Narrative Policy Framework

• Narrative Policy Framework is an interdisciplinary approach that defines policy narratives through structured elements such as setting, characters, plot, and moral.
• It integrates formal models like Bayesian networks and game-theoretic methods with computational techniques such as SRL and event-chain modeling.
• The framework is applied across domains like politics, media, and economics to analyze narrative shifts and their impact on policy outcomes.

The Narrative Policy Framework (NPF) is an interdisciplinary theoretical and computational approach that conceptualizes policy processes as fundamentally narrative in structure. It provides formal, empirical, and algorithmic methodologies to decompose, represent, and analyze the stories that shape collective beliefs, political contestation, and policy outcomes across diverse domains such as politics, economics, media, and social discourse.

1. Core Concepts and Structural Elements

At its foundation, the NPF posits that human communication about policy is conducted through narratives: structured accounts of events, characters, actions, settings, and causal relations that ascribe meaning and purpose to political and social phenomena. The minimal and widely adopted NPF schema includes four key components:

• Setting: The contextual background (temporal, spatial, or societal) in which policy stories occur.
• Characters: Stakeholders or agents cast in the roles of protagonists, antagonists, heroes, villains, or victims (e.g., "scientists," "government," "the public").
• Plot: The sequence of events and causal or temporal linkages connecting actions and outcomes among characters.
• Moral: The explicit or implicit value judgment, policy lesson, or prescribed action embedded in the story.

This formalism underpins both theoretical typologies and computational implementations, serving as the blueprint for narrative detection, classification, and comparative analysis across media types and time periods.

2. Formal Models and Mathematical Representations

Recent advances have formalized the NPF within precise mathematical frameworks:

• Causal Models and Bayesian Networks: Narratives are modeled as causal graphs (directed acyclic graphs, DAGs) that connect actions, intermediate variables, and consequences. Each narrative instantiates a specific structure (e.g., "lever" or "opportunity/threat" narratives), which yields subjective beliefs about policy efficacy through factorization of joint probability distributions (Eliaz et al., 2018). The belief derived from a narrative's Bayesian network is computed as:

$$P_R(y \mid a) = \sum_{x_2, ..., x_{n-1}} \prod_{i \in N} p(x_i \mid x_{R(i)})$$

where $a$ is the action, $y$ the outcome, and $R(i)$ the parent set in the DAG.

• Game-Theoretic and Dynamic Bayesian Models: The NCC (Narratives-Construct-Commitment) framework models narratives as signals in a multi-agent game, affecting beliefs, expectations, and institutionalization of commitment (Jiang et al., 29 Apr 2025). Market participants update beliefs on government credibility via recursive Bayesian updating, e.g.,

$$\theta_{t+1} = \theta_t + \eta (c_t p_t - \theta_t)$$

where $c_t$ is implementation consistency and $p_t$ narrative precision.

• Network and Interaction Maps: Policy actors are represented as nodes in a dynamically unfolding network, with edges denoting co-appearances or interactions, and edge attributes encoding sentiment and topical content (see below) (Min et al., 2016).

3. Computational and Empirical Methodologies

The NPF has been operationalized through multiple computational pipelines:

• Semantic Role Labeling (SRL): Extracts "who does what to whom" triplets from text (Agent, Verb, Patient), enabling low-dimensional, interpretable representations of narrative structure (e.g., RELATIO (Ash et al., 2021), COVID/French Election corpus analysis (Zhao et al., 2023)).
• Event-Centric and Causal Chain Modeling: Events are extracted as verb-object pairs; temporal and causal relations among events are inferred using models trained on structured knowledge graphs (e.g., ASER). Chains or clusters of events yield high-level "narrative themes," enabling mapping of frames in media analysis (Das et al., 4 Oct 2024).
• Sentiment and Topic Analysis: Sentiment is calculated at the level of narrative units, actors, or actor pairs (using indices such as the Sentiment Polarity Index, SPI). Topic modeling applies NNMF or LDA to recover the core policy issues and thematic shifts associated with individual actors, interactions, or phases in a narrative (Min et al., 2016).
• Change Point Detection: Statistical algorithms (e.g., MtChD (Zhao et al., 2023), bootstrap-based topical changes (Lange et al., 25 Jun 2025)) segment corpora into periods reflecting significant shifts in narrative structure, enabling longitudinal and comparative studies of narrative evolution.
• LLMs: LLMs are prompted to extract, classify, and compare structural NPF elements from text, describe narrative shifts, and distinguish between content/topic and full narrative change. Annotators and LLMs jointly code for the presence and configuration of setting, characters, plot, and moral (Lange et al., 25 Jun 2025).

