摘要
针对现有关系三元组抽取方法由于忽略关系本身的关系语义信息以及三元组中元素的相互依赖和不可分性所导致的抽取效果不佳问题,提出了一种基于关系提示的实体关系抽取方法。在构建单模块单步关系三重抽取模型(RPSS)的基础上,考虑不同层次的关系语义信息和符号级和特征级的关系提示信息,对实体和关系提示符进行联合编码,得到统一的全局表示;同时通过注意力机制挖掘不同嵌入之间的深层关联,构建三重交互矩阵,可在一个步骤中直接从单个模块中提取所有三元组。结果表明:所提方法在NYT、WebNLG两个基准数据集上实现了最佳的表现,F_(1)分别达到了93.3%和94.9%。
To address the issue of suboptimal performance in existing relation triplet extraction methods,caused by the oversight of relationship semantic information and the interdependence and indivisibility of elements in triplets,a novel entity relation extraction method based on relational prompts is proposed.Building upon a single-module single-step relation triple extraction model(RPSS),this method incorporates relationship semantic information at varying levels and symbol-level as well as feature-level relational prompt information.By jointly encoding entities and relational prompt symbols to derive a unified global representation,and employing attention mechanisms to delve into deep associations among different embeddings,a triple interaction matrix is constructed.This approach enables the extraction of all triplets directly from a single module in a single step.The proposed method demonstrates superior performance on benchmark datasets like NYT and WebNLG,achieving F_(1) scores of 93.3% and 94.9%,respectively.
作者
刘辉
张智
王启源
LIU Hui;ZHANG Zhi;WANG Qiyuan(College of Electronic Engineering,National University of Defense Technology,Hefei 230037,China;School of Computer Science and Technology,Harbin Institute of Technology,Harbin 150001,China)
出处
《西安交通大学学报》
北大核心
2025年第3期222-234,共13页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(62201601)
安徽省自然科学基金资助项目(2008085QF283)。
关键词
实体关系抽取
注意力机制
联合编码
entity relation extraction
attention mechanism
joint encoding
作者简介
刘辉(1983-),男,副教授。
