Transformer-based emotion profiling and clustering of cybersickness narratives in Steam reviews of first-person virtual reality games

Abstract

Cybersickness remains a persistent barrier in virtual reality (VR) gaming, yet how players emotionally frame discomfort in user-generated content is poorly understood. Prior large-scale analyses of Steam reviews have characterized emotional experience with a single polarity score. This study applies transformer-based emotion classification and unsupervised clustering to reviews of first-person VR games. A distilroberta-base model fine-tuned for emotion detection was applied to a cybersickness-focused subset of 1898 reviews and to a stratified sample of 50,000 drawn from a corpus of 1,375,125. Emotion profiles were compared across review valence with Mann-Whitney U tests, K-means clustering was applied to the stratified sample, and the cybersickness subset was compared with 9443 controls matched on application and valence, using ordinary least squares with cluster-robust standard errors and equivalence testing. Joy and neutral dominated. Sadness and disgust were higher in negative reviews and joy in positive ones, with large effects. Clustering identified six emotion profiles. The cybersickness subset initially appeared over-represented in the fear-dominant profile, but this difference disappeared under matching, and no dimension separated cases from controls, with equivalence established at 0.1 control standard deviations. Because the subset is defined at topic level, where prior validation places explicit symptom reports at 18.67%, the null applies to cybersickness-associated discourse rather than to verified symptom reports. In this corpus, emotional expression is organized by evaluation rather than topic, which constrains emotion-based screening and shows how platform-wide baselines can misattribute product composition to discourse.