Title: Document-level event extraction via gated cross-sentence attention network with global memory

Authors: Shunyu Yao; Dan Liu; Jie Hu

Addresses: Big Data and Artificial Intelligence Institute, China Telecom Research Institute, Beijing, China ' Big Data and Artificial Intelligence Institute, China Telecom Research Institute, Beijing, China ' Big Data and Artificial Intelligence Institute, China Telecom Research Institute, Beijing, China

Abstract: Event extraction aims to detect events from text, and identify arguments of the event and its role. When event-related information is distributed in a document, event arguments are always scattered across different sentences, and multiple events may co-exist in the same document. In order to meet these challenges, we propose a novel end-to-end model gated cross-sentence attention network with global memory called GCANGM. To address the multi-events challenge, we introduce a global memory unit to store the current path state and event information extracted in the past. In order to handle the argument-scattering challenge, we construct a gated cross-sentence attention layer so that the entity embedding can obtain the cross-sentence contextual information in the document. To prove the effectiveness of GCANGM, we conduct experiments on the widely used large-scale document-level event extraction dataset. The experimental results show that our method is effective in the challenging document-level event extraction task.

Keywords: event extraction; neural network; attention mechanism.

DOI: 10.1504/IJAHUC.2026.153342

International Journal of Ad Hoc and Ubiquitous Computing, 2026 Vol.51 No.4, pp.235 - 243

Received: 27 Feb 2025
Accepted: 25 Apr 2025
Published online: 01 May 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article