Few-shot cross domain event discovery in narrative text

Abstract

Cross-domain event detection presents notable challenges in the form of data scarcity, and existing few-shot algorithms only consider events whose types are predefined, resulting in low coverage or excessive trivial identification results. To address this issue, this paper proposes the task Few-shot Cross Domain Event Discovery, which includes two subtasks: Domain Event Discovery and Few-shot Domain Adaptation. The former aims to identify the type-agnostic event triggers, and the latter completes domain adaptation with only a few annotated domain samples. Additionally, we introduce a positive–negative balanced sampling mechanism and a novel domain parameter adapter for these two subtasks, respectively. Extensive experiments on the DuEE dataset and the ACE2005 dataset show that our proposed method outperforms the current state-of-the-art method by 6.3% in Mix-F1 score on average. Moreover, we achieve SOTA performance in all domains of the DuEE dataset.