[Paper Review] Deep Feelings: A Massive Cross-Lingual Study on the Relation between Emotions and Virality

Marco Guerini, Jacopo Staiano|arXiv (Cornell University)|Mar 16, 2015

Misinformation and Its Impacts32 references16 citations

TL;DR

This study investigates the relationship between emotions evoked by news articles and their virality across English and Italian social media platforms using a massive, crowdsourced dataset of 1.5 million emotional annotations. It finds that while cultural differences affect specific emotions, the Valence-Arousal-Dominance (VAD) circumplex model provides a consistent, cross-linguistic framework explaining virality, with high arousal linked to narrowcasting and high dominance to broadcasting effects.

ABSTRACT

This article provides a comprehensive investigation on the relations between virality of news articles and the emotions they are found to evoke. Virality, in our view, is a phenomenon with many facets, i.e. under this generic term several different effects of persuasive communication are comprised. By exploiting a high-coverage and bilingual corpus of documents containing metrics of their spread on social networks as well as a massive affective annotation provided by readers, we present a thorough analysis of the interplay between evoked emotions and viral facets. We highlight and discuss our findings in light of a cross-lingual approach: while we discover differences in evoked emotions and corresponding viral effects, we provide preliminary evidence of a generalized explanatory model rooted in the deep structure of emotions: the Valence-Arousal-Dominance (VAD) circumplex. We find that viral facets appear to be consistently affected by particular VAD configurations, and these configurations indicate a clear connection with distinct phenomena underlying persuasive communication.

Motivation & Objective

• To investigate how emotions influence different facets of virality in news articles across languages.
• To compare the impact of specific emotions and their deep affective components (Valence, Arousal, Dominance) on virality metrics.
• To test whether a generalized, language-invariant model based on the VAD circumplex can explain virality across cultures.
• To evaluate the explanatory power of emotion-based models versus VAD-based models in predicting virality.
• To explore implications for content marketing and native advertising by identifying emotion-virality patterns.

Proposed method

• Collected a bilingual corpus of 65,000 news articles from Rappler (English) and Corriere della Sera (Italian), each annotated with reader-evoked emotions.
• Used crowdsourced affective feedback (mood tags) to derive emotion and VAD (Valence-Arousal-Dominance) scores for each article.
• Measured virality through multiple indices: Google+ shares, tweets, and email shares (G+, Tweets, Emails).
• Binarized virality indices using a threshold based on mean + standard deviation to identify highly viral articles.
• Applied logistic regression models to predict virality using both emotion categories and VAD dimensions as predictors.
• Mapped emotion labels to VAD values using an established resource to enable cross-linguistic comparison and model generalization.

Experimental results

Research questions

• RQ1 How do specific emotions relate to different viral facets (e.g., sharing on Google+, Twitter, email) in English and Italian news content?
• RQ2 To what extent do cultural differences in emotional responses affect the relationship between emotions and virality?
• RQ3 Can the Valence-Arousal-Dominance (VAD) circumplex model serve as a generalized, cross-linguistic framework for explaining virality?
• RQ4 How does the explanatory power of emotion-based models compare to VAD-based models in predicting virality?
• RQ5 What is the relative impact of emotions on narrowcasting (e.g., email) versus broadcasting (e.g., social media) virality?

Key findings

• The VAD circumplex model provides a consistent, cross-linguistic explanation for virality, with high arousal linked to narrowcasting and high dominance to broadcasting phenomena.
• Despite cultural differences in emotional responses—particularly the discordant influence of sadness in English versus Italian datasets—the VAD model maintains strong explanatory power across both languages.
• VAD-based models achieved McFadden’s R² scores of up to 0.0840 on Corriere.it and 0.0569 on Rappler.com, outperforming emotion-based models and the best model from Berger et al. (2012), which had an R² of 0.0700.
• The study found that emotions have a stronger impact on narrowcasting (e.g., email shares) than on broadcasting (e.g., social media shares), a pattern consistent across both languages.
• Even after mapping emotions to VAD dimensions, the models retained strong explanatory power, indicating that VAD captures essential affective structure underlying virality.
• The results suggest that VAD configurations, particularly the alternation between dominance and arousal, are systematically linked to distinct persuasive communication phenomena underlying virality.