cs.CLApr 11, 2026
SaveA Structured Clustering Approach for Inducing Media Narratives
Organizations: University of Colorado Boulder · Independent Researcher
Abstract
Media narratives wield tremendous power in shaping public opinion, yet computational approaches struggle to capture the nuanced storytelling structures that communication theory emphasizes as central to how meaning is constructed. Existing approaches either miss subtle narrative patterns through coarse-grained analysis or require domain-specific taxonomies that limit scalability. To bridge this gap, we present a framework for inducing rich narrative schemas by jointly modeling events and characters via structured clustering. Our approach produces explainable narrative schemas that align with established framing theory while scaling to large corpora without exhaustive manual annotation.
Figures & tables
Figure 1: Law enforcement spending framed as worker protection versus government waste across policy domains. Verbs are in bold . Objects are underlined . (Verb, Object) tuples make up events. Causal markers are in blue . Characters portrayed as heroes are highlighted in green , while threats are highlighted in pink . Character groups are denoted in subscripts. The Economic frame is from Boydstun et al., 2018 ’s Policy Framework.
Figure 2: For a large scale news corpus, we first construct narrative event chains and obtain character and role annotations for them. We then cluster these narrative chains using the character and role information as constraints. The generated narrative clusters are representative of fine-grained and nuanced narrative schemas.
| Causal Narrative Chain | Verbalization and Character Roles |
|---|---|
| ((withhold, permit), CAUSAL, (introduce, law)) | Mayor Lou Barletta [POLITICIANS] introduced a law that withholds permits from companies hiring illegal immigrants [IMMIGRANTS] in response to concerns over illegal immigration in Hazleton. |
| ((exempt, police), CAUSAL, (turn, police)) | The New York law’s failure to exempt police [LAW ENFORCEMENT] from the magazine ban caused them to become criminals when carrying standard firearms, as noted by Mayor Michael Bloomberg [POLITICIANS] . |
Table 1: Examples of causal narrative chain verbalizations and their character role annotations. Characters portrayed as heroes are highlighted in green , threats are highlighted in pink , and victims are highlighted in yellow . Character groups are denoted in subscripts.
| Canonical Form of Chain A | Canonical Form of Chain B | Constraint |
|---|---|---|
| Immigrants: Victim Law Enforcement: Threat Stance: Pro | Immigrants: Threat Law Enforcement: Hero Stance: Pro | Cannot-Link |
| Government: Threat Politicians: Hero Stance: Anti | Gun Rights Advocates: Hero, Politicians: Hero Gun Control Advocates: Threat | May-Link |
Table 2: Illustrations of canonical representations of character-role configurations of narrative chains. We show cases where chains cannot be linked and may be linked. We only consider cannot-link constraints. See App. B.1 for more details.
| Domain | Clusters | Model | Top 25% Narrative Chains | All Narrative Chains | ||||
|---|---|---|---|---|---|---|---|---|
| Frame F1 | Exact Match Purity | Avg. Role Purity | Frame F1 | Exact Match Purity | Avg. Role Purity | |||
| Immigration | 150 | k-means | 33.22 | 32.31 | 86.88 | 41.19 | 26.90 | 80.79 |
| Structured Clustering | 36.96 | 37.83 | 87.77 | 42.32 | 32.79 | 81.48 | ||
| Gun Control | 200 | k-means | 32.86 | 35.16 | 86.40 | 37.65 | 29.22 | 81.18 |
| Structured Clustering | 36.45 | 42.71 | 89.24 | 41.68 | 36.66 | 82.90 | ||
Table 3: Performance evaluation across three metrics: frame prediction F1-score, exact match purity, and overall role purity. Metrics are computed for all narrative chains and for the top 25% chains closest to cluster centroids. Results compare our method against k-means baseline on immigration and gun control domains. Number of clusters for each domain was selected based on the process outlined in App. E.4 . Bold values indicate superior performance.
Figure 3: SHAP feature importance comparison between RoBERTa embeddings and narrative features for frame prediction on a single test instance, showing both approaches capture similar semantic patterns. Red/blue indicate positive/negative influence; deeper shades indicate higher influence. Full length narrative schemas in App. F.2 .
