Pith. sign in

REVIEW 4 major objections 5 minor 74 references

Fine-Tuned LLMs are "Time Capsules" for Tracking Societal Bias Through Books

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Decade-trained LLMs turn bestselling fiction into measurable bias records.

desk verdict A genuinely new corpus-and-probe method for tracking bias in books, but the headline causal claim outruns the evidence; worth engaging with seriously. read the letter →

arxiv 2502.05331 v2 pith:QAI3HPK3 submitted 2025-02-07 cs.CL

Sangmitra Madhusudan , Robert Morabito , Skye Reid , Nikta Gohari Sadr , Ali Emami This is my paper · ORCID
classification cs.CL
keywords societalbiasLLMfine-tuningtemporalbestsellerfictionBookPAGEcorpusdemographicstereotypestimecapsulemeasurement
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 introduces a method for turning bestselling fiction into a measurable record of social bias. It builds BookPAGE, a corpus of 593 fictional books split into seven decade subsets from 1950 to 2019, fine-tunes three large language models on each decade, and probes the resulting models with sentence-completion prompts about gender, sexual orientation, race, and religion. The central claim is that the fine-tuned models manifest biases reflective of their times, and that these biases come mainly from the books' content rather than from model architecture or prior training. If true, the method gives historians and social scientists a scalable way to ask how stereotypes in popular literature changed over time.

What carries the argument

The load-bearing object is BookPAGE, a decade-stratified corpus of 593 bestselling fictional books grouped into seven subsets, one per decade from the 1950s through the 2010s. Each book's full text is segmented into roughly 8,000-token chunks and formatted as sentence-completion tasks, so fine-tuning on a decade subset produces a decade-specific model. Bias is measured with Role-Specific Elicitation Prompts (REPs), sentence-completion scenarios that avoid explicit bias cues so that any skew in who completes the role comes from learned associations rather than detectable triggers. Responses are categorized into demographic subcategories by GPT-4, and named-entity-recognition overlap checks confirm that the fine-tuned models are actually recalling the books' content. The consistency of the resulting bias profiles across very different model architectures is what carries the claim that content, not architecture, drives the result.

What would settle it

If the same temporal trends appear when the decade labels are shuffled, for instance when a model fine-tuned on 2000s fiction still shows the post-9/11 Islam spike when labeled as 1950s, then the trends come from model priors rather than book content. More directly, replacing the 2000s bestseller subset with a random sample of 2000s fiction and seeing the Islam association disappear would show that the bestseller-selection step is doing the work.

Watch

Extended reading notes

Core claim

The paper claims that decade-specific fine-tuning makes LLMs act as time capsules: a model trained on 1950s bestsellers absorbs the demographic associations of that decade, while a model trained on 2010s bestsellers absorbs those of the 2010s. Across three models of different architectures and sizes, the fine-tuned variants show consistent temporal trends, such as women appearing as CEOs in 8% of 1950s-trained responses versus 22% in 2010s-trained responses, same-sex relationship references rising from near zero in earlier decades to 10-12% by the 2010s, and negative portrayals of Islam jumping in the 2000s, which the authors connect to post-9/11 attitudes. Because models with very different architectures converge toward similar decade-specific bias profiles after fine-tuning, the paper argues the biases originate in the training books, not in the models' initial training or design.

Load-bearing premise

The argument assumes that the bestseller lists used to build each decade's corpus, after removing duplicates and inaccessible titles, represent the era's prevailing social attitudes closely enough that changes in model responses track changes in society rather than changes in which books become bestsellers or are publicly available.

Editorial extensions

If this is right

  • Book corpora can be converted into quantitative, decade-by-decade bias timelines without relying on static word-embedding methods.
  • Temporal trends in popular fiction can be aligned with historical events, as with the 1990s uptick in women in leadership and the post-9/11 rise in negative Islam portrayals.
  • Fine-tuning on the same book data converges different LLMs toward similar bias profiles, implying that training-data composition matters more than model choice for these biases.
  • The same fine-tune-and-probe pipeline could be extended to other demographics, roles, or finer-grained time periods without changing the core method.

