Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Anger to Joy: How Nationality Personas Shape Emotion Attribution in Large Language Models

Mahammed Kamruzzaman, Abdullah Al Monsur, Gene Louis Kim, Anshuman Chhabra

TL;DR

This work analyzes whether nationality-based personas in large language models induce emotion attributions that reflect cultural norms or reinforce stereotypes. Using the ISEAR dataset, Hofstede cultural dimensions, four LLMs, and 110 nationality personas across UN regions, the authors conduct a large-scale audit of emotion attribution and model alignment, including region, model, gender intersectionality, and abstention patterns. Key findings show region-specific biases and partial alignment with human responses—especially for positive emotions like joy and sadness—and reveal that adding gender information can amplify stereotypes. The study highlights the importance of finer-grained, culturally aware alignment strategies to reduce representational harm while preserving useful personalization capabilities.

Abstract

Emotions are a fundamental facet of human experience, varying across individuals, cultural contexts, and nationalities. Given the recent success of Large Language Models (LLMs) as role-playing agents, we examine whether LLMs exhibit emotional stereotypes when assigned nationality-specific personas. Specifically, we investigate how different countries are represented in pre-trained LLMs through emotion attributions and whether these attributions align with cultural norms. To provide a deeper interpretive lens, we incorporate four key cultural dimensions, namely Power Distance, Uncertainty Avoidance, Long-Term Orientation, and Individualism, derived from Hofstedes cross-cultural framework. Our analysis reveals significant nationality-based differences, with emotions such as shame, fear, and joy being disproportionately assigned across regions. Furthermore, we observe notable misalignment between LLM-generated and human emotional responses, particularly for negative emotions, highlighting the presence of reductive and potentially biased stereotypes in LLM outputs.

From Anger to Joy: How Nationality Personas Shape Emotion Attribution in Large Language Models

TL;DR

This work analyzes whether nationality-based personas in large language models induce emotion attributions that reflect cultural norms or reinforce stereotypes. Using the ISEAR dataset, Hofstede cultural dimensions, four LLMs, and 110 nationality personas across UN regions, the authors conduct a large-scale audit of emotion attribution and model alignment, including region, model, gender intersectionality, and abstention patterns. Key findings show region-specific biases and partial alignment with human responses—especially for positive emotions like joy and sadness—and reveal that adding gender information can amplify stereotypes. The study highlights the importance of finer-grained, culturally aware alignment strategies to reduce representational harm while preserving useful personalization capabilities.

Abstract

Emotions are a fundamental facet of human experience, varying across individuals, cultural contexts, and nationalities. Given the recent success of Large Language Models (LLMs) as role-playing agents, we examine whether LLMs exhibit emotional stereotypes when assigned nationality-specific personas. Specifically, we investigate how different countries are represented in pre-trained LLMs through emotion attributions and whether these attributions align with cultural norms. To provide a deeper interpretive lens, we incorporate four key cultural dimensions, namely Power Distance, Uncertainty Avoidance, Long-Term Orientation, and Individualism, derived from Hofstedes cross-cultural framework. Our analysis reveals significant nationality-based differences, with emotions such as shame, fear, and joy being disproportionately assigned across regions. Furthermore, we observe notable misalignment between LLM-generated and human emotional responses, particularly for negative emotions, highlighting the presence of reductive and potentially biased stereotypes in LLM outputs.

Paper Structure

This paper contains 26 sections, 4 figures, 24 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Experimental Setup
  4. Results and Discussions
  5. The emotional attribution patterns of LLMs reveal distinct regional disparities
  6. Different LLMs have different emotional attributional patterns
  7. Analyzing Cultural and Emotional Dimensions through Hierarchy, Uncertainty, and Time Orientation (PDI, UAI, LTO)
  8. LLM refusal rates vary across regions
  9. Analyzing Cultural and Emotional Dimensions through Individualism
  10. Assessing LLM alignment with actual nationality-specific human responses
  11. Intersectionality analysis using additional gender-related information
  12. Conclusion
  13. Limitations
  14. Ethics Statement
  15. Models
  16. ...and 11 more sections

Figures (4)

  • Figure 1: An example showcasing our approach. We examine Gemma2 LLM's responses for the same emotional scenario: After a big fight with my parents. When adopting a persona from the USA, the model responds with Anger, whereas, for a Zambian persona, it expresses Sadness.
  • Figure 2: The top 10 most frequent emotions by region for each LLM, averaged across all three prompting templates.
  • Figure 3: The top 10 most frequent emotions for 10 countries across each LLM, averaged over all three prompting templates.
  • Figure 4: Results for intersectional experiments for five regions averaged across all four models and prompting templates. The results reveal notable region and gender disparities in emotion attribution. Specifically, LLMs tend to associate anger and shame more frequently with male-country, and sadness and fear to female-country intersections.