Improving the quality of predictions using textual information in online user reviews

Online reviews are often accessed by users deciding to buy a product, see a movie, or go to a restaurant. However, most reviews are written in a free-text format, usually with very scant structured metadata information and are therefore difficult for computers to understand, analyze, and aggregate. Users then face the daunting task of accessing and reading a large quantity of reviews to discover potentially useful information. We identified topical and sentiment information from free-form text reviews, and use this knowledge to improve user experience in accessing reviews. Specifically, we focus on improving recommendation accuracy in a restaurant review scenario. We propose methods to derive a text-based rating from the body of the reviews. We then group similar users together using soft clustering techniques based on the topics and sentiments that appear in the reviews. Our results show that using textual information results in better review score predictions than those derived from the coarse numerical star ratings given by the users. In addition, we use our techniques to make fine-grained predictions of user sentiments towards the individual topics covered in reviews with good accuracy.

► We implement a new quadratic regression model on the review text to derive text-based ratings.
► We compare the predictive power of star and text ratings using baseline average-based strategies.
► Popular rating-based methods do not improve over baseline predictions for our sparse data.
► We implement a text-based soft clustering of users and design a personalized prediction approach.
► We predict not just a rating, but the user sentiments towards individual restaurant features.