Learning Semantic Representation from Restaurant Reviews: A Study of Yelp Dataset

Abstract

Users' preference such as rating only provides uni-dimension information, but reasons behind users' preference may be related to various aspects of an item, such as the types, certain attributes. By observing user-generated review always provides such rich information, we proposed an item representation based on review data. This approach supports semantic operation, which could potentially enables more recommendation scenarios. Our experiments further demonstrated that this approach gained much better performance than classical item representation methods.