Multimodal Semantics-Based Supervised Latent Dirichlet Allocation for Event Classification

Abstract

Social event classification has always been a research topic of great interest in the field of social event analysis. In existing social event classification methods, although some researchers consciously use external semantics to improve model performance, they ignore the more easily available internal semantics. In this article, we propose a multimodal supervised topic model based on semantic weighting (Sem-MMSTM), which uses two kinds of internal semantics, namely part of speech semantics and category semantics. Our Sem-MMSTM model is capable of mining and making use of the semantics of the text itself and the category semantics of multimodal supervised corpus. The experimental results show that, compared with the state-of-the-art model, our proposed Sem-MMSTM yields significant performance improvement both on the metrics of classification accuracy (ACC) and interpretability of topics (PMI) due to the introduction of effective semantic information.