Fine-Grained Identification of Manipulative Rhetoric

34 concepts26 papers1 proposal

When we read a news article or a social media post, we rarely encounter a text that is entirely fabricated from start to finish. Instead, modern misinformation operates through subtle manipulation: an article might be 90% factually accurate, yet rely on isolated phrases or sentences designed to mislead, provoke fear, or discredit an opponent.

Traditional computational approaches to combating misinformation rely heavily on document-level classification—labeling an entire article as "propaganda" or "fake news." However, as noted by Barrón-Cedeño et al. and Zhang et al., this coarse granularity acts as a blunt instrument. It fails to pinpoint the exact linguistic triggers being deployed and risks over-labeling otherwise truthful content. To truly understand and unmask manipulation, natural language processing has shifted toward fine-grained rhetorical analysis: identifying the exact spans of text where manipulation occurs and classifying the specific persuasive techniques employed.

The Dual Challenge: Spans and Techniques

Fine-grained analysis transforms propaganda detection from a simple binary classification problem into a complex, joint sequence-tagging and multi-class classification task. As formalized by Da San Martino et al. and popularized in the SemEval-2020 Task 11 benchmark by Barrón-Cedeño et al., the objective is twofold:

  1. Span Identification: Pinpointing the exact start and end character offsets of a manipulative fragment.
  2. Technique Classification: Labeling that specific fragment with its precise rhetorical device (e.g., loaded language, straw man, flag-waving, or appeal to fear).

To achieve this, researchers fine-tune transformer-based language models (such as BERT or RoBERTa) to make token-level predictions, as demonstrated by Altiti et al. and Arsenos et al.. Instead of processing the text as a single vector representation, the model evaluates the contextual embeddings of individual words to draw boundaries around deceptive fragments.

Why Local Context is Not Enough

While token-level sequence taggers excel at spotting explicit, highly localized techniques like loaded language or name-calling, they frequently struggle with more complex logical fallacies. A phrase like "the sky is blue" is benign on its own, but within a specific context, it could be part of a red herring or a straw man argument. Identifying these techniques requires understanding how a specific fragment relates to the broader narrative.

To bridge this gap, researchers have developed multi-granular architectures that combine micro-level token features with macro-level contextual signals:

  • Contextual Gating: Khosla et al. introduced a gating mechanism that uses sentence-level context to filter token-level predictions. If the surrounding sentence lacks a manipulative tone, the gate suppresses isolated token false positives.
  • Multi-Granularity Networks (mMGN): Work by Mittal et al. leverages joint architectures that simultaneously optimize for sentence-level classification and token-level span extraction, ensuring that the model's local boundaries align with global semantic themes.
  • Bottom-Up Aggregation: Conversely, Jáñez-Martino et al. demonstrated that identifying fine-grained techniques first allows a system to "climb the tree" of propaganda types, using the presence of specific local fragments to make more accurate macro-level assessments of a document's intent.

Guiding Models with Discourse Structure

The reason multi-granular approaches work is that propaganda rarely occurs at random; it is deeply intertwined with how an author structures an argument. Lei et al. advanced this intuition by incorporating discourse structures into the identification process.

By analyzing the logical and narrative roles that sentences play within a story—such as whether a sentence serves as an "evaluation" or a "causal deduction"—their framework teaches models that certain rhetorical contexts are far more likely to harbor manipulative language than purely descriptive ones. This structural guidance helps the model differentiate between standard journalistic reporting and calculated persuasion.

Beyond Detection: Fostering Media Literacy

The ultimate value of fragment-level analysis extends far beyond improving F1-scores; it provides interpretability. When a system flags an entire article as propaganda, the user is left to wonder why. By transitioning to sub-sentence annotations, computational tools can offer concrete evidence of deception.

Systems like Prta, developed by Da San Martino et al., visually highlight the exact words and phrases designed to bypass critical thinking. By showing readers exactly how they are being influenced—whether through emotional triggers or logical misdirection—fine-grained analysis transforms automated detection from a passive filtering mechanism into an active tool for media literacy.

Key Foundational Work

Fine-Grained Analysis of Propaganda in News Articles (2019) — contributed 4 concept elements to this direction.
YNU-HPCC at SemEval-2020 Task 11: LSTM Network for Detection of Propaganda Techniques in News Articles (2020) — contributed 3 concept elements to this direction.
SemEval-2020 Task 11: Detection of Propaganda Techniques in News Articles (2020) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_fine_grained_rhetorical_analysis_c12a7a61,
  title = {fine-grained rhetorical analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fine-grained-rhetorical-analysis/rdr_c12a7a61259742a399a810ea8960e04f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_fine_grained_rhetorical_analysis_c12a7a61_v1,
  title = {fine-grained rhetorical analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fine-grained-rhetorical-analysis/rdr_c12a7a61259742a399a810ea8960e04f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake