面向小样本关系抽取的自适应胶囊网络  

Adaptive capsule network for few-shot relation extraction

作　　者：张晓明 窦全胜[1,3] 陈淑振 唐焕玲 Zhang Xiaoming;Dou Quansheng;Chen Shuzhen;Tang Huanling(School of Information&Electronic Engineering,Shandong Technology&Business Univer-sity,Yantai Shandong 264000,China;School of Computer Science&Technology,Shandong Technology&Business Univer-sity,Yantai Shandong 264000,China;Shandong Future Intelligent Computing Collaborative Innovation Center,Yantai Shandong 264000,China)

机构地区：[1]山东工商学院信息与电子工程学院,山东烟台264000 [2]山东工商学院计算机科学与技术学院,山东烟台264000 [3]山东省高等学校未来智能计算协同创新中心,山东烟台264000

出　　处：《计算机应用研究》2022年第8期2357-2362,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61976124,61976125)。

摘　　要：小样本关系抽取任务是自然语言处理中的热点问题,旨在使用低成本的标注数据训练关系抽取模型。目前广泛使用的原型网络存在类原型表达不准确、不完整等问题。为了克服该问题,提出一种自适应胶囊网络(adaptive capsule network,ACNet),ACNet借助胶囊网络的归纳能力生成类原型,并在此基础上对动态路由过程进行评估,使其面对不同样本能自适应调整网络参数。同时,在ACNet中引入一种记忆迭代机制,帮助模型快速确定类表示。在小样本关系数据集FewRel上进行实验验证得出,ACNet能够有效处理小样本关系抽取任务。The few-shot relationship extraction task is a hot issue in natural language processing.It aims to train the relationship extraction model using low-cost label data.The widely used prototype network has some problems,such as inaccurate and incomplete expression of class prototypes.This paper proposed an adaptive capsule network to solve this problem.ACNet generated a class prototype with the inductive capability of the capsule network.On this basis,it evaluated the dynamic routing process so that it could adjust network parameters adaptively to different samples.At the same time,it introduced a memory iteration mechanism in ACNet to help the model determine the class representation quickly.Experiments on a few-shot relational dataset FewRel show that ACNet can handle few-shot relational extraction tasks.

关 键 词：关系抽取 小样本 自适应 胶囊网络 动态路由 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]