GENDIS contains YouTube comments that include sexist, misogynistic, transphobic and sexualized language, and references to gender-based violence. Access is granted for research purposes. By requesting access you agree to use the data for research only, not to attempt to re-identify commenters or channels, and not to redistribute the raw text.

GENDIS

Gender-Related Discourse and Stance in Multilingual YouTube Comments from India, Kenya, and Nigeria

GENDIS is a human-annotated dataset of 9,600 YouTube comments (3,200 each from India, Kenya and Nigeria), labelled with a two-level scheme that separates whether a comment references gender from how it does so.

Most resources for gender-related online content are built around harm, so their non-harmful class mixes together comments that are unrelated to gender and comments that invoke gender neutrally or positively. GENDIS separates these, which makes it possible to estimate how common gender-related discourse is and to characterise its full evaluative range rather than only its abusive pole.

Annotation scheme

Level 1 — is gender referenced?

value	meaning
`0`	not gender-related
`1`	gender-related

Level 2 — evaluative stance (only for gender-related comments)

value	meaning
`0`	neutral — descriptive, non-evaluative reference
`1`	negative — harmful, stereotyping, objectifying, including benevolent sexism
`2`	positive — affirming, empowering, or equality-promoting

Stance is not surface tone: benevolent sexism counts as negative despite a positive tone, and a comment condemning violence counts as positive.

Every comment was labelled independently by three annotators from the relevant country panel, shown the comment text only (no video title, description or surrounding comments). level1 and level2 are majority labels; the individual ratings are retained in the *_annotator1/2/3 columns so disagreement can be studied directly.

level2 is null when the comment is not gender-related (5,157 comments) or when it is gender-related but no two annotators agreed on a stance, or the majority was "unsure" (379 comments). 4,064 comments carry a resolved stance.

Fields

column	description
`id`	stable identifier, joins to the paper's benchmark splits
`country`	`india`, `kenya`, `nigeria`
`text`	comment text
`stratum`	`keyword`, `random_uniform`, `legacy_nonkeyword` (see Sampling)
`keyword_present`	whether the comment matches a keyword-list term
`level1`	majority Level-1 label
`level2`	majority Level-2 label, nullable
`level1_annotator1..3`	individual Level-1 ratings
`level2_annotator1..3`	individual Level-2 ratings
`languages`	languages detected, `\|`-separated (automatic, gpt-5-nano)
`code_mixed`	whether two or more languages are mixed (automatic)
`language_group`	`English`, `Local`, `Code-mixed`, `Other` (automatic)
`fold`	fold index of the fixed stratified 5-fold split used in the paper

Sampling

Per country, 3,200 comments drawn from a corpus of comments on the 5,000 highest-subscriber channels for that country:

keyword (1,500) — sampled via locally reviewed multilingual keyword lists, to concentrate gender-related speech.
random_uniform (1,500) — drawn uniformly at random from the full corpus with no keyword condition. Prevalence estimates should use this stratum only.
legacy_nonkeyword (200) — sampled from comments matching no list term, so it describes the keyword-free subpopulation rather than the corpus. Retained for model evaluation, excluded from prevalence.

The full set is keyword-enriched by construction, so its label distribution describes the annotated sample, not the corpus.

Label distribution

	not gender-related	negative	neutral	positive	no resolved stance
India	1,730	631	704	113	22
Kenya	1,850	503	680	125	42
Nigeria	1,577	734	237	337	315
All	5,157	1,868	1,621	575	379

In the uniform random stratum, 23.8% (Kenya), 30.3% (India) and 35.0% (Nigeria) of comments are gender-related, and among those with a resolved stance 82% are neutral or positive.

Inter-annotator agreement

Fleiss' κ / Krippendorff's α, from the individual annotations before aggregation:

	Level 1	Level 2	Combined
India	.80 / .80	.82 / .83	.77 / .77
Kenya	.79 / .79	.71 / .72	.72 / .72
Nigeria	.55 / .54	.49 / .49	.47 / .47
All	.71 / .71	.70 / .68	.66 / .65

Agreement is markedly lower in Nigeria, where 19% of gender-related comments reached no Level-2 majority. The paper treats this as genuine normative contestation over what counts as evaluative rather than as annotation noise.

Ethics and limitations

Identifiers are not released. Comment IDs, video IDs, channel identifiers and usernames are excluded, so comments cannot be traced back to their authors. Do not attempt re-identification.
Content warning. The data contains sexism, misogyny, transphobia, sexualized language and references to gender-based violence.
Labels are the aggregated judgments of small country panels under a specific operational scheme, not a universal account of what is positive, negative or neutral. Other informed panels could label differently, especially for implicit or contested cases.
Country panels did not cover every language present in the sample, so nuance may be lost in slang and code-mixed content.
Language fields are automatic (LLM-assigned) and approximate; treat them as descriptive.
The sampling frame is Social Blade channel rankings, which do not verify creator nationality and are time-specific. Estimates characterise the sampled country-associated channel corpora, not national populations or YouTube at large.

Citation

To appear at the 10th Workshop on Online Abuse and Harms (WOAH 2026), co-located with EMNLP.

@inproceedings{koch-etal-2026-gendis,
  title     = {{GENDIS}: Gender-Related Discourse and Stance in Multilingual
               {Y}ou{T}ube Comments from {I}ndia, {K}enya, and {N}igeria},
  author    = {Koch, Luise and Tonneau, Manuel and Malhotra, Niyati and
               Rai, Sunny and Munoz Boudet, Ana Maria and Orozco-Olvera, Victor and
               Fraiberger, Samuel P. and Guntuku, Sharath Chandra},
  booktitle = {Proceedings of the 10th Workshop on Online Abuse and Harms (WOAH)},
  year      = {2026}
}

Downloads last month: 3

Total file size:

2.31 MB