A0448
Title: Variational inference for the keyword assisted topic models
Authors: Kiyoshi Inoue - Doshisha University Graduate School (Japan) [presenting]
Shintaro Yuki - Doshisha University (Japan)
Yoshikazu Terada - Osaka University; RIKEN (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: Latent Dirichlet allocation (LDA) is often applied to discover latent topics in documents. Sometimes, some keywords for some topics are known in advance. As an extension of LDA, the keyword assisted topic model (KeyATM) has been proposed to improve the quality of the estimated topic and to provide more interpretable results by incorporating keywords for each topic as prior information. The KeyATM uses the collapsed Gibbs sampling for estimation. However, it is known that the Gibbs sampler is slow, and thus, the current algorithm of the KeyATM is not scalable for large-scale data. Therefore, a variational inference algorithm is developed for the KeyATM, which is faster than the Gibbs sampling algorithm. The proposed algorithm of the KeyATM can be applied to large-scale data. The advantages of the proposed algorithm are validated through numerical experiments and real data application.
