PaperlayerView exact revision on arXiv

Source-linked AI summary

High prevalence of antagonizing rhetoric in trending topics on Twitter/X

Armin Pournaki, Eckehard Olbrich

arXiv:2610.22455v1cs.SIcs.CL

TL;DR

Prior research often treats political out-group negativity as a single dimension, leaving unclear whether social-media discourse contains antagonization or criticism. The paper uses an LLM-based approach to distinguish these modes and identify their targets in German trending-topic tweets. Antagonization is widespread, more prevalent among right-leaning accounts, and the opposing camps construct distinct out-groups.

  • Problem

    Large-scale studies often treat out-group negativity as one dimension, leaving unclear whether it reflects antagonization that denies legitimacy or criticism that preserves it.

    Source: §1 · p. 2

  • Method

    The paper applies an LLM-based annotation approach with statistical correction to measure antagonization and criticism and identify their targets in German trending-topic tweets.

    Source: §1 · p. 2

  • Results

    More than 30% of tweets contained antagonization versus about 20% criticism; rates were 48.8% for right-leaning and 28.4% for left-leaning accounts.

    Source: §7 · p. 11

  • Takeaways & Limitations

    The opposing political camps construct distinct out-groups, with the AfD and the Green party emerging as the main antagonists of the opposing sides.

    Source: §1 · p. 2

    An underlying antagonistic relation may exist without manifesting in the rhetoric the paper’s operationalization detects.

    Source: §2.2 · p. 3

Abstract

from arXiv · show

Disagreement between political opponents is a central element of democracy. This disagreement can be voiced in different ways. An opponent can be antagonized, i.e. portrayed as an enemy that needs to be overcome, or they can be criticized, i.e. treated as a legitimate adversary with whom one can engage. While the latter is a fundamental part of democratic exchange, the former can be detrimental to public discourse. Political disagreement increasingly takes place on social media, where negativity towards political out-groups has been found to drive engagement. However, existing research has largely treated such negativity as a single dimension, leaving open whether it is antagonization or criticism that circulates on these platforms. This work addresses this gap by introducing a computational approach to detect these two types of rhetoric in text and identify its targets. On a corpus of over 17 million tweets from trending topics in the German Twittersphere, the method reveals that about a third of all tweets contain antagonizing rhetoric. Comparing tweets from trending topics to a random sample shows that the discourse in trending topics is systematically more antagonizing than ordinary Twitter communication. Combining antagonization detection with a large-scale estimation of users' political leaning, we find that right-leaning users employ antagonizing rhetoric significantly more than their left-leaning counterparts. Furthermore, both forms of oppositional rhetoric are strongly amplified within political camps, with multipliers retweeting them up to twice as often as tweets without oppositional rhetoric. Finally, patterns of antagonized targets reveal how different political camps construct their out-groups, with the right-wing populist party AfD and the Green party as the central antagonists of the opposing camps.

1 Introduction

The paper distinguishes criticism, which preserves an opponent’s legitimacy, from antagonization, which denies it, and measures both in German trending-topic tweets. It addresses the gap left by research that treats out-group negativity as one dimension, finding that antagonization is widespread and varies across political camps.

1 Introduction

  • Criticism treats opponents as legitimate participants in political exchange, whereas antagonization portrays them as threats with no legitimate place in politics.The distinction turns on whether opponents’ legitimacy is upheld or denied; either mode can be harsh.Source: §1 · p. 1
  • Antagonization may erode mutual respect needed for deliberation and compromise, while criticism leaves opponents within an ongoing exchange.A shared antagonist may also help define group boundaries, particularly for heterogeneous in-groups.Source: §1 · p. 1
  • Out-group animosity is associated with greater online sharing and predicts virality more strongly than negative affect or moral-emotional language.More generally, negative, high-arousal content tends to spread further than neutral or positive content.Source: §1 · pp. 1–2
  • Right-leaning accounts antagonize substantially more than left-leaning accounts, and the opposing camps construct distinct out-groups.The AfD and the Green party emerge as the main antagonists of the opposing sides.Source: §1 · p. 2
  • Existing large-scale analyses measure out-group negativity as one dimension, leaving unclear whether it reflects delegitimizing antagonization or legitimacy-preserving criticism.The paper measures both modes in over 17 million tweets from German trending topics between 2021 and 2023.Source: §1 · p. 2

