QueerNews Corpus: Bias Mitigation in LLMs

• QueerNews Corpus is a large English-language, U.S.-centered collection of LGBTQIA+ news articles used for fairness adaptation in language models.
• The corpus is built from a 2015–2022 URL manifest and lightly cleaned to sentence-level text, preserving topical breadth despite minimal curation.
• Fine-tuning with parameter-efficient techniques like LoRA on this corpus has been shown to reduce anti-queer bias in LLMs significantly.

Searching arXiv for the cited papers to ground the article in current research. {} {"query":"arXiv (Menke et al., 18 Jul 2025) PRIDE Parameter-Efficient Reduction of Identity Discrimination for Equality in LLMs"} The QueerNews Corpus is a large English-language, U.S.-centered news collection focused on LGBTQIA+ topics and used as an un-annotated fine-tuning corpus in Menke and Hagendorff’s "PRIDE -- Parameter-Efficient Reduction of Identity Discrimination for Equality in LLMs" (Menke et al., 18 Jul 2025). It was recovered from a 2015–2022 URL manifest and lightly cleaned into sentence-level text. Within PRIDE, its primary function is not benchmark evaluation but targeted adaptation: the corpus provides queer-focused textual exposure for parameter-efficient fine-tuning, while bias is measured separately with WinoQueer. The corpus combines substantial scale with minimal annotation, and its methodological importance lies in demonstrating that a topical news corpus can support measurable anti-queer bias reduction in LLMs despite substantial coverage, authorship, and licensing constraints.

1. Origin, scope, and intended role

The source basis is the Felkner et al. (2024) QueerNews manifest of 169,298 URLs. These URLs point to US-based online news sites, including mainstream and LGBT-focused outlets, covering LGBTQIA+ topics. The publication window is 2015–2022, the language is English, and the geographic scope is the United States (Menke et al., 18 Jul 2025).

Selection is deliberately broad. Inclusion is defined by URLs tagged by Felkner et al. as “QueerNews” items, with no further manual topic filtering. This makes the corpus a topical collection derived from prior manifest construction rather than a hand-curated set of articles narrowed by a later editorial pass. In PRIDE, this design supports the use of queer-related journalism as a fine-tuning signal for fairness-oriented adaptation rather than as a manually labeled benchmark.

A common misunderstanding is to treat QueerNews as an identity-labeled corpus. It is not. All identity labels and splits, including gender versus sexual-orientation subsets, apply only to WinoQueer, not to QueerNews (Menke et al., 18 Jul 2025).

2. Recovery and curation pipeline

Corpus construction begins with URL re-hydration via HTTP GET requests to all 169,298 links. From this manifest, 91,805 pages were retrieved, which is approximately 54.22% of the original set. Pay-walled, geo-blocked, or dead links were omitted. The recovered HTML was converted to text through boilerplate removal to extract the main article body. Sentence splitting used a standard NLP splitter, exemplified by newline and punctuation heuristics, and noise filtering removed one-token “sentences,” such as headings and captions (Menke et al., 18 Jul 2025).

The final unit is the sentence. After cleaning, the corpus contains 3,101,705 clean sentences. No further topic filtering, de-duplication, or sentence-level annotation was performed.

Measure	Value
Total URLs in manifest	169,298
Pages successfully scraped	91,805 (54.22%)
Total sentences	3,101,705
Average sentences/article	~33.8

This processing regime is intentionally light. It preserves scale and topical breadth, but it also leaves substantial heterogeneity in genre, framing, and relevance. This suggests that QueerNews is best understood as a broad domain-adaptation corpus rather than a tightly controlled analytical dataset.

3. Annotation status, formalization, and data structure

QueerNews has no annotation schema for gender or sexual-orientation labels. It is used as an un-annotated “queer-focused” fine-tuning corpus. No formal definitions or LaTeX-formatted splits are provided for QueerNews; the paper’s formal definitions in LaTeX appear only in its description of WinoQueer’s gender-identity versus sexual-orientation subsets (Menke et al., 18 Jul 2025).

The absence of sentence-level labels has direct methodological consequences. The corpus does not specify whether a sentence is supportive, neutral, stigmatizing, policy-oriented, lifestyle-oriented, or otherwise. It also provides no topic master label. As a result, the corpus does not support supervised fairness learning in the narrow sense of labeled bias exemplars. Instead, its structure aligns with unsupervised or weakly structured adaptation, where relevance is conveyed by corpus provenance and topical concentration rather than by explicit per-instance labels.

Another misconception is that “queer-focused” implies community authorship or community validation. The corpus is queer-focused only in the sense that its URLs were tagged as QueerNews items. It carries no annotation of authorship, stance, or community membership.

4. Function within PRIDE and empirical effect on LLM bias

PRIDE evaluates two parameter-efficient fine-tuning techniques as alternatives to full-model fine-tuning: Low-Rank Adaptation (LoRA) and soft-prompt tuning. Using the WinoQueer benchmark, the study quantifies bias in three open-source LLMs and reports baseline bias scores reaching up to 98 out of 100 across a range of queer identities defined by gender and/or sexual orientation, where 50 indicates neutrality. Fine-tuning with LoRA, using less than 0.1% additional parameters, on the curated QueerNews corpus reduces those scores by up to 50 points and raises neutrality from virtually 0% to as much as 36%. Soft-prompt tuning with 10 virtual tokens delivers only marginal improvements (Menke et al., 18 Jul 2025).

These results locate the corpus within a specific fairness workflow. QueerNews is not the evaluation instrument; WinoQueer performs that role. QueerNews instead supplies the adaptation data that conditions the model toward less discriminatory behavior. The main empirical claim associated with the corpus is therefore indirect but operationally central: topical fine-tuning on queer-related journalism, when implemented through LoRA, can deliver meaningful fairness gains with minimal computation.

PRIDE further advocates broader adoption of community-informed PEFT, the creation of larger queer-authored corpora, richer evaluation suites beyond WinoQueer, and ongoing audits to keep LLMs inclusive (Menke et al., 18 Jul 2025). This places QueerNews in a transitional position: useful as a large-scale fairness adaptation resource, but not presented as a complete or final representation of LGBTQIA+ language.

5. Limitations, biases, and interpretive cautions

Several limitations are explicit. First, the corpus has a coverage bias: only about 54% of URLs were recoverable, and missing content may systematically differ. Second, it has a source bias characterized as “outsider” journalism, with few or no queer-authored voices; framing may therefore skew toward non-community viewpoints. Third, topical drift is substantial: articles vary widely in focus, including policy, lifestyle, and crime, and no further curation is applied. Fourth, language and locale are restricted to U.S. English, excluding global or multilingual queer press. Fifth, there is no sentence-level topic or master label, so raw text may contain irrelevant or neutral content. Sixth, the corpus does not capture intersectionality and nuance through metadata for race, socioeconomic status, or non-binary descriptors beyond those in WinoQueer (Menke et al., 18 Jul 2025).

These limitations affect interpretation in at least two ways. One is representational: a corpus of LGBTQIA+-topic reporting is not necessarily a corpus of LGBTQIA+ self-representation. The other is methodological: fairness improvements obtained through fine-tuning may reflect exposure to broader queer-related discourse, but they are not equivalent to correction against a comprehensive or community-authored normative standard. A plausible implication is that adapters trained on QueerNews may reduce specific forms of anti-queer discrimination while still inheriting the framing limits of the source press.

6. Access conditions and relation to adjacent LGBTQ+ media corpora

The underlying articles remain copyrighted by their original news outlets. No public redistribution of scraped plaintext is permitted. Researchers may re-scrape via the original URL manifest, subject to each site’s terms of service. Usage is limited to non-commercial research under “fair use” principles, and the corpus cannot be republished as a standalone corpus (Menke et al., 18 Jul 2025).

Within the broader landscape of LGBTQ+ media datasets, QueerNews occupies the position of a large, un-annotated article corpus. Birla and Akhtar’s "Visibility vs. Engagement: How Two Indian News Websites Reported on LGBTQ+ Individuals and Communities during the Pandemic" constructs a corpus from The Times of India and The Indian Express across pre-pandemic and pandemic periods, captures publication date, source, headline, full text, and sometimes section metadata, and analyzes the material using supervised sentiment analysis, BERTopic, and manual qualitative coding of transphobic and obsolete language (Birla et al., 20 Jul 2025). That resource is outlet-bounded and analysis-oriented. By contrast, QueerNews is broader in source base and explicitly used for fine-tuning rather than for headline-level sentiment supervision.

A second neighboring resource is the corpus in "Hope vs. Hate: Understanding User Interactions with LGBTQ+ News Content in Mainstream US News Media through the Lens of Hope Speech," which contains 1,419,047 YouTube comments on 3,161 LGBTQ+-relevant news videos from Fox News, CNN, and MSNBC, paired with a four-label hope speech taxonomy and PEFT-based classifiers (Pofcher et al., 13 Feb 2025). This corpus shifts the unit of analysis from source articles to audience response.

Taken together, these resources indicate three distinct empirical layers in LGBTQ+ media research: source-article corpora for adaptation, outlet-specific news corpora for representational analysis, and comment corpora for audience interaction analysis. In that landscape, QueerNews is notable for scale, topical relevance, and its demonstrated utility in parameter-efficient bias mitigation, while remaining constrained by recoverability, outsider framing, and non-redistributable copyrighted source text.