Spatial-Temporal Aware Disaster Topic Recognition in Social Media Text Using SageMaker

Abstract

With the popularization and widespread use of social media, a large amount of information is shared and disseminated by users on the platforms, including disaster-related text messages. The disaster information contained in these social media text messages is of great value for emergency response, disaster research, and the understanding of public opinion. In the task of text classification of social media disaster topics, previous approaches generally only semantically analyze the text itself to determine its relevance to the disaster, and this paper considers a new perspective that combines the spatio-temporal attributes of social media texts and textual information to complete the text classification task. Based on the above idea, this paper constructs a heterogeneous graph on the dataset using the temporal and spatial attributes of blog posts, initializes the blog nodes using BERT features, and learns the blog features through relational graph convolution operation. Experiments on the SageMaker platform use BERT as a baseline model, while blog post features are initialized using this model. The experimental results show that the spatio-temporal model achieves better performance compared to the baseline model.