Pith. sign in

REVIEW 3 major objections 5 minor 29 references

From Measurement to Mitigation: Exploring the Transferability of Debiasing Approaches to Gender Bias in Maltese Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that Counterfactual Data Augmentation—swapping gendered words in training sentences—is the only debiasing method tested that reliably lowers measured gender bias in Maltese language models, though the effect is partial…

desk verdict Useful new Maltese debiasing resources and a plausible but unproven headline claim: no fine-tuning-only control weakens the CDA result. read the letter →

arxiv 2507.03142 v1 pith:LI3BVV22 submitted 2025-07-03 cs.CL

Melanie Galea , Claudia Borg This is my paper · ORCID
classification cs.CL
keywords genderbiasMalteselow-resourcelanguagescounterfactualdataaugmentationdebiasinglanguagemodelsCrowS-PairsSEAT
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 asks whether debiasing methods invented for English can be transplanted to Maltese, a low-resource language with complex gendered grammar. It measures gender bias in two BERT-style Maltese models with translated versions of standard benchmarks, then applies four mitigation techniques. Its central finding is that Counterfactual Data Augmentation—swapping gendered words to opposite-gender versions in training sentences—is the only method that consistently reduces both explicit and implicit bias scores for Maltese. The reduction is real but partial: the gender-swapped sentences often break Maltese grammar, and embedding visualisations show male–female word pairs remaining distant after debiasing. The authors read this as evidence that bias mitigation must be rethought for morphologically rich languages rather than copied from English.

What carries the argument

The load-bearing mechanism is Counterfactual Data Augmentation: a gender wordlist of 193 male–female Maltese word pairs is used to rewrite about 17.4% of a 411k-sentence corpus into opposite-gender versions, and the model is fine-tuned on the combined original-plus-swapped data. The method's effectiveness is judged with CrowS-Pairs, a sentence-pair benchmark where higher scores mean more bias, and SEAT, an implicit-association test in sentence form, both adapted to Maltese, plus t-SNE visualisations of gendered word embeddings. CDA works by forcing the model to see male and female forms in the same contexts, but because Maltese marks gender on verbs and adjectives beyond the swapped noun, the counterfactual sentences are often ungrammatical (25.5% of a sampled 200), which the paper identifies as the main obstruction to fuller debiasing.

What would settle it

Fine-tune BERTu with default hyperparameters on the same 411k-sentence corpus but with no gender swapping; if CrowS and SEAT scores drop by roughly the same amount as they do under CDA, the debiasing effect is a fine-tuning artifact rather than evidence for CDA transfer.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that Counterfactual Data Augmentation transfers to Maltese better than the alternatives, though imperfectly. Across CrowS-Pairs and SEAT scores, CDA reduced measured bias for both BERTu and mBERTu, with monolingual BERTu improving from 55.40 to 49.19 CrowS and 0.530 to 0.460 SEAT, and mBERTu from 51.20 to 48.83 CrowS and 0.540 to 0.462 SEAT. Dropout regularization, GuiDebias, and Auto-Debias produced mixed or adverse results, sometimes lowering one score while raising the other. The paper also finds that the multilingual mBERTu starts out less biased than the monolingual BERTu, and that t-SNE visualisations corroborate partial debiasing while showing persistent gendered clustering for the female forms of doctor, lawyer, and teacher.

Load-bearing premise

The claim rests on the assumption that the observed score drops come from the gender-swapping rather than from the fine-tuning itself, since no control condition fine-tunes on the same data without swapping.

Editorial extensions

If this is right

  • If CDA is the most transferable method, future Maltese debiasing efforts should start from gender-balanced data augmentation rather than dropout or prompt-based methods.
  • The persistence of male–female distance in BERTu's embeddings after CDA implies that surface-level word swapping alone is insufficient in Maltese; verb and adjective agreement must also be counterfactually altered.
  • The finding that mBERTu is less biased than BERTu suggests multilingual pretraining dampens language-specific gender stereotypes, pointing to pretraining data diversity as a mitigation lever.
  • The conflict between CrowS and SEAT results for dropout, GuiDebias, and Auto-Debias means single-metric evaluations can invert conclusions about which method works.
  • The translated Maltese SEAT subset and CrowS-Pairs materials are reusable assets for benchmarking future debiasing attempts.

Reading between the lines

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

  • A testable extension of the paper's logic: a morphologically aware counterfactual generator that also swaps verb inflection and article gender would likely reduce the 25.5% grammatical-error rate and could push CrowS and SEAT scores below what CDA achieved.
  • The paper's finding that multilingual exposure reduces bias suggests a training-side implication the authors leave implicit: adding more diverse gendered-language data to pretraining corpora may be a more robust bias intervention than fine-tuning debiasing for low-resource languages.
  • The authors' own limitation note implies that their evaluation cannot separate bias reduction from language-model degradation; a downstream-task battery such as named-entity recognition or sentiment analysis would be the natural next experiment.
  • Because CrowS-Pairs for Maltese already exists from prior work and SEAT was translated here, subsequent studies can use these same benchmarks to test whether newer debiasing methods outperform CDA on Maltese.
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 / 5 minor

Summary. The paper adapts a suite of English-centric gender-bias measurement and mitigation tools to Maltese, a low-resource, morphologically rich language. Using BERTu and mBERTu as base models, the authors measure bias with a Maltese version of CrowS-Pairs, a translated subset of SEAT, and template-based probes; they then compare four debiasing methods: Counterfactual Data Augmentation (CDA), Dropout Regularization, GuiDebias, and Auto-Debias. The headline empirical claim is that CDA is the most effective debiasing technique for Maltese, reducing CrowS-Pairs and SEAT scores for both models, while the other methods show mixed or negative results. The authors also contribute translated evaluation data, a Maltese Debiasing Dataset, and code.

Significance. If the central comparison were adequately controlled, the paper would be a useful contribution to the under-studied area of bias mitigation in low-resource and gendered languages. Its explicit multi-metric approach, public release of evaluation data, and careful adaptation of benchmarks to Maltese are praiseworthy and align with current best practice in bias evaluation. The finding that CDA transfers only partially and with morphological side effects would be an informative result for the multilingual NLP community. However, the current evidence does not fully support the headline claim because the effect of CDA is not separated from the effect of continued fine-tuning, and the reported differences are not accompanied by uncertainty estimates.

major comments (3)
  1. [Section 3.2, Table 4] The headline claim that CDA is the most effective debiasing method for Maltese is not supported because the experimental design lacks a fine-tuning-only control. CDA fine-tunes on a corpus where 17.4% of sentences were gender-swapped, while the Dropout Regularization condition uses the same un-augmented corpus but deliberately changes dropout rates; neither condition isolates the effect of gender swapping from the effect of further training on a large corpus. A control that fine-tunes BERTu and mBERTu on the same 411k-sentence corpus with identical hyperparameters (five epochs, batch size 16, gradient accumulation 16, learning rate 2e-5) but without any gender swap is necessary to attribute the observed CrowS reductions (55.40 to 49.19 for BERTu; 51.20 to 48.83 for mBERTu) to the counterfactual augmentation rather than to continued pretraining. This missing control is load-bearing because the paper's central recommendation is specifically that CDA 'proved to be the most effective.'
  2. [Section 3.2, Limitations] The paper reports that a sample of 200 counterfactually generated Maltese sentences contained 25.5% grammatical errors due to Maltese's gendered morphology. Since CrowS-Pairs and SEAT are acceptability-sensitive sentence-level probes, the observed score reductions after CDA could partly reflect degraded language modeling or a distribution shift toward ungrammatical input, rather than a genuine reduction in stereotypical associations. The authors should provide auxiliary evidence that CDA improves bias without harming language-model utility, for example by reporting masked-language-model perplexity or accuracy on a clean Maltese held-out set before and after CDA, or by showing that the CrowS/SEAT reductions persist when the evaluation is restricted to grammatically correct counterfactual examples.
  3. [Section 4.2, Tables 4–7] All bias-mitigation comparisons are reported as single point estimates, with no standard errors, confidence intervals, significance tests, or multiple-seed variation. Several differences are small and could easily arise from training noise; for instance, mBERTu CrowS drops from 51.20 to 48.83 after CDA, and BERTu SEAT changes from 0.530 to 0.460, while GuiDebias changes BERTu CrowS from 55.40 to 55.46. Without repeated runs or a statistical test, the ranking of methods, including the claim that CDA is the most effective, is not statistically established. Additionally, the Dropout Regularization configuration (h=0.2, a=0.15) was selected from a set of configurations evaluated on the same bias metrics, introducing selection bias into the comparison; a validation split or a clearly stated hyperparameter-selection protocol is needed.
minor comments (5)
  1. [Throughout] The model name is rendered inconsistently as both 'mBERTu' and 'mBERTU' (e.g., Section 5 and Figure captions versus Tables 1, 4–7); please standardize to one spelling.
  2. [Section 3.2] The Auto-Debias method is attributed to Guo and Caliskan (2021), but the cited reference, 'Detecting emergent intersectional biases,' describes a different method for detecting biases, not Auto-Debias; please cite the correct source (Guo, Yang, and Caliskan, EMNLP 2022).
  3. [Tables 2–3] The tables contain spacing and diacritic artifacts, such as 'ja¯hdem b ¯hala', 'messa ˙g˙gier', and 'maxtrudaxxa'; these should be cleaned and typeset correctly for the camera-ready version.
  4. [Section 4.1] The template-based analysis in the main text reports only one template applied to BERTu; since the full results are hosted in the repository, presenting at least one additional template or a compact aggregate summary in the main text would improve reproducibility and readability.
  5. [Figures 1–6] The captions state 'BERTu after debiasing' and 'mBERTu after debiasing' without naming the debiasing method; because the paper compares several methods, each caption should explicitly state that CDA was used.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the bias evaluations and debiasing comparisons are empirical and externally benchmarked, with no fitted parameter renamed as a prediction.

full rationale

This is an empirical study with no derivation chain whose outputs are built from its inputs. Gender bias is measured with external benchmarks (CrowS-Pairs as extended by Fort et al. 2024, SEAT translated for Maltese) and the debiasing methods (CDA, Dropout Regularization, GuiDebias, Auto-Debias) are applied with fixed, stated hyperparameters (e.g., five epochs, batch size 16, lr 2e-5 for CDA). None of the debiasing procedures optimize toward CrowS or SEAT scores, so the reported reductions in Tables 4-7 are not fitted quantities. The central claim that CDA 'proved to be the most effective debiasing technique' is a comparison of before/after scores on these external metrics, not a quantity equivalent to its own definition. Self-citations to BERTu/mBERTu (Micallef et al. 2022) and to the Maltese CrowS-Pairs resource (Fort et al. 2024) provide the base models and evaluation dataset, but they are independent inputs rather than the paper's conclusion; the cited models and corpus are not derived from the paper's debiasing results. The absence of a fine-tuning-only control is a legitimate experimental confound regarding whether CDA's improvement comes from the gender-swapping or from continued training, but that is a causal attribution issue, not a circular reduction: CDA is not defined in terms of the evaluation metrics, and the paper does not claim to predict those metrics from its own fitted parameters. No equation or parameter is defined in terms of the target result, and no load-bearing argument reduces to a self-citation. Hence the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entities or parameters; its load-bearing assumptions are about the validity of translated bias metrics and the interpretability of fine-tuning experiments without a control condition.

free parameters (3)
  • dropout rates (hidden h=0.2, attention a=0.15) = h=0.2, a=0.15
    Selected via search over configurations using CrowS/SEAT evaluation scores; effectively tuned on the evaluation data.
  • CDA augmentation ratio = 17.4% (Maltese), 18.3% (English)
    Fraction of gender-swapped sentences in the fine-tuning corpus; a design choice that sets the strength of the intervention.
  • fine-tuning hyperparameters = 5 epochs, batch 16, grad accum 16, lr 2e-5
    Standard settings from prior work; not demonstrated as optimal for Maltese but affect all methods equally.
assumptions (3)
  • domain assumption CrowS-Pairs and SEAT, translated into Maltese, validly measure gender bias in Maltese LMs.
    The paper uses these metrics without validating construct equivalence in Maltese; Section 3.1.
  • domain assumption The gender-swapped sentences produced by CDA, despite 25.5% grammatical errors, constitute a valid debiasing intervention.
    Section 3.2 reports a sample where 25.5% of counterfactual sentences contain grammatical errors, yet the analysis treats CDA as a working debiasing method.
  • domain assumption Fine-tuning alone does not change CrowS/SEAT scores, so any reduction is attributable to the debiasing technique.
    Implied by the absence of a fine-tuning-only control in the experimental design; Table 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Measurement to Mitigation: Exploring the Transferability of Debiasing Approaches to Gender Bias in Maltese Language Models." pith.science (2026). https://pith.science/paper/LI3BVV22

@misc{pith2026250703142,
  author       = {Pith},
  title        = {Pith review of: From Measurement to Mitigation: Exploring the Transferability of Debiasing Approaches to Gender Bias in Maltese Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LI3BVV22}},
  note         = {Machine review of arXiv:2507.03142}
}
read the original abstract

