fg-zettelkasten

On narrative: The rhetorical mechanisms of online polarisation

3 min read

On narrative: The rhetorical mechanisms of online polarisation

Summary

This paper introduces narrative polarisation as a distinct dimension of online polarisation, defined as the divergent ways partisan groups assign actants — heroes, villains, helpers, victims — to key actors in a contested issue. The authors operationalise Greimas’ Actantial Model via a large language model and apply it to 212 YouTube videos and ~90,000 comments on the Israeli-Palestinian conflict. They find that while videos encode sharply opposed narratives (especially about who perpetrates violence versus who claims security or rights), comments substantially flatten this surface-level divergence. However, underlying narrative motifs — convergent, adversarial, and dependent configurations of actors — continue to carry partisan structure, suggesting that narrative polarisation can persist beneath apparent consensus and may be partially orthogonal to user ideology.

Key Contributions

  • Formalises “narrative polarisation” as a construct complementing affective and ideological polarisation, grounded in structuralist narratology.
  • Scales Greimas’ Actantial Model from folktale analysis to large-scale social media discourse.
  • Releases an LLM-based annotation pipeline (DeepSeek-R1-Distill-Qwen-32B) with codebook, validation data, and an OSF repository for actantial role extraction.
  • Proposes quantitative measures — overlap coefficient, subject divergence, and narrative motif typology — for analysing narrative-level polarisation.
  • Empirically shows divergence between content-level and comment-level narrative polarisation, complicating echo chamber accounts.

Methods

The authors retrieve YouTube videos using partisan search queries derived from Crowd Counting Consortium offline protest claims, plus neutral baselines, yielding 107 Israeli-leaning and 105 Palestinian-leaning videos (Oct 2023–Oct 2024) and 90,029 top-level comments. Whisper-large-v3 transcribes videos, segmented at 150-word boundaries. DeepSeek-R1-Distill-Qwen-32B annotates actantial roles using 21 actor labels and 7 object labels, validated against two expert annotators on 292 comments (avg micro F1 0.73, Krippendorff’s α 0.59 — comparable to inter-human agreement). Analyses combine overlap coefficients, a subject-divergence statistic with permutation tests (Bonferroni/FDR corrected), and qualitative close reading of three identified narrative motifs.

Findings

  • Videos showed strong narrative polarisation: Israeli-leaning videos attributed violence to Palestinian actors, while Palestinian-leaning videos attributed security and rights/freedoms claims primarily to Israeli actors.
  • Comments substantially reduced surface divergence: between-group overlap rose from 0.63 (transcripts) to 0.80 (comments), and average absolute subject divergence fell from 0.19 to 0.07.
  • The largest convergence occurred around violence attribution (divergence collapsing from -0.43 to -0.05); a new IS-skewed peace narrative emerged in comments.
  • Convergent motifs: in-group critique (e.g., Israeli-leaning commenters criticising Israeli security failures).
  • Adversarial motifs: structured violence discourse, often justifying retaliation against the opposing side.
  • Dependent motifs: predominantly cast Palestinians as subjects dependent on Israeli control, especially in Israeli-leaning discussion threads.
  • LLM–human inter-coder agreement matched human–human agreement and exceeded prior narrative-annotation benchmarks.

Connections

This paper extends polarisation research beyond opinion- and interaction-based measures toward rhetorical structure, complicating echo chamber findings such as Bakshy2015-rn and resonating with work showing that exposure-level polarisation does not map cleanly onto audience attitudes, e.g. Tornberg2025-ir and Tornberg2026-lc. Its LLM-based annotation pipeline connects methodologically to recent uses of generative models for measuring partisan or ideological content like Le-Mens2025-qz and Hackenburg2025-dj, while the narrative-motif analysis of conflict discourse complements affective and identity-centred approaches in Mosleh2024-op and Knupfer2025-vt.

Podcast

A research-radio episode discusses this paper: Listen

Graph View

Backlinks