Hybrid Semantic Type Representation for Zero-Shot Event Extraction

Abstract

Event extraction is a significant task in natural language processing. However, it is labor-intensive to get annotation when generalizing to new event types and ontologies. In this paper, we propose the HTR (Hybrid Type Representation) framework for zero-shot event extraction. We make a distinction of the abstraction level between events and roles, analyze role semantics, and propose a new representation approach, LRDB (label-related description-based), which is effective for both argument classification and collaboration with trigger extraction. We conduct extensive evaluation on ACE2005 dataset and achieve state-of-the-art.

Tasks

Reproductions