关系信息增强的少样本关系分类原型表示研究  

Relation Information Enhancement-based Prototype Representation Generating Approach for Few-shot Relation Classification

作　　者：王磊[1] 任鹏瑞 瞿剑峰 WANG Lei;REN Pengrui;QU Jianfeng(School of Computer Science and Technology,Soochow University,Suzhou 215000,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215000

出　　处：《小型微型计算机系统》2024年第7期1615-1621,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金青年基金项目(62102276)资助;江苏省青年基金项目(BK20210705)资助;江苏省高校面上项目(21KJD520005)资助;江苏省高等学校大学生创新创业训练计划项目(202110285110Y)资助;江苏高校优势学科建设工程项目资助.

摘　　要：少样本关系分类任务旨在只使用极少量样本来识别出句子中给定实体对的关系,现有工作在FewRel 1.0数据集上进行了大量研究,并没有考虑“以上都不是”(none-of-the-above,NOTA)关系的检测问题,而已有的考虑NOTA关系的工作都是在间接地计算查询样例与NOTA的相似度,这容易导致误差累积,模型性能因此表现不佳.为此,本文提出了一种基于关系信息增强的能够直接生成NOTA关系原型表示的方法.首先,本文提出了一种易混淆实例采样策略,通过挑选出信息丰富且易混淆的实例作为NOTA关系的支持数据并直接计算NOTA关系的原型表示;然后,本文使用了关系名和关系的描述信息作为外部关系信息,来提供给模型更多可用信息以生成更准确的原型表示.在广泛使用的FewRel 2.0数据集上的大量实验结果证明了本文提出的方法的优越性.Few-shot relation classification is to recognize the semantic relation between an entity pair with only a handful of samples.The existing work has carried out a lot of research on the FewRel 1.0 dataset.However,the problem of detecting the none-of-the-above(NOTA)relation in the few-shot relation classification is ignored.And the existing work considering the NOTA relation only indirectly calculates the similarity between query samples and NOTA,which is prone to error accumulation and poor model performance.In this paper,a novel relation information enhancement-based method for directly generating NOTA relation prototype representation is proposed.Firstly,a hard instances sampling strategy is employed to select out hard but informative instances as the supporting data of NOTA and compute the NOTA relation prototype representation directly.Then,we use relation name and relation description as external relation information to provide more available information for the model to generate more accurate prototype representation.Extensive experimental results on the widely-used FewRel 2.0 dataset demonstrate the superiority of our proposed method.

关 键 词：少样本学习 关系分类 原型网络 易混淆实例 关系信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]