Revisiting Queer Minorities in Lexicons

Krithika Ramesh, Sumeet Kumar, Ashiqur KhudabukhshNAACL 2022

TL;DR

This paper investigates how queer-related terms are represented in popular English lexicons used for content moderation. The authors find significant inconsistencies and a lack of distinction between pejorative and non-pejorative terms, which could negatively impact the visibility of queer individuals in online discourse. They propose a framework for evaluating lexicon consistency and call for more nuanced approaches to content moderation.

What Problem Does This Paper Address?

Content moderation systems often rely on lexicons to identify and flag harmful language. However, there's a critical gap in understanding how these lexicons represent marginalized communities, particularly the queer community. The authors highlight that while lexicons are crucial for content moderation, they often lack quality control and detailed methodology for their creation. This can lead to the inclusion or exclusion of terms in ways that might inadvertently silence or misrepresent certain groups. Specifically, the paper addresses the problem of how "queer-related words" (terms referring to gender and sexual minorities) are handled in these lexicons, questioning whether non-pejorative terms should even be present and how their inclusion or exclusion impacts the broader discourse around queer identities.

Methods

The researchers developed a "queer lexicon" ($L$) comprising 115 terms, categorized into pejorative ($L_p$) and non-pejorative ($L_{np}$) subsets. They then analyzed twelve well-known English lexicons used for content moderation. Their analysis focused on:

  • Coverage: Measuring the overlap between their queer lexicon and the terms present in the twelve lexicons.
  • Annotation: Examining the methods used to create the lexicons, noting a significant number with unspecified annotation processes.
  • Pejorative vs. Non-Pejorative Distinction: Assessing whether lexicons differentiate between harmful slurs and identity terms.
  • Consistency: Developing and applying a set of simple constraints to quantify how consistently lexicons handle queer-related terms. These constraints include:
  • If a non-pejorative term from $L_{np}$ is present, all other terms in $L_{np}$ should also be present.
  • If a pejorative term from $L_p$ is present, all other terms in $L_p$ should also be present.
  • If a non-pejorative term from $L_{np}$ is present, its corresponding pejorative term from $L_p$ should also be present (with a penalty if not).

Datasets

The paper does not introduce new datasets for training. Instead, it analyzes twelve existing English lexicons used for content moderation. These lexicons are detailed in Table 1 and include: HurtLex, Rezvan et al. (2018), Wiegand et al. (2018), Palomino et al. (2021), Kwon and Gruzd (2017), CMU Lexicon, LDNOOBW, HateBase, Google Profanity Words, NoSwearing, Wiktionary, and Surge AI. The authors also created their own queer lexicon ($L$) of 115 terms, split into pejorative ($L_p$) and non-pejorative ($L_{np}$) subsets, based on online resources and consensus labeling by two annotators.

Key Insights and Ideas

  • Poor Overlap: The analyzed lexicons have minimal overlap with the developed queer lexicon, suggesting that many queer-related terms are not captured by existing content moderation tools.
  • Lack of Distinction: A significant concern is that many lexicons fail to distinguish between pejorative slurs and non-pejorative identity terms (e.g., "gay," "queer"). This conflation can lead to problematic content moderation decisions.
  • Potential for Harm: The unfettered inclusion of non-pejorative queer terms alongside pejorative ones might inadvertently hinder the visibility and acceptance of queer individuals in mainstream discourse.
  • Inconsistent Annotation: A substantial number of lexicons lack clear annotation details, making it difficult to assess their quality and potential biases.
  • Call for Nuance: The paper argues for a more nuanced approach to lexicon development, emphasizing the need for transparency, clear methodologies, and careful consideration of the impact on marginalized communities.

Key Results

  • The CMU Lexicon showed the highest overlap (23.48%) with the queer lexicon ($L$), while others had significantly lower coverage. When all lexicons were combined, the overlap increased to 40.87%.
  • Ten out of the twelve lexicons contained more pejorative queer-related words than non-pejorative ones.
  • Figure 2 shows that terms like "gay" and "queer" are among the most frequently occurring queer-related words in these lexicons.
  • Consistency analysis (Table 2) revealed varying degrees of consistency across lexicons based on the proposed constraints. Notably, lexicons that did not contain any terms from the non-pejorative set ($L_{np}$) were found to be the most consistent.

Limitations and Future Work

  • Limited Scope of Queer Lexicon: The authors acknowledge that their queer lexicon ($L$) of 115 terms is not exhaustive and may not encompass all identities and terms across the queer spectrum.
  • Contextual Ambiguity: Some terms in the non-pejorative set ($L_{np}$), like "gay," can be considered derogatory depending on the context, a nuance that lexicons often fail to capture.
  • Future Work: The authors suggest several avenues for future research:
  • Expanding lexicons with more comprehensive lists of queer-related terms.
  • Developing more transparent annotation guidelines.
  • Explicitly distinguishing between pejorative and non-pejorative queer-related terms.
  • Improving lexicon consistency regarding queer-related terms.
  • Conducting multi-lingual and multi-cultural analyses of queer presence in lexicons.
  • Performing "in-the-wild" impact assessments to understand how lexicon variations affect content moderation outcomes.

Summary updated

BibTeX
@inproceedings{ramesh2022revisiting,
  author = {Krithika Ramesh and Sumeet Kumar and Ashiqur Khudabukhsh},
  title = {{Revisiting Queer Minorities in Lexicons}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure