PaperlayerView exact revision on arXiv

Source-linked AI summary

Crossing the Gender Binary: Modelling Gendered Reference in Science Fiction and General Fiction

Yuxuan Liu, Haim Dubossarsky, Ruth Ahnert

arXiv:2610.19028v1cs.CL

TL;DR

The paper asks whether science fiction’s imagined rearrangements of gender also appear in linguistic expectations, a question less examined than language-model bias around supposedly stable gender categories. It masks man, woman, and person in same-period science fiction and general-fiction corpora and classifies likely substitutes. Both genres show uneven crossover, while science fiction shows stronger neutralization and distinct gender-permeability patterns.

  • Problem

    Research on language-model gender bias has largely treated gender categories as stable, leaving their linguistic permeability and possible genre differences comparatively underexamined.

    Source: §1 · p. 2

  • Method

    The study masks man, woman, and person in literary corpora and classifies likely substitutes to compare gendered, neutral, and socially marked reference patterns.

    Source: §1 · p. 1

  • Results

    Both corpora show uneven gender crossover, while science fiction has greater movement from masculine to feminine reference and stronger movement of gendered terms toward neutral reference.

    Source: §6 · p. 14

  • Takeaways & Limitations

    The findings provide linguistic evidence for examining gender in science fiction and show how masked prediction can trace relations among social categories in language.

    Source: §6 · p. 14

    The social-category candidate sets are too infrequent to support a genre-level claim, and the three sexual-minority terms found for man/men are too few for positive claims about their contexts.

    Source: §5.4 · p. 13

Abstract

from arXiv · show

Science fiction is widely credited with unsettling gender. This article asks whether that holds at the linguistic level, in how readily gendered reference gives way to its counterpart, a neutral term, or other social category, and whether it separates the genre from general fiction. It does so by masking man/men, woman/women, and person/people across same-period science fiction and general fiction corpora and classifying RoBERTa's five likeliest substitutes by gender. Retention and displacement are permutation-tested over five resampling rounds, and the underlying sentences are read closely. Both corpora cross the gender binary in both directions, but unequally: where woman is masked, the model predicts a masculine term about twice as often as it predicts a feminine term where man is masked. Each mismatch marks a statistical surprise, a word the corpus did not expect there. Against this shared pattern, science fiction shows greater gender permeability than the general fiction corpus, placing man in feminine contexts more often than general fiction while the reverse rate holds steady. Its degendering tendency is stronger too, drawing both gendered terms toward the neutral. Beyond the binary, a masked woman also draws substitutes naming a social category and no gender far more often than a masked man does. In measuring the gender permeability of man, woman, and person in linguistic distribution and testing whether genre changes it, the study gives science fiction studies linguistic evidence for thematic claims and computational literary studies a pipeline applicable to other category vocabularies, such as race and class.

1. Introduction

The article asks whether science fiction’s imagined rearrangements of gender also appear in linguistic expectations around human-reference terms. It uses masked language modelling to compare those expectations with general fiction and examine how gendered terms give way to neutral or socially marked alternatives.

  • Masked language modelling hides a word and ranks likely replacements, making accumulated usage patterns measurable as expectations for a masked term.Source: §1 · pp. 1–2
  • Prior work on language-model bias has largely treated gender categories as stable, leaving their internal linguistic coherence and genre-dependent permeability comparatively underexamined.Source: §1 · p. 2
  • The analysis asks how often man, woman, and person are retained or displaced, which directions substitutions take, and whether science fiction differs from general fiction.Source: §1 · p. 2

2. Theoretical Framework

The framework treats gendered reference as a distributional pattern built through repeated language use, with a genericity axis for comparing gender-specific and neutral terms. It asks whether genre alters that pattern while cautioning that substitutions require contextual interpretation.

2.1 Gendered Intelligibility and Human-Reference Terms

  • Gendered meanings can accumulate and stabilize through repeated linguistic use, so substitutions among man, woman, and neutral reference can reveal how those meanings are organized.Source: §2.1 · p. 2
  • The genericity axis runs from gender-specific reference to generic human reference, defined here as identifying a human without specifying sex or social gender.Source: §2.1 · p. 2

2.2 The Genericity Axis

  • The axis plots word distributions rather than lived gender identity and reduces intersecting categories to examine gendered reference as one linguistic dimension.Source: §2.2 · p. 3
  • Because term configurations accumulate from usage contexts, general English provides a baseline for testing whether literary genre alters them.Source: §2.2 · p. 3

2.3 Genre as a Condition of the Axis

  • Earlier scholarship addresses gender in science fiction through theme and character, while this study tests whether the genre also has a distinct distribution of gendered linguistic expectations.Source: §2.3 · pp. 3–4
  • A departure from expected gendered reference is an opening for interpretation, not by itself evidence of subversion; its meaning depends on context and close reading.Source: §2.3 · p. 4

3. Materials and Methods

The study compares a science-fiction corpus with same-period general fiction, samples sentences containing three human-reference lemmas, and classifies masked-word predictions. It measures retention, substitution, and prediction uncertainty, then tests differences across corpora.

