Bayesian Non-parametric Inference of Multimodal Topic Hierarchies

Research on multimodal data analysis such as annotated image analysis is becoming more important than ever due to the increase in the amount of data. One of the approaches to this problem is multimodal topic models as an extension of Latent Dirichlet allocation (LDA). Symmetric correspondence topic models (SymCorrLDA) are state-of-the-art multimodal topic models that can appropriately model multimodal data considering inter-modal dependencies. Incidentally, hierarchically structured categories can help users find relevant data from a large amount of data collection. Hierarchical topic models such as Hierarchical latent Dirichlet allocation (hLDA) can discover a tree-structured hierarchy of latent topics from a given unimodal data collection; however, no hierarchical topic models can appropriately handle multimodal data considering inter-modal mutual dependencies. In this paper, we propose hSymCorrLDA to discover latent topic hierarchies from multimodal data by combining the ideas of the two previously mentioned models: multimodal topic models and hierarchical topic models. We demonstrate the effectiveness of our model compared with several baseline models through experiments with three datasets of annotated images.