We analyze how disinformation leverages are affected by combining lexicon- and model-based emotion detection with topic modeling across two English corpora: (i) a balanced “true vs. fake” news collection and (ii) a Twitter/X set labeled hate/offensive/neutral. Using the NRC Emotion Intensity Lexicon, then layering Structural Topic Modeling (STM) with emotion covariates, we uncover a consistent positivity paradox: harmful content often contains abundant joy tokens while remaining more negative overall and less neutral than controls. Lexicon results show hate/offensive tweets markedly increase anger/disgust/fear without reducing joy; both real and fake news retain trust/anticipation/surprise typical of news style. Model results indicate fake news is more negative and less neutral than real news, yet exhibits higher joy. STM reveals emotion–topic couplings: negative affect boosts ostensibly neutral or technocratic topics and depresses contentious political/legal ones; ‘fake’ labels increase the prevalence of immigration/violence, and U.S. election topics. Findings caution against detectors that equate harm with purely negative valence and motivate models integrating topic–emotion interactions.
The Affective Profile of Disinformation
Sciandra A.;
2026
Abstract
We analyze how disinformation leverages are affected by combining lexicon- and model-based emotion detection with topic modeling across two English corpora: (i) a balanced “true vs. fake” news collection and (ii) a Twitter/X set labeled hate/offensive/neutral. Using the NRC Emotion Intensity Lexicon, then layering Structural Topic Modeling (STM) with emotion covariates, we uncover a consistent positivity paradox: harmful content often contains abundant joy tokens while remaining more negative overall and less neutral than controls. Lexicon results show hate/offensive tweets markedly increase anger/disgust/fear without reducing joy; both real and fake news retain trust/anticipation/surprise typical of news style. Model results indicate fake news is more negative and less neutral than real news, yet exhibits higher joy. STM reveals emotion–topic couplings: negative affect boosts ostensibly neutral or technocratic topics and depresses contentious political/legal ones; ‘fake’ labels increase the prevalence of immigration/violence, and U.S. election topics. Findings caution against detectors that equate harm with purely negative valence and motivate models integrating topic–emotion interactions.
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
