Zero-Shot Event Detection Based on Prompt and Deep Prototype Clustering

Current event detection methods mainly focus on detecting known event types with annotated data, which are not suitable for detecting new event types in real-world scenarios where a large amount of unannotated information exists. Existing methods for new event detection primarily focus on the contextual information of the current word in a sentence, without fully utilizing the contextual semantic information of the entire event. Furthermore, the boundaries of marginal event types are often blurred and lack clarity, which hampers the effectiveness of new event type detection. To overcome these limitations, this paper proposes a zero-shot event detection model based on prompt and deep prototype clustering (ZEPC). By incorporating position encoding and prompt learning methods, the performance of trigger word prediction is optimized. Simultaneously, contrastive learning is employed to dynamically adjust the boundaries of marginal event types and combined with a deep prototype clustering network to improve the recognition performance of new event types. Our proposed zero-shot event detection method enhances the capability of representing zero-shot events and achieves good detection performance on both the known event detection and new event detection tasks in the FewShotED and ACE-2005 datasets.