Rating Estimation from Review Texts Using Long Short-Term Memory

Abstract

A lot of reviews of products have been posted on various web sites and services because of the spread of the Internet, and the estimation of ratings from review texts is actively performed. However, there are few such studies on Japanese review texts without limiting the product genre. In this paper, we propose a neural network model that takes as input a general Japanese product review text and estimates rating for it without limiting the product genre. By using Long Short-Term Memory (LSTM), which is one of the regression type neural network models that can handle sequential data, we analyze words in sentences considering their order. The rating estimation model is realized mainly by segmentation of texts, conversion to distributed representations, an LSTM layer, and a fully connected layer. In addition, we conduct evaluation experiments of the created model and consider the results.