A Structured Clustering Approach for Inducing Media Narratives

This paper introduces a structured clustering framework that jointly models events and characters to induce scalable, explainable narrative schemas, effectively bridging the gap between computational analysis and the nuanced storytelling structures emphasized in communication theory.

The Gist

Imagine the news as a giant, chaotic library where millions of stories are being written every day about the same big problems, like immigration or gun control. Sometimes, the stories seem to agree, but often, they tell the exact same facts in completely different ways. One story might say, "The government is protecting us," while another says, "The government is wasting our money."

The authors of this paper asked: How can we use computers to understand these different "stories" without a human reading every single article?

Here is a simple breakdown of their solution, using some everyday analogies.

1. The Problem: Computers Are Too Literal

Most computer programs that analyze news are like a very strict librarian who only looks at the title of the book. They might see the word "Immigration" and group all those books together. But they miss the story inside. They don't realize that one book is a tragedy about a family fleeing war, while another is a thriller about border security.

The authors wanted a computer that could read the plot, not just the title.

2. The Solution: Building "Story Skeletons"

The team built a system that acts like a detective assembling a crime scene. Instead of just reading the whole article, they break every story down into its smallest, most important parts:

• The Events (The "What"): They look for actions, like "Mayor passed a law" or "Police arrested someone."
• The Cause (The "Why"): They connect the dots. Did the arrest cause the law to pass? Or did the law cause the arrest?
• The Characters (The "Who"): This is the secret sauce. They don't just see "a person." They ask: Is this person a Hero, a Villain (Threat), or a Victim?
  • Example: In one story, "Immigrants" might be the Victims needing help. In another story about the same topic, "Immigrants" might be painted as the Threat causing problems.

3. The Magic Trick: The "No-Go" Zones

Once the computer has built these little "story skeletons" (Event + Cause + Character Roles), it needs to group similar stories together.

Usually, computers group things that sound alike. But the authors taught their computer a special rule: The "No-Go" Zone.

Imagine you are sorting a pile of mixed-up playing cards.

• Normal Computer: "These two cards both have a 'King' on them, so they go in the same pile."
• This New Computer: "Wait! This King is wearing a white cape (Hero), and that King is wearing a black mask (Villain). Even though they are both Kings, they belong in different piles because their stories are opposites."

By using these "No-Go" rules, the computer separates stories that look similar on the surface but have totally different moral messages underneath.

4. The Result: Finding the Hidden Patterns

After sorting millions of stories this way, the computer starts to see Narrative Schemas. These are like "story templates" that the media uses over and over again.

• Template A: "The Government (Hero) protects the People (Victim) from the Criminals (Threat)."
• Template B: "The Government (Threat) hurts the People (Victim) to help the Corporations (Hero)."

The system can now tell you, "Hey, 80% of the news about this topic is using Template A, but the other 20% is using Template B." This helps researchers see exactly how different groups are trying to shape public opinion.

Why Does This Matter?

Think of media narratives as glasses. Different groups wear different colored glasses to look at the same world.

• One pair of glasses makes a policy look like a "Safety Measure."
• Another pair makes it look like "Government Overreach."

This paper gives us a tool to take those glasses off, look at the raw data, and see exactly how the story is being constructed. It helps us understand not just what happened, but how the story is being sold to us, and who is being painted as the hero or the villain in the process.

In short: They taught a computer to stop just reading words and start understanding the drama, the moral, and the plot twists in the news, allowing us to see the hidden patterns in how we are told what to think.

Technical Summary

1. Problem Statement

Media narratives are powerful tools that shape public opinion by constructing interpretative frameworks, often through specific character roles (heroes, threats, victims) and causal event structures. However, current computational approaches to media analysis face a dichotomy:

• Coarse-grained approaches: Rely on broad labels (e.g., partisanship, general topics like "economy" or "security") which miss nuanced storytelling structures and thematic depth.
• Domain-specific taxonomies: Capture ideological subtleties but lack scalability and generalizability across different policy domains.

The paper addresses the need for a method that can automatically induce rich, structured narrative schemas from large-scale, domain-agnostic corpora. The goal is to bridge communication theory (which emphasizes framing and character archetypes) with Natural Language Processing (NLP) to create explainable, scalable models of media narratives without exhaustive manual annotation.

2. Methodology

The authors propose a multi-stage pipeline that moves from raw text to structured narrative schemas via Structured Clustering. The process involves four main stages:

A. Constructing Narrative Chains

1. Event Extraction: Events are extracted as (verb, object) tuples using dependency parsing and the ETypeClus framework.
2. Causal Relation Extraction: The system identifies causal links between event pairs. It uses a teacher-student distillation approach: an instruction-tuned LLM (Llama 3.3 70B) generates "silver labels" for causal relationships, which are used to train a smaller, efficient model (DAPrompt) for inference.
3. Verbalization: Extracted causal triplets (event, causal, event) are converted into coherent, concise sentences using a reasoning model (DeepSeek-R1-Distill-Llama-70B) in a zero-shot setting. These sentences encapsulate the causal sequence and relevant context.

B. Character and Role Extraction

1. Character Group Identification: The system identifies salient entities in the text and clusters them into predefined character groups (e.g., "Immigrants," "Law Enforcement," "Politicians") using k-means clustering initialized with domain-specific seeds.
2. Role Assignment: Using the verbalized chains and the full article context, the LLM assigns narrative roles to characters based on the Narrative Policy Framework:
   • Hero: Agents driving positive outcomes.
   • Threat: Agents creating conflict or harm.
   • Victim: Agents suffering consequences.
   • Stance: The overall policy stance (Pro/Anti/Neutral).

C. Structured Clustering (The Core Innovation)

Instead of clustering solely based on textual similarity (e.g., standard k-means on SBERT embeddings), the authors employ Constrained Clustering.

• Constraint Generation: The system generates "cannot-link" constraints. If two narrative chains have conflicting character-role configurations (e.g., Chain A portrays "Immigrants" as Victims while Chain B portrays "Immigrants" as Threats), they are forbidden from being clustered together.
• Algorithm: A modified k-means algorithm is used.
  • Initialization: Uses a constraint-aware variant of k-means++ to select initial centroids that satisfy constraints.
  • Objective Function: Minimizes the distance to cluster centroids while penalizing assignments that violate "cannot-link" constraints.
  • Result: This produces clusters that are semantically coherent not just in topic, but in framing logic (i.e., how the story is morally and causally constructed).

D. Narrative Schema Attribution

Finally, the system uses an LLM to analyze the clusters and generate a high-level Narrative Schema. This schema summarizes the cluster's overarching theme and explicitly identifies Entman's (1993) framing elements:

• Issue: The problem definition.
• Evaluation: Moral judgments or causal analysis.
• Resolution: Proposed solutions or remedies.

3. Key Contributions

1. Unified Framework: The first computational work to combine detailed character-level information (roles/groups) with event-level information (causal chains) to induce media narratives.
2. Structured Clustering: A novel approach to clustering that uses character-role configurations as hard constraints, ensuring that superficially similar texts with different moral framings are separated.
3. Scalability & Explainability: The method scales to large corpora without manual annotation of every story, yet produces schemas that align with established communication theories (framing, narrative policy framework).
4. Domain Agnosticism: The pipeline is designed to work across different policy domains (tested on Immigration and Gun Control) without requiring domain-specific re-engineering of the taxonomy.

4. Results and Evaluation

The framework was evaluated on the Media Frames Corpus (Immigration and Gun Control domains).

• Narrative Chain Quality:
  • Causal relation prediction achieved an F1 of 58.46 (retaining ~85% of the teacher model's performance).
  • Human evaluation of verbalizations scored 3.3–3.49/4 on completeness and fidelity.
  • Character/Role annotations scored 4.0–4.73/5 on accuracy.
• Clustering Performance:
  • Purity: Structured clustering significantly outperformed standard k-means in Exact Match Purity (up to +5-7 points) and Role Purity, indicating that the constraints successfully grouped narratives with consistent moral framings.
  • Frame Prediction: Clusters derived from the structured approach improved the F1 score of predicting high-level media frames by 3–4 points compared to baselines.
  • Intrusion Test: The method improved intrusion detection accuracy for Gun Control (85% vs. 75%), demonstrating that the clusters are semantically tight. (Note: Immigration results were mixed, attributed to lower constraint weights and dataset artifacts).
• Explainability:
  • SHAP analysis revealed that narrative features (cluster frequencies, role counts) captured predictive patterns equivalent to token-level RoBERTa embeddings but offered superior interpretability.
  • The generated schemas successfully identified Entman's framing elements (Issue, Evaluation, Resolution) in 94–96% of cases, with "Evaluation" being the most prominent element.

5. Significance

This paper represents a significant step forward in computational media analysis by moving beyond topic modeling to narrative modeling.

• Theoretical Alignment: It operationalizes abstract communication theories (like Entman's framing and the Narrative Policy Framework) into a concrete, scalable NLP pipeline.
• Methodological Advancement: By introducing character-role constraints into clustering, the authors demonstrate that "how" a story is told (the framing) is as distinct and important as "what" the story is about (the topic).
• Practical Application: The ability to automatically generate explainable schemas allows researchers to track how different groups construct narratives around controversial issues, identify reinforcement of stereotypes, and understand the evolution of public discourse without relying on manual coding of thousands of articles.

The authors acknowledge limitations regarding potential bias amplification through LLMs and the static nature of the analysis (lack of temporal modeling), but position this work as a foundational step toward more sophisticated, interpretable narrative analysis systems.