Unifying Algorithmic and Theoretical Perspectives: Emotions in Online Reviews and Sales

Abstract

Emotion artificial intelligence, the algorithm that recognizes and interprets various human emotions beyond valence (positive and negative polarity), is still in its infancy but has attracted attention from industry and academia. Based on discrete emotion theory and statistical language modeling, this work proposes an algorithm to enable automatic domain-adaptive emotion lexicon construction and multidimensional emotion detection in texts. Using a large-scale dataset of China’s movie market from 2012 to 2018, we constructed and validated a domain-specific emotion lexicon and demonstrated the predictive power of eight discrete emotions (i.e., surprise, joy, anticipation, love, anxiety, sadness, anger, and disgust) in online reviews on box office sales. We found that representing overall emotions through discrete emotions yields higher prediction accuracy than valence or latent emotion variables generated by topic modeling. To understand the source of the predictive power from a theoretical perspective and to test the cross-culture generalizability of our prediction study, we further conducted an experiment in the U.S. movie market based on theories on emotion, judgment, and decision-making. We found that discrete emotions, mediated by perceived processing fluency, significantly affect the perceived review helpfulness, which further influences purchase intention. Our work shows the economic value of emotions in online reviews, generates insight into the mechanism of their effects, and has managerial implications for online review platform design, movie marketing, and cinema operations.