Abstract
Online Gender-Based Violence (GBV), such as misogynistic abuse, is an increasingly prevalent problem that technological approaches have struggled to address. Through the lens of the GBV framework, which is rooted in social science and policy, we systematically review 63 available resources for automated identification of such language. We find the datasets are limited in a number of important ways, such as their lack of theoretical grounding and stakeholder input, static nature, and focus on certain media platforms. Based on this review, we recommend development of future resources rooted in sociological expertise and centering stakeholder voices, namely GBV experts and people with lived experience of GBV.
Introduction: The Conceptual Gap in NLP
While the Natural Language Processing (NLP) community has made significant strides in general "hate speech" detection, systems specifically designed to identify sexism and misogyny continue to suffer from high failure rates. This research identifies a fundamental "conceptual gap": most technical research treats online abuse as a purely linguistic problem, ignoring the complex social structures of power and violence that define it.
Currently, terms like "sexism" and "misogyny" are used loosely and interchangeably in dataset taxonomies. Furthermore, these datasets are often built in a vacuum, without the input of those most affected—victims and GBV experts. This leads to models that can catch overt slurs but fail to recognize the "non-linear," overlapping, and often subtle forms of abuse that characterize real-world gender-based violence.
Methodology
This study follows the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) protocol to ensure a rigorous and reproducible selection process. We analyzed 63 resources (datasets and papers) sourced from the DBLP Computer Science Bibliography and hatespeechdata.com.
Beyond standard metrics like F1 scores, resources were evaluated against the GBV framework—a comprehensive theorization of endemic violence used by the UN and EU. The evaluation covered six dimensions: conceptualization, representation (annotator background), sourcing (platforms and languages), sampling methods, and temporality.
Analysis of Existing Resources
The review acts as a meta-resource, providing a curated repository of the 63 reviewed items. The analysis reveals several critical biases in current data collection practices:
- Platform Bias: Over half (51.8%) of datasets rely exclusively or partially on Twitter.
- Language Bias: 41.3% are exclusively English, and 77.8% are limited to Indo-European languages.
- Sampling Bias: Most rely on keyword searches, which inherently misses "implicit" abuse that avoids specific slurs.
Key Insights and Results
The "Expert" Illusion and Stakeholder Participation
A critical finding is the lack of genuine stakeholder involvement. In many papers, "experts" are simply graduate students. Only one paper reported whether their annotators had lived experience with online abuse. Almost no datasets involved GBV survivors or professional advocates in the design or annotation process.
Theoretical Grounding
Drawing on feminist philosophy, we distinguish between sexism (the ideology justifying patriarchy) and misogyny (the enforcement mechanism). Most NLP datasets fail to make this distinction. Only a tiny fraction of resources were rooted in established sociological or feminist theory.
Data Perspectivism and Temporality
Only 10% of datasets preserved multiple labels to allow for the study of differing viewpoints. Most aggregate labels into a single "ground truth," ignoring that a victim’s perception of harm is often more relevant than a third-party crowdworker's. Furthermore, almost all resources are static snapshots, meaning their effectiveness decays as abuse tactics evolve to bypass filters.
Future Directions
To move from merely "detecting keywords" to actually mitigating the harm of online GBV, we propose the following recommendations:
- Participatory Design: Collaborate with GBV experts and survivors from the project's inception.
- Beyond Keywords: Develop new sampling methods to identify subtle abuse, such as micro-aggressions and gendered stereotypes.
- Dynamic Resources: Implement "human-in-the-loop" systems where datasets can be regularly updated with evolving abuse tactics.
- Multimodality: Address images, memes, and videos, as online GBV is rarely restricted to text.
Summary updated