Reading between the lines

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

  • A testable extension the authors do not run: fine-tune models on each decade's books with author demographics matched across decades; if the gender and race trends persist, they track the fiction itself rather than who wrote it.
  • Because the corpus is built from bestseller lists and online availability, part of a decade's apparent bias could come from which books became bestsellers or remained accessible; comparing against a random sample of published fiction from each decade would separate those effects.
  • The GPT-4 classifier used to label responses is itself a possible source of bias, so re-labeling a sample of responses with human annotators or a different model would test how much of the measured trend is classifier-driven.
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 BookPAGE, a corpus of 593 bestselling fiction books split into seven decades (1950–2019), and fine-tunes three LLMs (Llama 3 70B, Gemini 1.0 Pro, Mixtral 8x7B) on each decade's subset. The resulting models are probed with role-specific sentence-completion prompts and the responses are classified by GPT-4 into demographic subcategories for gender, sexual orientation, race, and religion. The paper reports temporal trends such as increasing female CEO portrayal, rising same-sex fiancé references, and increasing negative Islam associations after the 2000s, and argues that these biases stem mainly from the book content rather than model architecture or initial training. The central methodological claim is that fine-tuned LLMs can serve as 'time capsules' that reflect the societal biases of the books' eras.

Significance. If the causal claim were established, the method would be a scalable complement to close reading and embedding-based diachronic analysis, and BookPAGE would be a useful public resource for cultural analytics. The multi-architecture design, the comparison against pre-trained models, and the availability of the corpus and code are clear strengths. However, the headline causal attribution currently outruns the evidence: the paper does not directly measure the target biases in the book texts, and the observed trends are reported without uncertainty quantification. The significance of the contribution is therefore conditional on the additional validation and robustness analyses that a revision would need to provide.

major comments (4)
  1. [§5, Table 4] The headline temporal trends rest on very small counts. In the main experiments each prompt is run 50 times (Section 5), so the 4% to 12% increase in CEO-women responses in Table 4 corresponds to a shift from 2 to 6 responses out of 50. The reported Kruskal-Wallis tests compare subcategories within a single prompt and decade; they do not test whether the across-decade differences are statistically reliable. Without confidence intervals, bootstrap intervals, or an explicit test for trend (e.g., a logistic regression of category on decade), the temporal trajectories that are the paper's central contribution cannot be distinguished from sampling noise. This should be addressed with appropriate uncertainty quantification and significance tests for the decade trends.
  2. [Abstract, §6.5] The causal claim that the observed biases 'stem mainly from the books' content and not the models' architecture or initial training' is not established by the evidence presented. The convergence across architectures (Table 5, Appendix Tables 29–31) and the divergence from pre-trained models (Figures 12–14) do not exclude alternative explanations: the shared REP and GPT-4 classification pipeline, a generic fine-tuning shift toward narrative or stereotypical completions, or the high overlap between BookPAGE and the models' pre-training corpora (Table 1 shows Books3 overlap ranging from 59% to 92%). The NER validation in Table 3 demonstrates entity recall, not that the specific role-demographic biases are encoded in the fine-tuned models. No direct measurement of the target biases in the BookPAGE texts is reported; the GloVe analysis in Section 6.6 is limited to 1950s gender roles and is framed as a contrast rather than a validation. The causal claim requires a direct text-based validation showing that the same role-demographic associations can be measured in the books and that their temporal trajectories match the model-response trajectories.
  3. [§2.1, Limitations] The corpus construction confounds time period with corpus composition. The decade subsets differ not only in era but in which books became bestsellers, which titles were accessible online, and how much the corpus overlaps with LLM pre-training data (Books3 overlap rises from 59% in the 1950s to 92% in the 2000s, Table 1). The Limitations section concedes that bestsellers 'may overrepresent mainstream perspectives,' but the paper does not assess whether the observed temporal trends could be driven by these compositional changes rather than by societal bias. Because the 'time capsule' interpretation requires that the corpus be representative of each era's prevailing attitudes, the authors should provide robustness analyses, such as reweighting by author demographics, replicating on an independently collected corpus, or examining whether accessibility and overlap patterns correlate with the model-response trends.
  4. [Footnote 8, §5] The religion subcategory retention rule—keeping only religions with over 15% of model responses in a preliminary analysis—is a post-hoc, data-dependent selection. Since religion is one of the four headline demographics, this filter can inflate or distort the reported trends (e.g., Table 4's defacing-monument Islam association rising from 22% to 48%). The threshold should be pre-specified, the full set of predefined religions (including Hinduism and Atheism, as listed in Appendix Table 10) should be reported, or a sensitivity analysis should be provided showing that the trends are robust to alternative thresholds.
minor comments (5)
  1. [§2.1] The exclusion and supplementation procedures are difficult to follow: it is not clear why the 1950s and 2010s specifically had 30 or more exclusions, or how the New York Times titles were matched to the decade. Please clarify the criteria and the arithmetic (647 selected, 69 excluded, 15 added = 593).
  2. [§3.2] The entity-overlap validation is reported only for the 1950s subset (Table 3 and Appendix Table 12). Please state explicitly whether the validation was performed for all decades or justify restricting the check to one decade.
  3. [§6.6] The statement that the roles Top Student and CEO 'did not exist' in the 1950s subset is imprecise; presumably the relevant terms were absent from the training data or from the GloVe vocabulary. Please rephrase to avoid the implication that the roles were historically nonexistent.
  4. [Table 4 caption] The caption says the Kruskal-Wallis test shows subcategories 'statistically significantly different from others in the same prompt'; as written this is a within-prompt comparison across subcategories, not a test of decade trends, and no correction for multiple comparisons is mentioned. Please clarify the statistical procedure.
  5. [Abstract vs. Limitations] The abstract's causal phrasing ('these biases stem mainly from the books' content') is stronger than the Limitations section's statement that 'we cannot definitively establish causal relationships.' Please align the claims with the evidence presented.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical measurement pipeline, and the central causal claim, while underdetermined, is not definitionally forced by the paper's own construction.

full rationale

The paper's derivation chain is empirical rather than definitional: BookPAGE texts are collected, LLMs are fine-tuned on decade-specific subsets, and bias measures are obtained by prompting the resulting models and classifying responses with GPT-4. No parameter of the bias trends is fitted from the trend itself and then reported as a prediction, and no equation equates the book-content claim to the measured outputs by construction. The central causal attribution (Abstract: 'these biases stem mainly from the books' content'; Section 6.5) is defended by convergence across three architectures and by divergence from pre-trained variants, which is evidence, not a tautology. The strongest concerns are external validity and underdetermination, and the Limitations section itself flags them: fine-tuning 'may inadvertently introduce biases beyond those present in the original texts,' GPT-4 classification 'introduces potential for additional bias or errors,' and 'we cannot definitively establish causal relationships.' These are correctness/interpretation risks, not circularity: the outputs do not equal the inputs by definition even if the causal interpretation is too strong. The 15% religion retention criterion (footnote 8) is a data-dependent category choice that could affect statistical claims, but it does not make the reported trends equivalent to a fitted parameter. No load-bearing self-citations or imported uniqueness theorems appear. Accordingly the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The analysis has no free parameters in the sense of fitted constants, but several methodological choices (response count, religion retention cutoff, hyperparameters) are hand-picked and could influence the reported trends. The key axioms are domain assumptions about corpus representativeness, the transfer of textual associations to prompted outputs, and the reliability of GPT-4 as the annotator.

free parameters (3)
  • Religion subcategory retention threshold = 15% of model responses
    In Section 5, the authors state 'We retained in main experiments the religions that had over 15% of model responses during a preliminary analysis.' This data-dependent cutoff determines which subcategories appear in the reported religion distributions, affecting all religion trends.
  • Number of responses per prompt = 50
    Section 5: 'we used all fine-tuned models ... to generate 50 responses per prompt.' The choice of 50 is arbitrary and small enough that percentage shifts of 4-8 points correspond to 2-4 responses.
  • Fine-tuning hyperparameters = Open models: batch size 4, LR 1e-5, weight decay 0.01, warmup 100; Gemini: epochs 4, LR multiplier 1.0
    These are standard choices but not fitted; they affect the degree to which the model internalizes book content. Listed for completeness.
assumptions (4)
  • domain assumption Bestseller lists of each decade are a valid proxy for mainstream societal attitudes and biases of that era.
    Section 2.1: 'Popular fiction, particularly bestsellers, often reflects and shapes mainstream societal values and attitudes of their time.' If bestseller composition is driven by marketing or by which titles are accessible online, the temporal trends may not represent societal bias evolution.
  • domain assumption Fine-tuning a pre-trained LLM on a text corpus causes the model's open-ended completions to reflect the statistical associations in that corpus, and these associations can be elicited by sentence-completion prompts.
    Sections 3 and 4. The entire method rests on this linking premise; the entity-overlap validation shows entity-level learning, but not that semantic bias associations transfer from corpus to prompted outputs.
  • domain assumption GPT-4's classification of responses into demographic subcategories is accurate enough for the reported percentages.
    Section 4.2 and Appendix A.3. They report 92% agreement with human labels on 50 cases, but systematic errors in the 8% could be correlated with the trends being measured.
  • domain assumption The base models (Llama-3-70B, Mixtral, Gemini) have no pre-existing temporal bias differences across decades that interact with the fine-tuning.
    The comparison to pre-trained models is used to attribute changes to the books, but if the base model's prior biases are not uniform across the prompt categories, the fine-tuned/pre-trained difference may misattribute effects. This is partially controlled by using three different base models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fine-Tuned LLMs are "Time Capsules" for Tracking Societal Bias Through Books." pith.science (2026). https://pith.science/paper/QAI3HPK3

@misc{pith2026250205331,
  author       = {Pith},
  title        = {Pith review of: Fine-Tuned LLMs are "Time Capsules" for Tracking Societal Bias Through Books},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QAI3HPK3}},
  note         = {Machine review of arXiv:2502.05331}
}
read the original abstract

Books, while often rich in cultural insights, can also mirror societal biases of their eras - biases that Large Language Models (LLMs) may learn and perpetuate during training. We introduce a novel method to trace and quantify these biases using fine-tuned LLMs. We develop BookPAGE, a corpus comprising 593 fictional books across seven decades (1950-2019), to track bias evolution. By fine-tuning LLMs on books from each decade and using targeted prompts, we examine shifts in biases related to gender, sexual orientation, race, and religion. Our findings indicate that LLMs trained on decade-specific books manifest biases reflective of their times, with both gradual trends and notable shifts. For example, model responses showed a progressive increase in the portrayal of women in leadership roles (from 8% to 22%) from the 1950s to 2010s, with a significant uptick in the 1990s (from 4% to 12%), possibly aligning with third-wave feminism. Same-sex relationship references increased markedly from the 1980s to 2000s (from 0% to 10%), mirroring growing LGBTQ+ visibility. Concerningly, negative portrayals of Islam rose sharply in the 2000s (26% to 38%), likely reflecting post-9/11 sentiments. Importantly, we demonstrate that these biases stem mainly from the books' content and not the models' architecture or initial training. Our study offers a new perspective on societal bias trends by bridging AI, literary studies, and social science research.

Figures

Figures reproduced from arXiv: 2502.05331 by the authors.

Figure 1
Figure 1. Temporal trends in Llama 3 70B’s responses [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of methodology from corpus creation to bias analysis, with section references [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Entity extraction and comparison for a single [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Standard deviation of subcategory frequencies [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Temporal trends in Llama 3 70B’s responses [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Temporal trends in Llama 3 70B’s responses to racial associations with mathematicians. The lines [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Number of male and female authors over decades 1950-59 1960-69 1970-79 1980-89 1990-99 2000-09 2010-19 0 20 40 60 Number of Authors Heterosexual Homosexual Bisexual Unknown [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Number of authors by sexual orientation over decades 1950-59 1960-69 1970-79 1980-89 1990-99 2000-09 2010-19 0 20 40 60 Number of Authors White Black Asian Mixed [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 12
Figure 12. Figure 12: Average percentage responses, over decades and prompts, categorized into each subcategory for the [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Average percentage responses, over decades and prompts, categorized into each subcategory for the [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 14
Figure 14. Figure 14: Average percentage responses, over decades and prompts, categorized into each subcategory for the [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 42 canonical work pages

  1. [1]

    Abubakar Abid, Maheen Farooqi, and James Zou. 2021. http://arxiv.org/abs/2101.05783 Persistent anti-muslim bias in large language models

    arXiv 2021

  2. [2]

    Fabbri, Simeng Han, Wojciech Kryscinski, Faisal Ladhak, Bryan Li, Kathleen McKeown, Dragomir Radev, Tianyi Zhang, and Sam Wiseman

