面向用水结构研究领域的多阶段实体关系联合抽取方法
Multi-stage entity and relation joint extraction method for water usage structure research field
陶天然1
作者信息
- 1. 华北水利水电大学信息工程学院,郑州 450046
- 折叠
摘要
以往的知识抽取模型忽略了实体关系间的内在语义关联,并且在处理具有复杂关系的用水数据集时会产生大量的冗余信息.针对以上问题,提出一种融合语义信息的实体关系联合抽取模型.模型包括三个阶段:第一阶段,将经过BERT-wwm编码后的文本信息投影到关系检测空间中,过滤掉关系集合中的冗余数据;第二阶段,利用多头注意力机制将关系信息融合进文本编码,获取对应关系下的头实体和尾实体集合;第三阶段,引入融合上下文语义信息的实体相关矩阵,完成对三元组的准确提取.实验结果表明,所设计的模型在用水结构研究数据集上取得了较好的实体关系抽取效果.
Abstract
Previous knowledge extraction models overlooked the intrinsic semantic relationships between entities and,in par-allel,resulted in a significant amount of redundant information when dealing with water usage datasets with complex relationships.To address these issues,this paper proposes a novel entity-relation joint extraction model that integrates semantic information.The model consists of three stages:In the first stage,text information encoded through BERT-wwm is projected into the relation detec-tion space to filter out redundant data in the relation set.In the second stage,a multi-head attention mechanism is employed to fuse relation information into text encoding,obtaining sets of head and tail entities corresponding to each relation.In the third stage,an entity-related matrix incorporating fused contextual semantic information is introduced to accurately extract triplets.Experimental results demonstrate that the designed model achieves a notable entity-relation extraction performance on a water usage structure re-search dataset.
关键词
用水结构研究/联合抽取/多阶段抽取/语义信息/知识图谱Key words
water usage structure research/joint extraction/multi-stage extraction/semantic information/knowledge graph引用本文复制引用
出版年
2024
NETL
NSTL
万方数据