基于Span方法和多叉解码树的实体关系抽取  

Entity Relation Extraction Based on Span Method and Multi-fork Decoding Tree

作　　者：张鑫 冼广铭 梅灏洋 周岑钰 刘赢方 ZHANG Xin;XIAN Guang-ming;MEI Hao-yang;ZHOU Cen-yu;LIU Ying-fang(School of Software,South China Normal University,Foshan 528225,China)

机构地区：[1]华南师范大学软件学院,广东佛山528225

出　　处：《计算机技术与发展》2023年第5期152-158,166,共8页Computer Technology and Development

基　　金：国家自然科学基金(61070015)。

摘　　要：实体关系抽取作为自然语言处理领域的一项关键技术,在构建知识图谱、信息检索等领域有着极为重要的意义。然实体关系抽取模型普遍存在词与词之间依赖性运用不足、实体识别效果低下以及单解码带来的三元组强行执行某种不必要顺序的问题。为了解决这三个方面的问题,提升模型的性能,提出了一种新的实体关系抽取模型。该模型首先运用提取特征能力更强的BERT预训练模型获取句子表征,然后采用图卷积神经网络来增强实体与关系之间的依赖关系,再使用对实体提取能力更强的Span方法(识别实体的神经网络方法)进行实体抽取,最后采用深度多叉解码树实施并行解码得到相应的关系三元组。在CoNLL04、ADE数据集上的实验结果表明,与其他的关系抽取基线模型相比,该模型的F1值具有较好的提升,同时也验证了该文模型的有效性与泛化能力。As a key technology in the field of natural language processing,entity relation extraction is of great significance in the construction of knowledge graphs,information retrieval and other fields.However,the entity relation extraction model generally has the problems of insufficient application of dependencies between words,low entity recognition effect,and the forced execution of an unnecessary order of triples brought by single decoding.In order to solve three problems and improve the performance of the model,a new entity relation model is proposed.The model first uses the BERT pre-training model with stronger feature extraction ability to obtain sentence representation,and then uses graph convolutional neural network to enhance the dependency between entities and relationships.The Span method(Neural Network Methods for Recognizing Entities),which has stronger entity extraction ability,is used for entity extraction.Finally,a deep multi-fork decoding tree is used to implement parallel decoding to obtain the corresponding relationship triples.The experiments on the CoNLL04 and ADE datasets show that compared with other relation extraction baseline models,the F1 value of the proposed model has a better improvement.And it also verifies the effectiveness and generalization ability of the proposed model.

关 键 词：实体识别 关系抽取 深度学习 预训练模型 多叉解码树 图神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]