Automatic topic discovery of online hospital reviews using an improved LDA with Variational Gibbs Sampling
Richard de Groof, Haiping Xu · 2017
E-commerce websites such as Yelp.com, allow users to write online reviews of products and services, so new customers can have quick access to user experiences, covering everything from auto-repair to hospitals. However, a typical user may find it difficult to identify a topic of interest due to the overwhelming amount of review information. To deal with this issue, the Latent Dirichlet Allocation (LDA) model can be used to associate meaningful terms with text-based reviews, permitting keyword retrieval of individual documents. LDA is a powerful unsupervised learning approach, which has been widely used for topic modeling as well as in other related fields. A conventional implementation of LDA is through the Markov-Chain Monte-Carlo methodology, called Collapsed Gibbs Sampling (CGS). However, due to the usage of random numbers in the CGS approach, results from multiple trials on the same data are usually inconsistent. To avoid this tendency, we revise the conventional LDA approach using Variational Gibbs Sampling (VGS). VGS eliminates random numbers, and thus leads to consistent results as well as better performance. Our case study shows that our improved LDA can be used to automatically identify keywords and topics in online hospital reviews. Due to the usage of VGS, the accuracy of topic identification has been consistently improved.