Online customer reviews and especially star ratings are a crucial information source for consumers and an important performance indicator for businesses. Here, the language model BERT enables to predict the associated star rating to a review text (i.e., also for non-rated reviews). However, BERT’s mapping for star rating predictions is not interpretable, which evokes a need for explanation methods to gain actionable insights. Thus, we present an approach to explain BERT’s rating predictions with a global decision tree based on a generalized gini impurity. Aspect-based sentiments and customer personality traits based on BERT’s text representation are used as features. We find that these features are suited to explain BERT’s star rating predictions and that severe misalignments can be reduced by up to 30% using the proposed approach compared to alternatives. Thereby, the proposed approach enables to convey valuable analyses of customer assessments also to platforms where reviews are not rated.
Binder, Markus, "But How does it Work? Explaining BERT’s Star Rating Predictions of Online Customer Reviews" (2021). PACIS 2021 Proceedings. 28.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.