Seeing the Same Story Differently: Framing‐Divergent Event Coreference for Computational Framing Analysis

Jin Zhao, Xinrui Hu, Nianwen XueEMNLP 2025

TL;DR

This paper introduces FRECO, a novel task and dataset for identifying pairs of event mentions that describe the same real-world occurrence but are framed differently across news sources. By focusing on event-level framing divergence, FRECO offers a more nuanced and scalable approach to understanding media bias and narrative construction than traditional topic-based methods.

What Problem Does This Paper Address?

News media often present the same factual events in vastly different ways, shaping public perception through "framing" rather than outright falsehoods. Traditional computational framing analysis often relies on coarse topic classification or predefined frame taxonomies, which struggle to capture subtle, event-level differences in how events are portrayed. This limitation hinders our ability to analyze how media narratives diverge, especially across different issues or domains. The authors aim to bridge this gap by developing a method to computationally identify and analyze these framing differences at the event coreference level.

Methods

The core of this work is the Framing-Divergent Event Coreference (FRECO) task. This task identifies pairs of event mentions that refer to the same underlying real-world occurrence but differ in their framing. This divergence can manifest in various ways, including:

  • Lexical choice: Using different words to describe the same action (e.g., "hunted down" vs. "pursued").
  • Connotation: Employing loaded terms versus neutral language.
  • Causal attribution: Emphasizing different causes or consequences.
  • Perspective: Highlighting different actors or viewpoints.

To support this task, the authors propose a bootstrapped mining pipeline. This pipeline starts with a small set of high-quality, manually annotated FRECO pairs and iteratively expands the training data by identifying high-confidence FRECO instances from a larger corpus. This semi-supervised approach allows for scalable extraction of framing-divergent events.

The authors also explore various fine-tuning strategies for Large Language Models (LLMs), including Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), and their sequential combinations. They also investigate the impact of incorporating Structured Semantic Role Labeling (SRL) information to enhance event representations.

Datasets

The authors introduce the FRECO Corpus, a new benchmark dataset specifically designed for this task. It comprises approximately 3,800 annotated event mention pairs, sourced from ideologically diverse news coverage of four contentious topics: the 2024 Putin's election win (PUTIN), the Al-Shifa hospital raid (AL-SHIFA), the 2019 July 1 Hong Kong protest (HONGKONG), and the Kyle Rittenhouse shooting (RITTENHOUSE). The corpus is built upon the Richer Event-CorefBank (RECB) dataset.

Key Insights and Ideas

  • Event-level framing is crucial: Focusing on coreferential events, rather than just topics, allows for a more granular and accurate understanding of how narratives diverge.
  • Divergent framing is a signal, not noise: Unlike traditional coreference tasks that aim for exact matches, FRECO treats variations in argument structure, granularity, and perspective as the key signal for identifying framing differences.
  • LLMs can be effectively fine-tuned for FRECO: Both SFT and DPO, especially in combination, significantly improve performance on the FRECO task, outperforming zero-shot baselines.
  • SRL enhances event understanding: Incorporating structured semantic role information leads to better event representations and improved FRECO classification.
  • Bootstrapping scales analysis: The proposed bootstrapped mining pipeline effectively expands the annotated dataset, enabling large-scale analysis of framing divergence.

Key Results

  • The combined SFT and DPO fine-tuning strategies consistently achieved the best performance across various models and topics, outperforming standalone SFT or DPO.
  • Larger LLMs (Llama-3.1-8B) generally outperformed smaller ones (Llama-3.2-3B).
  • The best-performing fine-tuned LLMs consistently outperformed fine-tuned RoBERTaBASE cross-encoder classifiers.
  • GPT-4, when used in a zero-shot setting, underperformed fine-tuned models, highlighting the need for task-specific adaptation.
  • The bootstrapped mining pipeline successfully added 6,693 new positive FRECO pairs, augmenting the original 1,765 gold pairs, and achieved an estimated 70.5% precision.
  • FRECO events were shown to be effective framing devices for Media Attitude Detection, improving classification accuracy compared to random event selection.

Limitations and Future Work

  • Limited Topical Coverage: The FRECO corpus currently spans only four contentious topics. Future work should expand to a wider range of topics, including less polarized domains and alternative media types (e.g., social media, international news).
  • Characterizing Framing Differences: The current approach detects whether events differ in framing but not how. Future work could categorize framing contrasts along dimensions like tone, moral attribution, or causal focus to enhance interpretability.
  • Ethical Considerations: The authors acknowledge that framing analysis tools could be misused to discredit narratives or manipulate public perception. They advocate for responsible use.
  • Interpretable Framing Categorization: Future work could enhance interpretability by categorizing framing contrasts along established dimensions, such as tone, moral attribution, causal focus, or frame theme.

Summary updated

BibTeX
@inproceedings{zhao2025seeing,
  author = {Jin Zhao and Xinrui Hu and Nianwen Xue},
  title = {{Seeing the Same Story Differently: Framing‐Divergent Event Coreference for Computational Framing Analysis}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure