Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Stereotype Detection as a Catalyst for Enhanced Bias Detection: A Multi-Task Learning Approach

T0 review · 4 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 Training language models to detect stereotypes alongside bias improves their bias detection, according to experiments on the new StereoBias dataset.

desk verdict Useful new dataset and a plausible MTL result, but the abstract's universal claim needs a caveat and the controls are not as clean as advertised. read the letter →

arxiv 2507.01715 v1 pith:PEQ2OF3E submitted 2025-07-02 cs.CL

Aditya Tomar , Rudra Murthy , Pushpak Bhattacharyya This is my paper
classification cs.CL
keywords biasdetectionstereotypemulti-tasklearningStereolanguagemodelfairnessnaturalprocessingQLoRAcross-datasetgeneralization
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

This paper argues that stereotype detection can act as a catalyst for bias detection: a model trained jointly on both tasks catches biased statements that the same model trained on bias alone misses. To test this, the authors built StereoBias, a dataset of 5,012 sentences labeled for both bias and stereotype, and compared single-task fine-tuning against two multi-task setups across encoder-only and decoder-only models. The multi-task models consistently achieved higher macro-F1 scores on bias detection, with the largest relative gain around 14 percent. A control experiment pairing bias detection with sentiment analysis produced smaller improvements, which the paper reads as evidence that the gain comes from the conceptual link between bias and stereotypes rather than from multi-task learning in general. If true, this suggests that stereotype information is a practical resource for building fairer language models.

What carries the argument

The load-bearing object is StereoBias, a 5,012-sentence dataset where every sentence is labeled independently for bias (bias/no bias) and stereotype (stereotype/no stereotype) across religion, gender, socio-economic status, race, profession, and an 'others' category. These dual labels make possible two multi-task architectures: Shared-MTL, which feeds a shared transformer representation into separate bias and stereotype heads, and Full-MTL, which treats the four bias–stereotype combinations as a single classification task. The four-way contingency is the central identity that carries the argument: it lets the model learn when a statement is biased without being stereotypical, stereotypical without being biased, both, or neither, and it is this disentangling that the paper credits with improving bias detection. The control pairing with sentiment analysis works by contrast: it tests whether any auxiliary task helps, and the smaller gain there is the evidence that the bias–stereotype pairing is doing the work.

What would settle it

Run the same STL versus Shared-MTL comparison on a corpus of naturally occurring text whose sentences were annotated for bias and stereotype in two independent annotation passes, so that the labels are not derived from stereotype-selected sentences; if the multi-task bias gain drops to near zero or reverses, the catalyst effect is an artifact of dataset construction rather than a property of the bias–stereotype relationship.

Watch

Extended reading notes

Core claim

The central claim is that jointly learning to detect stereotypes improves a model's ability to detect bias, and that the improvement is specific to the bias–stereotype relationship. The paper's evidence is a set of fine-tuning experiments on StereoBias, a newly constructed dataset in which each sentence carries both a bias label and a stereotype label. Across RoBERTa-large, ALBERT-xxlarge-v2, BERT-large-uncased, and three decoder-only models, Shared-MTL (one transformer with two classification heads) and Full-MTL (a single four-way classifier over the bias–stereotype combination) beat single-task baselines on bias macro-F1 in nearly every configuration, with gains up to about 14 percent relative. Cross-dataset runs pairing StereoSet with ToxicBias and BABE showed the same pattern, and a control pairing bias with sentiment analysis improved bias detection less, supporting the claim that stereotype information, not multi-task learning per se, drives the gain.

Load-bearing premise

The link between stereotypes and bias in StereoBias may be stronger than in ordinary language, because the dataset was assembled by taking sentences already known to contain stereotypes (the stereotypical completions from StereoSet and CrowS-Pairs), so the multi-task gain might shrink when applied to naturally occurring text where the two signals are not so tightly coupled.

Editorial extensions

If this is right

  • Bias detectors can be improved by adding stereotype detection as an auxiliary task, even when the stereotype training data comes from a separate dataset with different annotation guidelines.
  • The four-way label format (no bias/no stereotype, bias only, stereotype only, both) is a workable alternative to two separate heads and captures the joint relationship explicitly.
  • The benefit is not generic multi-task learning: pairing bias detection with a conceptually distant task such as sentiment analysis yields smaller gains, so task alignment matters for deployment choices.
  • There may be a small trade-off: in some StereoBias configurations stereotype detection scores dip slightly when the model is trained for both tasks at once.
  • Decoder-only models fine-tuned with QLoRA, a memory-efficient adaptation method, reach competitive bias-detection performance, so the approach is not limited to encoder-only architectures.

Reading between the lines

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

  • The authors leave inference-time prompting untested; a plausible corollary is that asking an LLM to label stereotypes first would sharpen bias detection without fine-tuning.
  • Because StereoBias selects stereotypical completions, the measured bias–stereotype link may be inflated; an out-of-distribution test on naturally occurring comments would show whether the multi-task gain transfers.
  • If the catalyst effect is real, weakly supervised stereotype labels from a large model could bootstrap bias detection in languages or domains where bias annotations are scarce.
  • The paper's own limitation about all annotators sharing one cultural background implies the bias–stereotype mapping may be culture-specific; re-annotating the same sentences with annotators from other regions would reveal how much of the gain is universal.
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

4 major / 5 minor

Summary. The paper introduces StereoBias, a 5,012-sentence dataset with both bias and stereotype labels derived from StereoSet and CrowS-Pairs, and evaluates single-task (STL), shared-encoder multi-task (Shared-MTL), and four-class joint (Full-MTL) classifiers using encoder-only models (RoBERTa-large, ALBERT-xxlarge-v2, BERT-large-uncased) and decoder-only models fine-tuned with QLoRA (Llama-3.1-8B, Gemma-7B, Mistral-7B-v0.3). The central claim is that joint training on stereotype detection significantly improves bias detection over single-task training, with cross-dataset experiments on ToxicBias+StereoSet and BABE+StereoSet and a sentiment-control experiment (ToxicBias+sst2) intended to show that the gains come from task relatedness rather than from multi-task learning alone. The paper reports macro-F1 improvements in most configurations and a maximum relative improvement of about 13.92%.

Significance. If the central claim holds, the paper makes a useful practical contribution: bias detection is an important moderation task, and the paper shows that a relatively cheap auxiliary stereotype-detection signal can improve it. The release of StereoBias with dual labels is genuinely valuable, and the cross-dataset experiments in Tables 7–8 go beyond a single benchmark by evaluating bias on external datasets (ToxicBias and BABE) while using StereoSet as the auxiliary task. The error analysis, annotation details, and explicit discussion of limitations are strengths. However, the evidence does not currently support the headline claim in its unqualified form: one configuration (Gemma-7B Shared-MTL) shows a decrease, several gains are not marked as statistically significant, and the sentiment-control comparison lacks a matched setup and significance testing. This is an empirical paper, and the main issue is the strength of the evidence rather than an internal inconsistency.

