Multi-Grained Ranking Network for Event Extraction

Abstract

Event Extraction is a complicated task in Information Extraction including three subtasks: Entity Mention Detection, Event Detection, and Argument Role Prediction. Previous work for Event Extraction is usually done in the form of pipelined approach. Recently, end-to-end models for Event Extraction are getting more attention, but two problems still exist. First, different subtasks focus on different granularities of context information, but in the existing methods, all three subtasks share a unified representation, thus the information interested by a specific task may be underestimated or even ignored. Second, the interrelationship among events and entities has a strong influence in figuring out the meaning of a specific trigger or entity, but the existing methods are suffering in issues like time-dependency that a decision made later can only affect the rest of the sentence but has no way to update the previous one. Such limitations lead to insufficient use of the correlation between event elements in the sentence. In this paper, we propose a Multi-Grained Ranking Network for Event Extraction, which performs the three subtasks simultaneously in a single model. To handle the first problem, we design a dynamic Span Generator to construct multi-grained span representations from three granularities: sentence, neighbors, and content. Task-specific BERT-Rescale Gate units are applied in the Generator to capture the features of interest for each subtask respectively. As for the second challenge, we implement two attention mechanisms to explore the association of event-event and event-entity. The Events-Aware Attention is assigned to capture the association between events. While the Arguments-Aware Co-Attention is designed to model the relationship among entities in a specific event. The experiments on ACE-2005 prove that the proposed method outperforms the state-of-the-art methods in all subtasks of end-to-end Event Extraction.