DuEE-Fin: A Large-Scale Dataset for Document-Level Event Extraction

Cuiyun Han, Jinchuan Zhang, Xinyu Li, Guojin Xu, Weihua Peng, Zengfeng Zeng

Abstract: To tackle the data scarcity problem of document-level event extraction, we come up with a large-scale benchmark, DuEE-Fin, which consists of 15,000+ events categorized into 13 event types, and 81,000+ event arguments mapped in 92 argument roles. We constructed DuEE-Fin from real-world Chinese financial news, which allows one document to contain several events, multiple arguments to share the same argument role and one argument to play different roles in different events. Therefore, it presents some considerable challenges in document-level event extraction task such as multi-event recognition and multi-value argument identification, that are referred to as key issues for document-level event extraction task. Along with DuEE-Fin, we also hosted an open competition, which has attracted 1,690 teams and achieved exciting results. We performed experiments on DuEE-Fin with most popular document-level event extraction systems. However, results showed that even some SOTA models performed poorly with our data. Facing these challenges, we found it necessary to propose more effective methods.