Abstract
Hate speech detection in Natural Language Processing (NLP) is plagued by subjectivity and inconsistent definitions. This paper introduces a modular framework comprising five distinct criteria: Target Groups, Target Group Dominance, Speaker Characteristics, Type of Negative Reference, and Potential Consequences. By applying this framework to a meta-analysis of over 20 widely used English hate speech datasets, the authors demonstrate how current research often overlooks critical contextual factors like social dominance and speaker identity. This modular approach provides a roadmap for researchers to develop more precise, task-specific annotation guidelines and select datasets that align with their specific safety or legal goals.
Introduction: The Problem of Subjectivity
In the field of NLP, "hate speech" is often treated as a monolithic concept, yet it remains a moving target. Terms such as "toxic," "abusive," and "offensive" are frequently used interchangeably without rigorous definition. This lack of specificity leads to inconsistent datasets and biased models that struggle to generalize, particularly in "borderline cases" where context is paramount.
The core challenge is the inherent subjectivity of hate speech, which is context-dependent and culturally sensitive. Most current datasets fail to provide clear guidelines on fundamental questions: Does an attack on a dominant group (e.g., "men") constitute hate speech? Does the speaker's membership in the targeted group change the classification? Without addressing these variables, models simply inherit the unstated biases of their human annotators, leading to unpredictable and often harmful outcomes in real-world applications.
A Modular Framework for Definition
Rather than proposing a single, universal definition of hate speech, the authors advocate for a meta-prescriptive setup. They decompose the definition into five modular components that researchers can "toggle" based on the requirements of their specific task—whether they are developing a social media safety policy or a model for legal enforcement.
The five criteria are:
- Target Groups: Identifying specific protected characteristics such as race, religion, gender, or disability.
- Target Group Dominance: Deciding whether the definition applies only to marginalized groups or includes dominant groups as well.
- Speaker/Perpetrator Characteristics: Considering the identity, societal role, or intent of the speaker (e.g., distinguishing between a slur used in "re-appropriation" by a group member versus its use by an outsider).
- Type of Negative Reference: Requiring explicit links to stereotypes, group characteristics, or slurs.
- Potential Consequences: Evaluating the severity of the utterance, ranging from general group insults to direct incitement to violence, hate, or discrimination.
The framework distinguishes between "standard" aspects (commonly accepted in legal and NLP contexts, such as incitement to violence) and "optional" aspects that vary by task.
Meta-Analysis of Existing Datasets
The paper provides a comprehensive meta-analysis of more than 20 widely used English hate speech datasets sourced from hatespeechdata.com. By mapping datasets from major studies—including Davidson et al. (2017), Founta et al. (2018), and Zampieri et al. (2019)—against the five criteria, the authors reveal significant variation in how hate speech is operationalized.
For instance, some datasets like Davidson et al. (2017) do not specify target groups in their core definition, while others like Basile et al. (2019) focus narrowly on specific groups like immigrants and women. Furthermore, most datasets leave "explicit references" (such as the use of slurs or stereotypes) unspecified, making it difficult to determine if the resulting models are capturing implicit or explicit forms of hate.
Key Insights: Gaps in Current Research
The mapping process highlighted two major gaps in the current NLP landscape:
- The "Dominance" Gap: Most existing definitions do not explicitly account for social dominance. The analysis found that only one dataset (Mathew et al., 2021) explicitly included a dominant group (Caucasians) as a potential target.
- Perpetrator Blindness: Almost no current NLP datasets take the characteristics of the speaker into account. This is a critical oversight, as the same words can have vastly different impacts depending on the speaker's identity or position of power.
Limitations and Future Work
While the modular framework offers a path toward standardization, several limitations remain. The current meta-analysis is restricted to English-language datasets; linguistic and cultural nuances in other languages may necessitate additional variables. Furthermore, while clear definitions reduce subjectivity, they cannot entirely eliminate the personal biases of annotators. The authors suggest that gathering annotator demographics remains an essential practice.
Future work will involve applying these criteria to non-English datasets and encouraging the broader NLP community to adopt this modular approach when documenting new datasets (e.g., through "Data Statements"). By making the "points of subjectivity" explicit, researchers can build models that are more transparent, accountable, and aligned with specific societal needs.
Summary updated