2 Theoretical background

The theoretical framework distinguishes antagonization from criticism by whether opponents’ legitimacy is denied or upheld. It relates this distinction to agonistic theory and to accounts of group identity formed through shared out-groups.

2.1 Defining antagonization

  • Antagonization portrays a target as an enemy or threat, or uses demeaning personal attacks; criticism objects to actions or ideas without denying the opponent’s legitimacy.The distinction centers on enemy framing or personal disparagement versus disagreement focused on actions, decisions, policies, or ideas.Source: §2.1 · pp. 2–3

2.2 Relations to the theory of agonistic pluralism

  • The framework shares Mouffe’s legitimacy threshold between antagonism and agonism, but measures expressed rhetoric rather than her ontological political relation.An underlying antagonistic relation may exist without appearing in detectable rhetoric, and the paper’s criticism category captures only one register of Mouffe’s agonistic exchange.Source: §2.2 · p. 3

2.3 Negational identities and chains of equivalence

  • Negational identities define an in-group through what it rejects, while common antagonists can unite heterogeneous groups and link otherwise separate struggles.This mechanism is not tied to one ideology; the authors expect antagonization to target a range of entities that together form an in-group’s “constitutive outside.”Source: §2.3 · pp. 3–4

3 Related methodological work

Prior computational approaches measure political negativity through dictionaries, sentiment or emotion, supervised classifiers, and more recently language models. The paper’s approach adds distinctions between antagonization and criticism and identifies targets without prespecifying out-groups.

Dictionaries and sentiment analysis

  • Dictionaries offer interpretability but may not transfer across contexts, while sentiment and emotion tools miss nuanced forms of hostility such as irony, sarcasm, and ridicule.Sentiment-based approaches can treat negative language as polarization even though polarization does not necessarily involve negative emotion.Source: §3 · p. 4

Supervised classifiers

  • Existing supervised classifiers generally treat out-party negativity as one dimension, collapsing substantive criticism and hostile antagonization.Prior work used labeled examples to track engagement, polarization over time, or differences in out-group animosity across topics.Source: §3 · pp. 4–5

Large language models

  • Recent studies use LLMs to classify partisan animosity or combine stance, emotional tone, and agreement patterns to analyze polarization.These approaches include feed-based reranking and large-scale social-media text analysis.Source: §3 · p. 5

Contribution

  • The paper uses LLMs without predefined keywords and distinguishes delegitimizing antagonization from criticism while extracting targets without assuming out-groups in advance.Its target analysis lets collectively constructed out-groups emerge inductively.Source: §3 · p. 5

4 Data

The study draws on German Twitter data from trending topics and a separate German-language sample of ordinary tweets. The collections overlap in time, enabling comparison of the two settings.

4 Data

  • The first Twitter collection, D1, contains 17,666,756 tweets associated with trending topics in Germany, collected over the period from April 2021 through July 2023.The authors also describe D1 as over 17 million tweets.Source: §4 · p. 5
  • The second collection is a subset of Twitter’s 1% Spritzer dataset, covering four months across 2022 and 2023.Those months were chosen to overlap with the trending-topic collection.Source: §4 · p. 5
  • Restricting the Spritzer sample to German-language tweets yields over 2 million tweets for the four overlapping months.This restriction was applied for comparability with the trending-topic data.Source: §4 · p. 5

5 Methods

The study builds an annotated-text and language-model workflow to classify antagonization, criticism, and targets, then combines those estimates with topic and retweet-network analyses of political camps. It corrects classifier-based category estimates for error and uses retweet patterns to examine issue and camp structures.

5.1 LLM-based extraction of antagonization

  • Two native German-speaking authors independently annotated 1,000 tweets, reconciled 256 disagreements (κ = 0.57), and established the ground-truth dataset.Tweets were labeled for antagonization, criticism, or neither toward a specific target.Source: §5.1 · p. 5
  • The selected qwen3:32b classifier achieved macro F1 0.7 and weighted F1 0.75 after prompt refinement and screening out models below 7B parameters.The structured prompt was based on the codebook and evaluated against human annotation.Source: §5.1 · p. 5
  • Prediction-powered inference uses the human-annotated validation set to correct category proportions for classifier error and construct uncertainty intervals.The authors use it because applying the imperfect classifier to the full corpus would yield biased category-proportion estimates.Source: §5.1 · p. 6
  • Target extraction agreed with human annotation in 84% of cases, after which the authors manually merged different mentions of the same target.The classifier identifies potential targets of antagonizing or criticizing rhetoric.Source: §5.1 · p. 6

5.2 Structural analysis of issue polarization

  • BERTopic-derived tweet topics were manually aggregated into broader issues, and each trend received the issue label of its majority of constituent tweets.The pipeline uses pre-trained sentence embeddings and clustering to surface fine-grained topics.Source: §5.2 · p. 6
  • Retweet networks and stochastic block modeling support a two-camp user alignment; the analysis also tracks issue patterns and the most-retweeted influencers and most-retweeting multipliers.The study’s figures display estimated category shares by dataset and political leaning, and monthly antagonization shares, with confidence intervals.Source: §5.2 · pp. 6–7, Figure 1 · p. 7, Figure 2 · p. 7

6 Results

Antagonizing rhetoric is common in German Twitter trends, particularly among right-leaning users, and is more prevalent there than in a random sample. Its prevalence varies by issue, oppositional tweets are amplified, and political camps target different out-groups.

6.1 Antagonization statistics

  • 32.7% of trending-topic tweets contain antagonization, and the share is 48.8% among right-leaning tweets versus 28.4% among left-leaning tweets.Criticism appears in 21.1% overall and is more common on the left (23.8%) than the right (18.9%).Source: §6.1 · p. 7
  • 32.5% of trending-topic tweets are antagonizing versus 18.2% in the random sample; criticism is 20.2% versus 12.0%.Both forms are more prevalent in trending topics in every compared month.Source: Figure 3 · p. 8, §6.1 · p. 8
  • Antagonization rose from 28% in October 2021–July 2022 to 37% in October 2022–July 2023.The paper relates lower-share months to trending issues with more informational or apolitical content, including Covid, Christmas, and sports.Source: §6.1 · p. 8
  • 64% of tweets about Right-Wing Extremism are antagonizing, compared with 54% about Racism and 50% about Migration; several issues show camp-specific differences.Migration’s antagonizing content is over 73% attributable to right-leaning accounts, while left-leaning accounts predominantly antagonize on Right-Wing Extremism, Abortion, and Social Media; even nominally apolitical issues contain ideological antagonization.Source: §6.1 · p. 9, Figure 4 · p. 9

6.2 Amplification of antagonizing and criticizing rhetoric

  • Overall amplification factors are 1.62 for antagonization and 2.18 for criticism, relative to tweets without oppositional rhetoric.Source: Table 1 · p. 10, §6.2 · p. 10
  • Same-camp multipliers amplify antagonizing and critical tweets strongly, while cross-camp amplification is weaker or below the none-tweet baseline.Cross-camp amplification estimates use relatively few retweets because multipliers rarely retweet opposing-camp influencers (<1% on the left and 3% on the right). Right-leaning influencers supply more antagonizing originals (44% versus 22%), although relative amplification is slightly weaker than on the left.Source: §6.2 · p. 10, Table 1 · p. 10

6.3 Target analysis

  • The Green party and AfD are the central antagonized targets of opposing camps; left-leaning accounts also target Querdenker and Putin, while right-leaning accounts target media, government, migrants, and refugees.Source: §6.3 · p. 10, Table 2 · p. 11
  • The AfD is the left’s most antagonized target but is absent from its top 20 criticized targets; the same distinction holds for Russia(ns).The AfD also ranks sixth among targets within the right-leaning camp, partly reflecting self-victimization narratives; genuine intra-camp antagonization is also present, including left-wing attacks on the Green party over climate goals.Source: §6.3 · pp. 10–11

7 Discussion

The discussion finds that antagonization is common in German trending topics and differs from criticism in its democratic implications and political distribution. It also highlights amplification, target patterns, temporal change, and limitations of measuring rhetoric tweet by tweet.

7 Discussion

  • The framework separates criticism that preserves an opponent’s legitimacy from antagonization that portrays the opponent as an illegitimate threat, addressing a limitation of measures that conflate policy disagreement with personal hostility.Antagonization may undermine mutual toleration, but the classification does not distinguish justified from unjustified cases.Source: §7 · pp. 12, 14
  • 48.8% of right-leaning tweets are antagonizing, compared with 28.4% of left-leaning tweets; the authors suggest Germany’s governing coalition may partly explain this asymmetry.Among influencers, the corresponding rates are 44% on the right and 22% on the left.Source: §7 · p. 12
  • Right-leaning users also antagonize mainstream media and migrants, consistent with populist communication that constructs enemies beyond political elites.The study’s target patterns identify the Green party and AfD as central antagonists of opposing camps.Source: §7 · pp. 11–12
  • Multipliers retweet antagonization from their own camp’s influencers about twice as often as unclassified tweets, and criticism’s overall amplification factor exceeds antagonization’s (2.19 vs 1.62).Cross-camp amplification factors are higher for criticism, while cross-camp retweeting remains rare.Source: §7 · p. 13
  • Trending topics contain 32.5% antagonizing rhetoric versus 18.2% in the random sample, while antagonization also rises after September 2022 in both datasets.The increase is only partly consistent with changes following Twitter’s ownership transition; it is already visible before most took effect.Source: §7 · p. 13
  • The estimates rely on an LLM classifier with macro F1 of 0.70 and corrections that assume its errors remain stable over time and transfer to the random sample.Criticism is difficult for both annotators and classifiers to identify; isolated-tweet judgments may miss context or antagonization built across tweets.Source: §7 · p. 13
  • About 50% of classified tweets contain either antagonization or criticism, leaving other discursive strategies outside the framework’s two categories.The authors note solidarity rhetoric as a possible counterpart to antagonization for promoting in-group cohesion, especially on the left.Source: §7 · pp. 13–14
  • More than half of tweets contain antagonizing or criticizing rhetoric, with antagonization accounting for about a third and appearing more often among right-leaning accounts.The study identifies the AfD and the Green party as the opposing camps’ central antagonists.Source: §7 · p. 14

Funding statement

The authors acknowledge support from the European Union’s Horizon Europe program.

Funding statement

  • AP and EO acknowledge Horizon Europe funding under grant agreement 101094752 for the Social Media for Democracy (SoMe4Dem) project.The project concerns understanding the causal mechanisms of digital citizenship.Source: Funding statement · p. 17

S.1.1 Trending tweets in Germany

The trending-tweet dataset covers German trends over more than two years, collecting tweets around the five most frequent daily topics.

S.1.1 Trending tweets in Germany

  • From March 29, 2021, to July 12, 2023, the dataset contains 19,105,532 German-speaking tweets across 2,693 trends, collected for each trend on its trending day and the following day.The five most frequent daily trends were selected using repeated checks of Germany’s Twitter Trend API listing.Source: §S.1.1 · p. 18

S.1.2 1% random sample

The study uses a platform-level random sample of German-language public tweets to compare with the trending-topic data.

S.1.2 1% random sample

  • The four-month German-language Spritzer sample contains 2,186,054 tweets from a uniform 1% stream of public tweets.The selected months overlap with the trending-topic collection period.Source: §S.1.2 · p. 18

S.1.3 Data completeness

The supplement documents data-collection and annotation limitations, classifier and target-validation procedures, and rules for labeling rhetoric and targets. It also describes prediction-powered inference assumptions and uncertainty accounting, target normalization, and supplementary topic-level tweet visualizations.

S.1.3 Data completeness

  • The Search API captured 80% of tweets collected via the Stream API in a cited comparison, while trend collection lasted at most 24 hours and had 96% total uptime.The collection was inactive from 2022-07-30 to 2022-09-03, and the 24-hour limit could omit tweets at a trend’s onset.Source: §S.1.3 · p. 18

S.2 Human annotation procedure

  • The two annotators’ agreement was κ = 0.57, with implicit contextual criticism and the boundary between criticism and ridicule particularly difficult to judge.Disagreements were reconciled through discussion to create the ground-truth dataset.Source: §S.2 · p. 19, Figure S1 · p. 19

S.3 Prompt

  • The prompt distinguishes antagonization from criticism by enemy framing or personal disparagement versus objections without those features, and identifies targets only for the author’s own stance.It treats attributed or quoted rhetoric as “none” and requests directly stated target names.Source: §S.3 · p. 20

S.4 Model selection

  • The model comparison evaluated open-weight models on the 1,000 annotated tweets and sorted them by macro F1 score.The supplied table caption does not report model-specific scores.Source: Table S1 · p. 22

S.5 Validation of target extraction

  • 84% of extracted targets were correct among 400 tweets where human annotation and the classifier both identified antagonization or criticism.This agreement-restricted evaluation assesses target extraction independently of classification errors.Source: §S.5 · p. 24, Table S4 · p. 24

S.6 Target merging

  • The authors manually merged variant strings for the 50 most frequent targets in each political camp to consolidate references to the same entity.They first removed trailing spaces and hashtag symbols and replaced Twitter handles with account names.Source: §S.6 · p. 24

S.7 Targets of criticism

  • The supplement presents a table of the top criticism targets by political leaning, but the supplied passage gives no target names or values.Source: Table S5 · p. 26

S.8 Statistical uncertainty quantification via prediction-powered inference (PPI)

  • Across all tweets, the classifier understates criticism by 4.3 percentage points and overstates “none” by 2.8 and antagonization by 1.5; for unassigned accounts, it overstates antagonization by 7.9 and understates criticism by 9.3.Prediction-powered inference corrects classifier shares using a random human-annotated validation set of 1,000 tweets; within political camps, corrections do not exceed 4 percentage points. The method assumes camp assignments are error-free and, for monthly estimates, that classifier errors are stationary.Source: §S.8 · pp. 27–28, Table S6 · p. 28
  • PPI intervals include corpus-sampling variance, while classifier measurement error dominates total uncertainty; inference assumes representative validation annotations, an error-free leaning classifier, and stationary classifier errors over time.Bootstrap intervals alone would understate uncertainty by orders of magnitude because they omit classifier measurement error; reporting them alongside PPI intervals would present only a subset of the accounted-for uncertainty.Source: §S.8 · p. 28

S.9 Number of tweets per topic

S.10 Codebook for identifying antagonization and criticism in tweets

  • The codebook distinguishes antagonization—hostile or demeaning framing—from criticism of actions, policies, or ideas that treats opponents as legitimate; neutral or merely attributed stances are labeled none.When criticism and antagonization coexist, antagonization takes precedence; ambiguous stance or target defaults to none.Source: §S.10 · p. 30
  • For oppositional tweets, annotators identify direct targets and quote them exactly as written, without inferring specificity; attribution, endorsement, hashtags, mentions, and demeaning emojis follow explicit stance rules.The rules distinguish attacks on people from criticism of policies and treat ambiguous stance or target conservatively as none.Source: §S.10 · pp. 31–32

Related papers

Loading 2610.22455v1…