Event detection based on the label attention mechanism

Qing Cheng, Yanghui Fu, Jincai Huang, Guangquan Cheng, Hang Du

Abstract: Event detection is an important subtask of event extraction. The goal of event detection (ED) is to detect the occurrences of events and categorize their descriptions in the text. Previous work solved this task by recognizing and classifying event triggers. As a result, existing approaches to the task of event detection require both manually annotated triggers and event types in training data. However, some texts have no triggers, and some triggers are ambiguous. Moreover, triggers are nonessential to event detection, and annotation of the training corpus is expensive and time-consuming. To address this problem, we propose a novel framework called the event detection model based on the label attention mechanism (EDLA), which does not depend on triggers but rather models the task as a text multilabel classification task. Additionally, our model considers the semantic information of event labels, which increases the model's semantic understanding of labels. Experimental results using the DuEE dataset demonstrate its effectiveness, including increasing the F1-score of event classification to 95.8% and providing an increase of 6.0 in the F1-score over the traditional pipeline methods. It obtains new state-of-the-art results on the event detection task.