Learning to rank products based on online product reviews using a hierarchical deep neural network

Abstract

Product ranking based on online product reviews is a task of inferring relative user preferences between different products as a variant of entity-level sentiment analysis. Despite the complex relationship between the overall user's preference and individual diverse opinions, existing approaches generally employ empirical assumptions about sentiment features of the products of interest. In this paper, we propose a novel unified approach for learning to rank products based on online product reviews. Unlike existing approaches, it uses deep-learning techniques to extract the high-level latent review representation that contains the most semantic information in the learning process. For this approach, we extend the recently proposed hierarchical attention network to operate in the ranking domain. This network hierarchically learns optimal feature representations of the products and their reviews through the use of two-level attention-based encoders. To construct a more advanced ranking model, several features were added to give sufficient information about the relative user preferences, and two representative ranking loss functions, RankNet and ListNet, were applied. Furthermore, we demonstrate that this network outperforms the existing methods in sales rank prediction based on online product reviews.