It's Not Just What You Say, But When You Say It

Abstract

Most research on online product reviews focuses on content, such as topics, while largely ignoring the order in which they appear. This paper presents an integrated model with three components to examine the positional effects of valence-driven topics on product ratings. First, we introduce the Valence-Based Sentence-Constrained Latent Dirichlet Allocation model, which incorporates sentence-level valences to improve topic inference. Second, we propose the Beta-transformation model, which captures the importance of topic positions for predicting online reviews that accommodate varying review lengths. Finally, we introduce lift parameters to measure topic-specific position effects, providing deeper insight into how different topics exert distinct influences when placed in their most important positions. We illustrate our model with online movie and airline reviews and find that the last sentence carries the greatest weight in conveying the rationale behind the ratings.

Keywords: Topic Modeling, Valence, Positional Effect, Narrative Structure, Natural Language Processing (NLP)