Belief Is All You Need: Modeling Narrative Archetypes in Conspiratorial Discourse

ACL ARR 2026 January Submission1887 Authors

Keywords: Conspiratorial discourse, Belief modeling, Narrative archetypes, Signed graphs

Abstract: Conspiratorial discourse is increasingly embedded in digital communication, yet its structure remains poorly understood. Analyzing Singapore-based Telegram groups, we show that conspiratorial content is integrated into everyday discussion rather than isolated echo chambers. We propose a two-stage framework: (1) RoBERTa-large classifies messages as conspiratorial or non-conspiratorial (F1 = 0.866 on 2,000 expert-annotated messages); (2) a signed belief graph models belief alignment via signed, similarity-weighted edges and is learned using a Signed Belief Graph Neural Network (SiBeGNN) with Sign-Disentanglement Loss to separate ideological alignment from narrative style. Hierarchical clustering of 553,648 messages reveals seven narrative archetypes: General Legal Topics, Medical Concerns, Media Discussions, Banking and Finance, Contradictions in Authority, Group Moderation, and General Discussions. SiBeGNN substantially outperforms standard methods (cDBI = 8.38 vs. 13.60–67.27), with 88\% inter-rater validation. Findings show conspiratorial discourse permeates mundane domains such as finance, law, and daily life, challenging assumptions of isolated online radicalization. The framework advances belief-aware discourse modeling for low-moderation platforms and informs stance detection, political discourse analysis, and content moderation policy.

Paper Type: Long

Research Area: Computational Social Science, Cultural Analytics, and NLP for Social Good

Research Area Keywords: language/cultural bias analysis, policy and governance

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis, Theory

Languages Studied: English

Submission Number: 1887