  • The comparison uses 336 Gollancz SF Masterworks works and 700 randomly sampled Anglophone NovelTM novels from the same period.Source: §3 · p. 4
  • For each target lemma, the researchers randomly sampled 500 sentences per round across five rounds, removing generic-mankind uses from man/men before sampling.Source: §3 · p. 4
  • RoBERTa produced each masked position’s top five lexical predictions, which Gemini-based classification assigned to gender categories or other-category and unknown labels.Source: §3 · p. 4
  • Retention, substitution, and entropy summarize category stability, replacement direction, and prediction certainty; corpus differences use permutation tests with bootstrap intervals, while entropy uses ANOVA and corrected post-hoc tests.Source: §3 · p. 4

4. Results

Across both corpora, woman has the lowest within-category retention, and gender crossover is bidirectional but uneven. Science fiction strengthens neutral reference and shifts some prediction patterns, while several changes for man and woman remain descriptive after correction.

  • 57.4% versus 52.4% Neutral-Gender retention for person/people, 39.8% versus 41.7% Male retention for man, and 34.4% versus 34.3% Female retention for woman preserve the same ordering across Gollancz and NovelTM.Source: §4 · pp. 4, 6
  • 16.8% versus 8.2% in NovelTM and 17.2% versus 9.5% in Gollancz show woman-to-masculine crossover exceeds man-to-feminine crossover in both corpora.The between-direction differences are significant in both corpora (Gollancz ∆= −7.7 pp, p < .001; NovelTM ∆≈−8.5 pp, p < .001).Source: §4 · p. 6
  • Gollancz person/people has lower entropy than NovelTM (H = 1.5931 vs. H = 1.6504; ∆= −.057, p < .001), alongside higher Neutral-Gender retention (57.4% vs. 52.4%; ∆= +5.0 pp, p < .001).Source: §4 · p. 6
  • Neutral-Gender crossover is higher in Gollancz for man (30.1% vs. 27.7%) and woman (26.9% vs. 24.2%), while their entropy differences do not survive Bonferroni correction.Source: §4 · p. 6

5. Discussion

The discussion considers how science fiction and general fiction differ in neutral reference and gendered crossings, while noting persistent asymmetries and women’s wider social classifications. It also examines contrasting political passages and uses the model’s full candidate distribution to connect sentence-level predictions with broader linguistic relations.

5.1 The Masculine Neutral: Generalization, Typification, and Degendering

  • Person and people remain neutral in 52.4% of NovelTM and 57.4% of Gollancz cases, but when gendered they tilt masculine at roughly twice the feminine rate.The masculine neutral is associated with technical origination and group action in Gollancz, and with honour, professional standing, and gentlemanly virtues in NovelTM.Source: §5.1 · p. 6
  • Science fiction predicts neutral terms more often than general fiction for man/men (30.1% vs. 27.7%) and woman/women (26.9% vs. 24.2%), without erasing their different routes to neutrality.Man typically reaches person through generalization, while woman reaches it through typification; the asymmetry persists even when gender falls away from reference.Source: §5.1 · p. 8

5.2 Woman as Man: Capability on Masculine Ground

  • Masked woman/women receives a masculine prediction at nearly identical rates in NovelTM (16.8%) and Gollancz (17.2%), often in contexts of expertise, professional activity, authority, or practical capability.A fictional society governed by women still draws strongly masculine predictions, and masculine-coded capability reaching woman does not by itself show that associations around woman have changed. Estimates pool corpora rather than decades because the smaller period samples would be unstable.Source: §5.2 · pp. 8–10

5.3 Man as Woman: Feminization within and beyond Gendered Conventions

  • Feminine-aligned candidates appear for masked man/men less often than masculine candidates do for woman/women: 8.2% in NovelTM and 9.5% in Gollancz; male retention falls from 41.7% to 39.8%.Some crossings feminize men through associations such as frailty or emotional dependence; others arise when fictional conditions redistribute gendered positions, yet the model still often restores familiar associations.Source: §5.3 · pp. 11–12

