A Structured Clustering Approach for Inducing Media Narratives

Conference: ACL 2026
arXiv: 2604.10368
Code: Yes (mentioned in paper)
Area: Interpretability
Keywords: Media narratives, structured clustering, causal event chains, role analysis, framing theory

TL;DR

The paper proposes a framework to automatically induce media narrative patterns from large-scale news corpora. By jointly modeling causal event chains and role information (Hero/Threat/Victim), it utilizes a role-constrained clustering algorithm to organize narrative chains into semantically coherent patterns. It generates interpretable narrative patterns consistent with framing theory in the domains of immigration and gun control.

Background & Motivation

Background: Media narratives exert significant influence in shaping public opinion. NLP research on media analysis has accumulated through two main approaches: (1) coarse-grained labeling (e.g., "Left/Right" stance, "Political/Economic/Security" thematic frames), which is scalable but loses nuances of narrative structure; (2) domain-specific taxonomies (e.g., specialized labels for immigration or economic issues), which capture nuances but lack cross-domain generalization.

Limitations of Prior Work: Coarse-grained methods ignore the delicate narrative structures emphasized in communication studies—how readers are guided toward specific conclusions through role settings and causal constructions. Domain-specific methods require extensive manual annotation and cannot scale to new domains. The fragmentation between these two approaches limits consistent narrative analysis across domains.

Key Challenge: The contradiction between scalability and depth of interpretation—either sacrificing narrative detail for scale or sacrificing scale for depth.

Goal: Design a narrative induction framework that maintains deep narrative structure (event causality + functional role positioning) while scaling to large corpora without requiring domain-specific taxonomies.

Key Insight: Drawing from the Narrative Policy Framework (NPF) in communication science, this work views roles (Hero/Threat/Victim) as key structural elements of narrative analysis. Role constraints are used to distinguish event chains that share surface similarities but possess different narrative meanings.

Core Idea: Construct atomic narrative representations using causal event chains and role labels, then automatically induce high-level narrative patterns across articles via role-constrained clustering (cannot-link constraints).

Method

Overall Architecture

The framework aims to automatically induce semantically coherent and interpretable narrative patterns from large-scale news corpora without relying on domain taxonomies. Taking a single article as input, it first decomposes the text into causal event chains expressed in natural language. It then labels the narrative functions of roles (Hero/Threat/Victim) within the chains. These two signals are combined into atomic narrative representations. Subsequently, role-constrained clustering organizes thousands of narrative chains into several high-level patterns. Finally, an LLM outputs narrative descriptions (issue definition, evaluation, solution) for each cluster according to Entman’s framing theory, resulting in a list of interpretable narrative patterns.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Input: Single News Article"]
    A --> CHAIN
    A --> ROLE
    subgraph CHAIN["Causal Narrative Chain Construction"]
        direction TB
        C1["Dependency parsing extracts (verb, object) event tuples"] --> C2["DAPrompt identifies causality<br/>(Llama 3.3 70B silver label distillation)"] --> C3["DeepSeek-R1 natural language realization into coherent sentences"]
    end
    subgraph ROLE["Role and Narrative Function Labeling"]
        direction TB
        R1["Llama 3.3 70B extracts role mentions"] --> R2["k-means clusters into role groups"] --> R3["DeepSeek-R1 labels functions<br/>(Hero/Threat/Victim) + Stance"]
    end
    CHAIN --> M["Atomic Narrative Representation<br/>(Causal Chain + Role Function)"]
    ROLE --> M
    M --> CLU["Role-constrained Structured Clustering<br/>Penalty for conflicting chain pairs (cannot-link)"]
    CLU --> OUT["LLM outputs Entman frame narrative descriptions<br/>(Problem Definition / Evaluation / Solution)"]

Key Designs

1. Causal Narrative Chain Construction: Compressing articles into "Because X, therefore Y" atomic units

Pure event sequences only show what happened, missing the core manipulation in media narratives: causal construction—guiding readers to specific conclusions through the logic of "Because X, therefore Y." First, dependency parsing extracts \((verb, object)\) event tuples. Then, the DAPrompt method identifies causal relationships between event pairs. To manage the cost of large-scale LLM calls, Llama 3.3 70B generates silver labels for 20K event pairs, which are then distilled into a lightweight DAPrompt model. Once causal triples are obtained, DeepSeek-R1 realizes them into coherent natural language sentences, allowing subsequent clustering to utilize both structural information and semantic representations.

2. Role and Narrative Function Labeling: Key signals determining narrative ideological direction

In identical event chains, different role assignments can convey opposite messages—"immigrant=victim, officer=threat" versus "immigrant=threat, officer=hero" involve the same entities but point to opposing stances. To capture this, Llama 3.3 70B extracts role mentions under a 5-shot setting, which are then clustered into groups (e.g., "Immigrants," "Law Enforcement") via k-means. Subsequently, DeepSeek-R1 labels each role with its narrative function (Hero/Threat/Victim) and overall stance (Support/Oppose) in a zero-shot manner. This layer explicates the evaluative dimension of the narrative, serving as the basis for constrained clustering.

