Emotions in Argumentation: an Empirical Evaluation

M. Sahbi Benlamine, Maher Chaouachi, Serena Villata, Elena Cabrio, Claude Frasson, Fabien GandonIJCAI 2015

Abstract

Traditional models of argumentation often assume participants are purely rational agents. This paper challenges that assumption by providing empirical evidence that human debates are deeply intertwined with emotional states. By synchronizing EEG brainwave data and facial expression analysis with a bipolar argumentation framework, the researchers demonstrate that negative argument relations (attacks) are strongly correlated with anger and disgust, while sadness leads to conversational withdrawal. This work provides a new multimodal dataset and insights essential for developing emotionally-intelligent AI moderators for online platforms.

Introduction

Most formal logic models treat argumentation as a sequence of rational moves. However, human interaction—from political debates to online community forums—is rarely devoid of emotion. The authors address the lack of empirical data linking the structural components of an argument (support vs. attack) with the physiological and psychological state of the arguer. Understanding this link is critical for the design of hybrid human-AI societies and e-democracy platforms.

Methodology

The study involved 20 students participating in 12 debates on controversial topics such as nuclear energy and animal testing. The researchers employed a multimodal data collection strategy:

  • Bipolar Argumentation Framework: Arguments were categorized as either supporting or attacking previous statements.
  • Facial Expression Analysis: FaceReader 6.0 tracked six basic emotions (Happy, Sad, Angry, Surprised, Scared, Disgusted) and valence.
  • EEG Brain Sensing: Emotiv EPOC headsets measured a "Mental Engagement Index" based on theta, alpha, and beta frequencies.
  • Data Synchronization: A novel post-processing phase synchronized message timestamps with the sender’s physiological and facial data.

The Dataset

The authors contributed a significant new dataset to the research community, consisting of 598 unique arguments and 263 annotated pairs across diverse topics.

Table 1: Summary of the debate topics and argument statistics in the collected dataset.

The dataset includes rich metadata for each argument, including participant demographics, personality types, and synchronized emotional/engagement metrics. The reliability of the argument annotations was high, with a Cohen’s kappa of 0.82.

Key Insights and Results

The empirical analysis revealed several significant patterns in how emotions drive argumentative behavior:

The Negativity Effect

Negative emotions, particularly anger and disgust, were far more prevalent than positive ones. Happiness was observed in less than 1% of the total debate time, suggesting that conflict is primarily driven by and results in negative affect.

Engagement and Aggression

High mental engagement did not correlate with positive "flow" but rather with defensive or offensive frustration. Specifically, engagement showed a positive correlation with anger ($r=0.306$). Furthermore, in specific sessions, there was a near-linear correlation between disgust and the frequency of "attack" arguments.

Table 2: Correlation matrix showing the relationship between emotional states and argument types (Number of Arguments, Attacks, and Supports).

Sadness and Withdrawal

Sadness acted as a "retreat" signal. The researchers found a strong negative correlation between sadness and the volume of arguments produced ($r$ between $-0.25$ and $-0.70$). When participants felt sad or unpleasant, they tended to stop engaging in the debate entirely, exhibiting retreating behavior.

Conclusion

This research proves that argumentation is not merely a linguistic or logical exercise but a physiological one. The findings suggest that the "rational" structure of a debate is heavily influenced by the emotional signatures of its participants. Future work will explore "emotion persistence"—how long an emotional state from one exchange influences subsequent arguments—and the potential for automated sentiment analysis to predict argument polarity.

Summary updated

BibTeX
@inproceedings{benlamine2015emotions,
  author = {M. Sahbi Benlamine and Maher Chaouachi and Serena Villata and Elena Cabrio and Claude Frasson and Fabien Gandon},
  title = {{Emotions in Argumentation: an Empirical Evaluation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure