Military News Event Extraction Based on MRC

Due to the problems of diverse event types, complex argument roles and frequent emergence of new event types in the military news domain. For the classification-based approach in the event extraction task in the military news domain, there is the problem of not being able to explicitly model the semantics of event labels, which makes it difficult to adequately capture the deep correlation between event types and argument roles, and thus leads to the problem of insufficient generalization ability of the model as well as the problem of error propagation. In this paper, we propose an event extraction framework based on MRC (Machine Reading Comprehension). The framework divides the event extraction process is divided into two components: event detection and recognizing event argument extraction, and designs specialized problem templates for different subtasks. Specifically, the first TriEE model extracts event trigger words and their event types from the input sentences, while the second ArgEE model identifies candidate event theses in the sentences by means of multiple rounds of question and answer. Our experimental comparative analysis with the other four baseline models on the CMNEE dataset shows that the MRC-based event extraction framework achieves better results than the other models in both event detection and thesis element recognition tasks. There is a certain improvement in the recall rate and F1 value. Furthermore the framework effectively leverages the dependency between event arguments and event types, demonstrating enhanced generalization capabilities.