Research On Few-shot Event Extraction Methods

Event extraction is an important task in the field of information extraction.It is widely used in the fields of public opinion monitoring,finance,and risk control,and has received extensive attention in recent years.Current event extraction methods require a large amount of labeled data.However,due to issues such as data privacy and labeling costs,a large amount of labeled data cannot be obtained in many scenarios.Few-shot event extraction is concerned.The task of few-shot event extraction has high research value and is also a new task with greater difficulty.This article focuses on the problem of few-shot event extraction,mainly from the following aspects:(1)Research on event extraction method based on multi-task learning.This paper studies two different labeling methods,including sequence labeling and multi binary tagging.Multi-task learning combines the two subtasks of event extraction including event detection and argument extraction to achieve the effect of mutual promotion.(2)Research on few-shot event extraction method based on QA.This paper attempts to introduce QA model for event extraction,and summarizes two guiding principles for Query construction: introducing more information;as close as possible to natural language.The question answering model has strong generalization and migration ability.In the case of few shot,the QA model achieves the best.(3)Research on the method of few-shot event extraction based on two stages.This paper proposes to decompose the task of event detection and argument extraction into two tasks of boundary recognition and type classification respectively.It is found that the two-stage method has a great improvement compared with the one-stage method.In addition,this article also tries to combine the two-stage method with metric learning.Compared with simple metric learning and two-stage method,twostage metric learning has a great improvement.