Framing attributions

Introduction

Explanations are fundamental part of the social world.

Overarching theories across paradigms consider why attributing that “A causes B” matters in our understanding of the social world (Stone 1989; Shiller 2019; Entman 1993)

Examples

Causal reasoning in politics

Selective attribution helps explain how people assign causes of social issues, and how to solve them (Sahar 2014; Rudolph 2006; McCabe 2016).

In politics, Partisan Motivated Reasoning is the mechanism behind why we tend to assign responsibility for positive outcomes to our own party, while blaming bad outcomes on opponents.

Causal reasoning in politics

However, attribution is not just a psychological process, but also a discursive practice:

• Politician’s use of justifications can alter accountability (Hansson 2015; Grose, Malhotra, and Parks Van Houweling 2015)

• Media uses causal stories to allocate responsibility and this affects citizens’ blame attributions (Kim 2015; Rozenas and Stukal 2019; Iyengar 1989)

• Elite’s attributions of responsibility influences public attitudes and choices (Hobolt and Tilley 2014; Bisgaard 2015)

Attributions as empirical objects

Graham and Singh (2023)

Attributions as empirical objects

Understanding the mechanisms behind selective attribution is being addressed by improving experimental design (Graham and Singh 2023; Tappin, Pennycook, and Rand 2021).

However, questions of external validity persist.

Attributions as empirical objects

When analyzing attributions as they are made in political discourse, content analysis relies on domain specific coding.

Attributions as empirical objects

These two approaches have limited the understanding of the broader structure of causal attribution.

We need a general, scalable method to answer the question:

“who is claiming what and about whom?”

How to measure attributions?

Text-as-data methods are well-suited to capture discursive objects in natural occurring text data.

Causal Language Modeling (CLM) offers a systematic approach to identifying and analyzing causal claims.

Research question

Methodological:

How can causal attributions be used to empirically study political discourse?

Conceptual:

How are causal attribution employed in media framing?

Challenges

• Causal language is very hard to model.

• CLM has not been tested in political text data.

• Rhetorical, strategic, and often contested, and highly contextual.

• Consensus about basic causal stories is fragmented.

• The task is not to determine if statements are true, but to parse as expressed.

Operationalize attributions

Break down causal attributions to its linguistic components And get structured data from unstructured text

Instrumentalize attributions

Causal Language Modeling:

• Define what causal language looks like in text

  • Annotation codebook

  • Corpus creation

• Train a model to capture these expressions

  • Classification and Span Detection tasks

PolitiCAUSE

• First corpus of political text annotated for causality (Corral et al. 2024).

• Annotated ~18,000 sentences of the UNGD corpus (Jankin, Baturo, and Dasandi 2017).

PolitiCAUSE results

Large Language Models

Larger contexts promise better performance.

This raises a memorization v.s. generalization question between linguistic cues and causal mechanisms (Kıcıman et al. 2023; Hobbhahn, Lieberum, and Seiler 2022; Takayanagi et al. 2024).

Large Language Models

Large Language Models Evaluation

Designed an evaluation to determine if models can parse causal expressions: Linguistic Causal Disambiguation.

Used adversarial testing to begin understanding the mechanism used to perform a “causal classification” task when using token generation.

Large Language Models results

Lessons learned: Language Modeling

As a method for political analysis

Causal Language Modeling Framework for the study of attribution in political discourse.

Transformer-based models provide a balance between transferability, scalability, and accuracy with modest training data needs.

Provide a probabilistic output perfect for downstream analysis

Method

1. BERT (Devlin et al. 2018) for Sequence Classification

Probabilistic outcomes

Strengths and opportunities

• Adaptable for specific constructions.
• Flexible to model choice.
• Downstream analysis agnostic.
• Based on semantics.

• Annotation is cognitively taxing.
• Sensitive to preprocessing choices.
• Computational requirements.

Application: Media frames

Causal attributions are seldom operationalized as frames, despite being central to the theory (Entman 1993).

Using causal attributions as frames, I seek answer the question:

How are causal attribution employed in media framing?

To answer this question, I propose non-attribution as a unique strategy.

Non-attribution

Events that have a known cause and a known outcome such as:

Non-attribution

Case study: Corpus and data

• AJ, BBC and CNN
• October 2023 to June 2025

Causal Language Model ft. on news corpus

Is there a difference in use of causal headlines?

How is causality assigned across actors?

Observable Implications

Conclusions

Methodological: How can causal attributions be used to empirically study political discourse?

• Introduce a semi-supervised, NLP-based framework for identifying and measuring causal attributions in political discourse.

Theoretical: How are causal attribution strategically used in media framing?

• Non-attribution as a framing strategy

By advancing a methodological question I show how we can extend theory of framing and attribution.

My contributions

• Create a corpus and annotation scheme to detect causal attributions in political text.

• Develop an evaluation framework for causal language extraction for LMs.

• Introduce a supervised, NLP-based framework to identify and measure causal attributions in political texts.

• Analyze how attributions are used in media and present non-attribution as a framing strategy for neutrality performance.

Thank you

Paulina García-Corral

corral@hertie-school.org

Appendix

Paper 1: Results

Paper 1: Error Analysis

Paper 1: Discussion

• Explicit causal sentences only: Implicit structures are also rhetorical and possibly intentional.

• Single sentence causal structures: Causal attributions also exist across sentences.

• Corpus external validity from UN debates: Elite rhetoric, audience is other elites, prepared.

Paper 2: Corpus and data

Paper 2: Error Analysis

Model’s with more parameters exhibit better performance overall:

• Following recall and precision smaller models may be relying on more explicit causal markers, while larger models can “infer from context”.

• Scale contributes not only to broader generalization, but also to more faithful alignment with syntactic and semantic cues in discourse.

• Smaller models appear more susceptible to heuristic pattern matching and are prone to overgeneralizing causal signals from pretraining data.

Paper 3: Background

• Causal Language Models (Tan, Zuo, and Ng 2023; Drury, Gonçalo Oliveira, and Andrade Lopes 2022)
• Attribution work more broadly (Sotirovic 2003; Rozenas and Stukal 2019, 2019; Hameleers, Bos, and Vreese 2017, 2017; Tilley and Hobolt 2011; Graham and Singh 2023; Malhotra and Kuo 2008; Iyengar 1987)

Paper 3: Validation

Paper 3: Discussion

• Model Selection: Model agnostic and potential to include LLMs
• Annotation rules can be further detailed for specific cases
• Additional information due to a known semantic relationship

Paper 4: Background

• Selective attribution always works under the assumption of framing works as a competitions between actors (Iyengar 1987; Rozenas and Stukal 2019; Bisgaard 2015)

• Media bias has also shown the omission can also affect responsibility (Gentzkow and Shapiro 2006) .

Paper 4: Corpus and data

• AJ, BBC and CNN
• October 2023 to June 2025

Paper 4: Agent Annotation

Paper 4: Discussion

• Low used tokens render very big CI

• News events are not matched nor weighted

• NER might be more robust than “actor = subject” analysis

• Mechanism might be a function of maximizing content production and minimizing ideological inconsistency