Decomposition and recombination. A soft cascade model for event detection

Abstract

Event detection (ED) necessitates the execution of two distinct sub-tasks: Trigger Identification (TI) and Trigger Classification (TC). Existing ED methods employ multi-class classification to simultaneously identify triggers and classify trigger types. Such a paradigm only entangles TI and TC, exacerbating the type imbalance issue due to the preponderance of None events. This study proposes a simple yet effective decomposition and recombination (D&R) decoding framework for ED, by first decomposing it into independently processed sub-tasks, TI and TC, and then allowing them to be recombined in an explicitly interactive manner. Particularly, we first decompose the ED prediction into two probabilistic predictions of TI and TC. In this way, None events and no-None events will be treated separately to avoid exacerbating the data imbalance problem as a whole. Subsequently, we introduce an event graph encoder to inject association knowledge for events, alleviating the imbalance problem in no-None events. Additionally, we propose a probabilistic supervised contrastive learner that aligns the input features of the two sub-tasks and captures their commonalities. Finally, the two output probabilities are combined in a soft cascading manner, enabling our model to effectively select event types and distinguish between None and no-None types compared to prior studies. Experimental results reveal that our model achieves sota result on the MAVEN leaderboard and outperforms all competitive baselines on ACE-2005.