What Drives Consumer Choices? Mining Aspects and Opinions on Large Scale Review Data Using Distributed Representation of Words

With the increasing popularity of online review sites, developing methods to mine and analyze information contained in the vast amounts of noisy user-generated reviews becomes a necessity. In this work, we develop a method to uncover the various aspects of a product or service reviewed by a user, and the opinions associated with them, in an automated fashion. We use the neural network model Word2Vec to build a vector space representation of a large corpus of user-generated, online restaurant reviews, and harness these distributed representations for aspect-based sentiment analysis. User generated text data is intrinsically noisy, with misspellings, informal language, and digressions. Because of the many variations in spelling and expression, the data is also very sparse. Despite these inherent challenges we are able to represent the reviews by key drivers of consumer sentiment, allowing for highly accurate sentiment prediction using a method that is both scalable and human interpretable.