Review Rating Prediction aims to predict a user's numeric rating (usually from 1 to 5 stars) in a review from the text of the review. Rating Prediction task can be effectively used to:
Review Rating Prediction has traditionally been a model of a multi-class classification or regression task.
This technique can achieve 70-78% in rating prediction accuracy.
Our prediction algorithm is based on vector space model. In this model, user reviews are represented as vectors in multidimensional space, which components are opinion words, intensifiers and opinion objects with associated sentiment phrases. We apply the improved term frequency/inverse document frequency approach, which uses semantic features of opinion words to enhance the accuracy of rating prediction. The similarity between user review and rating category is determined as cosine angle between appropriate vectors.
To evaluate the effectiveness of our prediction algorithm we’ve created a dataset of hotel and restaurant reviews from TripAdvisor.com (25000 training and 25000 testing reviews for each rating category). The experiments have shown that our rating prediction technique achieves 75-83% in accuracy and can be successfully applied in review filtering systems for suspicious opinions detected.
For the implementation of this use case the following components from the Intellexer SDK were used: