Understanding Public Perceptions of AI Conversational Agents: A Cross-Cultural Analysis

Title of the Paper

Understanding Public Perceptions of AI Conversational Agents: A Cross-Cultural Analysis

Paper Information

• Field of Study: Human-Computer Interaction and Cross-Cultural Comparison
• Keywords: Conversational agents, Cultural differences, Public perceptions, Topic modeling, Word embedding, AI acceptance, Emotional valence, Warmth perception, Competence perception, Social cognition

Research Background and Issues

• Identified Issues and Challenges: Despite the widespread application of conversational agents (CAs), there are still limitations in understanding their social cognition, particularly in cross-cultural contexts. These technologies, due to differences in design features and users' cultural backgrounds, may elicit vastly different emotional and cognitive responses.
• Significance: Understanding public perceptions of CAs not only helps optimize design and enhance user experience but also provides guidance for ethical frameworks and transparent practices in future AI development.
• Motivation and Related Work: Current studies mainly focus on CAs in specific domains (e.g., customer service, healthcare) or user perceptions of specific products (e.g., Alexa and Siri). There is a lack of comprehensive research across multiple platforms and countries, especially detailed analyses of warmth perception, competence evaluation, and emotional valence.

Proposed Solution

• Methods or Solutions:
  • Utilize BERTopic and word embedding techniques to analyze nearly one million social media posts from Twitter and Weibo, exploring public discussion topics and emotional cognition regarding CAs.
  • Propose an analytical framework that integrates cultural backgrounds and technical features influencing user perceptions.
  • Investigate three key dimensions (warmth perception, competence evaluation, and emotional valence) and conduct cross-cultural and cross-technical feature comparisons.
• Innovations:
  • Combine large-scale social data analysis with cross-cultural comparison, offering a new perspective on CA acceptance and design directions.
  • Conduct a detailed analysis of the interaction between technical features of CAs (e.g., human-like appearance, physical embodiment, conversational modes) and cultural factors.
• Implementation Steps:
  1. Collect public posts from Twitter and Weibo between 2017 and 2023.
  2. Use BERTopic modeling to cluster discussion topics.
  3. Apply word embedding techniques to analyze emotional dimensions (e.g., "warmth vs coldness," "competence vs incompetence").
  4. Categorize CAs into virtual companions, intelligent assistants, and other technological feature groups, studying their performance differences across cultures.

Research Findings

• Specific Findings:
  • Highlighted thematic differences in CA-related discussions between China and the U.S.: Chinese users focus more on socialization and emotional expression, while U.S. users emphasize task-driven interactions.
  • U.S. participants perceive higher warmth and competence in CAs but exhibit a more neutral emotional valence overall. In contrast, Chinese participants hold more positive emotional attitudes toward CAs but provide lower competence evaluations.
  • Technical features (e.g., virtual companions vs intelligent assistants) significantly influence users' perceptions of warmth and competence. Chinese users tend to attribute higher competence to companion-like CAs with high warmth.
  • Social media discussions on industry and socio-cultural topics may be structurally influenced by national policies and media environments.
• Advantages Over Existing Solutions:
  • Introduced multi-dimensional semantic and emotional analysis models capable of capturing the broad diversity of social media data.
  • Strengthened understanding of cultural adaptability in technology design through quantitative cross-cultural analysis.
• Experimental or Evaluation Results:
  • Themes extracted by BERTopic (Top 5 themes in China: individual interaction and experience, technical features, industrial applications, socio-cultural events, law and ethics; additional political discussions in the U.S.).
  • Word embedding results show a strong correlation between warmth and emotional valence, while competence perception is relatively weaker.
• Limitations and Future Directions:
  • Limitations:
    1. Data sources are limited to single social media platforms (Weibo and Twitter), which may reduce generalizability.
    2. Social media data may be influenced by censorship or user self-censorship, failing to fully reflect public opinion.
    3. Insufficient sample sizes for certain technical feature categories during classification may lead to oversimplification.
  • Future Directions:
    1. Expand to multi-platform data analysis, supplement with surveys and interviews to establish a more balanced picture of public perceptions.
    2. Deepen research on technical forms (e.g., the potential impact of physical embodiment on interaction experiences) and focus on culturally localized design principles.

By focusing on cultural contexts and technical details, this study provides important theoretical insights into understanding human interaction with conversational AI and optimizing its design. It also calls on developers to balance personalization and globalization, considering warmth, competence, and emotional dimensions in design.