💬 NLP

ExTax: Explainable Disinformation Detection via Persuasion, Emotion, and Narrative Role Taxonomies

ExTax is a novel, explainable disinformation detection framework that unifies 17 dimensions of persuasive rhetoric, emotional manipulation, and narrative roles into a taxonomic space to outperform state-of-the-art baselines in accuracy and robustness while providing human-auditable manipulation profiles.

Original authors: Shang Luo, Yingguang Yang, Zhenchen Sun, Yang Liu, Bin Chong, Jingru Chen, Yancheng Chen, Jiayu Liang, Kefu Xu, Hao Peng, Philip S. Yu

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shang Luo, Yingguang Yang, Zhenchen Sun, Yang Liu, Bin Chong, Jingru Chen, Yancheng Chen, Jiayu Liang, Kefu Xu, Hao Peng, Philip S. Yu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet is a massive, noisy marketplace. For a long time, spotting a fake product (disinformation) was easy: you just looked for obvious typos or nonsense words. But now, with the rise of powerful AI, scammers are selling "perfectly wrapped" fake products. They look, sound, and feel exactly like the real thing. They don't just lie; they use clever tricks to make you feel a certain way or believe a story without you even realizing it.

The paper introduces ExTax, a new tool designed to catch these sophisticated fakes. Think of ExTax not as a simple "True/False" stamp, but as a forensic detective that breaks a story down into three specific "crime scenes" to see how it was manipulated.

Here is how ExTax works, using simple analogies:

1. The Three "Crime Scenes" (The Taxonomy)

Instead of just asking, "Is this fake?", ExTax asks three deeper questions, looking for specific patterns in the text:

The Rhetoric Trap (Persuasion): Imagine a salesperson who doesn't just sell a product but attacks your neighbor to make you feel safe buying theirs. ExTax looks for these "sales tricks." It checks if the text is attacking someone's reputation, oversimplifying a complex problem, or using confusing language to distract you.
The Emotional Trigger (Emotion): Have you ever shared a post just because it made you furious or terrified? ExTax acts like an emotion detector. It scans for specific high-intensity feelings like Anger, Fear, or Anxiety that are often used to bypass your logical brain and make you act impulsively.
The Character Roles (Narrative): Every story has heroes and villains. In fake news, the "villains" and "victims" are often assigned unfairly. ExTax looks at how people are framed. Are they painted as "Deceptive Subversives" (secret bad guys) or "Institutional Toxins" (corrupt systems)? Or are they "Ethical Stabilizers" (good guys)? Fake news often forces people into these extreme, negative roles.

2. The "Council of Experts" (The Method)

One problem with using AI to find fake news is that the AI itself might be confused or biased. To solve this, ExTax doesn't rely on just one AI.

The Divergent Council: It asks four different top-tier AI models to analyze the same text. It's like asking four different detectives to read a suspect's alibi.
The "Entropy" Filter: Sometimes the detectives disagree. One says, "This is definitely fear-mongering," while another says, "I'm not sure." Instead of forcing a yes/no answer, ExTax uses a clever math trick called Entropy-driven Label Smoothing.
- The Analogy: If all four detectives agree, ExTax is 100% confident. If they are split down the middle, ExTax says, "Okay, this is a gray area," and gives a "maybe" score instead of a hard "yes." This prevents the system from getting tricked by the AI's own confusion.

3. The "Spotlight" (The Detection)

Once ExTax has gathered these clues (the persuasion tricks, the emotional triggers, and the character roles), it uses a special Heterogeneous Multi-Head Attention mechanism.

The Analogy: Imagine a spotlight with three different colored lenses (Red for Persuasion, Blue for Emotion, Green for Roles). ExTax shines these lights on the text simultaneously. It doesn't just look at the whole picture; it zooms in to see exactly where the red light hits (the attack on reputation) and where the blue light hits (the fear). By combining these specific views, it builds a complete "manipulation profile" of the text.

Why This Matters (The Results)

The paper tested ExTax on five different types of content, from short social media posts to long news articles.

The "Genre" Problem: Many previous tools were great at spotting fakes in short tweets but terrible at spotting them in long news articles (or vice versa). It's like a security guard who is good at checking backpacks but bad at checking suitcases.
ExTax's Success: ExTax remained strong across both types. While other tools saw their accuracy drop by nearly 33% when switching from tweets to articles, ExTax stayed consistent. It achieved a top score of 0.8456, beating the current best methods.

The Bottom Line

ExTax doesn't just tell you "This is fake." It explains why. It gives you a report card showing: "This text tried to manipulate you by attacking a reputation, triggering your anger, and painting a politician as a villain."

This makes the detection process transparent and auditable, helping humans understand the specific psychological tricks being used, rather than just trusting a black-box computer verdict.

Note on Limitations: The authors admit that because they use expensive, commercial AI models to do the initial analysis, it costs money to run. Also, the system was trained mostly on English texts about Western politics and health, so it might not work as well on stories in other languages or from different cultures. Finally, they emphasize that this tool is for human decision support—it helps humans make better judgments, but it shouldn't be the sole judge of what gets deleted or kept online.

Technical Summary: ExTax – Explainable Disinformation Detection via Persuasion, Emotion, and Narrative Role Taxonomies

Problem Statement

The democratization of Large Language Models (LLMs) has accelerated the generation of highly fluent disinformation that relies less on surface-level falsity and more on sophisticated cognitive manipulation. Such content often combines persuasive rhetoric, emotional manipulation, and narrative role construction to influence reader interpretation through multiple pathways. Existing detection methods typically operate in isolation, focusing on single signals such as syntax, external knowledge, or isolated affective cues. Consequently, they struggle to capture the multi-faceted manipulative intents underlying modern disinformation and fail to provide human-auditable explanations. Furthermore, no existing corpus jointly annotates these complementary dimensions, and attributes elicited from a single LLM often inherit model-specific biases.

Methodology

The authors propose ExTax, a taxonomy-aligned framework designed to unify three complementary facets of manipulation into a single, interpretable detection system. The architecture consists of three core stages:

1. Taxonomic Attribute Elicitation

To address the lack of joint annotations and mitigate single-model bias, ExTax employs a Divergent Attribute Synthesis strategy.

Multi-LLM Querying: The system queries four frontier LLMs (GPT-5 mini, DeepSeek-V4-Flash, Qwen3.6 Flash, and Grok 4.1 Fast Reasoning) using taxonomy-specific prompts to extract binary features for three dimensions:
- Persuasion: 6 strategies (e.g., Attack on Reputation, Justification, Simplification).
- Emotion: 5 manipulation methods (e.g., Fear, Anger, Hope, Anxiety, Sadness).
- Narrative Roles: 6 categories (e.g., Ethical Stabilizers, Deceptive Subversives, Institutional Toxins).
Entropy-Driven Dynamic Label Smoothing: To reconcile disagreements among the LLMs, the system calculates the proportion of positive votes ( $p_{i,c}$ ) for each category. It then computes binary entropy ( $H_{i,c}$ ) to measure inter-model uncertainty. The final soft label ( $y^{(i,c)}_{tax}$ ) is a weighted combination of the vote proportion and a neutral value (0.5), where high disagreement (high entropy) pulls the label toward 0.5, and high agreement preserves the vote proportion. This transforms noisy hard decisions into calibrated soft supervision.

2. Taxonomy-Aligned Representation Learning

The framework maps input text into a unified 17-dimensional taxonomic space.

Adaptive Gated Pooling: Instead of standard mean pooling, the model uses a learnable gating vector to assign importance weights to individual tokens from a pre-trained LLM encoder (RoBERTa-large).
Facet-Specific Projection: Three independent Multi-Layer Perceptron (MLP) heads project the pooled representation into the respective taxonomic spaces (Persuasion, Emotion, Narrative Roles). The model is trained using Binary Cross-Entropy loss against the smoothed soft labels derived in the first stage.

3. Taxonomy-Based Disinformation Detection

The final detection phase integrates the learned taxonomic signals with contextual text encodings.

Feature Transformation & Prompting: Token representations are projected into the 17-dimensional taxonomic space. A Learnable Taxonomic Prompt mechanism concatenates the predicted taxonomic vector with learnable semantic prompts to recalibrate feature importance.
Heterogeneous Multi-Head Attention: A specialized attention mechanism is employed where the number of attention heads and their dimensions strictly correspond to the three taxonomic facets ( $d_1=6, d_2=5, d_3=6$ ). This ensures a rigorous one-to-one alignment between the attention mechanism and the specific manipulation strategies being modeled.
Prediction: The aggregated attention output is processed through a weighted sum and a compact MLP to generate the final binary classification (Disinformation vs. Reliable).

Key Contributions

Multifaceted Taxonomic Modeling: The paper reframes disinformation detection as reasoning over three complementary cognitive facets (persuasion, emotion, narrative roles), instantiated as a unified 17-dimensional taxonomic representation.
Noise-Robust Attribute Synthesis: The authors introduce a method to aggregate outputs from multiple frontier LLMs and regularize their disagreements via entropy-driven dynamic label smoothing, effectively converting inter-LLM disagreement into reliable synthetic supervision.
Taxonomy-Aligned Detection and Explanation: By integrating taxonomic signals with contextual encodings via heterogeneous attention, the framework enables robust detection while exposing the specific rhetorical, affective, and narrative evidence underlying each prediction.

Experimental Results

ExTax was evaluated across five cross-domain and cross-genre benchmarks (CoAID, ECTF, ISOT Fake News, MultiDis, and EUDisinfo), covering both short-form posts and long-form articles.

Overall Performance: ExTax achieved an overall Macro F1 of 0.8456, outperforming the strongest LLM-based baseline (RAEmoLLM) by +2.01 points and the strongest deep-learning baseline (RoBERTa-large) by +3.13 points.
Robustness to Genre Imbalance: The framework demonstrated significant robustness under severe genre imbalance (70% posts, 30% articles). While the strongest deep baseline (DRES) degraded from a Macro F1 of 0.9454 on posts to 0.6194 on articles, ExTax maintained consistent performance across both genres, suggesting that taxonomy-aligned features capture genre-invariant manipulation patterns.
Ablation Studies: The hybrid approach of divergent attribute synthesis and entropy-driven smoothing was shown to be critical. While single-LLM synthesis yielded lower performance in the representation learning stage, the hybrid approach provided better-calibrated supervision for the final detection stage, leading to superior overall F1 scores.

Significance and Claims

The paper claims that ExTax moves disinformation detection beyond binary "fake/real" verdicts toward an interpretable and auditable diagnosis of how deceptive content manipulates readers. By grounding predictions in a structured manipulation profile (e.g., identifying "Attack on Reputation" paired with "Anger" and "Deceptive Subversives"), the system offers transparency that resembles the analytical process of human fact-checkers.

The authors note that while ExTax improves detection robustness and explainability, it relies on commercial frontier LLMs for attribute elicitation, which introduces inference costs and potential reproducibility concerns as APIs evolve. Additionally, the evaluation is currently limited to English and Western-centric corpora, and the structured explanations have not yet been validated through human user studies with fact-checkers. Consequently, the authors position ExTax as a decision-support tool rather than a fully automated moderation system, intended to be used with human oversight.