Tracking Perspectives on Event Participants: a Structural Analysis of the Framing of Real-World Events in Co-Referential Corpora

Levi Remijnse, Pia Sommerauer, Antske Fokkens, Piek T.j.m. VossenCOLING 2024

TL;DR

This paper introduces a novel method for analyzing how news media frames real-world events and their participants over time. By combining referential grounding (linking mentions to real-world entities) with FrameNet's semantic role analysis, they track shifts in participant focus and perspective in news streams, revealing how narratives evolve.

What Problem Does This Paper Address?

News reporting on significant events isn't static; it evolves as new information emerges and the narrative develops. This evolution can lead to changes in which participants are highlighted (focalization) and how they are described (framing). Existing NLP tools often handle reference resolution or semantic role labeling separately, but lack a unified approach to track how these aspects interact within a dynamic, co-referential news corpus. This paper aims to bridge this gap by analyzing how the framing of event participants changes over time as the surrounding narrative unfolds, using a referentially grounded corpus.

Methods

The core of this research lies in a structural linguistic analysis of a specially curated dataset. This dataset consists of Dutch news articles co-referencing specific real-world events (the Utrecht shooting and the MH17 crash). The key methodological innovations are:

  • Referential Grounding: Mentions in news articles are linked to structured data (Wikidata) representing real-world entities and events. This allows tracking who is being talked about.
  • Frame Semantics (FrameNet): Mentions are annotated with FrameNet frames and frame elements. This captures how participants are described and what conceptual roles they play in the event.
  • Temporal Distance Clustering (TDC): News articles are grouped into temporal clusters based on their distance from the anchor event. This allows for analyzing changes in framing across different stages of the event's reporting lifecycle.
  • Participant Categorization: Participants are categorized as Directly Related Participants (DpA - directly involved) or Indirectly Related Participants (IpA - indirectly involved).

By combining these elements, the researchers can analyze shifts in both the frequency of mentions (focalization) and the types of semantic roles assigned (framing) to participants across different temporal stages of the narrative.

Datasets

The study utilizes a referentially grounded corpus collected by Remijnse et al. (2022). This corpus was specifically designed for this type of analysis and includes:

  • Utrecht Shooting Corpus: 42 Dutch news articles related to the 2019 Utrecht shooting, annotated with entity-links to Wikidata and FrameNet frames.
  • MH17 Corpus: 117 Dutch news articles related to the Malaysia Airlines Flight 17 (MH17) crash, similarly annotated.

The paper does not introduce a new benchmark dataset but rather leverages and analyzes this existing, specially constructed resource.

Key Insights and Ideas

  • Narrative Drives Framing: The study strongly suggests that changes in the news narrative directly influence how participants are framed. As the story evolves, different participants become focal, and their perspectives shift accordingly.
  • DpAs vs. IpAs Behave Differently: Directly involved participants (DpAs) are consistently referenced to anchor the narrative, even if their role diminishes. Indirectly involved participants (IpAs) can emerge as focal points in later stages of the narrative, especially in complex events like MH17.
  • Anchoring vs. Active Role: Participants frequently mentioned to anchor a story (e.g., to link a current event to a past anchor incident) might retain a fixed framing even if their active role in the narrative wanes. This highlights the dual function of mentions: anchoring and active participation.
  • Event-Specific Narratives: The patterns of focalization and framing are unique to each event, indicating that the specific nature of the incident and its unfolding story significantly shape the media's portrayal of participants.

Key Results

The analysis revealed distinct patterns for the two anchor incidents:

  • Utrecht Shooting: The narrative evolved around DpAs, with shifts in focus from the perpetrator (Gökmen Tanis) to victims, and then back to the perpetrator in later stages. Framing also varied, with elements like "Trial@DEFENDANT" increasing as the narrative progressed.
  • MH17: IpAs played a more prominent role over time. The Dutch Safety Board was initially focal, but later, the Russian government and suspects took over the narrative. The victims, while still mentioned for anchoring, showed a fixed framing, indicating their active narrative role had diminished.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Scope: The study is limited to two specific incidents and one incident type. To find stronger, generalizable patterns, more event types and sources of reference text are needed.
  • Manual Annotation: The current corpus was manually annotated. Future work should focus on developing automated techniques for frame annotation, coreference resolution, and entity-linking to scale up the research.
  • TDC Selection: The temporal distance clusters (TDCs) were hand-picked. Automating this process is crucial for larger-scale analysis.
  • Discourse-Level Tools: Current tools are not fully equipped to annotate frame and frame element relations at the discourse level, which could provide deeper insights.

Future work will involve scaling up the corpus collection and annotation using automated methods and exploring more diverse event types to identify broader patterns in narrative framing.

Summary updated

BibTeX
@inproceedings{remijnse2024tracking,
  author = {Levi Remijnse and Pia Sommerauer and Antske Fokkens and Piek T.J.M. Vossen},
  title = {{Tracking Perspectives on Event Participants: a Structural Analysis of the Framing of Real-World Events in Co-Referential Corpora}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure