Wide-grained capsule network with sentence-level feature to detect meteorological event in social network

Abstract In recent years, frequent meteorological disasters have caused great concern to people. It is particularly important to timely detect the meteorological events and release early warning information. Most traditional meteorological event detection methods rely on physical sensors, but such practice is usually costly and inflexible. As a new form of lightweight social sensor, social networks make up for the shortcomings of traditional physical sensors. In this paper, we propose a sentence-level feature-based meteorological event detection model to detect 14 types of meteorological events defined by the China Meteorological Administration (CMA) in Sina Weibo. Our joint model consists of two modules: a fine-tuned BERT as the language model and a wide-grained capsule network as the event detection network. The design of our model considers the correlation among meteorological events and achieves the best results on all metrics compared with other baseline models. Moreover, as a practical application, our model has been applied to the meteorological event monitoring platform in the CMA Public Meteorological Service Center to provide online services.