Graph-based event schema induction in open-domain corpus

Abstract

An event schema provides a formal language for representing events and modeling knowledge about the world. Existing event schema induction methods often only applies text features to the cluster, restricting its cluster capabilities. This article presents a Graph-Based Event Schema Induction model to extract structural features from our constructed graph. Inspired by in-context learning, we propose a way to conceptualize clusters to generate event schemas. We evaluated the clustering experiment using the Adjusted Rand Index (ARI), normalized mutual information (NMI), accuracy (ACC), and BCubed-F1 metrics and generated event schemas based on overlap ratio and acceptable ratio. The experimental results show that our method has shown improvement in terms of clustering effectiveness, and the generated event schemas achieved highly acceptable ratio.