Technical Summary: From Emotion to Expression: Theoretical Foundations and Resources for Fear Speech

Problem Statement

Despite fear's profound capacity to mobilize societies, distort communication, and reshape collective behavior, the field of computational linguistics lacks a coherent theoretical framework and resource base for studying "fear speech" as a distinct phenomenon. Currently, fear is primarily analyzed as a general emotion within sentiment analysis or argumentation mining, rather than as a specific strategic communicative act. While hate speech has garnered significant attention and resource development, fear speech remains theoretically diffuse and empirically under-resourced. This gap is critical because fear speech often outperforms hate speech in reach and engagement on social media, partly because it appears "civiler" and evades current moderation systems. The challenge lies in the fragmented nature of existing research, which spans multiple disciplines (Psychology, Political Science, Communication Science, Linguistics) without a unified definition or consolidated dataset landscape.

Methodology

The authors employ a cross-disciplinary approach to bridge theoretical perspectives and audit existing empirical resources. The methodology consists of three primary phases:

1. Theoretical Synthesis: The paper reviews and compares theories of fear from four key disciplines:

   • Political Science: Focusing on securitization theory, where elites use speech acts to frame issues as existential threats to justify extraordinary measures.
   • Communication Science: Analyzing fear appeals (e.g., Protection Motivation Theory, Extended Parallel Process Model) as tools for persuasion, contingent on perceived threat and efficacy.
   • Psychology: Examining fear as a basic emotion tied to survival, appraisal, and physiological responses.
   • Linguistics: Investigating how metaphors, discourse, and framing construct fear narratives.
     These perspectives are synthesized into a multi-level framework (Source, Message, Reception, Outcome) to understand fear in public discourse.

2. Definition Formulation: Building on existing definitions from Buyse (2014), Saha et al. (2023), and Greipl et al. (2024a), the authors propose an expanded definition of fear speech. They define it as a "deliberate communicative act" that uses strategic language, rhetoric, and framing to portray a targeted entity as an inherent or imminent threat on cultural, societal, or existential levels, with the primary aim of evoking fear to influence attitudes or policy.

3. Resource Audit and Taxonomy Construction:

   • Discovery: The authors utilized an AI-supported prompt-based paper discovery method (using Allen AI's Paper Finder) to overcome vocabulary mismatches across disciplines. They generated 27 prompts covering fear speech, propaganda, framing, populism, fallacies, and abusive speech, yielding 810 candidate papers.
   • Screening: Using active learning (ASReview), they screened 317 papers, resulting in 37 relevant datasets. Papers were categorized into two buckets: B1 (directly addressing fear speech) and B2 (addressing adjacent phenomena like propaganda or framing that encode fear dimensions).
   • Taxonomy: A taxonomy was developed to consolidate different dimensions of fear (e.g., "Appeal to fear," "Security framing," "Threat," "Crisis rhetoric") across the identified datasets.
   • Analysis: The authors conducted a structured analysis of these 37 datasets based on four research questions (RQs): disciplinary orientation, data sources, conceptualization/labeling, and accessibility/reusability.

Key Contributions

The paper makes three specific contributions:

1. Comparative Theoretical Overview: It provides a systematic comparison of how fear is conceptualized across Political Science, Communication Science, Psychology, and Linguistics, offering a multi-level framework for analyzing fear in public discourse.
2. Refined Definition: It proposes a consolidated definition of fear speech that integrates strategic communicative intent with the portrayal of existential threats, distinguishing it from general emotion detection or hate speech.
3. Resource Landscape and Taxonomy: It presents the first systematic survey of datasets relevant to fear speech, identifying 37 resources. It introduces a taxonomy that maps diverse labels (e.g., "fear mongering," "security framing") to core fear dimensions, highlighting the potential for reusing resources from adjacent fields.

Results

The audit of the resource landscape reveals several critical findings:

• Disciplinary Imbalance: There is a strong concentration of fear-related datasets in computer science-oriented venues (32 out of 37 datasets come from Computational Linguistics, CS, ML, AI, etc.), with significant underrepresentation from social science perspectives.
• Data Sources and Topics: News media and Twitter are the dominant data sources. While fear speech datasets often originate from less-moderated platforms (WhatsApp, Telegram, Gab), the broader landscape is dominated by news and Twitter. Topics are heavily clustered around immigration (10 datasets) and politics (8 datasets), with moderate diversity in other areas like climate change and war.
• Conceptualization: The most frequent conceptualizations in existing datasets correspond to "propaganda" and "framing." Labels like "Appeal to fear" and "Appeal to fear/prejudice" are common in propaganda studies. Conversely, phenomena like "fallacy" and "populism" are underrepresented in the current dataset landscape regarding fear dimensions.
• Accessibility and Reusability: Significant barriers to reusability exist. Only 21 of the 26 accessible links were active at the time of auditing. Annotation guidelines are scarce (present in only 13 datasets), and inter-annotator agreement scores are reported in only 22 of the 33 manually annotated datasets. Furthermore, the specific count of fear-related instances is often unspecified.

Significance

The paper argues that fear speech warrants distinct research attention comparable to hate speech, given its growing prevalence and ability to evade moderation. By bridging theoretical insights from social sciences with computational resources, the authors aim to establish fear as a valid, operationalizable construct for computational modeling.

The significance of this work lies in its provision of a strategic starting point for "modularizing" fear. Rather than treating fear as a monolithic emotion, the paper suggests decomposing it into identifiable linguistic and rhetorical signals (strategic language, threat portrayal, intent) that can be detected using existing NLP techniques. The proposed taxonomy and the identified resource landscape offer a foundation for informed, collaborative data development. The authors emphasize that this work does not merely catalog existing resources but invites the computational linguistics community to view fear not as a marginal emotional category, but as a core analytic dimension that connects emotion, social influence, and linguistic form. The authors also acknowledge the "dual-use" risk, noting that their survey could theoretically be used to develop systems that manipulate users via fear, though the primary goal is to build a solid empirical foundation for understanding and modeling the phenomenon.