Social Narrative Coalitions

• Social narrative coalitions are groups of actors whose shared interpretative frameworks create coherent, collective narratives across diverse digital platforms.
• They are analyzed using computational linguistics, network science, and bibliometrics to quantify narrative alignment, valence, and cross-issue coherence.
• Applications include mitigating misinformation, managing political polarization, and understanding digital protest dynamics through systematic coalition detection.

A social narrative coalition is a set of actors, communities, or issues in which interpretative lenses, narrative structures, or shared references align to produce a coherent, group-level framing of events, identities, or themes. These coalitions emerge across human and artificial contexts, ranging from online political spheres to social movement communities, multilingual conspiracy ecologies, and even populations of LLM agents. The formalization, detection, and quantitative analysis of social narrative coalitions require integration of methodologies from computational linguistics, network science, bibliometrics, and agent-based game theory to map how narrative elements bind coalitional structure at multiple scales.

1. Foundational Definitions and Conceptual Frameworks

Across diverse empirical contexts, social narrative coalitions are grounded in the operationalization of “narrative” as a structured, meaning-making device. In the German Twittersphere, a narrative is formalized as an actantial network $G(V,E)$, where $V$ are named actors and $E$ are weighted, signed relations $e = (i \to j, w(i,j), \sigma(i,j))$ encoding frequency and valence (supportive: $\sigma \approx +1$; conflictive: $\sigma \approx -1$; neutral: $\sigma \approx 0$) with respect to narrative action (Pournaki, 21 Jul 2025). Conflicting narratives are defined as $(G_l, G_r)$ pairs derived from rival opinion camps, where conflicting relationships bear opposite valences: $$\operatorname{sign}(\sigma_l(i,j)) \neq \operatorname{sign}(\sigma_r(i,j))$$.

Social narrative coalitions arise via narrative alignment—a discursive strategy by which overlapping actants, aligned valence patterns, and meta-narratives induce a cross-issue interpretive bond. More generally, narrative coalitions may be instantiated as communities in a shared narrative map (Reddit: events as DAG, communities as subreddits (Norambuena et al., 2022); Telegram: channel–channel bibliographic coupling via shared hyperlinks (Willaert, 2024)), or as user clusters defined by co-participation in latent narrative clusters (embedding + clustering methodology) across fragmented social platforms (Gerard et al., 22 May 2025).

In multi-agent artificial systems, narrative coalitions manifest as emergent groupings of agents whose shared narrative priming conditions joint strategic behavior, leading to stable (or unstable) collaborative equilibria (Großmann et al., 6 May 2025).

2. Methodological Approaches for Extracting and Quantifying Coalitions

2.1 Actantial Network and Conflict Analysis

On X/Twitter, social narrative coalitions are extracted by:

• Harvesting trending topics over time windows, clustering issues (e.g., war, Covid, climate) with techniques such as BERTopic.
• Partitioning the user retweet graph using stochastic block models into competing clusters (left vs. right).
• Parsing tweets into AMR graphs, extracting predicate–argument triples (actant signals), and aggregating into weighted actantial networks.
• Classifying narrative relations (support/conflict/neutral) with LLM-driven prompt labeling.

Issue alignment is numerically captured with user-alignment scores and an issue-alignment matrix $IA_{i,j} = \cos(a_i, a_j)$, computing the cosine similarity between user co-assignment vectors for issues $I_i$ and $I_j$. High $IA$ indicates that the same user cleavage underpins narratives across multiple issues, evidencing a coalition (Pournaki, 21 Jul 2025).

2.2 Graphical and Embedding-based Approaches

Reddit narrative coalitions employ a graph-theoretic LP maximization that extracts a high-acceptance, topically coherent DAG per community (subreddit), measured using content embeddings (USE), clustering (UMAP+HDBSCAN), and Reddit-native acceptance metrics (upvote ratio, score percentile). Overlapping or diverging narrative graphs provide a metric space for coalition comparison (Norambuena et al., 2022).

Cross-platform analysis constructs user–narrative affiliation matrices based on cluster participation (e.g., DP-means over MPNet embeddings). Pairwise cosine similarity produces the user–user social graph (CANE/t-CANE), with community detection (Louvain) recovering persistent narrative coalitions—often spanning platforms (X, Truth Social) and held together by “bridge users” who broker the migration of narrative themes (Gerard et al., 22 May 2025).

2.3 Bibliometric and Hyperlink-Driven Methods

Telegram narrative coalitions are detected using bibliographic coupling. Channels are nodes; edge weights are the number of shared references to scientific or credible domains (OpenAlex sources). Louvain modularity partitioning on the coupling graph yields coalitional structure. Coalition properties are elucidated via the structural density, edge weights, and thematic labeling of core channel cliques (Willaert, 2024).

2.4 Experimental and Agent-based Models

In controlled network experiments, human participants are assigned to fixed topologies (complete vs. ring networks) and incentivized for hashtag coordination and causal framing. The emergence of dominant narrative coalitions, entropy reduction, and alignment in causal language vary by network modularity, path length, and clustering—a direct demonstration of how network parameters gate coalition topology (Priniski et al., 2024).

For LLM-agents, narrative priming is operationalized via prompt-embedded story fragments. Repeated public goods games yield collaboration scores as a function of narrative homogeneity or heterogeneity. When all agents are aligned via a prosocial narrative prime, stable cooperative coalitions form; with mixed priming, prosocial coalitions collapse and self-interested agents dominate (Großmann et al., 6 May 2025).

3. Structural, Thematic, and Dynamical Properties of Narrative Coalitions

Table 1: Empirical Manifestations of Social Narrative Coalitions

Platform/Context	Coalition Definition	Quantitative Markers
X/Twitter (political)	Alignment across actantial networks/issues	$IA_{i,j}$, overlap of salient actants
Reddit (protests)	Subreddit-based accepted narrative DAGs	Acceptance ≥ 0.85, DAG shape/landmarks
Telegram (conspiracies)	Channel–channel bibliographic coupling	BCF ≥ 2, modularity $Q \approx 0.42$
Cross-platform (CANE)	User–user narrative-affiliation graph	Sim(u,v), bridge user entropy $H$
LLM-Agent Games	Agents sharing the same narrative prime	Collaboration score, equilibrium structure
Online network expt	Network-induced local narrative convergence	Hashtag/causal-topic dominance, entropy $H(t)$

The coalitional structure is reinforced by thematic and linguistic patterns: recurring meta-narratives (e.g., government overreach vs. solidarity (Pournaki, 21 Jul 2025)), ideological or linguistic subcommunities (e.g., German-language anti-vax, transhumanist, QAnon clusters (Willaert, 2024)), or technical coordination devices (dominant tags in homogeneous networks (Priniski et al., 2024)).

Dynamical properties include coalition stability, bridge zone formation (users/channels linking coalitions, enabling narrative transfer (Gerard et al., 22 May 2025, Willaert, 2024)), and shifting valence in response to new events or actors (narrative trace evolution (Pournaki, 21 Jul 2025)).

4. Exemplary Empirical Findings Across Domains

4.1 Political Polarization and Issue Alignment

In the German Twittersphere, quantitative analysis reveals high alignment across multiple issues: $IA_{\text{Covid,Climate}} \approx 0.82$, $IA_{\text{Ukraine,Climate}} \approx 0.78$. Left and right coalitions display consistent actantial interpretations (e.g., “NATO→Ukraine” as protector vs. warmonger; “vaccination→public-health” as life-saving vs. harmful), with conflictive traces demarcating coalition boundaries (Pournaki, 21 Jul 2025).

4.2 Social Movements and Community Narratives

During the 2021 Cuban protests on Reddit, each of five subreddits forms a distinct narrative coalition, with tree-like, linear, or branched DAGs highlighting cause, repression, or external commentary. Average upvote ratios in extracted narratives range from 0.87 to 0.96, reflecting high community acceptance (Norambuena et al., 2022).

