On Narrative: The Rhetorical Mechanisms of Online Polarisation

TL;DR

This paper formalizes narrative polarisation by applying Greimas' Actantial Model to two partisan YouTube information environments studying the Israeli–Palestinian conflict. It combines an LLM-based actant annotation pipeline with human validation to extract surface-level narratives and deeper narrative motifs from 212 videos and 90,029 comments, revealing that videos are highly polarised while comments partially depolarise on the surface but preserve deeper, motif-level differences. The study shows that narrative structures—beyond simple topic differences—shape perceptions of central actors and their interests, with motif congruence and opposition influencing how polarisation manifests across platforms. The proposed framework enables scalable, actor-centric analysis of polarisation in online discourse and offers a pathway to compare narratives across issues and media.

Abstract

Polarisation research has demonstrated how people cluster in homogeneous groups with opposing opinions. However, this effect emerges not only through interaction between people, limiting communication between groups, but also between narratives, shaping opinions and partisan identities. Yet, how polarised groups collectively construct and negotiate opposing interpretations of reality, and whether narratives move between groups despite limited interactions, remains unexplored. To address this gap, we formalise the concept of narrative polarisation and demonstrate its measurement in 212 YouTube videos and 90,029 comments on the Israeli-Palestinian conflict. Based on structural narrative theory and implemented through a large language model, we extract the narrative roles assigned to central actors in two partisan information environments. We find that while videos produce highly polarised narratives, comments significantly reduce narrative polarisation, harmonising discourse on the surface level. However, on a deeper narrative level, recurring narrative motifs reveal additional differences between partisan groups.

Figures (9)

• Figure 1: Greimas' Actantial Model greimas_structural_1984, containing six actants. These functional characters are arranged along three axes: desire, communication, and power. Arrows indicate the direction of the relationship.
• Figure 2: Subject–Object co-occurrence patterns across all comments. Each square represents the frequency of a given Subject (x-axis) and Object (y-axis) combination in the full dataset (n = 90,029 comments). Square size corresponds to the relative prevalence of each combination, with the most frequent being Israel/Israelis–violence (n = 6,078). Combinations with fewer than 60 comments are omitted for clarity. Comments referring to the meta-actors audience/commenter or video creator as Subject (28%) are excluded. Colour indicates the log ratio between observed and expected counts under the assumption of independence: positive values (red) denote combinations occurring more often than expected, and negative values (blue) denote combinations occurring less often.
• Figure 3: Subject divergence in comments and transcripts. Subject divergence refers to differences in how partisan groups attribute the Subject role to different conflict actors. Shown are the attribution patterns for various Objects. Values near zero indicate low divergence, reflecting similar attribution patterns across groups. Negative values reflect greater attribution by the Israeli-leaning group (relative to the Palestinian-leaning group) to Palestinian actors, whereas positive values reflect greater attribution to Israeli actors. Error bars represent 95% bootstrap confidence intervals (n=3000).
• Figure 4: Difference in narrative motif shares between partisan groups. Shown are three types of narrative motifs based on the Actantial Model (top left), with each representing a unique constellation of Israeli (IS) and Palestinian (PA) actors illustrated next to the motif titles. Markers indicate the relative motif difference between partisan groups for different surface-level narratives. Motif actors are determined by the surface-level subjects (IS or PA) shown on the x-axis. Negative values indicate higher prevalence in the Palestinian-leaning group, positive values indicate higher prevalence in the Israeli-leaning group. Only narrative-motif combinations with an average prevalence above 5% across partisan groups are shown. Error bars represent 95% bootstrap confidence intervals (n=3000).
• Figure 5: Subject divergence in comments and transcripts. Shown are the attribution patterns of the Subject role for different actors, varying the actor groupings used in the main text. The main-text grouping (Fig. \ref{['fig:narrative_polarisation']}) categorises Israeli actors as IDF, Israel/Israelis, Jews and Zionists, and Palestinian actors as Arabs, Hamas, Muslims and Palestine/Palestinians. Panels (a)–(c) vary these groupings by dropping actor pairs: (a) IDF and Hamas, (b) Zionists and Arabs, and (c) Jews and Muslims. Error bars represent 95% bootstrap confidence intervals (n=3000).