Multi-Round Extraction and Dynamic Role Selection Framework For Document-Level Event Extraction

Kaizhou Zhang, Pengfei Wang, Lei Zhang

Abstract: Document-level Event Extraction (DEE) aims to extract structured event information from a document, which is an indispensable downstream task for many NLP applications. Argument-scatter and multi-event are its two main challenges. Recent work decomposes this challenging DEE task into multiple steps such as entity recognition, contextual information modeling, and event arguments extraction. Besides, they all extract event arguments in a predefined fixed role order. Though it is effective, its cumbersome steps and fixed extraction order will bring about the problem of error propagation. To address this issue, we propose a Multi-round Extraction and Dynamic Role Selection (MREDRS) Framework for the DEE task. We model the DEE task in an end-to-end manner to avoid these cumbersome steps. Multi-round extraction can deal with multi-event problem. In order to avoid the error propagation problem caused by the fixed role extraction order, we dynamically select the next role according to the current extraction state. We conducted experiments on the commonly used DEE dataset and extensive experimental results demonstrated the effectiveness of our method.