Semantic-based topic model for public opinion analysis in sudden-onset disasters

Abstract

Sudden-onset disasters have put forward more stringent requirements for the government to carry out public opinion analysis work. However, most existing topic models ignore the contextual semantics of disaster texts, and fail to balance the robustness and the training cost. To address these issues, a neural clustering topic model is proposed in this work. The topic probability distribution of the LDA model is integrated with the distribution semantic vector generated by a lite BERT. The fused vectors are reconstructed by a nonlinear manifold learning algorithm, and re-clustered into topics by a mini-batch based k-means++ algorithm. Compared to state-of-the-art models on three sudden-onset disaster datasets, the proposed model shows an increase of 1.79 % in average topic coherence and 33.87 % in topic diversity. Meanwhile, the inference time is reduced by 84.09 % on average. The visual study of the latent process of the proposed model reflects that its ability to compact intra-cluster vector distances and sparse inter-cluster vector distances is the potential reason for its better performance. It can be considered that the application of the proposed model can help the government enhance its ability to manage negative public opinions in sudden-onset disasters.