989). “Demarginalizing the Intersection of Race and Sex ”. University of Chicago Legal Forum: Vol. 1989: Iss. 1, Article 8

    1989

  20. [28]

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan

  21. [30]

    Census of India. (2011). Language Data Report

    2011

  22. [31]

    Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings

    Bolukbasi, T., et al. (2016). “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings” NeurIPS 16

    2016

  23. [32]

    Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith

  24. [34]

    Eojin Jeon, Mingyu Lee, Juhyeong Park, Yeachan Kim, Wing -Lam Mok, and SangKeun Lee

  25. [36]

    The Pursuit of Fairness in Artificial Intelligence Models: A Survey

    Tahsin Alamgir Kheya, et al. (2024). “The Pursuit of Fairness in Artificial Intelligence Models: A Survey” https://doi.org/10.48550/arXiv.2403.17333

    2024 doi arXiv

  26. [37]

    Demarginalizing the Intersection of Race and Sex: A Black Feminist Critique of Antidiscrimination Doctrine, Feminist Theory and Antiracist Politics

    Crenshaw, Kimberlé (1989). “Demarginalizing the Intersection of Race and Sex: A Black Feminist Critique of Antidiscrimination Doctrine, Feminist Theory and Antiracist Politics”. The University of Chicago Legal Forum 140:139-167

    1989

  27. [38]

    John Lalor, Yi Yang, Kendall Smith, Nicole Forsgren, and Ahmed Abbasi

  28. [40]

    Debiasing multilingual word embeddings: A case study of three indian languages

    Bansal S, Garimella V , Suhane A, Mukherjee A. “Debiasing multilingual word embeddings: A case study of three indian languages ”. In: Proceedings of the 32nd ACM Conference on Hypertext and Social Media.2021, 27–34

    2021

  29. [41]

    Worst of Both Worlds: Biases Compound in Pre -trained Vision-and-Language Models

    Srinivasan, R., et al. (2021). “Worst of Both Worlds: Biases Compound in Pre -trained Vision-and-Language Models”. GeBNLP 2022

    2021

  30. [2020]

    RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models

    “RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models”, In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3356–3369, Online. Association for Computational Linguistics

    2020

  31. [2021]

    A Survey on Bias and Fairness in Machine Learning

    “A Survey on Bias and Fairness in Machine Learning” . ACM Comput. Surv. 54, 6, Article 115 (July 2022), 35 pages. https://doi.org/10.1145/3457607

    2022 doi

  32. [2022]

    Benchmarking Intersectional Biases in NLP

    “Benchmarking Intersectional Biases in NLP ”. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages 3598 –3609, Seattle, United States. Association for Computational Linguistics

    2022

  33. [2023]

    Improving Bias Mitigation through Bias Experts in Natural Language Understanding

    “Improving Bias Mitigation through Bias Experts in Natural Language Understanding” . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages 11053 –11066, Singapore. Association for Computational Linguistics

    2023

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.