Although celebrity avatar virtual streamers are rapidly adopted in live streaming commerce, there is limited evidence on how users form emotional evaluations in natural interactive settings, particularly regarding the joint effects of usage motivations and risk concerns. This study aims to identify the key motivational and risk factors that shape users' emotional evaluations and to compare their relative importance.
Drawing on uses and gratifications theory and perceived risk theory, we analyze 95,650 user comments from Weibo, Douyin, Bilibili and Xiaohongshu. BERTopic is used to identify discussion topics, which are then manually coded into six motivation dimensions and four risk dimensions to construct a dual-path analytical framework. SnowNLP is applied to the original Chinese texts to generate a continuous sentiment score E(0–1) as a measure of users' emotional evaluations. Finally, regression analysis with platform and celebrity fixed effects, together with dominance analysis, is used to estimate the effects of each dimension on the sentiment score and compare their relative importance.
Risk dimensions are more concentrated than motivations. Authenticity risk (R1) is the strongest negative predictor and ranks first in relative importance. Among motivations, all dimensions except informational motivation (M1) significantly increase E; entertainment motivation (M2) and celebrity-oriented social motivation (M3) are the most influential. Platform dummies are significant, whereas celebrity fixed effects are not. Dominance analysis confirms that authenticity risk contributes most to the explained variance, followed by M2 and M3.
Focusing on AI celebrity avatar streamers modeled on real public figures, this study shows how motivational and risk-related factors jointly shape users' emotional evaluations within a unified framework, and identifies authenticity risk as the key constraint on positive evaluations. Methodologically, it proposes a text-mining framework based on natural comments, which links topic modeling, theory-driven dimension quantification and sentiment scoring to test motivation and risk mechanisms in celebrity avatar live streaming.