    Divyansh Agarwal, Alexander R. Fabbri, Simeng Han, Wojciech Kryscinski, Faisal Ladhak, Bryan Li, Kathleen McKeown, Dragomir Radev, Tianyi Zhang, and Sam Wiseman. 2022. https://aclanthology.org/2022.creativesumm-1.10 CREATIVESUMM : Shared task on automatic summarization for creative writing . In Proceedings of The Workshop on Automatic Summarization for Cr...

    work page 2022

  3. [3]

    Lochan Basyal and Mihir Sanghvi. 2023. http://arxiv.org/abs/2310.10449 Text summarization using large language models: A comparative study of mpt-7b-instruct, falcon-7b-instruct, and openai chat-gpt models

    arXiv 2023

  4. [4]

    Paul Batesel. 1989. http://www.jstor.org/stable/23414450 Best sellers and the public attitude . Studies in Popular Culture, 12(1):15--27

    work page arXiv 1989

  5. [5]

    Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. 2016. http://arxiv.org/abs/1607.06520 Man is to computer programmer as woman is to homemaker? debiasing word embeddings

    arXiv 2016

  6. [6]

    Nadav Borenstein, Karolina Stańczak, Thea Rolskov, Natália da Silva Perez, Natacha Klein Käfer, and Isabelle Augenstein. 2023. http://arxiv.org/abs/2305.12376 Measuring intersectional biases in historical documents

    work page Pith review arXiv 2023

  7. [7]

    Boutwell, Joseph L

    Brian B. Boutwell, Joseph L. Nedelec, Bo Winegard, Todd Shackelford, Kevin M. Beaver, Michael Vaughn, J. C. Barnes, and John P. Wright. 2017. https://doi.org/10.1371/journal.pone.0183356 The prevalence of discrimination across racial groups in contemporary america: Results from a nationally representative sample of adults . PLOS ONE, 12(8):1--8

    work page doi:10.1371/journal.pone.0183356 2017

  8. [8]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

Show all 74 references
  1. [9]

    Marc-Etienne Brunet, Colleen Alkalay-Houlihan, Ashton Anderson, and Richard Zemel. 2019. http://arxiv.org/abs/1810.03611 Understanding the origins of bias in word embeddings

    2019 arXiv

  2. [10]

    Jose Camacho-Collados, Luis Espinosa Anke, and Steven Schockaert. 2019. https://doi.org/10.18653/v1/P19-1318 Relational word embeddings . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3286--3296, Florence, Italy. Association ...

    2019 doi

  3. [11]

    Michael Cart and Christine Jenkins. 2006. The Heart Has Its Reasons: Young Adult Literature with Gay/Lesbian/Queer Content, 1969-2004. Studies in Young Adult Literature. Scarecrow Press

    2006

  4. [12]

    Chloe Comeau. 2024. Diversity in publishing: Does author identity affect author treatment in the north american fiction publishing industry? Academic Leadership Journal in Student Research, 7(2)

    2024

  5. [13]

    Alison Cook and Christy Glass. 2014. https://doi.org/10.1111/gwao.12018 Women and top leadership positions: Towards an institutional analysis . Gender, Work & Organization, 21(1):91--103. First published online: 21 March 2013

    2014 doi

  6. [14]

    Rob Cover. 2022. https://doi.org/10.1177/1363460720982924 Populist contestations: Cultural change and the competing languages of sexual and gender identity . Sexualities, 25(5-6):660--675

    2022 doi

  7. [15]

    Kate Crawford. 2017. https://neurips.cc/virtual/2017/invited-talk/8742 The trouble with bias . NeurIPS invited talk

    2017

  8. [16]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. http://arxiv.org/abs/1810.04805 Bert: Pre-training of deep bidirectional transformers for language understanding

    2019 arXiv

  9. [17]

    Manaal Faruqui, Yulia Tsvetkov, Pushpendre Rastogi, and Chris Dyer. 2016. https://doi.org/10.18653/v1/W16-2506 Problems with evaluation of word embeddings using word similarity tasks . In Proceedings of the 1st Workshop on Evaluating Vector-Space Representations for NLP , page...

    2016 doi

  10. [18]

    Jonathan Fox. 2017. https://doi.org/10.1007/s41682-017-0009-3 Religious discrimination in european and western christian-majority democracies . Zeitschrift für Religion, Gesellschaft und Politik, 1(2):185--209

    2017 doi

  11. [19]

    Giorgio Franceschelli and Mirco Musolesi. 2023. http://arxiv.org/abs/2304.00008 On the creativity of large language models

    2023 arXiv

  12. [20]

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. http://arxiv.org/abs/2101.00027 The pile: An 800gb dataset of diverse text for language modeling

    2020 arXiv

  13. [21]

    Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. https://doi.org/10.1073/pnas.1720347115 Word embeddings quantify 100 years of gender and ethnic stereotypes . Proceedings of the National Academy of Sciences, 115(16):E3635--E3644

    2018 doi

  14. [22]

    John Gaston. 2022. Invisibility in the academy: Representation of black scientists in stem literature. Journal of STEM Education, 23(1):35--48

    2022

  15. [23]

    Gee, Annie Ro, Salma Shariff-Marco, and David Chae

    Gilbert C. Gee, Annie Ro, Salma Shariff-Marco, and David Chae. 2009. https://doi.org/10.1093/epirev/mxp009 Racial Discrimination and Health Among Asian Americans: Evidence, Assessment, and Directions for Future Research . Epidemiologic Reviews, 31(1):130--151

    2009 doi

  16. [24]

    Sandra M Gilbert and Susan Gubar. 2020. The madwoman in the attic: The woman writer and the nineteenth-century literary imagination. Yale University Press

    2020

  17. [25]

    Yoav Goldberg and Jon Orwant. 2013. https://aclanthology.org/S13-1035 A dataset of syntactic-ngrams over time from a very large corpus of E nglish books . In Second Joint Conference on Lexical and Computational Semantics (* SEM ), Volume 1: Proceedings of the Main Conference a...

    2013

  18. [26]

    Hila Gonen and Yoav Goldberg. 2019. https://doi.org/10.18653/v1/N19-1061 Lipstick on a pig: D ebiasing methods cover up systematic gender biases in word embeddings but do not remove them . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association ...

    2019 doi

  19. [27]

    Angela M Gooden and Mark A Gooden. 2001. Gender representation in notable children's picture books: 1995--1999. Sex roles, 45:89--101

    2001

  20. [28]

    Larry Gross. 2001. Up from invisibility: Lesbians, gay men, and the media in America. Columbia University Press

    2001

  21. [29]

    Patricia Snell Herzog, David P King, Rafia A Khader, Amy Strohmeier, and Andrew L Williams. 2020. Studying religiosity and spirituality: A review of macro, micro, and meso-level approaches. Religions, 11(9):437

    2020

  22. [30]

    Leslie Heywood and Jennifer Drake. 1997. Third wave agenda: Being feminist, doing feminism. U of Minnesota Press

    1997

  23. [31]

    Matthew Honnibal and Ines Montani. 2017. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. Software available from https://spacy.io

    2017

  24. [32]

    Hoobler, Courtney R

    Jenny M. Hoobler, Courtney R. Masterson, Stella M. Nkomo, and Eric J. Michel. 2018. https://doi.org/10.1177/0149206316628643 The business case for women leaders: Meta-analysis, research critique, and path forward . Journal of Management, 44(6):2473--2499

    2018 doi

  25. [33]

    Md Naimul Hoque, Bhavya Ghai, and Niklas Elmqvist. 2022. https://doi.org/10.1145/3532106.3533526 Dramatvis personae: Visual text analytics for identifying social biases in creative writing . In Proceedings of the 2022 ACM Designing Interactive Systems Conference, DIS '22, page...

    2022

  26. [34]

    Ynda Jas. 2020. https://doi.org/10.3224/insep.si2020.05 Sexuality in a non-binary world: redefining and expanding the linguistic repertoire . INSEP -- Journal of the International Network for Sexual Ethics & Politics, 8:71--92

    2020 doi

  27. [35]

    Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. 2024. Mixtral of experts. arXiv preprint arXiv:2401.04088

    2024 arXiv

  28. [36]

    Mali Jin, Yida Mu, Diana Maynard, and Kalina Bontcheva. 2023. http://arxiv.org/abs/2309.14146 Examining temporal bias in abusive language detection

    2023 arXiv

  29. [37]

    Muhammed Yusuf Kocyigit, Anietie Andy, and Derry Wijaya. 2023. http://arxiv.org/abs/2310.15847 A novel method for analysing racial bias: Collection of person level references

