Amazon Review Analysis Using a Neural Topic Model with External Information

Topic models have been proposed and widely used to find topics in documents. Most topic models use only the document term frequency matrix, and variational inference is used to estimate parameters. Therefore, it has been known that these models cannot use external information of documents and also have difficulties in model extension. To solve these problems, Scholar (sparse contextual hidden and observed language autoencoder) model, a neural topic model that uses external information, was proposed. Scholar model is a neural topic model combining a linear model, thus can reflect external information and also can be easily extended. In this paper, we introduce Scholar model and apply it to gift card review data traded on Amazon. The analysis data consists of 182,625 reviews with 21,231 vocabularies, and includes product and purchase information. For the analysis, various combinations of hyper parameters were considered, and the selected model was compared to previous topic models. The analysis results showed that Scholar model performed better than other models, and the topics were interpreted through top words for each topic. In addition, top words according to each gift card were also investigated.