Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 58 of the 58 references it cites were found in a registry or an index; 13 of the 51 sources its claims cite were checked for support.

“Real men in the battlefield”: a hybrid AI–discourse analysis of online homophobia and transphobia

Open access
Baider, Fabienne, Mesaritou, Orthodoxia · 2026 · Frontiers in Communication

Abstract

This qualitative study advances research on homophobic and transphobic discourse by combining critical discourse analysis with AI-assisted methods. While previous work has identified the topos of threat as a common feature of both, the present study shifts the focus to emotional strategies, examining how contempt and disgust operate alongside argumentative framing. The analysis draws on a corpus of approximately 2,000 annotated online comments (1,000 homophobic and 1,000 transphobic). AI-based and lexicon-driven sentiment analysis, supported by statistical testing, complement a fine-grained qualitative comparison of affective and discursive patterns. The findings reveal that contempt overwhelmingly structures discourse targeting both communities, whereas disgust appears more prominently in homophobic comments. Rather than functioning as a reactive response to perceived threat, contempt emerges as a stabilizing emotion that reinforces social hierarchies, moral superiority, gender policing, and—in our data—forms of institutional exclusion. In terms of framing, the misfit social frame dominates across both datasets, while the medical frame is more strongly associated with transphobic discourse. Overall, the results point to a shared affective architecture centered on contempt, with disgust playing a more marginal but differentiated role. At the same time, homophobic discourse exhibits a higher proportion and intensity of negative content, suggesting subtle differences in tone and interactional dynamics. Discussions of transphobic speech, by contrast, show a greater presence of counter-narratives. Finally, we explore the extent to which these features support the automatic detection of homophobic versus transphobic discourse. While classification remains moderately accurate, emotional and framing features prove complementary to lexical cues when combined. To our knowledge, this is the first study to systematically contrast homophobic and transphobic discourse through the joint lens of emotion and framing, offering insights into the development of more targeted counter-discursive strategies.

Overview

CoChat Summary

TL;DR

This mixed-methods study combines critical discourse analysis with AI-assisted sentiment analysis and machine learning to compare homophobic and transphobic Facebook comments. It finds that contempt, not disgust, dominates both discourses, while medical framing is more characteristic of transphobia and misfit framing of homophobia.

Key findings

  • Contempt was the dominant emotion in both datasets: 34.3% (332/969) of homophobic comments and 25.4% (288/1,133) of transphobic comments expressed contempt, versus 2.3% (22) and 1.1% (12) for disgust; a McNemar test showed contempt occurred significantly more often than disgust (χ² = 538.09, p < 0.001).
  • Framing differed significantly by dataset: the medical frame accounted for 78.54% of transphobic comments versus 6.16% of homophobic comments, while the misfit frame accounted for 93.84% of homophobic comments versus 21.46% of transphobic comments (χ² = 16.67, p < 0.001).
  • Homophobic discourse had more negative content (34.16%) than transphobic discourse (27.63%; χ² = 10.19, p = 0.0014, φ = 0.07), while counter-narratives appeared more often in transphobic discussions (25.95%) than homophobic ones (16.82%).
  • Automatic classification was moderate: text-only logistic regression achieved macro F1 = 0.766 and accuracy = 0.768; text + emotions achieved F1 = 0.767 and accuracy = 0.770; the full model achieved F1 = 0.744 and accuracy = 0.744.

Methods

  • The corpus contained 2,102 English Facebook comments from 2017: 1,133 transphobic comments from one BBC article (“UK military chiefs praise transgender troops”) and 969 homophobic comments from two articles (“Australia rugby team same-sex marriage spat,” 468 comments; “Sydney Uni marriage protest: ‘Yes’ voters clash with ‘No’ voters,” 501 comments).
  • Two expert annotators manually coded emotions (disgust, contempt), frames (medical, misfit, other), and IMSyPP categories (acceptable, inappropriate, offensive, violent); reliability was assessed with percent agreement, Krippendorff’s alpha, and Gwet’s AC1.
  • Sentiment analysis used VADER, TextBlob, and a HuggingFace transformer model, with only HuggingFace reported in detail; statistical analysis was conducted in R and Python, and classification used logistic regression with TF-IDF and feature ablations A1–A6 under 5-fold stratified cross-validation.

Limitations

  • The paper states the sample is small, the dataset cannot be shared, the transphobia data derive from a single article while homophobia aggregates two sources, and this asymmetry may introduce topical and contextual discrepancies; it also notes that 2017 Facebook data may be challenged by later changes in online discourse and platform norms.
  • The authors state results should be interpreted cautiously and confirmed with larger, more diverse datasets; they plan a follow-up with a larger annotated sample and other machine learning programs.
  • Evident from the text: disgust was very rare, and despite high percent agreement, Krippendorff’s alpha for disgust was 0.54 in transphobia and 0.61 in homophobia, indicating chance-corrected reliability concerns for that low-frequency category; classification performance was only moderate and added features did not improve over text-only.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fcomm.2026.1748828

Cites

58 of 58 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
From the PDF
Image 2
From the PDF
Image 3
From the PDF
Image 4
From the PDF

Verification

Claims the paper makes

13 Valid · 1 Not valid · 16 Not checked · 129 Own results and interpretations

75 more claims · 1 with their cited sources checked

Create a free account to see all 159 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 44
  • It may state two separate claims as one. 49

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 53 references

Research works printed but not linked: 4

  1. Allport, 1954, The nature of prejudice
  2. Fillmore, 1982, Linguistics in the morning calm: Selected papers from SICOL- 1981, 111
  3. Lillian, 2005, Homophobic discourse: A “popular” Canadian example, Sky Journal of Linguistics, 18, 119
  4. Rozin, 2008, Handbook of emotions, 757

Metadata

Authors
Baider, F., & Mesaritou, O.
Affiliations
University of Cyprus (CY)
Corresponding author
Baider, Fabienne
Venue
Frontiers in Communication
Published
2026-04-10
Volume
11
ISSN
2297-900X (ISSN-L)
OpenAlex
W7153037528
Cited by
0
References
53
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Communication (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Social Sciences › Psychology › Social Psychology › LGBTQ Health, Identity, and Policy
  • Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
  • Social Sciences › Gender Studies › Gender and Feminist Studies
Keywords
homophobic discourse, online comments, contempt, disgust, emotional strategies, counter-narratives, gender policing, social hierarchy, counter-discourse

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.