UEE: A Unified Model for Event Extraction

Zhenzhen Duan, Yi Guo, Chunyu Yao, Xue Chen

Abstract: Event extraction (EE) can be divided into four subtasks: trigger identification, trigger classification, argument identification and argument classification. Most the previous studies focused on extracting flat events while neglecting overlapped events. Few models can handle both flat EE and overlapped EE. Sequence labeling models cannot resolve the event overlap problem. Multi-stage pipeline models introduce error propagation. One-stage joint extraction models cannot fully leverage the potential event information and their high complexity make them unsuitable for datasets with a wide variety of event-argument types. Therefore, we propose a unified model for event extraction, called UEE. Our method can use the potential event information for the argument classification subtask and utilize the predefined type-restricted decoding strategy to improve the model’s performance. We conduct experiments on three flat and overlapped EE benchmarks, namely FewFC, ACE05-CN and DuEE, and show that UEE achieves the state-of-the-art (SoTA) results. Moreover, the parameter number and inference speed of UEE are better than those of baselines in the same condition.