Cross-media hot topic auto-tracking model based on semantics and temporal context

Abstract

As the Internet multimedia information grows explosively, seeking an automatic technology to realize the effective organization and management of crossmedia emergency information is significantly necessary. A novel cross-media hot topic auto-tracking model based on semantics and temporal context is proposed in this paper. According to the semantic correlations of cross-media information, we learn the image visual semantics by the text semantics based on the Latent Dirichlet Allocation probability model, and establish the unified cross-media information description on the same semantic level. Also a semantics-based two-step feature dimension-reduction scheme is proposed to establish the efficient semantic feature space. The self-adaptive learning of topic model is realized to track the dynamic changes in the topic. Experimental results demonstrate that the proposed method outperforms the existing methods, which further improves the effect of hot topic auto-tracking.