Web(er) of Hate: A Survey on How Hate Speech Is Typed

Luna Wang, Andrew Caines, Alice HutchingsACL 2025

TL;DR

This paper critically examines how hate speech datasets are curated, arguing that a universal definition of hate speech is elusive. Instead, it proposes a "reflexive approach" where researchers explicitly document their own value judgments and frames of reference, fostering transparency and rigor in dataset creation.

What Problem Does This Paper Address?

The field of online hate speech detection is rapidly growing, with a proliferation of datasets designed to train models. However, these datasets exhibit significant diversity in their curation processes, leading to challenges in cross-dataset generalization and a lack of transparency regarding the underlying assumptions. Previous surveys have highlighted this heterogeneity but often overlook crucial aspects like the design goals and quality assurance measures. This paper aims to address this gap by providing a comprehensive analysis of hate speech dataset curation, emphasizing the inherent subjectivity and the need for researchers to be more explicit about their methodological choices.

Methods

The authors draw upon Max Weber's concept of "ideal types" as an analytical framework. Ideal types, in this context, are not meant to be exhaustive definitions but rather heuristic constructs that help researchers make sense of complex social phenomena like hate speech. By applying this lens, the paper analyzes how researchers' frames of reference (cultural norms, historical perspectives, etc.) influence their operationalization of hate speech. The study also involves a systematic review and categorization of existing hate speech datasets based on various criteria, including their definitions, goals, languages, data collection methods, annotation tasks, annotator composition, and ethical considerations.

Datasets

The paper analyzes a large number of publicly available hate speech datasets. The primary source for dataset identification was the community-maintained Hate Speech Dataset Catalogue, supplemented by targeted searches on Google Scholar. The authors identified 135 distinct datasets across 36 languages. While specific new datasets are not introduced, the paper provides a comprehensive overview and categorization of existing ones, highlighting their diversity and the methodologies employed in their creation.

Key Insights and Ideas

  • The Elusive Nature of Hate Speech Definitions: The paper strongly argues that a single, universally applicable definition of hate speech is practically impossible. Instead, researchers' "frames of reference" inevitably shape how they define and categorize hate speech.
  • The Value of a "Reflexive Approach": To combat the inherent subjectivity, the authors advocate for a reflexive approach. This means researchers must critically examine and explicitly document their own value judgments, assumptions, and perspectives throughout the dataset creation process. This transparency is crucial for understanding potential biases and improving the reliability of datasets.
  • Ideal Types as Analytical Lenses: Max Weber's ideal types are proposed not as rigid categories for classifying datasets, but as interpretive lenses to understand the socio-political motivations and underlying assumptions that drive the creation of these datasets.
  • Annotator Composition Matters: The paper highlights the trade-offs between smaller, curated annotator pools (suited for prescriptive guidelines) and larger, crowdsourced pools (better for descriptive approaches that capture diverse opinions).
  • Critique of Annotation Aggregation: The authors question the common practice of aggregating annotations through majority voting, arguing that it can oversimplify complex issues and remove valuable diversity of opinion, especially in descriptive paradigms.

Key Results

The paper presents a detailed breakdown of the analyzed datasets across various dimensions:

  • Definitions: 30% of datasets state their own definitions, with the "basis for hate" being the most frequently highlighted aspect (60%).
  • Goals: A significant proportion of research focuses on automation and model development (39%), exploring new directions (34%), and addressing research gaps (28%).
  • Languages: English dominates, followed by Italian and German, though efforts towards multilingual datasets and code-switching are noted.
  • Data Collection: X/Twitter is the most prevalent data source (70%), followed by Facebook and YouTube.
  • Annotation Tasks: Binary classification is the most common task (34%), but multi-class (24%) and hierarchical (54%) approaches are also used, each with trade-offs in granularity and nuance.
  • Annotators: Most datasets use multiple annotators, with crowdsourcing being a common method. However, a significant number (78%) do not report annotator demographics, hindering bias assessment.
  • Quality Assurance: Around half of the datasets do not report their QA procedures, with those that do employing a mix of pre-annotation selection, onboarding training, and post-annotation validation.
  • Ethics: While a positive trend of addressing ethical matters is observed, only 14% of papers explicitly revealed IRB approval. Anonymization is the most discussed ethical concern.

Limitations and Future Work

The authors acknowledge several limitations:

  • Focus on Publicly Available Datasets: The study primarily examines publicly accessible datasets, which may not fully represent the diversity of methodologies used in industry or private research.
  • No Empirical Evaluation: The paper does not conduct empirical evaluations of annotation quality or dataset performance in downstream tasks.
  • Interpretive Nature: The discussion of ideal types and annotation paradigms is necessarily interpretive, and alternative theoretical frameworks could offer different perspectives.

Future work could involve empirical evaluations of annotation quality and dataset performance. The authors also suggest that the application of Weber's ideal types of social action could be further explored to interpret the socio-political underpinnings of hate speech content itself. The increasing awareness of LLMs and their environmental footprint may also lead to more consideration of environmental impacts in future dataset creation.

Summary updated

BibTeX
@inproceedings{wang2025web,
  author = {Luna Wang and Andrew Caines and Alice Hutchings},
  title = {{Web(er) of Hate: A Survey on How Hate Speech Is Typed}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure