The Text-Score Allocation Model : Finding Latent Topics of Online Review Documents and Multi-Item Ratings

This study focuses on online review data in which comments are written in natural languages and evaluations are attached as integers. This study develops a topic model incorporating both natural languages and evaluation scores, expanding latent Dirichlet allocation (LDA). The model consists of two components: LDA and a Dirichlet-binomial clustering model. The latter assumes binomial distributions for the review scores. Since the model assumes conjugate distributions, we can apply a fast and stable estimator based on collapsed Gibbs sampling to estimate the parameters. Further,the model enables us to examine the relationship between vocabulary words and review scores based on the topic allocation results