5.4 Woman and the Wider Field of Social Difference

  • Woman/women candidate sets include sexuality-related terms at 13 Gollancz positions and 8 NovelTM positions, and other social categories at 27 and 19 positions, respectively.These counts are too rare to support a genre-level claim; the shared pattern is that such classifications are largely absent from corresponding man/men sets.Source: §5.4 · pp. 12–13
  • Lower-ranked candidates around woman can reveal a wider discourse of classification and exclusion, beyond the model’s most likely gendered completion.The candidate sets place sexuality-related terms alongside women and other socially marked groups, showing neighboring categories in the local linguistic space rather than equating those groups.Source: §5.4 · pp. 12–13
  • Despite contrasting political purposes, Wells’s and Jones’s candidate sets reveal a shared categorical grammar connecting woman with sexual and racial minorities.Wells’s exclusionary prose pushes the model toward homosexuals, while Jones’s advocacy-oriented prose pushes it toward gays.Source: §5.4 · p. 14
  • The full candidate distribution traces the linguistic field around a word, including near-displacing terms, recurring categories, and associations implicit in the text.It can connect an unexpected sentence-level prediction to wider discourse and historically sedimented associations, rather than uncovering a hidden meaning.Source: §5.4 · p. 14
  • Close reading identifies a passage’s explicit argument, while its candidate set shows how its language sits within a wider network of associations.The method links local textual friction to systemic linguistic relations that make an unexpected prediction possible.Source: §5.4 · p. 14

6. Conclusion

Both literary corpora show uneven movement across gendered reference, while science fiction changes some predictive patterns without erasing the broader asymmetries. The study brings computational distributions and close reading together to make relations among linguistic expectations and social categories perceptible.

6. Conclusion

  • Woman admits masculine reconstruction roughly one in six times, versus fewer than one in ten masked man occurrences admitting a feminine one.Neutral reference also differs by route: person/people tends masculine, man generalizes, and woman tends toward typification.Source: §6 · p. 14
  • Science fiction increases movement from masculine to neutral reference and permeability from the masculine side, while woman-to-man movement remains stable across corpora.Neutral human reference becomes more predictable and predictive concentration shifts away from man toward person/people, even as overall linguistic weighting remains largely unchanged.Source: §6 · p. 14
  • Woman’s prediction neighbourhood includes sexuality, race, religion, and other marked categories largely absent from man’s.These associations are not evidence that the categories are equivalent or measures of authorial intention.Source: §6 · p. 14
  • The model’s distributions complement interpretation by linking individual passages with aggregate linguistic patterns and historical contexts.Mask and Classify is presented as a framework for tracing conceptual relations through language, including how categories approach, displace, or resist one another.Source: §6 · pp. 14–15

AI Disclosure Statement

Large language models supported specified annotation steps in the study; the statement distinguishes those uses from the creation of the manuscript.

AI Disclosure Statement

  • Gemini 2.5 Pro, DeepSeek-chat, and Llama 3.3 70B annotated word senses for man/men occurrences, with a sample checked against independent human annotation.The procedure is described in Section 3 and note 1.Source: AI Disclosure Statement · p. 15
  • Gemini 2.5 Pro assigned category labels to predictions returned by roberta-base.The procedure is described in Section 3 and note 1.Source: AI Disclosure Statement · p. 15
  • The authors state that no generative AI beyond the specified method uses was used to create the manuscript.Source: AI Disclosure Statement · p. 15

Notes

The note describes a multi-model annotation procedure for man/men senses, including deduplication, majority labeling, and a human reliability check.

Notes

  • The 4,928 unique man/men sentences comprised 2,444 from Gollancz SF and 2,484 from NovelTM.Repeated sentences from resampling were deduplicated before annotation, then labels were mapped back to sampled occurrences.Source: Notes · p. 15
  • Three language models annotated shuffled batches in five independent rounds, with majority labels across rounds and then models determining each final sense assignment.Instances classified as generic-mankind were removed before the main sampling step.Source: Notes · p. 15
  • Two human annotators independently labeled 100 sentences to estimate automated annotation accuracy across the full sentence pool.Chi-square tests assessed whether the three models produced consistent sense distributions.Source: Notes · p. 15
Loading 2610.19028v1…