3. Role-constrained Structured Clustering: Forcing the separation of "same topic, opposing narrative" chains

Clustering based solely on textual similarity might group "immigrants are victims" and "immigrants are threats" together since both discuss immigration. To prevent this, cannot-link constraints are generated for chain pairs with conflicting role function configurations (where the same role group is assigned different functions). The k-means objective function is modified to \(\mathcal{R} = \frac{1}{2}\sum \|x_i - \mu_{l_i}\|^2 + \sum w_c \mathbf{1}[l_i = l_j]\), imposing a penalty \(w_c\) on same-cluster assignments that violate constraints. Initialization is replaced with a constraint-aware k-means++ variant to prevent conflicting chains from falling into the same cluster from the start. These constraints ensure clusters are consistent in narrative structure, not just topical similarity.

Key Experimental Results

Main Results (Structured Clustering vs. Standard k-means)

Domain	Method	Frame F1	Exact Match Purity	Avg. Role Purity
Immigration	k-means	41.19	26.90	80.79
Immigration	Ours	42.32	32.79	81.48
Gun Control	k-means	37.65	29.22	81.18
Gun Control	Ours	41.68	36.66	82.90

Ablation Study (Top 25% chains closest to centroid)

Domain	Method	Frame F1	Exact Match Purity
Immigration	k-means	33.22	32.31
Immigration	Ours	36.96	37.83
Gun Control	k-means	32.86	35.16
Gun Control	Ours	36.45	42.71

Key Findings

• Structured clustering consistently outperforms standard k-means across all metrics, particularly in Exact Match Purity (gain of +7.44pp in Gun Control).
• Role constraints are vital for distinguishing nuanced narrative differences: In Gun Control, narratives of "police as heroes protecting the public" and "police as threats violating rights" were successfully separated.
• Natural language realization of causal chains is of high quality: 3.3/4 for Immigration, 3.49/4 for Gun Control.
• Role labeling accuracy is excellent: 4.73/5 for Gun Control, 4.0/5 for Immigration.
• Quality is higher for chains near the cluster centers (purity metrics are better for Top 25%), suggesting that clustering identifies meaningful core structures.

Highlights & Insights

• Using narrative functions (Hero/Threat/Victim) as clustering constraints serves as an ingenious bridge connecting communication theory and computational methods. This ensures clusters share structural consistency rather than just topical overlap.
• The pipeline design balances cost and quality effectively: Using large models to generate silver labels followed by distillation to lightweight models for inference results in only a 15% performance loss in causality prediction while significantly reducing costs.
• Domain-agnosticism is a major advantage: Only minimal role group labeling (at the cluster level rather than sample level) is required to extend the method to new domains.

Limitations & Future Work

• Evaluation was limited to two English policy domains (Immigration, Gun Control); generalization across more languages and domains remains to be verified.
• The F1 for causal relationship prediction is only 58.46, which may cause error propagation to downstream clustering.
• Role group identification still requires minor manual labeling (cluster-level); fully unsupervised role discovery is a possible future direction.
• Temporal dynamics are not considered—how narrative patterns evolve over time (e.g., shifts before and after elections).
• Choosing the number of clusters \(k\) remains an open problem.

Related Work & Insights

• vs. Chambers & Jurafsky (2008): Classical narrative schema induction focuses on event sequences without considering the evaluative dimension of roles. This work adds role function labeling, aligning more closely with communication science theories.
• vs. Frame Detection (Card et al., 2015): Frame detection uses coarse-grained labels; this work achieves more granular and interpretable narrative patterns through structured clustering.
• vs. LLM-based Analysis: While direct LLM analysis is an end-to-end solution, it lacks interpretability and structure. This pipeline allows for validation and explanation at every step.

Rating

• Novelty: ⭐⭐⭐⭐ Solid theoretical motivation by using role narrative functions as clustering constraints.
• Experimental Thoroughness: ⭐⭐⭐⭐ Includes human evaluation and multiple metrics, though limited to two domains.
• Writing Quality: ⭐⭐⭐⭐⭐ Clear motivation, intuitive diagrams, and complete methodological flow.
• Value: ⭐⭐⭐⭐ Practical value for computational communication and media analysis; the method is transferable.

Related Papers

• [NeurIPS 2025] SpEx: A Spectral Approach to Explainable Clustering
• [ACL 2026] Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse
• [ACL 2026] Style over Story: Measuring LLM Narrative Preferences via Structured Selection
• [ICML 2026] Learning Coherent Representations: A Topological Approach to Interpretability
• [ICLR 2026] Semantic Regexes: Auto-Interpreting LLM Features with a Structured Language