Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (15): 27-37.DOI: 10.3778/j.issn.1002-8331.2301-0105

• Research Hotspots and Reviews • Previous Articles     Next Articles

Survey on Few-Shot Relation Extraction

LIU Bei, XU Zhuoming, TAO Wan, LIU Sanmin   

  1. 1.School of Computer and Information, Anhui Polytechnic University, Wuhu, Anhui 241000, China
    2.School of Computer and Information, Hohai University, Nanjing 211100, China
  • Online:2023-08-01 Published:2023-08-01

少样本关系抽取研究综述

刘蓓,许卓明,陶皖,刘三民   

  1. 1.安徽工程大学 计算机与信息学院,安徽 芜湖 241000
    2.河海大学 计算机与信息学院,南京 211100

Abstract: Relation extraction is a vital subtask of information extraction, and it is also an important component in the construction of the knowledge graph. The goal of relation extraction is to extract semantic relationships between entities from the natural language text, so as to discover the connections between the data. The process of relation extraction depends on a large number of labeled training data, though there are often being existed some cold start problems in practical applications. How to learn to extract relation through a small number of samples has become one of the hot topics in this field. Based on the investigation of a large number of literatures, the recent research achievements of the problems on few-shot relation extraction are summarized. Firstly, starting from the definition of few-shot relation extraction tasks, the mechanism of training and the methods of classification are introduced. Secondly, the paper introduces the siamese network, graph neural network and prototypical network from perspective of metric learning. And then the initial network parameters and pre-training network parameters are introduced from the perspective of parameter optimization learning too. Next, the paper introduces the common data sets, evaluation indexes and experimental results of representative methods. Finally, the possible development trends of the future research on few-shot relation extraction are prospected.

Key words: few-shot learning, relation extraction, meta-learning, metric learning, prototypical networks

摘要: 关系抽取是信息抽取的一项重要子任务,也是构建知识图谱的重要环节,其目标是从自然语言文本中抽取出实体之间的语义关系,从而更好地挖掘数据之间的联系。关系抽取的过程需要依赖大量标注的训练样本,而实际应用中却经常存在冷启动问题,如何通过少量样本进行关系抽取已成为该领域关注的热点之一。在调研大量文献的基础上对少样本关系抽取的近期研究现状进行总结,先从少样本关系抽取任务的定义出发,介绍了少样本关系抽取任务的训练机制与分类情况;从度量学习和参数优化学习两个角度分别介绍了基于孪生网络、图神经网络和原型网络,以及基于初始化网络参数和预训练网络参数在少样本关系抽取问题上的研究成果;介绍了少样本关系抽取的常用数据集、评价指标及代表性方法的实验结果;总结了现有研究存在的问题,并展望了少样本关系抽取未来可能的发展趋势。

关键词: 少样本学习, 关系抽取, 元学习, 度量学习, 原型网络