Label-Related/Unrelated Topic Switching Model: A Partially Labeled Topic Model Handling Infinite Label-Unrelated Topics

Abstract

We propose a Label-Related/Unrelated Topic Switching Model (LRU-TSM) based on Latent Dirichlet Allocation (LDA) for modeling a labeled corpus. In this model, each word is allocated to a label-related topic or a label-unrelated topic. Label-related topics utilize label information, and label-unrelated topics utilize the framework of Bayesian Nonparametrics, which can estimate the number of topics in posterior distributions. Our model handles label-related and -unrelated topics explicitly, in contrast to the earlier model, and improves the performances of applications to which is applied. Using real-world datasets, we show that our model outperforms the earlier model in terms of perplexity and efficiency for label prediction tasks that involve predicting labels for documents or pictures without labels.