Event Extraction for Military Target Motion in Open-source Military News

Open-source military news is the data source of open-source intelligence analysis, which can assist intelligence analysis by event extracting. Military events include the military target motion events which involve changes in multiple places. It is very important to infer the movement trajectory of a military target from the location-changing process of multiple place arguments. However, the traditional event extraction cannot reflect the location-changing process of multiple place arguments during the target moving. To solve this problem, this paper designs multiple place roles to learn location semantics relationship among place arguments for reflecting the location-changing process of multiple place arguments. We adopt the method of sequence labeling to convert military target moving event extraction into a multi-classification problem. This paper uses the pre-trained model Roberta for feature learning and a linear layer for a trigger word (or argument) classification. Since there is no standard military event dataset, this paper constructs an event dataset of military target moving. On this dataset, the F1 of event extraction reaches 72.7%, which is a certain improvement compared with each benchmark model.