HAEE: Low-Resource Event Detection with Hierarchy-Aware Event Graph Embeddings

Guoxuan Ding, Xiaobo Guo, Gaode Chen, Lei Wang, Daren Zha

Abstract: The event detection (ED) task aims to extract structured event information from unstructured text. Recent works in ED rely heavily on annotated training data and often lack the ability to construct semantic knowledge, leading to a significant dependence on resource. In this paper, we propose a hierarchy-aware model called HAEE by constructing event graph embeddings. We utilize two relations (cause and subevent) to help model events on two dimensions of polar coordinates, so as to distinguish events and establish event-event relations. Specifically, events under the cause relation are constructed at the same level of the hierarchy through rotation, while events under the subevent relation are constructed at different levels of the hierarchy through modulus. In this way, coexistence and interactions between relations in time and space can be fully utilized to enhance event representation and allow the knowledge to flow into the low-resource samples. The experiments show that HAEE has high performance in low-resource ED task, and the analysis of different dimensions of embeddings proves that HAEE can effectively model the semantic hierarchies in the event graph.