A network-based matching design for text mining of hyper-polarised online reviews

Online reviews provide users with the opportunity to rate various types of items such as movies, music, and video games using a combination of numeric scores and textual comments. The study proposes a novel method that combines network modeling with statistical matching to estimate the unbiased association between words and hyper-polarized items in online reviews. The application of this method to a sample of 40,665 items from the website Metacritic detects 218 hyper-polarized items; these are matched with an equal number of items using 8 covariates of item quality and network centrality. Application of the method reveals an unbiased association between hyper-polarization and semantics indicating reactive social action in online reviews, especially related to controversial political issues in the USA.