4.3 Conspiracy, Knowledge, and Epistemic Coupling

On Telegram, six major narrative coalitions emerge around shared (recontextualized) hyperlinks to scholarly and news sources. Coalition sizes vary (Science & Technology: ≈400 channels; Far-Right Conspiracies: ≈150). Thematic drives—regional language, political ideology, or shared scientific authorities—determine coalition boundaries. Central channels evidence high degree (up to 182); modularity $Q \approx 0.42$ quantifies coalition distinctness (Willaert, 2024).

4.4 Cross-Platform Migration and Bridge Users

Analysis of discourse graph structures between Truth Social and X identifies a bridging community (entropy $H=0.72$) consisting of only 0.33% of users, yet responsible for 67–69% of narrative migrations. Bridge users, despite median activity, play a structurally disproportionate role in seeding narratives across ecosystem boundaries (Gerard et al., 22 May 2025).

4.5 Network-driven Coalition Formation in Human Groups

Network experiments demonstrate that homogeneously-mixed topologies rapidly yield a single dominant narrative coalition (entropy $H \rightarrow 0$), while spatially-embedded networks support persistent local coalitions (entropy $H>0.5$, modularity $Q \gg 0$). Coordination is quantifiable through adoption/dominance rates and pairwise matching frequency (Priniski et al., 2024).

4.6 Narrative Priming in Artificial Agent Societies

Repeating public goods games among LLM agents exhibit stable cooperative equilibria ($\mathrm{Collab} = 0.95$–$0.99$) under homogeneous, prosocial narrative priming, and collapse to Nash-type defection ($\mathrm{Collab} = 0.38$–$0.55$) under heterogeneous or self-interest priming. Robustness analysis reveals that prosocial coalitions are fragile to narrative heterogeneity (Großmann et al., 6 May 2025).

5. Applications, Implications, and Policy Relevance

The formalization and detection of social narrative coalitions provide a structural and discursive lens through which to understand polarization, social movement identity, misinformation transmission, and coordination dynamics. Applications include:

• Cross-platform governance and moderation, based on identification of bridge users and coalition brokers for early intervention (Gerard et al., 22 May 2025).
• Social movement and protest analysis, extracting coalition-specific narrative maps for journalism and intelligence (Norambuena et al., 2022).
• Misinformation and conspiracy tracking, identifying high-coupling communities around retracted or controversial sources (Willaert, 2024).
• Social system and platform design, recommending network architectures or prompt structures to foster cross-group coalition formation and reduce fragmentation (Priniski et al., 2024, Großmann et al., 6 May 2025).

A plausible implication is that interventions targeting structural or discursive weak ties (e.g., introducing or removing bridge accounts, promoting cross-coalitional actors) may be more effective than those aimed solely at high-volume actors.

6. Limitations, Extensions, and Open Challenges

Known limitations in the current detection and operationalization methodologies include:

• Binary or low-resolution valence in actantial networks may obscure nuanced emplotment and framing (Pournaki, 21 Jul 2025).
• Reductive partitioning (e.g., two-cluster splits) may miss intra-coalitional heterogeneity or the presence of sub-coalitions.
• Static graph analyses overlook the full temporal dynamics of coalition formation, decay, and boundary realignment.
• Reliance on explicit shared references (e.g., hyperlinks or hashtags) may undercount covert or multimodal narrative practices (Willaert, 2024).
• In agent-based systems, narrative priming’s effectiveness depends on the stability and interpretability of LLM behavioral dynamics; robustness to exploitation by self-interested actors remains a challenge (Großmann et al., 6 May 2025).

Future directions include tracking $\sigma_t(i,j)$ over time, extending beyond trending content to viral subcultures or marginalized debates, improved detection of sarcasm, integrating richer event taxonomies, and comparative, cross-national mapping of coalitional structures and their migration pathways (Pournaki, 21 Jul 2025, Gerard et al., 22 May 2025, Priniski et al., 2024). Cross-disciplinary synthesis—coupling computational, network, and narrative-theoretical approaches—will be required to bridge these open problems.