摘要
实体关系联合抽取是构建知识图谱的重要阶段,旨在抽取文本中存在语义关系的实体对。针对已有的实体关系联合抽取方法在抽取过程中存在的冗余关系预测、实体关系重叠以及上下文潜在语义信息捕捉不足的问题,提出联合多头自注意力机制和对抗训练的方法进行实体关系的抽取。该方法利用多头自注意力机制捕获潜在语义特征,以提升模型对上下文语义信息的感知能力;将对抗训练引入模型的训练阶段,以增强模型的泛化能力和鲁棒性。实验结果表明:与现有主流模型对比,提出的模型在NYT和WebNLG两个公共数据集上都取得了更优的F 1值,在处理实体关系重叠问题以及不定数量三元组抽取上都能保持稳定的性能表现,验证了模型的有效性。
Extracting entities and relations is an important stage in the construction of the knowledge graph,with the goal of extracting pairs of entities that have semantic relationships with one another from a given text.To solve the problems of redundant relation prediction,overlapping entities and relations,and the inadequate capture of semantic information by prevalent methods for the joint extraction of entities and relations,we propose the fusion of a multi-head self-attention mechanism with adversarial training in this study to jointly extract entities and relations from texts.The proposed method uses a multi-head self-attention mechanism to capture the potential semantic features of the text and enhance the ability of the model to identify contextual semantic information.Adversarial training is introduced to the training of the model to enhance its robustness and capability for generalization.The results of experiments showed that the proposed model achieved higher values of the F 1 score than prevalent models in the area on both the NYT and the WebNLG public datasets,and maintained stable performance when dealing with overlapping entities and relations as well as an indefinite number of triple extractions.This verifies the effectiveness of the proposed model.
作者
甘雨金
李红军
唐小川
王子怡
甘晨灼
胡正浩
GAN Yujin;LI Hongjun;TANG Xiaochuan;WANG Ziyi;GAN Chenzhuo;HU Zhenghao(College of Computer Science and Cyber Security(Pilot Software College),Chengdu University of Technology,Chengdu 610059,China)
出处
《成都理工大学学报（自然科学版）》
CAS
CSCD
北大核心
2024年第3期513-521,共9页
Journal of Chengdu University of Technology: Science & Technology Edition
基金
国家自然科学基金(42050104)
自然资源部深时地理环境重建与应用重点实验室开放基金项目(DGERA20221102)。
关键词
实体关系联合抽取
对抗训练
多头自注意力
知识图谱
joint extraction of entities and relations
adversarial training
multi-head self-attention mechanism
knowledge graph
