Document-Level Multi-event Extraction via Event Ontology Guiding

Abstract

Document-level Event Extraction (DEE) aims to extract event information from a whole document, in which extracting multiple events is a fundamental challenge. Previous works struggle to handle the Document-level Multi-Event Extraction (DMEE) due to facing two main issues: (a) the argument in one event can correspond to diverse roles in different events; (b) arguments from multiple events appear in the document in an unorganized way. Event ontology is a schema for describing events that contains types, corresponding roles, and their structural relations, which can provide hints to solve the above issues. In this paper, we propose a document-level Event Ontology Guiding multi-event extraction model (EOG), which utilizes the structural and semantic information of event ontology as role-orientated guidance to distinguish multiple events properties, thus can improve the performance of document-level multi-event extraction. Specifically, EOG constructs Event Ontology Embedding layer to capture the structural and semantic information of event ontology. A transformer-based Guiding Interact Module is then designed to model the structural information cross-events and cross-roles under the guidance of event ontology. Experimental results on the DMEE dataset demonstrate that the proposed EOG can achieve better performance on extracting multiple events from the document over baseline models.