摘要
关系抽取旨在识别非结构化文本中实体对之间的语义关系,现有方法无法灵活应用于开放域场景,如何在开集环境中自适应地进行关系抽取仍然是该领域的一项重要挑战。针对关系抽取场景中未知类别样本识别问题,提出了一种原型对比学习驱动的鲁棒性关系抽取方法。根据高斯分布为每个类别初始化可学习的原型中心,通过改进对比学习损失函数拉近同类样本到类别原型的距离,进一步通过增加正则化项约束样本输出概率分布与异类原型的差异。与对比方法相比,所提方法在3个数据集下的开集准确率比次优的模型分别提升了2.93%,3.16%,3.18%,且在闭集上的准确率没有降低,表明了模型能够在特征空间中拉近同类样本之间距离,推开异类样本之间距离,从而在不干扰已知关系类别的情况下,有效提升关系抽取模型对开放未知关系类别样本检测的鲁棒性。
Relation extraction aims to identify semantic relationships between entity pairs in unstructured text.However,existing methods cannot be flexibly applied to open-domain scenarios,and it remains a significant challenge in this field to adaptively perform relation extraction in an open set environment.To identify unknown class samples in relation extraction problem,a robust relation extraction method driven by prototype contrastive learning is proposed.Initializing a learnable prototype center for each class based on Gaussian distribution,the method improves the contrastive learning loss function to reduce the distance between samples of the same class and their class prototype.Further,it enhances the model by adding a regularization term to constrain the difference between the output probability distribution of samples and prototypes of different classes.Compared to baseline methods,the proposed approach achieves a respective improvement of 2.93%,3.16%,and 3.18%in open set accuracy on three datasets,without decreasing accuracy on closed sets.This demonstrates the model’s ability to reduce the distance between samples of the same class in feature space while pushing apart samples of different classes,effectively enhancing the model’s robustness in detecting samples of unknown relation categories in open set scenarios without affecting known relation categories.
作者
吴涛
徐敖远
田侃
先兴平
袁野
张姝
曹新汶
WU Tao;XU Aoyuan;TIAN Kan;XIAN Xingping;YUAN Ye;ZHANG Shu;CAO Xinwen(School of Security and Information Law,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;CQUPT-CCTGM Joint Laboratory of Intelligent Museum,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;School of Economics and Management,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China)
出处
《重庆邮电大学学报(自然科学版)》
北大核心
2025年第1期17-28,共12页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金项目(62376047,62106030)
重庆市自然科学基金创新发展联合基金重点项目(CSTB2023NSCQ-LZX0003,CSTB2023NSCQ-LMX0023)
重庆市教委科学技术研究计划重点项目(KJZD-K202300603)
重庆市技术创新与应用发展面上项目(CSTB2022TIAD-GPX0014)。
关键词
关系抽取
开集识别
鲁棒性分类
原型学习
relation extraction
open set recognition
robust classification
prototype contrastive learning