major comments (4)
  1. [Abstract; §5.2; Tables 1–2] The unqualified claim that joint training “significantly improves bias detection compared to training them separately” is not supported by the full set of results. In Table 2, Gemma-7B Shared-MTL gives 0.7338 bias macro-F1 versus 0.7409 for STL, a numerical decrease; in Table 1, BERT-large-uncased Shared-MTL (0.7507 vs 0.7391) and RoBERTa-large Full-MTL (0.7471 vs 0.7409) are not marked as significant even though the abstract states the improvement is significant. Please qualify the claim (e.g., “in most settings”), report the p-values for all STL/MTL pairs, and explicitly discuss the Gemma-7B negative case.
  2. [§5.3; §A.3.1] The claim that the improvement “stems from the connection between bias and stereotypes, not multi-task learning alone” is not established. Table 9 shows that bias+sentiment MTL also improves bias detection over STL (e.g., RoBERTa-large: 0.6362 vs 0.6224; BERT-large-uncased: 0.6222 vs 0.5805), so the control is not simply “related task helps, unrelated task does not.” Moreover, the bias+stereotype numbers in Tables 7–8 come from ToxicBias+StereoSet while the sentiment control uses ToxicBias+sst2, so the two MTL setups differ in both the auxiliary task and the auxiliary dataset. No significance test or confidence interval is reported for the difference between the two MTL gains, and Tables 7–8 do not report significance at all despite the procedure described in §4.3. A matched comparison (same bias dataset, comparable auxiliary data, and a statistical test on the difference between MTL configurations) is needed to support the causal claim.
  3. [§4.1; §A.1.1] StereoBias is constructed by taking stereotypical completions from StereoSet and sent_more sentences from CrowS-Pairs and then annotating the same sentences for bias and stereotype. This design can inflate the correlation between the bias and stereotype labels relative to naturally occurring text, making stereotype detection an unusually informative shortcut for the bias label on the StereoBias test set. The cross-dataset experiments partially mitigate this concern because they evaluate bias on external datasets, but they do not test whether a StereoBias-trained MTL model transfers to independent bias data, and they do not include a control (e.g., shuffled stereotype labels or a matched unrelated auxiliary task) that would rule out generic MTL regularization as the source of the gain. I request an external transfer evaluation of the StereoBias-trained model or a matched control-task comparison on StereoBias itself before accepting the general claim.
  4. [§3.2; Tables 1–2] Full-MTL is described as a four-class classification task, but Tables 1 and 2 report “Bias (Macro-F1)” in this setting without stating how the four-class predictions are converted to a binary bias score. If the reported number is the macro-F1 over all four classes, it is not comparable to the binary STL scores; if it is obtained by collapsing the four classes into bias/no-bias, that procedure should be stated explicitly. Please clarify this reporting choice, since it affects the interpretation of the Full-MTL results that contribute to the central claim.
minor comments (5)
  1. [§1, Figure 1] Figure 1 uses Llama-3.3-70B-Instruct, a model that is not included in the experiments; the illustration therefore does not demonstrate the paper's fine-tuned MTL approach and should be relabeled or replaced with an example from an evaluated model.
  2. [Abstract; §4.1] The dataset is described as spanning “five categories: religion, gender, socio-economic status, race, profession, and others,” which lists six items; please correct the count or the list.
  3. [§4.1; §A.3.2] There are inconsistencies in naming and cross-referencing: “Crows-Pairs” appears in §4.1 while “CrowS-Pairs” is used elsewhere, and §A.3.2 refers to “Figure 10” when the result is in Table 10.
  4. [§A.1; Table 4] Table 4 gives statistics for StereoSet and ToxicBias but not for StereoBias; since StereoBias is the main resource, the split sizes and label distributions should be reported in the table rather than only in Figure 2.
  5. [§A.3.2] The zero-shot and five-shot evaluation uses lowest-perplexity selection via lm-eval-harness, but the paper does not state whether the candidate sequences are scored as full sentences or as continuations; this detail is needed for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the reported MTL improvement is an empirical result, and the dataset-construction concern is a transfer/confound issue rather than a definitional reduction.

full rationale

This paper is an empirical multi-task-learning study; its central claim is a measured performance comparison rather than a derivation. The improvement of Shared-MTL and Full-MTL over STL for bias detection (Tables 1, 2, 7, and 8) is obtained by training models on held-out splits and is not a fitted parameter renamed as a prediction, nor does any equation define the output in terms of the input. The closest concern is the construction of StereoBias in Section 4.1 and Appendix A.1.1: the dataset is assembled by selecting stereotypical sentences from StereoSet and CrowS-Pairs and re-annotating them for both bias and stereotype labels, which could inflate the measured bias–stereotype correlation relative to naturally occurring text. That is an empirical validity and transfer concern—and the paper itself acknowledges related limitations (all annotators share one cultural background, and evaluation is confined to Western-context datasets)—but it does not make the MTL result true by construction and does not reduce the reported gain to the dataset-definitional choice. The sentiment-control experiment in Appendix A.3.1 is suggestive rather than conclusive because it uses a different bias dataset and reports no significance test, but this is an evidentiary weakness, not circularity. Self-citations appear only as related-work context (e.g., multi-task metaphor/hyperbole work and prior bias datasets) and are not load-bearing for the central empirical finding. I therefore find no significant circularity.