Figure 4: SHAP feature importance values (×1000) for the top 6 narrative schemas and top 2 role features that predict the Legality, Constitutionality, Jurisdiction frame in gun control news articles. Full length narrative schemas are listed in App. F.3 .
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Character Group | Extracted Character Examples |
|---|---|
| Immigrants | Documented Immigrants, Undocumented Aliens, Undocumented Minors, Peruvian Immigrant |
| Refugees | Asylees, Immigrant Refugee, Palastinian Refugees, Haitian Refugee Community |
| Asylum Seekers | Asylum Applicants, Asylum Seekers, Chinese Asylum Seekers, Haitian Asylum Seekers |
| Workers | Agricultural Guest Workers, Illegal Workers, Assembly Workers, Tech Workers |
| Politicians | Donald Trump, Barack Obama, Republican Political Leadership, Republican Gov. Jan Brewer |
| Law Enforcement | Alameda Police, City Police Chief, Immigration Police, Customs Agents, U.S. Border Patrol |
Table 4: Final character groups for the immigration domain and selected examples of characters extracted from the dataset. For our analysis, we merge Refugess, Asylum Seekers, and Workers into the Immigrants group.
| Character Group | Extracted Character Examples |
|---|---|
| Politicians | Republican State Senator, Texas Gov. George W. Bush, Joe Biden |
| Gun Control Advocates | Advocates of Gun Control, Anti-gun Movement, Anti-Gun-Violence Activists, Colorado Coalition Against Gun Violence |
| Gun Right Advocates | Gun Rights Activists, Gun Lobbyists, Pro-Gun Organizations, Rocky Mountain Gun Owners |
| Law Enforcement | Local Law Enforcement Agencies, D.C. Police Chief, Airport Security Officials, Hempstead Police Department |
| Judiciary | California Supreme Court, Federal Judge, District Court Judge, U.S. District Judge Phyllis Hamilton |
| Government | City Government, Federal Government, Governing Bodies, Government Agencies |
Table 5: Final character groups for the gun control domain and selected examples of characters extracted from the dataset.
| Domain | Articles | Narrative Chains | Average Chains per Article | Constraints | Frame Labels |
|---|---|---|---|---|---|
| Immigration | 1988 | 36,956 | 23 | 33.3 million | 14 |
| Gun Control | 1933 | 44,269 | 28 | 37.4 million | 10 |
Table 6: Final sampled dataset statistics for both Immigration and Gun Control domains.
| Dataset | Precision | Recall | F1 |
|---|---|---|---|
| EventStoryLine Corpus (ESC) | 55.3 | 96.5 | 70.3 |
| MavenERE | 54.1 | 88.2 | 67.1 |
Table 7: Performance of teacher model on popular Event Causality Identification (ECI) datasets (Dev F1)
| Seed | Precision | Recall | F1 |
|---|---|---|---|
| 7 | 50.3 | 78.9 | 61.4 |
| 14 | 46.5 | 75.6 | 57.6 |
| 21 | 48.9 | 75.6 | 59.4 |
| 42 | 42.1 | 82.9 | 55.9 |
| 63 | 46.7 | 76.4 | 58.0 |
| Mean ( ) | 46.9 | 77.9 | 58.5 |
Table 8: Multi-seed evaluation of the DAPrompt ECI setup on the dev set
| High Quality Schema | Salient Frame Elements |
|---|---|
| Charlton Heston leads the NRA in opposing stricter gun-control legislation by advocating for rigorous prosecution of federal gun law violations. | Issue, Evaluation, Resolution |
| Law enforcement efforts lead to the indictment and prosecution of individuals involved in immigration violations, highlighting immigrants as both victims and threats. | Issue, Evaluation |
Table 9: Examples of high-quality generated schemas, evaluated on their overall coherence, as well as on striking a balance between coverage of and specificity to the sentences included in the context.
