基于Transformer交叉集成特征的实体关系抽取  

Entity relationship extraction based on Transformer for feature cross integration

作　　者：陈金玉 王名扬[1] 刘旭[1] CHEN Jinyu;WANG Mingyang;LIU Xu(College of Computer and Control Engineering,Northeast Forestry University,Harbin 150036,China)

机构地区：[1]东北林业大学计算机与控制工程学院,黑龙江哈尔滨150036

出　　处：《东北师大学报(自然科学版)》2025年第1期74-81,共8页Journal of Northeast Normal University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(71473034);黑龙江省自然科学基金资助项目(LH2019G001)。

摘　　要：提出基于Transformer交叉集成特征的实体关系抽取模型ERonTCI,通过对不同层次特征信息的交叉集成来提升实体关系抽取的精度.ERonTCI模型将BERT预训练模型作为嵌入层,将BiGRU网络和MSCNN网络作为主干结构,分别提取文本中的全局特征和局部特征;使用Transformer实现对全局和局部特征的集成,以避免信息损失,提升模型对文本特征信息的保留度;将集成后的特征信息输入线性层完成实体关系的抽取工作.在公开数据集上进行实验,结果表明,ERonTCI模型取得了比基线模型更好的实体关系抽取效果.This paper proposes an entity relationship extraction model ERonTCI based on Transformer for feature cross integration,which aims to improve the accuracy of entity relationship extraction through the cross-integration of different levels of feature information.The ERonTCI model uses the BERT pre-training model as the embedding layer,uses the BiGRU network and the MSCNN network as the feature extraction backbone,and extracts the global features and local features in the text respectively.Transformer is used to integrate global and local features to avoid information loss and improve the model s retention of text feature information;the integrated feature information is input into the linear layer to complete the extraction of entity relations.The experimental results on the public data set show that the ERonTCI model has achieved better entity relationship extraction performance than the baseline models.

关 键 词：实体关系抽取 TRANSFORMER BiGRU MSCNN 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]