4. Applications Across Policy Domains

The NPF has demonstrated explanatory and predictive power in empirical policy studies:

• Media and Political Discourse: Analyses of legislative speeches, news coverage, and online discussions reveal how different parties construct competing causal stories, allocate praise/blame, and structure polarization through narrative framing (Ash et al., 2021, Otmakhova et al., 31 May 2025, Zhao et al., 2023).
• Policy Evolution and Rhetorical Themes: Historical content analysis has traced the evolution of core narrative themes (e.g., competition, prestige, collaboration, leadership, new paradigm) in the U.S. space exploration debate, revealing gaps between rhetorical ambition and policy implementation (Holland et al., 2018).
• Causal Micro-Narratives: Sentence-level, ontology-driven annotation of "micro-narratives" allows large-scale quantification of causal attributions in news about economic phenomena, such as inflation. Fine-tuned LLMs achieve human-comparable F1 scores in detection and classification of cause/effect relationships (Heddaya et al., 7 Oct 2024).
• Economic Expectations and Commitment: Dynamic Bayesian models link narrative communication, belief updating, and expectation management to measurable macroeconomic outcomes, showing how coordinated, credible narrative signaling is a prerequisite for shifts in investment, innovation, and productivity (Jiang et al., 29 Apr 2025).

5. Narrative Shift Detection and Dynamics

The NPF provides both conceptual and algorithmic tools for detecting, explaining, and comparing narrative shifts:

• Automated Narrative Shift Detection: Hybrid pipelines combine dynamic topic modeling (e.g., RollingLDA) with LLM-based extraction and annotation, using NPF as the structural test for a genuine narrative change (Lange et al., 25 Jun 2025). The NPF requirement (presence of setting, characters, plot, and moral) enables large-scale, rigorous classification of shift types.
• Statistical Significance and Robustness: Log-odds scoring, bootstrapping, and time-series analysis are used to quantify the salience and duration of shifts, supporting reproducible findings.

6. Challenges, Limitations, and Future Directions

• Model Hallucination: LLMs can over-detect narrative shifts when none exist, revealing an ongoing challenge in operationalizing the boundary between topical and narrative changes within the NPF (cf. (Lange et al., 25 Jun 2025)).
• Annotation and Ontology Curation: While component-wise approaches improve annotation quality, defining exhaustive and relevant ontologies of causes, effects, and narrative elements remains non-trivial and context-dependent (Heddaya et al., 7 Oct 2024).
• Computational Requirements: Full-scale LLM annotation of large corpora is resource-intensive, necessitating pre-filtering and integration with scalable models (e.g., topic models).
• Generalizability: Frameworks developed for one domain (e.g., climate discourse, COVID-19 crisis communication) have been shown to generalize, but the robustness of stakeholder typologies and narrative frame components warrants further paper (Otmakhova et al., 31 May 2025).
• Human-in-the-Loop Methods: Iterative, active-learning approaches hold promise for further integrating expert knowledge, automated extraction, and theory-driven coding.

7. Summary Table: Methods, Structures, and Application Domains

Approach	Structural Focus	Example Domains
Causal DAGs/Bayesian Ns	Actions, intermediates, outcomes	Political polarization, policy justification (Eliaz et al., 2018)
SRL-Based Extraction	Agent, action, patient triplets	Legislative speech, social media (Ash et al., 2021, Zhao et al., 2023)
Sentiment/Topic Analysis	Emotional valence and issue frames	Narrative phase detection (Min et al., 2016)
LLM-Based Coding	Setting, characters, plot, moral	Media shift detection, causal micro-narratives (Lange et al., 25 Jun 2025, Heddaya et al., 7 Oct 2024)

Conclusion

The Narrative Policy Framework integrates conceptual, mathematical, and computational methods to systematically dissect and analyze policy narratives. It makes possible rigorous, replicable, and scalable studies of narrative dynamics across domains, illuminating how the construction and evolution of stories fundamentally shape policy beliefs, contestation, and outcomes. Through advancements in formal modeling, network analysis, machine learning, and high-frequency empirical evidence, the NPF situates narratives as both the object and mechanism of collective decision-making in contemporary societies.