| Low Quality Schema | Comment |
|---|---|
| Politicians’ efforts to pass anti-crime legislation with stricter gun control measures face opposition from Gun Right Advocates, emphasizing public safety. | It is unclear who is advocating for ‘public safety’ |
| The U.S. government strengthens immigration enforcement through enhanced security measures and database implementation to combat illegal activities. | ‘Illegal activities’ lacks specificity to the context sentences, which reference falsified documents |
Table 10: Examples of low-quality generated schemas, evaluated on their overall coherence, as well as on striking a balance between coverage of and specificity to the sentences included in the context. We also include justification for the low-quality label.
| Error Type | Article Excerpt | Causal Event Pair | Verbalization | Character Annotation | Assigned Schema | Analysis |
|---|---|---|---|---|---|---|
| Incorrect event extraction | Sen. Larry Craig, R-Idaho, said Democrats would turn off voters if they used the Virginia Tech tragedy to roll back gun owners’ rights: “I could give you a litany of laws broken or lines crossed. Laws sometimes don’t protect the innocent, tragically enough…” | ((roll, tragedy), CAUSE, (roll, right)) | The Virginia Tech tragedy prompted gun rights advocates to roll out opposition against any potential rollback of gun owners’ rights. | Characters: gun rights advocates (Gun Right Advocates) – Hero Stance: Anti | Gun rights advocates heroically defend their constitutional right to bear arms against government and judicial threats, emphasizing its importance for freedom and safety. | The correct events should be (use, tragedy) and (roll back, rights). This produces a cascading error: the causality determination and verbalization are both incorrect and unfaithful to the article. Nevertheless, the assigned schema remains faithful to the article. |
| Incorrect causal relation prediction | But however many people are eventually able to gain legal status, there is one thing that should be required for all seeking to make a new life in this country: education. | ((require, thing), CAUSE, (gain, status)) | Requiring education for young immigrants helps them gain legal status in the U.S. | Characters: young immigrants (Immigrants) – Neutral Stance: Neutral | Immigrants face obstacles accessing education due to legal status and policy restrictions, with politicians influencing their access through supportive or restrictive legislation. | The causality here is subtle, if present at all. The article argues that education should be a legal requirement for immigrants seeking legal status. However, the LLM instead reasons that education ensures attainment of legal status, which is unfaithful to the article. The assigned schema is consequently incorrect, even though it is closely related to the topic being discussed. |
| Missing entity from the verbalization | Workers on the front lines of welfare reform, under new orders to deny food stamps to innumerable legal immigrants… are finding that translating Federal law into reality is not coming easily.Field workers are befuddled about what to do… and some are fighting their torment about the justness of it all. | ((deny, stamp), CAUSE, (fight, torment)) | The new rules denying food stamps to legal immigrants have caused significant emotional distress, leading many to fight against the torment of losing essential benefits. | Characters: legal immigrants (Immigrants) – Victim Stance: Pro | The denial of food stamps to legal immigrants due to welfare reform has led to high hunger levels, prompting advocacy groups to urge Congress to restore benefits. | Although factually plausible, the verbalization incorrectly omits the “Workers” entity. It is the workers who experience the “torment’,’ not the immigrants. The broader article discusses the plight of immigrants and the uncertainties they face, which explains the LLM’s misattribution. The assigned schema correctly addresses the hunger crisis but does not directly correspond to the event pair under consideration. |
| Unfaithful verbalization | As testimony continued Monday in Sami Al-Arian’s bail hearing, the question of whether he will get out of jail before trial became even murkier. Federal immigration officials have notified the lawyers for Al-Arian and one of the other defendants in the case of their intent to keep the two men in custody, even if the magistrate grants them bail. | ((notify, lawyer), CAUSE, (grant, bail)) | Federal immigration officials notified Al-Arian’s lawyers of their intent to keep him in custody, which could hinder the magistrate judge from granting bail. | Characters: Federal immigration officials (Law Enforcement) – Threat Al-Arian’s lawyers (Immigration Advocates) – Hero Stance: Neutral | Judges often act as heroes by making rulings that protect immigrants, who are frequently portrayed as victims, while government actions sometimes pose threats, leading to legal resolutions that address these challenges. | The verbalization incorrectly inverts the causal relationship between the events as described in the article excerpt. The assigned schema, however, is largely faithful to the article, correctly capturing the roles played by the different entities involved in such legal proceedings. |
| Incorrect character annotation | Over the past few years, the Palm Beach County School District has hired 71 teachers from around the world. …. Other international recruits include 49 who hail from the Philippines, …. But if the school board approves the plan tonight, the district will start offering its small but growing crop of foreign teaching recruits a chance for permanent residency … | ((include, 49), CAUSE, (offer, crop)) | The Palm Beach County School District’s inclusion of 49 teachers from the Philippines has led to their decision to offer these foreign educators a chance for permanent residency as an incentive to retain them. | Characters: Palm Beach County School District (Government) – Neutral teachers from the Philippines (Immigrants) – Hero Stance: Pro | Efforts by educators and government officials to support immigrant students’ educational needs despite enrollment challenges. | Arguably, the School District is the hero in this narrative by offering permanent residency to its foreign hires. The assigned schema, however, is incorrect, as the article discusses immigrant teachers rather than immigrant students. Since our algorithm treats immigrant teachers and students as the same character group, it is understandable why this schema was assigned based on semantic similarity alone. |
Table 11: Error analysis of the narrative extraction pipeline. Each row illustrates a distinct error type with the original article excerpt, extracted causal event pair, generated verbalization, character annotation, assigned narrative schema, and our analysis.
| Domain | Model | Train F1 | Dev F1 | Test F1 |
|---|---|---|---|---|
| Immigration | RoBERTa | 69.30 | 62.70 | 61.40 |
| Narrative Features | 65.97 | 43.46 | 38.57 | |
| Gun Control | RoBERTa | 75.41 | 62.98 | 64.19 |
| Narrative Features | 74.53 | 48.74 | 46.84 |
Table 12: Performance of RoBERTa and Narrative Feature Classifier on the Immigration and Gun Control Datasets. We use these models for our explainability study.
Figure 5: SHAP feature importance values (×1000) for the top 6 narrative schemas and top 2 role features that predict the Crime and Punishment frame in gun control news articles. Full length narrative schemas are listed in App. F.4.1 .
Figure 6: SHAP feature importance values (×1000) for the top 6 narrative schemas and top 2 role features that predict the Cultural Identity frame in immigration news articles. Full length narrative schemas are listed in App. F.4.2 .
Figure 7: SHAP feature importance values (×1000) for the top 6 narrative schemas and top 2 role features that predict the Economic Identity frame in immigration news articles. Full length narrative schemas are listed in App. F.4.3 .
Figure 8: Top 10 most important features for frame prediction in the immigration domain, showing SHAP importance values (×1000) for narrative schemas (blue) and character portrayals (red). Features are ranked by their average contribution across all frame classifications in immigration news articles. Full length narrative schemas are listed in App. F.5.1 .
Figure 9: Top 10 most important features for frame prediction in the immigration domain, showing SHAP importance values (×1000) for narrative schemas (blue) and character portrayals (red). Features are ranked by their average contribution across all frame classifications in gun control news articles. Full length narrative schemas are listed in App. F.5.2 .
| Domain | Model | Government | Immigrants | Immigration Advocates | Judiciary | ||||
|---|---|---|---|---|---|---|---|---|---|
| Purity@25 | Purity@100 | Purity@25 | Purity@100 | Purity@25 | Purity@100 | Purity@25 | Purity@100 | ||
| Immigration | k-means | 81.90 | 73.23 | 82.73 | 78.05 | 95.59 | 93.02 | 96.82 | 91.92 |
| Structured Clustering | 83.04 | 73.08 | 85.13 | 79.55 | 94.45 | 91.68 | 96.74 | 92.10 | |
Table 13: Role-based purity scores for Government, Immigrants, Immigration Advocates, and Judiciary character groups in the Immigration domain. Metrics evaluated using all narrative chains and top 25% chains nearest to cluster centroids.
