Thesis (M.S.)--State University of New York at Buffalo, 2020
Includes bibliographical references
There has been significant growth of internet users generating or looking for a product and experience reviews in the past decade. At the same time, business owners rely on these reviews to publicize offerings and to help influence the user's choices on sites like TripAdvisor and Yelp. One challenge is that users often provide information that may or may not result in a useful aggregate rating, or the platform may not offer the facility to specify aspect ratings of interest to users or businesses. For this reason, predicting numeric aggregate and aspect ratings has become a significant area of research with a focus on natural language processing. The problem is referred to as Review Rating Prediction (RRP), which provides an aggregate numeric rating for an offering, and what we will call is Aspect based Rating Prediction (ABRP), which provides a numeric rating for a specific aspect offering.Current approaches that make use of the text-based reviews to predict numeric reviews focus on creating predictive models that incorporate either the user's contextual information, the product's context information, or a hybrid of the two. While these approaches extract features from the text, the features that they obtain typically do not use the full context of the text, nor are they good at disambiguation, and few make use of modern machine learning techniques.In this thesis, we treat RRP & ABRP as multi-label, multi-class classification problems in machine learning, where the multiple classes are the numeric ratings, and the various labels are the aspects. Our goal is to automatically assign a numerical score for existing aspects that may be offering dependent. We propose a novel review rating prediction method, which improves the prediction accuracy by capturing deep semantics of the review.Our approach extracts feature vectors from user reviews using Bidirectional Encoder Representation from Transformer (BERT) language models and Deep network classifiers, including Artificial Neural Networks (ANNs), Convolutional Neural Network (CNNs), and Long Short Term Memory networks (LSTMs). Our first approach focuses on transfer learning, but we significantly improved results by using the contextual word representations from the post-trained language model together with a fine-tuning method on DNNs.To demonstrate our approach, we built a dataset using TripAdvisor as it allows the users to provide aspect wise ratings along with an aggregate rating for any restaurant. A total of 23,913 reviews has been collected from Buffalo, Syracuse & New York City. In addition to the aggregate rating, each review has three aspect numeric ratings for food, service, and value on a scale of one to five. We also compared our work to state-of-the-art techniques. Empirical experiments demonstrate that our model's performance is comparable with or even better than the methods used for RRP in literature in terms of different performance metrics for the Yelp 2013, 2014, and 2015 datasets
Electronic reproduction. Ann Arbor, Mich. : ProQuest, 2021
