基于并联残差膨胀卷积网络的短文本实体关系联合抽取  

Short text entity relation joint extraction based on parallel residual expansion convolutional network

作　　者：曾伟 奚雪峰 崔志明[1,2,3] ZENG Wei;XI Xuefeng;CUI Zhiming(Suzhou University of Science and Technology,Suzhou 215000,China;Suzhou Key Laboratory of Virtual Reality Intelligent Interaction and Application Technology,Suzhou 215000,China;Suzhou Smart City Research Institute,Suzhou 215000,China)

机构地区：[1]苏州科技大学电子与信息工程学院,江苏苏州215000 [2]苏州市虚拟现实智能交互及应用技术重点实验室,江苏苏州215000 [3]苏州科技大学智慧城市研究院,江苏苏州215000

出　　处：《现代电子技术》2025年第2期169-178,共10页Modern Electronics Technique

基　　金：国家自然科学基金项目(62176175);江苏省“六大人才高峰”高层次人才项目资助(XYDXX-086);苏州市科技计划项目(SGC2021078)。

摘　　要：关系抽取旨在从文本中提取出实体对之间存在的语义关系,但现有的关系抽取方法均存在关系冗余和重叠的不足,尤其是对于短文本,会因上下文信息不足而出现语义信息不足和噪声大等问题。此外,一般流水线式的关系抽取模型还存在误差传递问题。为此,文中提出一种基于并联残差膨胀卷积网络的短文本实体关系联合抽取方法。该方法利用BERT生成语义特征信息,采用并联残差膨胀卷积网络来捕获语义信息,从而提升上下文信息的捕获能力并缓解噪声。联合抽取框架通过抽取潜在关系来过滤无关关系,然后再抽取实体以预测三元组,从而解决关系冗余和重叠问题,并提高计算效率。实验结果表明,与现有的主流模型相比,所提模型在三个公共数据集NYT、WebNLG和DuIE上的F1值分别为90.9%、91.3%和73.5%,相较于基线模型均有提升,验证了该模型的有效性。Relationship extraction aims to extract semantic relationships between entity pairs from text,but existing relationship extraction methods suffer from the shortcomings of relationship redundancy and overlap,especially for short texts,which may result in insufficient semantic information and loud noise due to insufficient contextual information.Moreover,conventional pipeline based relation extraction models face error propagation issues.A method of short text entity relation joint extraction based on parallel residual expansion convolutional network is proposed.In this method,BERT(bidirectional encoder representations from transformers)is used to generate semantic feature information,and the parallel residual dilated convolutional network is employed to capture semantic information,thereby enhancing the ability to capture context information and alleviate noise.The joint extraction framework can be used to filter out irrelevant relationships by extracting potential relationships,and extract entities to predict triplets,thus solving the problems of relationship redundancy and overlap,and improving computational efficiency.The experimental results demonstrate that,in comparison with existing mainstream models,the F1 values of the proposed model on the three public datasets NYT,WebNLG and DuIE are 90.9%,91.3%and 73.5%,respectively,which are improved compared with the baseline model,which verifies the effectiveness of the model.

关 键 词：实体关系抽取 短文本 残差膨胀卷积网络 语义特征 联合抽取 BERT编码器 

分 类 号：TN919-34[电子电信—通信与信息系统] TP391.1[电子电信—信息与通信工程]