    2023 arXiv

  30. [38]

    Tomáš Kočiský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Edward Grefenstette. 2017. http://arxiv.org/abs/1712.07040 The narrativeqa reading comprehension challenge

    2017 arXiv

  31. [39]

    Wojciech Kryściński, Nazneen Rajani, Divyansh Agarwal, Caiming Xiong, and Dragomir Radev. 2022. http://arxiv.org/abs/2105.08209 Booksum: A collection of datasets for long-form narrative summarization

    2022 arXiv

  32. [40]

    Faisal Ladhak, Bryan Li, Yaser Al-Onaizan, and Kathleen McKeown. 2020. https://doi.org/10.18653/v1/2020.acl-main.453 Exploring content selection in summarization of novel chapters . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pag...

    2020 doi

  33. [41]

    Waterman, Emry Breedlove, Jarvis T

    Maddalena Marini, Pamela D. Waterman, Emry Breedlove, Jarvis T. Chen, Christian Testa, Sari L. Reisner, Dana J. Pardee, Kenneth H. Mayer, and Nancy Krieger. 2021. https://doi.org/10.1186/s12889-021-10171-7 The target/perpetrator brief-implicit association test (b-iat): an impl...

    2021 doi

  34. [42]

    McCarty and Anna H

    Megan K. McCarty and Anna H. Burt. 2024. https://doi.org/10.1007/s11199-024-01449-2 Understanding perceptions of gender non-binary people: Consensual and unique stereotypes and prejudice . Sex Roles, 90(3):392--416

    2024 doi

  35. [43]

    Jean-Baptiste Michel, Yuan Kui Shen, Aviva Presser Aiden, Adrian Veres, Matthew K Gray, Google Books Team, Joseph P Pickett, Dale Hoiberg, Dan Clancy, Peter Norvig, et al. 2011. Quantitative analysis of culture using millions of digitized books. science, 331(6014):176--182

    2011

  36. [44]

    Franco Moretti. 2013. Distant reading, volume 93. Verso

    2013

  37. [45]

    Peter Morey and Amina Yaqin. 2011. https://www.hup.harvard.edu/catalog.php?isbn=9780674048522 Framing Muslims: Stereotyping and Representation after 9/11 . Harvard University Press, Cambridge, MA

    2011

  38. [46]

    Yida Mu, Kalina Bontcheva, and Nikolaos Aletras. 2023. https://doi.org/10.18653/v1/2023.findings-eacl.55 It ' s about time: Rethinking evaluation on rumor detection benchmarks using chronological splits . In Findings of the Association for Computational Linguistics: EACL 2023,...

    2023 doi

  39. [47]

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. 2024. http://arxiv.org/abs/2307.06435 A comprehensive overview of large language models

    2024 arXiv

  40. [48]

    Malvina Nissim, Rik van Noord, and Rob van der Goot. 2020. https://doi.org/10.1162/coli_a_00379 Fair is better than sensational: Man is to doctor as woman is to doctor . Computational Linguistics, 46(2):487--497

    2020 doi

  41. [49]

    Michael Oliver and Guan Wang. 2024. http://arxiv.org/abs/2407.13906 Crafting efficient fine-tuning strategies for large language models

    2024 arXiv

  42. [50]

    Alexandra Olteanu, Carlos Castillo, Fernando Diaz, and Emre K c man. 2019. Social data: Biases, methodological pitfalls, and ethical boundaries. Frontiers in big data, 2:13

    2019

  43. [51]

    Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532--1543

    2014

  44. [52]

    Project gutenberg

    Project Gutenberg . Project gutenberg. https://www.gutenberg.org. Accessed: 2024-10-05

    2024

  45. [53]

    Ryan and S

    Michelle K. Ryan and S. Alexander Haslam. 2005. https://doi.org/10.1111/j.1467-8551.2005.00433.x The glass cliff: Evidence that women are over-represented in precarious leadership positions . British Journal of Management, 16(2):81--90

    2005

  46. [54]

    Margaret Scanlan. 2001. http://www.jstor.org/stable/j.ctt6wrpv0 Plotting Terror: Novelists and Terrorists in Contemporary Fiction . University of Virginia Press

    2001

  47. [55]

    Alessandro Scir \`e , Simone Conia, Simone Ciciliano, and Roberto Navigli. 2023. https://doi.org/10.18653/v1/2023.findings-acl.54 Echoes from alexandria: A large resource for multilingual book summarization . In Findings of the Association for Computational Linguistics: ACL 20...

