Discussion themes and emotional expressions on Chinese social media regarding AI virtual companions: a topic modeling...

As AI virtual companions become part of everyday life, they have attracted growing public and scholarly attention. However, evidence remains limited on how users understand, express, and evaluate these systems in real online contexts. Existing studies have relied largely on surveys or experiments, with relatively few examining discussion themes and emotional structures around AI virtual companions in authentic online settings. Against this background, this study examines discussion patterns surrounding AI virtual companions in Chinese social media and the structure of researcher-constructed higher-order experiential categories reflected in users' comments.

A total of 20,807 comments from 168 relevant Xiaohongshu posts were collected as raw data. After text cleaning, BERTopic was used for initial topic modeling, followed by manual review to refine and consolidate the topics. Sentiment analysis was conducted by fine-tuning a Chinese RoBERTa model on a manually labeled corpus.

The study identified 29 topics, further integrated into seven higher-order experiential categories. The most frequently discussed topics included Perceptions of Monetization and Platform Mechanisms, Relationship Memory and Continuity, Virtual Romance Experience, and Controversies Over Respect in Interaction, indicating that user discussions extended beyond simple functional evaluation. Significant differences were found across topics and higher-order experiential categories in sentiment distribution and sentiment scores. Emotional Gains and Losses, Positive Emotions, and Virtual Romance Experience showed more positive tendencies, whereas Usage Restrictions, Originality Disputes, Privacy Risks, and Discussions of Boundary Violations and Sensitive Content showed more negative tendencies. At the higher-order experiential category level, Virtual Intimacy and Boundary Negotiation and Emotional Engagement and Regulation showed more positive emotional tendencies overall, whereas Risk Perception and Ethical Boundaries and Platform Functionality and Monetization showed more negative tendencies.

Online discussions of AI virtual companions are no longer limited to technical functions or use evaluations, but increasingly involve relationship construction, emotional experience, and boundary negotiation. Users' emotional expressions are shaped by specific discussion contexts and relational meanings. By integrating topic modeling with sentiment analysis, this study provides structured empirical evidence for understanding user experiences with AI virtual companions in Chinese social media and offers implications for future research and design on relational AI.