Media Framing through the Lens of Event-Centric Narratives

Technical Summary: Media Framing through the Lens of Event-Centric Narratives

Problem Statement
Media framing analysis aims to understand how news organizations curate and organize aspects of issues to promote specific interpretations. Existing computational approaches predominantly rely on high-level topic markers, such as latent topic models (e.g., LDA) or manually constructed taxonomies of keywords. The authors argue that these methods are often too broad to capture the nuances of framing. By reducing frames to co-occurring keywords or broad themes (e.g., "economic"), these approaches fail to explain how different aspects are chosen and organized to construct an argument. Specifically, a single topic marker like "economic" can frame immigration as either a boost or a drain on the economy, a distinction that keyword-based models struggle to disambiguate without understanding the underlying narrative structure.

Methodology
The paper proposes a computational framework that shifts the focus from static topics to dynamic, event-centric narratives. The methodology consists of three primary stages:

1. Event Extraction:
   The system adopts a verb-centric view of events, extracting (verb, object) pairs from text. Using a dependency parser, the framework identifies non-auxiliary verbs and their corresponding object heads (distinguishing between active and passive voice) to form a list of event mentions for each document.

2. Relation Extraction and Chain Construction:
   A classifier is trained to predict relationships between pairs of extracted events within a document. The model focuses on two specific relation types:

   • Temporal: Chronological ordering (e.g., Precedence, Succession, Synchronous).
   • Causal: Cause-and-effect relationships (e.g., Reason, Result).
     The classifier is trained on a subset of the ASER (Activities, States, Events and their Relations) knowledge graph, utilizing RoBERTa to represent event pairs and their surrounding sentences. Only event pairs with predicted temporal or causal relations are retained to construct single-hop narrative chains represented as (event1, relation, event2).

3. Narrative Clustering and Expansion:
   To transform these chains into high-level themes, the framework employs a Large Language Model (LLM), specifically Llama 3.1 8B. The LLM is prompted in a zero-shot setting to expand each narrative chain into a descriptive sentence that articulates the causal or temporal sequence. These expanded sentences are then embedded using SBERT and clustered using K-means. The authors experiment with cluster counts ranging from 25 to 200 to identify distinct narrative themes.

Key Contributions

• A Narrative-Centric Framework: The authors propose a novel approach to media framing analysis that extracts event-centric narrative representations and groups them into higher-level themes, moving beyond simple keyword co-occurrence.
• Cross-Domain Generalizability: The framework is applied and evaluated across two distinct news domains: immigration and gun control, demonstrating its adaptability to different policy issues.
• Predictive and Explanatory Utility: The study provides a comprehensive evaluation showing that the induced narrative clusters serve as significant signals for predicting and explaining the Boydstun et al. (2014) policy frame taxonomy.

Results

• Relation Prediction: The relation classifier achieved a macro-F1 score of 0.60, outperforming baselines including Majority Class, Random, and Logistic Regression with GloVe embeddings. The model showed higher precision in identifying temporal relations compared to causal ones, though recall for causal relations was prioritized to ensure coverage.
• Cluster Quality: An intrusion test revealed high inter-annotator agreement (82.61%) and labeling accuracy (67.5%) for the immigration dataset, indicating well-defined, semantically coherent clusters. The gun control dataset showed lower accuracy (37.5%), which the authors attribute to a lack of variation in narrative themes and overlapping clusters in that domain.
• Frame Prediction:
  • Using only narrative cluster features (without raw text), the model outperformed baselines including Random, LDA topics, and Event Types. The optimal performance was achieved with 150 clusters for immigration and 50 for gun control.
  • When narrative cluster features were combined with RoBERTa-based text representations, the model showed a minor but consistent improvement in accuracy and F1 scores over text-only baselines, suggesting that narrative structures introduce useful inductive bias.
• Qualitative Findings: The analysis successfully isolated narrative themes corresponding to specific frames. For example, the "Crime and Punishment" frame in immigration was strongly associated with narratives about "smuggling of undocumented immigrants" and "arresting smugglers." Similarly, the "Legality, Constitutionality, Jurisdiction" frame in gun control was linked to narratives regarding "Second Amendment rights" and "court rulings."

Significance and Claims
The paper claims that its primary contribution lies in demonstrating that framing is fundamentally a narrative construction problem. By extracting and clustering event chains, the framework can capture the "story-telling patterns" that disambiguate high-level framing dimensions. The authors assert that their approach offers a richer and more concise representation of documents than traditional topic modeling, as it focuses on the salient narratives chosen to make arguments rather than just the topics discussed.

The authors maintain a modest tone regarding their claims, acknowledging that the framework is not designed to maximize performance through massive, complex models due to computational constraints. Instead, the work aims to highlight a research direction that underscores the importance of identifying key nuances in narrative chains. They note limitations in domain generalization (tested only on immigration and gun control) and the reliance on human annotation for cluster evaluation, suggesting these as areas for future exploration.