Date of Award

2-24-2016

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science

First Advisor

Ningning Wu

Abstract

Online reviews can easily be accessed from e-commerce websites, dedicated web portals, blogs, discussion groups, forums, and other social media. It provides affluent information about products or services for consumers, merchants, and business analysts. It brings convenience, and also confusion. People feel overwhelmed by the information they receive from online reviews. "Is this product real good, or just so so? What are the opinions of the other consumers who had experience on the item? Why the reviews from different sites read different?" These questions are frequently asked by the review readers. This dissertation answers the above questions by first investigating the consistency issue of review ratings across multiple review sources, and then further analyzing the item features being mentioned in the reviews, and finally designing a scalable multiple-source review integration and summarization system. A probabilistic topic model SentiLDA is proposed in the system to discover the item features and sentiments simultaneously from the reviews. This model exploits the review ratings as priors for sentiment distributions in the reviews, and outperforms the other state-of-the-art topic models. The system is implemented following the Map-Reduce paradigm in the Hadoop environment, which is scalable for big data applications.

Download

Share

COinS