| Domain | Model | Law Enforcement | Politicians | Stance | |||
|---|---|---|---|---|---|---|---|
| Purity@25 | Purity@100 | Purity@25 | Purity@100 | Purity@25 | Purity@100 | ||
| Immigration | k-means | 88.42 | 80.48 | 82.92 | 77.17 | 79.77 | 71.67 |
| Structured Clustering | 87.53 | 79.03 | 84.16 | 77.73 | 83.36 | 77.19 | |
Table 14: Role-based purity scores for Law Enforcement and Politicians character groups as well as Stance Purity in the Immigration domain. Metrics evaluated using all narrative chains and top 25% chains nearest to cluster centroids.
| Domain | Model | Government | Gun Control Advocates | Gun Crime Victims | Gun Right Advocates | ||||
|---|---|---|---|---|---|---|---|---|---|
| Purity@25 | Purity@100 | Purity@25 | Purity@100 | Purity@25 | Purity@100 | Purity@25 | Purity@100 | ||
| Gun Control | k-means | 85.23 | 77.73 | 91.37 | 86.98 | 95.34 | 93.62 | 79.57 | 72.24 |
| Structured Clustering | 80.05 | 90.40 | 84.05 | 96.45 | 90.68 | 84.59 | 75.32 | 92.10 | |
Table 15: Role-based purity scores for Government, Gun Control Advocates, Gun Crime Victims, and Gun Right Advocates character groups in the Gun Control domain. Metrics evaluated using all narrative chains and top 25% chains nearest to cluster centroids.
| Domain | Model | Judiciary | Law Enforcement | Politicians | Stance | ||||
|---|---|---|---|---|---|---|---|---|---|
| Purity@25 | Purity@100 | Purity@25 | Purity@100 | Purity@25 | Purity@100 | Purity@25 | Purity@100 | ||
| Gun Control | k-means | 92.84 | 88.27 | 85.66 | 80.64 | 84.82 | 77.69 | 76.38 | 72.24 |
| Structured Clustering | 92.33 | 89.54 | 91.35 | 82.34 | 86.66 | 80.5 | 85.11 | 80.68 | |
Table 16: Role-based purity scores for Judiciary, Law Enforcement, and Politicians character groups as well as Stance Purity in the Gun Control domain. Metrics evaluated using all narrative chains and top 25% chains nearest to cluster centroids.
| Domain | Model | Easy Accuracy | Medium Accuracy | Hard Accuracy | Overall | Inter Annotator Agreement | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Annotator A | Annotator B | Annotator A | Annotator B | Annotator A | Annotator B | Annotator A | Annotator B | Cohen’s kappa | 95% C.I. | ||
| Immigration | k-means | 96.00 | 98.00 | 92.00 | 88.00 | 76.00 | 74.00 | 88.00 | 86.67 | 80.99* | [67.50, 94.47] |
| Structured Clustering | 94.00 | 92.00 | 88.00 | 92.00 | 68.00 | 60.00 | 83.33 | 81.33 | |||
| Gun Control | k-means | 92.00 | 92.00 | 84.00 | 88.00 | 50.00 | 46.00 | 75.33 | 75.33 | 94.96* | [87.44, 102.48] |
| Structured Clustering | 94.00 | 92.00 | 88.00 | 86.00 | 70.00 | 78.00 | 84.00 | 85.33 | |||
Table 17: Intrusion task accuracy stratified by difficulty level (easy, medium, hard) comparing k-means and our algorithm across both domains. Cohen’s kappa inter-rater reliability reported. *p <0.001
List of Text Boxes L.1 Causal Relation Extraction (contd. on next page)
List of Text Boxes L.2 Causal Relation Extraction (contd.)
List of Text Boxes L.3 Causal Narrative Chain Verbalization
List of Text Boxes L.4 Character Extraction
List of Text Boxes L.5 Character Role Extraction (contd. on next page)
List of Text Boxes L.6 Character Role Extraction (contd.)
List of Text Boxes L.7 Role Descriptions for Immigration
List of Text Boxes L.8 Role Descriptions for Gun Control
List of Text Boxes L.9 Narrative Theme Attribution (contd. on next page)
List of Text Boxes L.10 Narrative Theme Attribution (contd.)
More in this topic
Clustering
Clustering groups data points by shared structure, developing methods to discover, validate, and interpret groups across diverse data and constraints.
See all 345 papers