Title:A Novel End-To-End Event Geolocation Method Leveraging Hyperbolic Space and Toponym Hierarchies

Abstract:Timely detection and geolocation of events based on social data can provide critical information for applications such as crisis response and resource allocation. However, most existing methods are greatly affected by event detection errors, leading to insufficient geolocation accuracy. To this end, this paper proposes a novel end-to-end event geolocation method (GTOP) leveraging Hyperbolic space and toponym hierarchies. Specifically, the proposed method contains one event detection module and one geolocation module. The event detection module constructs a heterogeneous information networks based on social data, and then constructs a homogeneous message graph and combines it with the text and time feature of the message to learning initial features of nodes. Node features are updated in Hyperbolic space and then fed into a classifier for event detection. To reduce the geolocation error, this paper proposes a noise toponym filtering algorithm (HIST) based on the hierarchical structure of toponyms. HIST analyzes the hierarchical structure of toponyms mentioned in the event cluster, taking the highly frequent city-level locations as the coarse-grained locations for events. By comparing the hierarchical structure of the toponyms within the cluster against those of the coarse-grained locations of events, HIST filters out noisy toponyms. To further improve the geolocation accuracy, we propose a fine-grained pseudo toponyms generation algorithm (FIT) based on the output of HIST, and combine generated pseudo toponyms with filtered toponyms to locate events based on the geographic center points of the combined toponyms. Extensive experiments are conducted on the Chinese dataset constructed in this paper and another public English dataset. The experimental results show that the proposed method is superior to the state-of-the-art baselines.