    2023 doi

  48. [56]

    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

  49. [57]

    Nikki Sullivan. 2003. http://www.jstor.org/stable/10.3366/j.ctvxcrwj6 A Critical Introduction to Queer Theory . Edinburgh University Press

    2003 doi

  50. [58]

    John Sutherland. 2007. https://doi.org/10.1093/actrade/9780199214891.001.0001 Bestsellers: A Very Short Introduction . Oxford University Press

    2007

  51. [59]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. https://arxiv.org/abs/2312.11805 Gemini: A family of highly capable multimodal models . ArXiv preprint arXiv:2312.11805

    2023 arXiv

  52. [60]

    Paulina Toro Isaza, Guangxuan Xu, Toye Oloko, Yufang Hou, Nanyun Peng, and Dakuo Wang. 2023. https://doi.org/10.18653/v1/2023.acl-long.359 Are fairy tales fair? analyzing gender bias in temporal narrative event chains of children ' s fairy tales . In Proceedings of the 61st An...

    2023 doi

  53. [61]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  54. [62]

    Rocco Tripodi, Massimo Warglien, Simon Levis Sullam, and Deborah Paci. 2019. https://doi.org/10.18653/v1/W19-4715 Tracing antisemitic language through diachronic embedding projections: F rance 1789-1914 . In Proceedings of the 1st International Workshop on Computational Approa...

    2019 doi

  55. [63]

    Ted Underwood, David Bamman, and Sabrina Lee. 2018. https://doi.org/10.22148/16.019 The Transformation of Gender in English - Language Fiction . Journal of Cultural Analytics, 3(2)

    2018 doi

  56. [64]

    Gerry Veenstra. 2011. https://doi.org/10.1186/1475-9276-10-3 Race, gender, class, and sexual orientation: intersecting axes of inequality and self-rated health in canada . International Journal for Equity in Health, 10(1):3

    2011 doi

  57. [65]

    Kara Wagner. 2016. https://doi.org/10.5621/sciefictstud.43.1.0089 The invisible scientist: Representations of black characters in science fiction . Science Fiction Studies, 43(1):89--104

    2016 doi

  58. [66]

    Suzanna Danuta Walters. 2003. All the rage: The story of gay visibility in America. University of Chicago Press

    2003

  59. [67]

    Yu, and Qingsong Wen

    Shen Wang, Tianlong Xu, Hang Li, Chaoli Zhang, Joleen Liang, Jiliang Tang, Philip S. Yu, and Qingsong Wen. 2024 a . http://arxiv.org/abs/2403.18105 Large language models for education: A survey and outlook

    2024 arXiv

  60. [68]

    Tiannan Wang, Jiamin Chen, Qingrui Jia, Shuai Wang, Ruoyu Fang, Huilin Wang, Zhaowei Gao, Chunzhao Xie, Chuou Xu, Jihong Dai, Yibin Liu, Jialong Wu, Shengwei Ding, Long Li, Zhiwei Huang, Xinle Deng, Teng Yu, Gangan Ma, Han Xiao, Zixin Chen, Danjun Xiang, Yunxia Wang, Yuanyuan ...

    2024 arXiv

  61. [69]

    Dominik Wurzer and Yumeng Qin. 2020. https://doi.org/10.1145/3397271.3401306 How umass-fsd inadvertently leverages temporal bias . In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’20, page 2097–2100. ACM

    2020

  62. [70]

    Huimin Xu, Zhang Zhang, Lingfei Wu, and Cheng-Jun Wang. 2019. The cinderella complex: Word embeddings reveal gender stereotypes in movies and books. PloS one, 14(11):e0225385

    2019

  63. [71]

    Zhixue Zhao, George Chrysostomou, Kalina Bontcheva, and Nikolaos Aletras. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.298 On the impact of temporal concept drift on model explanations . In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 4...

    2022 doi

  64. [72]

    Yukun Zhu, Ryan Kiros, Richard Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. june 2015. arXiv preprint arXiv:1506.06724

    2015 arXiv

  65. [73]

    URL: " 'urlintro :=

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

  66. [74]

    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 8, 2026 · model on record in the stance chip above.