The advancement of Large Language Models (LLMs) has transformed Natural Language Processing (NLP), enabling performance across diverse tasks with little task-specific training. However, LLMs remain susceptible to social biases, particularly reflecting harmful stereotypes from training data, which can disproportionately affect marginalised communities. We measure gender bias in Maltese LMs, arguing that such bias is harmful as it reinforces societal stereotypes and fails to account for gender diversity, which is especially problematic in gendered, low-resource languages. While bias evaluation and mitigation efforts have progressed for English-centric models, research on low-resourced and morphologically rich languages remains limited. This research investigates the transferability of debiasing methods to Maltese language models, focusing on BERTu and mBERTu, BERT-based monolingual and multilingual models respectively. Bias measurement and mitigation techniques from English are adapted to Maltese, using benchmarks such as CrowS-Pairs and SEAT, alongside debiasing methods Counterfactual Data Augmentation, Dropout Regularization, Auto-Debias, and GuiDebias. We also contribute to future work in the study of gender bias in Maltese by creating evaluation datasets. Our findings highlight the challenges of applying existing bias mitigation methods to linguistically complex languages, underscoring the need for more inclusive approaches in the development of multilingual NLP.

Figures

Figures reproduced from arXiv: 2507.03142 by the authors.

Figure 1
Figure 1. t-SNE visualization of BERTu’s word embeddings for the gendered pair tabib-tabiba (Maltese for ‘doctor’ [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of word embeddings for the gendered pair " [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. t-SNE visualization of BERTu’s embeddings [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: t-SNE visualization of mBERTu’s embed￾dings for ‘avukat-avukata’ (lawyer, m-f) before and after CDA. (a) BERTu t-SNE graph for ‘g¯halliem-g¯halliema’. (b) BERTu t-SNE graph for ‘g¯halliem-g¯halliema’ after debiasing [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 6
Figure 6. Figure 6: t-SNE visualization of mBERTu’s embeddings [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 5
Figure 5. Figure 5: t-SNE visualization of BERTu’s embeddings [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 6 canonical work pages

  1. [1]

    Marion Bartl, Malvina Nissim, and Albert Gatt. 2020. https://aclanthology.org/2020.gebnlp-1.1 Unmasking contextual stereotypes: Measuring and mitigating BERT ' s gender bias . In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 1--16, Barcelona, Spain (Online). Association for Computational Linguistics

    work page 2020

  2. [2]

    Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. https://doi.org/10.1145/3442188.3445922 On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21, page 610–623, New York, NY, USA. Association for Computing...

    arXiv 2021

  3. [3]

    Su Lin Blodgett and Brendan O'Connor. 2017. https://arxiv.org/abs/1707.00061 Racial disparity in natural language processing: A case study of social media african-american english . CoRR, abs/1707.00061

    arXiv 2017

  4. [4]

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. https://proceedings.neurips.cc/paper_files/paper/2016/file/a486cd07e4ac3d270571622f4f316ec5-Paper.pdf Man is to computer programmer as woman is to homemaker? debiasing word embeddings . In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc

    2016

  5. [5]

    Hudson, Ehsan Adeli, Russ B

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2021. https://arxiv.org/abs/2108.07258 On the opportunities and risks of foundation models . CoRR, abs/2108.07258

    arXiv 2021

  6. [6]

    Bryson, and Arvind Narayanan

    Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. https://doi.org/10.1126/science.aal4230 Semantics derived automatically from language corpora contain human-like biases . Science, 356(6334):183--186

    doi:10.1126/science.aal4230 2017

  7. [7]

    Rodrigo Alejandro Ch \'a vez Mulsa and Gerasimos Spanakis. 2020. https://aclanthology.org/2020.gebnlp-1.6 Evaluating bias in D utch word embeddings . In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 56--71, Barcelona, Spain (Online). Association for Computational Linguistics

    work page 2020

  8. [8]

    Pieter Delobelle, Ewoenam Tokpo, Toon Calders, and Bettina Berendt. 2022. https://doi.org/10.18653/v1/2022.naacl-main.122 Measuring fairness with biased rulers: A comparative study on bias metrics for pre-trained language models . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human La...

    doi:10.18653/v1/2022.naacl-main.122 2022

Show all 29 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  2. [10]

    Karen Fort, Laura Alonso Alemany, Luciana Benotti, Julien Bezan c on, Claudia Borg, Marthese Borg, Yongjian Chen, Fanny Ducel, Yoann Dupont, Guido Ivetta, et al. 2024. https://aclanthology.org/2024.lrec-main.1545 Your stereotypical mileage may vary: Practical challenges of eva...

