The Enemy from Within: A Study of Political Delegitimization Discourse in Israeli Political Speech

Naama Rivlin-Angert, Guy Mor-LanEMNLP 2025

TL;DR

This paper introduces a novel computational framework for detecting and analyzing "Political Delegitimization Discourse" (PDD) – symbolic attacks on the legitimacy of political opponents. By curating a large, annotated Hebrew corpus and developing a two-stage deep learning pipeline, they demonstrate a significant rise in PDD over three decades, with higher prevalence on social media and distinct patterns across political blocs and contexts.

What Problem Does This Paper Address?

Legitimacy is a cornerstone of democratic governance, but in recent years, democratic norms have eroded due to rising polarization and hostile rhetoric. A key manifestation of this erosion is "Political Delegitimization Discourse" (PDD), where political opponents are attacked not on their policies, but on their fundamental right to participate in the political system. While adjacent research has focused on incivility, hate speech, and toxicity, these frameworks often miss PDD because it can be subtle, relying on implication or framing rather than overt insults. Existing PDD research is limited to manual coding of small datasets, lacking scalability and longitudinal insight. This paper aims to bridge this gap by developing a computational approach to systematically identify and track PDD in political discourse.

Methods

The authors propose a two-stage deep learning pipeline for PDD detection and characterization.

  1. Binary PDD Classifier: This stage identifies sentences that express PDD. They experimented with both encoder models (mBERT, AlephBERT, HeRO, DictaBERT) and decoder LLMs (Gemma, Qwen, DictaLM 2.0).
  2. Multi-task Classifier for Properties & Span Classifier for Targets: For sentences identified as PDD, this stage further annotates them for intensity, incivility, target type (group, person, institute), and affective framing. They also identify the specific spans within the text that refer to the delegitimized target.

They fine-tuned these models on a novel, large-scale, manually annotated Hebrew corpus.

Datasets

The study utilizes a novel, manually annotated Hebrew-language corpus comprising 10,410 sentences from three sources:

  • Facebook posts: Official posts from verified Israeli political parties, incumbent MKs, and viable candidates (57K posts, 322K sentences from Dec 2018 - Apr 2021).
  • Parliamentary speeches: The IsraParlTweet dataset, covering Knesset floor speeches from 1992-2023.
  • News data: Articles from 38 leading Hebrew-language outlets (1.7M articles from Oct 2018 - Nov 2022).

Of the 10,410 sentences, 1,812 (17.4%) exhibit PDD, and 642 of these are further annotated for intensity, incivility, target type, and affective framing. The dataset was split into train (70%), validation (15%), and test (15%) sets.

Key Insights and Ideas

  • PDD is more than just incivility: The core innovation is defining PDD as a distinct analytical category focused on denying political legitimacy, rather than just rude or hateful language. This allows for the detection of subtler forms of discourse.
  • Computational analysis is feasible and valuable: The paper demonstrates that automated PDD analysis is not only possible but also yields significant insights into democratic discourse dynamics.
  • Social media amplifies PDD: PDD is more prevalent and intense on social media platforms compared to parliamentary debates, highlighting the role of unmediated communication.
  • Context matters significantly: PDD spikes around election campaigns and recedes during periods of coalition building, suggesting its strategic use in political competition.
  • Ideological differences in targeting: While all blocs use PDD, the right-wing bloc uniquely frames entire political groups as illegitimate, extending beyond individual targets.

Key Results

  • Upward Trend: PDD has shown a marked increase over the past three decades in Israeli parliamentary speeches, with a sharp rise since 2020.
  • Platform Differences: PDD prevalence is similar across Facebook and Knesset speeches (around 7%), but Facebook PDD is consistently higher in incivility, out-group references, harm to common good accusations, and intensity.
  • Gender Differences: Male politicians exhibit slightly higher average levels of PDD than their female counterparts, both in parliamentary speeches and on Facebook.
  • Bloc Differences: In the long term, right-wing politicians show the highest PDD. However, in the 2018-2021 period, left-wing politicians showed the highest PDD on both platforms, suggesting shifts in rhetorical strategies.
  • Contextual Spikes: PDD significantly spikes around national elections on Facebook, dropping sharply afterward. Opposition actors consistently use more PDD than coalition members.
  • Model Performance: The DictaLM 2.0 decoder model achieved the best performance, with an F1 score of 0.74 for binary PDD detection and a macro-F1 of 0.67 for classifying delegitimization characteristics.

Limitations and Future Work

  • Scope and Transferability: The findings are specific to the Hebrew language and Israeli political context. Adapting the framework to other languages and contexts requires linguistic expertise and domain knowledge due to varying political dynamics and social cleavages.
  • Sentence-Level Annotation: The current annotation scheme is sentence-level, potentially missing delegitimization that emerges across multiple sentences or conversational turns, especially subtle rhetorical strategies like irony or implication.
  • Single Annotator for Fine-grained Attributes: While the lead author, a senior expert, annotated the fine-grained characteristics, using a single annotator could introduce bias.
  • Descriptive and Correlational Analysis: The study documents variations in PDD but does not establish causal pathways.

Future work suggested includes integrating richer contextual signals, extending to diverse data sources, examining how delegitimizing rhetoric influences audience attitudes, and adapting the annotation scheme and modeling pipeline to other contexts of democratic erosion for comparative analyses.

Summary updated

BibTeX
@inproceedings{rivlinangert2025enemy,
  author = {Naama Rivlin-Angert and Guy Mor-Lan},
  title = {{The Enemy from Within: A Study of Political Delegitimization Discourse in Israeli Political Speech}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure