Narrative Media Framing in Political Discourse

Narrative frames is a type of media frames that use elements of narrativity to highlight some aspects of a complex issue and condense it into a simplified "story" that promotes a particular interpretation. These elements of storytelling, such as representing an issue through the lens of stakeholders and conflicts rather than direct description of the facts, make narrative frames a highly effective device, particularly in the context of contested issues such as climate change.

Despite that, automatic framing analysis still mostly disregards narrative frames and conceptualizes frames in a topic-like fashion. Recent work in NLP has studied elements of narratives such as characters or events in news reporting, but these studies often lose the link to the core mechanism of framing, and do not allow to abstract away from particular events and characters in order to differentiate between contrasting narrative frames.

In our work, we dissect narratives into components which allow to link them to mechanisms of framing, on one hand, and help to understand, distinsguish, and model them. We show how:

• Character roles (framing particular entities as heroes, villains, and victims) as well as focus on a particular character helps to resolve ambivalence of interpretation (which is necessary for framing to occur) and assign a particular moral evaluation.
• Conflict and resolution represent causes and solutions to complex issues, which are core elements of their framing.
• Cultural stories link to broader value systems, allowing the frames to evoke wider associations existing in the culture.

Let's take a deeper look at these components and see how they allow to uniquely represent narrative frames and distinguish between them.

Characters

Characters and their prototypical roles follow the simplified hero, villain, and victim (HVV) triad from the Narrative Policy Framework. By assigning entities to particular roles, we resolve issue ambivalence by conveying moral judgment. The framework distinguishes between main characters and other entities, using the single most central character fulfilling each role.

To differentiate narratives, we also identify the focus on either hero, villain, or victim, resulting in "heroic", "blaming", and "victimizing" narrative frames respectively.

Conflict and Resolution

This component encapsulates the "plot" element and links to framing criteria that point to causes and resolutions. We conceptualize this as a four-way distinction where characters can:

• Fuel Conflict: Perform actions that cause or exacerbate the issue
• Fuel Resolution: Perform actions that help resolve the issue
• Prevent Conflict: Oppose actions that cause or exacerbate the issue
• Prevent Resolution: Oppose actions that help resolve the issue

Cultural Stories

Cultural stories map narrative combinations to four larger interpretation schemata, defining the degree of external control and group belonging:

Understanding how subtle differences in narrative framing can completely change the interpretation of similar events is crucial. Scroll through these examples to see how changes in a narrative frame structure lead to contrasting ways of framing.

Test our narrative framing model with your own articles. Our trained model will identify heroes, villains, victims, focus, conflict type, and cultural story components.

On benefits of compositionality

To apply our framework, we rigorously annotate a dataset of climate-related articles published in the USA in 2017-2019, conduct extensive experiments with LLMs, and analyse the results in the context of political leanings. Please refer to our paper for details.

Notably, the component-focused approach to narrative frames demonstrates benefits in all these aspects:

Inter-Annotator Agreement Improvement

37%

Direct Narrative Labeling

63%

Labeling by Narrative Components

Narrative frame components can be reliably annotated, achieving Krippendorff α in the range of 0.67-0.82.

Deriving the resulting narrative frame label from component annotations results in a much higher agreement than predicting the narrative frame directly (63% vs 37%).

Improved Model Performance

While regular LLM tricks like few-shot learning, PEFT and reasoning do not help to improve narrative prediction, including their structural components - even when predicted by the model rather than provided - does.

Political Leaning Analysis

While there are some trends in narrative frames used by left-bias and right-bias sources, the paterns become more prominent when we look at the components that underly them. For instance, as the picture on the right shows, right-leaning outlets favor individualistic narratives with "preventing resolution" framing, while left-leaning outlets emphasize collective action.

‹ ›

Another country, another time

The beauty of our framework is that it can generalize across topics and context. In particular, it allows to discover new narrative frames which were previously not included in the framework, rather than force them into one of the pre-existing "bucket".

To demonstrate that, we additionally annotate climate-focused articles from Australian outlets published in 2024, with a similar distribution across left-centred, left and right-centered sources.

As the picture below shows, we are able not only to compare the relative frequency of the narrative frames included in the USA dataset in Australian media, but also to discover new combinations of frame componets.

We have presented a rigorous formalization and taxonomy of components of narrative framing, synthesizing the Narrative Policy Framework with Entman's components of a frame. Our method allows inductive detection of narratives from political texts in terms of their character roles, focus, conflict, and underlying cultural story. Our key contributions include:

1. Theoretical Framework: A comprehensive formalization connecting elements of narrativity with fundamental aspects of framing
2. Empirical Dataset: A high-quality annotated dataset of 100 climate change articles with reliable inter-annotator agreement
3. LLM Evaluation: Systematic testing of multiple LLMs on narrative frame prediction tasks, revealing current limitations
4. Cross-Domain Validation: Demonstration of framework generalizability from climate change to COVID-19 political discourse

Our framework results in promising improvements for automatic narrative prediction with LLMs and lays a foundation for large-scale studies of narrative frame manifestation and effects. The structured approach to annotation achieves substantially higher agreement (63% vs 37%) compared to direct narrative labeling.

This work opens important research directions for computational analysis of narrative framing in political discourse, providing tools for understanding how complex issues are communicated and perceived across different political contexts.

If you find our work useful, please cite our paper:

@article{otmakhova2024narrative,
  title={Narrative Media Framing in Political Discourse},
  author={Otmakhova, Yulia and Frermann, Lea},
  journal={arXiv preprint},
  year={2024},
  institution={School of Computing and Information Systems, The University of Melbourne}
}