Label graph augmented soft cascade decoding model for overlapping event extraction

Yiming Hei, Lihong Wang, Jiawei Sheng, Jianwei Liu, Qian Li, Shu Guo

Abstract: Event extraction (EE) is a fundamental information extraction task that aims to identify structured events, including event types, triggers and arguments, from unstructured texts. The overlapping EE, in which a trigger may be assigned to multiple event types, or an argument owns more than one role, is a more challenging task. Existing methods deconstruct the task in a pipeline-based cascade decoding paradigm, sequentially extracting overlapping event elements (triggers and arguments) based on the previous extracted results. On the one hand, this hard cascade decoding architecture is prone to serious error propagation in overlapping EE. On the other hand, they ignore the correlation between event types and argument roles, which is beneficial for this task. Facing these issues, we present an Event-Role label graph augmented cascade framework by modeling EE as a sub-task-dependent soft decoding architecture. To mitigate error propagation between sub-tasks, we propose a reliability-aware soft cascading module, which gradually transfers smoother features to downstream sub-tasks. To further enhance the model’s ability to classify overlapped elements, we design the Event-Role label graph representation learner to incorporate event-to-event and role-to-role association information into label space of event types and argument roles. It facilitates overlapped elements to be assigned to multiple labels with approximate feature space. We perform extensive experiments on two widely-used EE benchmark datasets, FewFC and ACE-2005, where our model outperforms state-of-the-art methods on the overlapping EE task and also adapts to the general EE without overlapped elements.