Analyzing the Use of Metaphors in News Editorials for Political Framing

Meghdut Sengupta, Roxanne El Baff, Milad Alshomary, Henning WachsmuthNAACL 2024 · DOI

TL;DR

This paper introduces a new dataset of news editorials annotated for metaphors and their persuasive effects on liberal and conservative readers. They found that liberals are significantly influenced by metaphors, while conservatives are more resistant, with the impact varying based on the metaphor's conceptual domain.

What Problem Does This Paper Address?

Political discourse heavily relies on metaphors to frame issues and influence public opinion. While linguistics and social sciences acknowledge the power of metaphors in shaping political perspectives, there's a significant gap in understanding how metaphors are used in political framing and their specific persuasive effects on audiences with different political ideologies. Existing research often focuses on metaphor identification or general framing effects, but rarely connects these directly to ideological differences in persuasion. This paper aims to bridge this gap by creating a dataset that links metaphorical language in news editorials to their impact on liberal and conservative readers.

Methods

The authors re-annotated an existing corpus of news editorials (Webis-News-Editorials18) with two levels of information:

  1. Metaphor Identification: They employed a procedure based on Steen's (2010) Metaphor Identification Procedure (MIPVU) to identify both single-word and composite metaphors. A key innovation here is the inclusion of composite metaphors, which are multi-word expressions that create metaphorical meaning, a feature often overlooked in previous datasets.
  2. Domain Annotation: For each identified metaphor, they annotated its source domain (where the meaning is derived from) and target domain (what is being described). They reused and combined domain lists from previous work (Shutova et al., 2010; Gordon et al., 2015).

To analyze the persuasive effects, they categorized the source and target domains into 14 ontological categories based on Gordon et al. (2015). Statistical tests (ANOVA, Kruskal-Wallis, t-tests, Mann-Whitney) were used to determine significant differences in metaphor ratios and domain counts across different persuasive effects (Challenging, Reinforcing, No Effect) for liberal and conservative readers.

Datasets

  • Source Data: The study utilized a subset of the publicly available Webis-News-Editorials18 dataset. This dataset originally contained 1000 New York Times editorials annotated for persuasive effects on liberal and conservative readers.
  • New Benchmark: The paper introduces a newly annotated corpus derived from 300 randomly sampled editorials from the Webis-News-Editorials18 dataset. This new corpus is annotated with:
  • Identification of single and composite metaphors.
  • Source and target domains for each metaphor.
  • The original persuasive effect annotations (Challenging, Reinforcing, No Effect) for liberal and conservative readers.

Key Insights and Ideas

  • Ideological Divide in Metaphor Susceptibility: The most striking finding is the differential impact of metaphors on liberals versus conservatives. Liberals are significantly swayed by metaphorical language, while conservatives appear more resistant. This suggests that metaphors act as a style feature that affects different ideological groups differently.
  • Domain-Specific Influence: The type of metaphor matters. For liberals, metaphors from "Darkness & Light" (e.g., death) and "Nature" domains have a significant impact. For conservatives, the "Nature" domain is more influential, though their overall resistance to metaphors is a more dominant finding.
  • Composite Metaphors Matter: The inclusion of composite metaphors is crucial, as they are abundant and contribute to the overall metaphorical landscape of political discourse.
  • Persuasion is Complex: The study highlights that persuasion isn't monolithic. The effectiveness of metaphors depends on the reader's ideology and the specific conceptual domain the metaphor draws from.

Key Results

  • The dataset contains 12,006 metaphors across 300 news editorials, with 70% being single metaphors and 30% composite.
  • Liberals: Showed significant differences in metaphor counts across persuasive effects (Challenging, Reinforcing, No Effect) with medium effect sizes (r ≈ 0.3-0.35). Reinforcing and Challenging editorials had significantly more metaphors.
  • Conservatives: No significant differences were observed in metaphor counts across the three persuasive effects, suggesting metaphors have little to no significant effect on them. This aligns with previous findings suggesting conservatives are less influenced by stylistic features of liberal-leaning newspapers.
  • Domain Analysis:
    • For liberals, "Darkness & Light" and "Nature" source domains were more prominent in effective editorials.
  • For conservatives, "Journey" source domains were more prominent in reinforcing editorials.
  • Target domains related to "Power & Control" were less prominent in effective editorials for conservatives.

Limitations and Future Work

  • Subjectivity of Metaphoricity: The authors acknowledge that metaphoricity exists on a spectrum, and their approach of considering all degrees of metaphoricity for simplicity might have missed some nuanced metaphorical utterances.
  • Context Window for Composite Metaphors: The context window for identifying composite metaphors was limited to two words on either side of the pivot word. In some cases, the entire sentence might be needed for proper identification.
  • Confounding Factors in Persuasion: The persuasive effect annotations were from a previous study, and the authors couldn't control for other human judgment factors (like personality) that might influence persuasion beyond ideology. Future work could involve direct comparison studies to validate these findings.
  • Dataset Size: While 300 editorials is substantial, the authors note that gaining deep insights into source and target domains is challenging with this size. Future work could expand the dataset.

Summary updated

BibTeX
@inproceedings{sengupta2024analyzing,
  author = {Meghdut Sengupta and Roxanne El Baff and Milad Alshomary and Henning Wachsmuth},
  title = {{Analyzing the Use of Metaphors in News Editorials for Political Framing}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.naacl-long.199},
}
Figure