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

The ledger is mostly about assumptions in the experimental protocol rather than fitted constants. There are no invented entities. The main burden is the assumption that the dataset's label correlation reflects a real-world relationship and that the statistical tests capture the relevant variance.

assumptions (3)
  • domain assumption Macro-F1 on the test split is a valid measure of bias and stereotype detection performance.
    The paper reports and compares Macro-F1 throughout (§5) without discussing whether this metric aligns with downstream fairness goals.
  • domain assumption The paired t-test on per-example correctness is appropriate despite using a single trained model per configuration.
    §4.3; this treats test examples as independent and ignores variance across training seeds.
  • domain assumption The human annotations in StereoBias constitute ground truth for bias and stereotype.
    §4.1 and §A.6; three annotators with shared Indian nationality, majority voting, Fleiss' kappa 0.6239 for bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stereotype Detection as a Catalyst for Enhanced Bias Detection: A Multi-Task Learning Approach." pith.science (2026). https://pith.science/paper/PEQ2OF3E

@misc{pith2026250701715,
  author       = {Pith},
  title        = {Pith review of: Stereotype Detection as a Catalyst for Enhanced Bias Detection: A Multi-Task Learning Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PEQ2OF3E}},
  note         = {Machine review of arXiv:2507.01715}
}
read the original abstract

Bias and stereotypes in language models can cause harm, especially in sensitive areas like content moderation and decision-making. This paper addresses bias and stereotype detection by exploring how jointly learning these tasks enhances model performance. We introduce StereoBias, a unique dataset labeled for bias and stereotype detection across five categories: religion, gender, socio-economic status, race, profession, and others, enabling a deeper study of their relationship. Our experiments compare encoder-only models and fine-tuned decoder-only models using QLoRA. While encoder-only models perform well, decoder-only models also show competitive results. Crucially, joint training on bias and stereotype detection significantly improves bias detection compared to training them separately. Additional experiments with sentiment analysis confirm that the improvements stem from the connection between bias and stereotypes, not multi-task learning alone. These findings highlight the value of leveraging stereotype information to build fairer and more effective AI systems.

Figures

Figures reproduced from arXiv: 2507.01715 by the authors.

Figure 1
Figure 1. Example showing how incorporating stereo [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distribution of non-stereotypical/stereotypical and unbiased/biased sentences in StereoBias Dataset. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Number of sentences per category in the StereoBias dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 4 canonical work pages

  1. [1]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  2. [2]

    Naveen Badathala, Abisek Rajakumar Kalarani, Tejpalsingh Siledar, and Pushpak Bhattacharyya. 2023. https://doi.org/10.18653/v1/2023.findings-acl.26 A match made in heaven: A multi-task framework for hyperbole and metaphor detection . In Findings of the Association for Computational Linguistics: ACL 2023, pages 388--401, Toronto, Canada. Association for Co...

    doi:10.18653/v1/2023.findings-acl.26 2023

  3. [3]

    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

  4. [4]

    Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. https://doi.org/10.18653/v1/2021.acl-long.81 Stereotyping N orwegian salmon: An inventory of pitfalls in fairness benchmark datasets . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conferenc...

    doi:10.18653/v1/2021.acl-long.81 2021

  5. [5]

    Dushyant Singh Chauhan, Dhanush S R, Asif Ekbal, and Pushpak Bhattacharyya. 2020. https://doi.org/10.18653/v1/2020.acl-main.401 Sentiment and emotion help sarcasm? a multi-task learning framework for multi-modal sarcasm, sentiment and emotion analysis . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4351-...

    doi:10.18653/v1/2020.acl-main.401 2020

  6. [6]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. https://arxiv.org/abs/2305.14314 Qlora: Efficient finetuning of quantized llms . Preprint, arXiv:2305.14314

    arXiv 2023

  7. [7]

    Jacob Devlin, Ming - Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. https://arxiv.org/abs/1810.04805 BERT: pre-training of deep bidirectional transformers for language understanding . CoRR, abs/1810.04805

    arXiv 2018

  8. [8]

    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 Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long a...

    doi:10.18653/v1/n19-1423 2019

Show all 31 references
  1. [9]

    Hilton and William von Hippel

    James L. Hilton and William von Hippel. 1996. https://doi.org/10.1146/annurev.psych.47.1.237 Stereotypes . Annual Review of Psychology, 47(Volume 47, 1996):237--271

    1996 doi

  2. [10]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. https://arxiv.org/abs/2106.09685 Lora: Low-rank adaptation of large language models . Preprint, arXiv:2106.09685

    2021 arXiv

  3. [11]

    Akshita Jha, Aida Mostafazadeh Davani, Chandan K Reddy, Shachi Dave, Vinodkumar Prabhakaran, and Sunipa Dev. 2023. https://doi.org/10.18653/v1/2023.acl-long.548 S ee GULL : A stereotype benchmark with broad geo-cultural coverage leveraging generative models . In Proceedings of...

    2023 doi

  4. [12]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  5. [13]

    Aditya Joshi, Sarvnaz Karimi, Ross Sparks, Cecile Paris, and C Raina MacIntyre. 2019. https://aclanthology.org/U19-1020/ Does multi-task learning always help?: An evaluation on health informatics . In Proceedings of the 17th Annual Workshop of the Australasian Language Technol...

    2019

  6. [14]

    Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. 2019. https://doi.org/10.18653/v1/W19-3823 Measuring bias in contextualized word representations . In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 166--172, Flor...

    2019 doi

  7. [15]

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. https://arxiv.org/abs/1909.11942 ALBERT: A lite BERT for self-supervised learning of language representations . CoRR, abs/1909.11942

    2019 arXiv

  8. [16]

    https://pubmed.ncbi.nlm.nih.gov/843571/ The measurement of observer agreement for categorical data

    JR Landis and GG Koch. https://pubmed.ncbi.nlm.nih.gov/843571/ The measurement of observer agreement for categorical data

  9. [17]

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

    2019 arXiv

  10. [18]

    Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. https://doi.org/10.18653/v1/2021.acl-long.416 S tereo S et: Measuring stereotypical bias in pretrained language models . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Inte...

    2021 doi

  11. [19]

    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. [20]

    Nihar Sahoo, Himanshu Gupta, and Pushpak Bhattacharyya. 2022. https://doi.org/10.18653/v1/2022.conll-1.10 Detecting unintended social bias in toxic language datasets . In Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 132--143, Abu...

    2022 doi

  13. [21]

    Nihar Sahoo, Pranamya Kulkarni, Arif Ahmad, Tanu Goyal, Narjis Asad, Aparna Garimella, and Pushpak Bhattacharyya. 2024. https://doi.org/10.18653/v1/2024.naacl-long.487 I ndi B ias: A benchmark dataset to measure social biases in language models for I ndian context . In Proceed...

    2024 doi

  14. [22]

    Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2021. https://doi.org/10.18653/v1/2021.acl-long.330 Societal biases in language generation: Progress and challenges . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the...

    2021 doi

  15. [23]

    Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2019. https://doi.org/10.18653/v1/D19-1339 The woman worked as a babysitter: On biases in language generation . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9...

    2019 doi

  16. [24]

    Hari Shrawgi, Prasanjit Rath, Tushar Singhal, and Sandipan Dandapat. 2024. https://aclanthology.org/2024.eacl-long.111/ Uncovering stereotypes in large language models: A task complexity-based approach . In Proceedings of the 18th Conference of the European Chapter of the Asso...

    2024

  17. [25]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. https://aclanthology.org/D13-1170/ Recursive deep models for semantic compositionality over a sentiment treebank . In Proceedings of the 2013 Conference on Em...

    2013

  18. [26]

    Timo Spinde, Manuel Plank, Jan-David Krieger, Terry Ruas, Bela Gipp, and Akiko Aizawa. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.101 Neural media bias detection using distant supervision with BABE - bias annotations by experts . In Findings of the Association for C...

    2021 doi

  19. [27]

    Yarden Tal, Inbal Magar, and Roy Schwartz. 2022. https://doi.org/10.18653/v1/2022.gebnlp-1.13 Fewer errors, but more stereotypes? the effect of model size on gender bias . In Proceedings of the 4th Workshop on Gender Bias in Natural Language Processing (GeBNLP), pages 112--120...

    2022 doi

  20. [28]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi \`e re, Mihir Sanjay Kale, Juliette Love, et al. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295

    2024 arXiv

  21. [29]

    Craig S Webster, Saana Taylor, Courtney Thomas, and Jennifer M Weller. 2022. Social bias, discrimination and inequity in healthcare: mechanisms, implications and recommendations. BJA education, 22(4):131--137

    2022

  22. [30]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  23. [31]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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