    2024

  3. [11]

    Gallegos, Ryan A

    Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. https://doi.org/10.1162/coli_a_00524 Bias and fairness in large language models: A survey . Computational Linguistics, 50(3):1097--1179

    2024 doi

  4. [12]

    Wei Guo and Aylin Caliskan. 2021. https://doi.org/10.1145/3461702.3462536 Detecting emergent intersectional biases: Contextualized word embeddings contain a distribution of human-like biases . In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, AIES '21,...

    2021

  5. [13]

    Hadas Kotek, Rikker Dockum, and David Sun. 2023. https://doi.org/10.1145/3582269.3615599 Gender bias and stereotypes in large language models . In Proceedings of The ACM Collective Intelligence Conference, CI '23, page 12–24, New York, NY, USA. Association for Computing Machinery

    2023

  6. [14]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://api.semanticscholar.org/CorpusID:198953378 Roberta: A robustly optimized bert pretraining approach . ArXiv, abs/1907.11692

    2019 arXiv

  7. [15]

    Kaiji Lu, Piotr Mardziel, Fangjing Wu, Preetam Amancharla, and Anupam Datta. 2018. https://arxiv.org/abs/1807.11714 Gender bias in neural natural language processing . CoRR, abs/1807.11714

    2018 arXiv

  8. [17]

    Bowman, and Rachel Rudinger

    Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. 2019 b . https://doi.org/10.18653/v1/N19-1063 On measuring social biases in sentence encoders . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computatio...

    2019 doi

  9. [18]

    Nicholas Meade, Spandana Gella, Devamanyu Hazarika, Prakhar Gupta, Di Jin, Siva Reddy, Yang Liu, and Dilek Hakkani-Tur. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.796 Using in-context learning to improve dialogue safety . In Findings of the Association for Computati...

    2023 doi

  10. [19]

    Kurt Micallef, Albert Gatt, Marc Tanti, Lonneke van der Plas, and Claudia Borg. 2022. https://doi.org/10.18653/v1/2022.deeplo-1.10 Pre-training data quality and quantity for a low-resource language: New corpus and BERT models for M altese . In Proceedings of the Third Workshop...

    2022 doi

  11. [20]

    Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.154 C row S -pairs: A challenge dataset for measuring social biases in masked language models . In Proceedings of the 2020 Conference on Empirical Methods in Na...

    2020 doi

  12. [21]

    Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, et al

    NLLB Team , Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, et al. 2022. https://arxiv.org/abs/arXiv:1902.01382 No language left behind: Scaling human-centered machine translation

    2022 arXiv

  13. [22]

    OpenAI. 2023. https://openai.com/research/gpt-4 Gpt-4 technical report . Technical report, OpenAI

    2023

  14. [23]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners

    2019

  15. [24]

    Mike Rosner and Claudia Borg. 2022. https://european-language-equality.eu/wp-content/uploads/2022/03/ELE___Deliverable_D1_25__Language_Report_Maltese_.pdf Report on the Maltese Language . Language Technology Support of Europe’s Languages in 2020/2021

    2022

  16. [25]

    Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. https://doi.org/10.18653/v1/P19-1163 The risk of racial bias in hate speech detection . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668--1678, ...

    2019 doi

  17. [26]

    Kellie Webster, Xuezhi Wang, Ian Tenney, Alex Beutel, Emily Pitler, Ellie Pavlick, Jilin Chen, Ed Chi, and Slav Petrov. 2020. Measuring and reducing gendered correlations in pre-trained models. arXiv preprint arXiv:2010.06032

    2020 arXiv

  18. [27]

    Chi, Quoc Le, and Denny Zhou

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed H. Chi, Quoc Le, and Denny Zhou. 2022. https://arxiv.org/abs/2201.11903 Chain of thought prompting elicits reasoning in large language models . CoRR, abs/2201.11903

    2022 arXiv

  19. [28]

    Tae-Jin Woo, Woo-Jeoung Nam, Yeong-Joon Ju, and Seong-Whan Lee. 2023. https://doi.org/10.1109/ICASSP49357.2023.10095658 Compensatory debiasing for gender imbalances in language models . In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Process...

    2023

  20. [29]

    Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. 2018. https://doi.org/10.1145/3278721.3278779 Mitigating unwanted biases with adversarial learning . In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, AIES '18, page 335–340, New York, NY, USA. Asso...

    2018

  21. [30]

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai - Wei Chang. 2018. https://arxiv.org/abs/1804.06876 Gender bias in coreference resolution: Evaluation and debiasing methods . CoRR, abs/1804.06876

    2018 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.