TL;DR
This paper argues that current offensive speech detection models fail to capture the nuances of minoritized language use by focusing on isolated utterances rather than the social relations that inform them. The authors propose a research agenda centered on "relationality" – understanding speech as embedded within social contexts and relationships – to build more robust and equitable NLP systems.
What Problem Does This Paper Address?
Current Natural Language Processing (NLP) models for classifying offensive speech often struggle with minoritized language, such as reclaimed slurs or queer speech. These models tend to misclassify such language as toxic or hateful because they focus on the literal meaning of words rather than the social context in which they are used. This leads to biased outcomes, potentially censoring legitimate forms of expression within marginalized communities. The paper highlights that offensiveness is not an inherent property of a word but rather a product of social relations, cultural context, and the dynamics between speakers and listeners. The core problem is the lack of a framework that effectively operationalizes social context, particularly relational aspects, for robust offensive speech classification.
Methods
The paper proposes a shift from analyzing isolated utterances to understanding speech through the lens of relationality. This involves focusing on the social relations that influence the production, meaning, reception, and outcomes of language. The authors suggest several avenues for operationalizing this relational approach:
- Linguistic Style Measurement: Leveraging existing NLP techniques for modeling linguistic style and dialect, which implicitly carry information about cultural context and community membership. This can help infer shared norms and social proximity.
- Annotation Task Design: Rethinking how data is annotated to provide richer social context to human annotators. This includes exploring the impact of providing conversational turns, website origins, or temporal information, and investigating how annotator diversity and social identity influence their judgments.
- Extra-Linguistic Features: Incorporating features beyond the text itself, such as user behavior (e.g., likes, blocks), conversation structure, and metadata that captures temporal and geocultural situatedness. This allows for a more holistic understanding of the interaction.
- Interaction Outcomes: Analyzing measurable outcomes of offensive interactions, such as downvotes or blocking user profiles, as proxies for the impact and nature of offensive language.
Datasets
Not discussed in the paper. The paper focuses on proposing a research agenda and theoretical framework rather than presenting empirical results from specific datasets.
Key Insights and Ideas
- Offensiveness is Relational, Not Intrinsic: The core insight is that whether speech is offensive is determined by the social relations and context surrounding it, not by the words themselves. This challenges the common approach of treating offensiveness as a property of a linguistic token.
- Minoritized Speech as a Motivating Example: The paper uses examples like queer speech and African American English (AAE) to illustrate how current classification systems fail due to their inability to grasp the nuanced, context-dependent use of language within these communities.
- Style as a Bridge to Relationality: Linguistic style is presented as a key artifact of social context that can reveal social relations and community norms, offering a way to distinguish minoritized speech from antagonistic forms.
- Ethical Imperative for Contextual Understanding: The authors emphasize that failing to account for minoritized language use risks increased censorship and perpetuates harm, making a relational approach an ethical necessity.
- Beyond "Common Sense": The paper critiques relying on generalized or "common sense" interpretations of language in classification, arguing that these often reflect dominant norms and can be exclusionary.
Key Results
Not discussed in the paper. This paper is a research agenda and does not present empirical results.
Limitations and Future Work
The authors acknowledge several limitations and propose future research directions:
- Limitations of Style: While style is a valuable indicator, it can be used in adversarial ways (e.g., mockery, impersonation) and may not always align with community norms or social identification processes. It's unclear how relationality can account for these complex uses of style.
- Inferring Social Identity: There's a tension regarding the ethical risks of inferring social identity from speech, as this can lead to misgendering harms and undermine autonomy.
- Operationalizing Relationality: The primary challenge is defining the scope and operationalizing the component parts of a relational approach for NLP tasks.
- Future Research Questions: The paper outlines several key research questions, including:
- What social relations are crucial for text classification?
- How can style measurements be complemented by sociological and ethical aspects?
- How can annotation tasks be improved to provide better context for raters?
- What influence does additional context have on annotation behavior and model performance?
- What extra-linguistic features can infer social context for offensive speech detection?
- What interaction outcomes can be measured to model offensive interactions?
Summary updated