Syntactic Enhanced Projection Network for Few-Shot Chinese Event Extraction

Abstract

Few-shot learning event extraction methods gain more and more attention due to their ability to handle new event types. Current few-shot learning studies mainly focus on English event detection, which suffering from error propagation due to the identify-then-classify paradigm. And these methods could not be applied to Chinese event extraction directly, because they suffer from the Chinese word-trigger mismatch problem. In this work, we explore the Chinese event extraction with limited labeled data and reformulate it as a few-shot sequence tagging task. To this end, we propose a novel and practical few-shot syntactic enhanced projection network (SEPN), which exploits a syntactic learner to not only integrate the semantics of the characters and the words by Graph Convolution Networks, but also make the extracted feature more discriminative through a cross attention mechanism. Differing from prototypical networks which may lead to poor performance due to the prototype of each class could be closely distributed in the embedding space, SEPN learns to project embedding to space where different labels are well-separated. Furthermore, we deliberately construct an adaptive max-margin loss to obtain efficient and robust prototype representation. Numerical experiments conducted on the ACE-2005 dataset demonstrate the efficacy of the proposed few-shot Chinese event extraction.