首页|基于异质图卷积神经网络的论点对抽取模型

基于异质图卷积神经网络的论点对抽取模型

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
  • 维普
针对论点对抽取任务中存在着评审段和反驳段之间交互信息难以捕获以及忽略了对句子间的相对位置信息进行建模问题，提出基于异质图卷积神经网络的论点对抽取模型。该模型在评审段和反驳段中构建异质图，定义 2 种不同类型的节点及 4 种不同类型的边，通过关系图卷积神经网络来更新图中节点的表示。提出位置感知的句子对生成器，利用旋转位置编码来建模评审段和反驳段句子间的相对位置信息。在RR-passage和RR-submis-sion-v2 数据集上进行实验，实验结果表明，提出模型的性能均优于所有的基线模型。这表明通过构建异质图区分不同的节点类型和边的类型，设计位置感知的句子对生成器，能够提升论点对抽取模型的效果。
Heterogeneous graph convolutional neural network for argument pair extraction
An argument pair extraction model based on heterogeneous graph convolutional neural network was proposed aiming at the issue of difficulty in capturing interactive information between review passage and rebuttal passage and neglecting to model relative positional information between sentences.Heterogeneous graphs were constructed within the review passage and rebuttal passage.Two types of nodes and four types of edges were defined.The relational graph convolutional neural network was utilized to update the representations of nodes within the graph.A position-aware sentence pair generator was introduced,and rotary position embedding was employed to model the relative positional information between sentences in review passage and rebuttal passage.Experimental evaluations on the RR-passage and RR-submission-v2 datasets demonstrate that the proposed model outperforms all baseline models.The performance of the argument pair extraction model can be enhanced by constructing heterogeneous graphs to distinguish between different types of nodes and edges and designing a position-aware sentence pair generator.

argument miningargument pair extractiongraph neural networkrotary position embeddingnatural language processing

刘议丹、朱小飞、尹雅博

展开 >

重庆理工大学计算机科学与工程学院,重庆 400054

论辩挖掘 论点对抽取 图神经网络 旋转位置编码 自然语言处理

国家自然科学基金重庆市自然科学基金重庆市教委科学技术研究计划重庆理工大学研究生教育高质量发展行动计划

62141201CSTB2022NSCQ-MSX1672KJZD-M202201102gzlcx20233230

2024

10.3785/j.issn.1008-973X.2024.05.003

浙江大学学报(工学版)
浙江大学

浙江大学学报(工学版)

CSTPCD北大核心
影响因子：0.625
ISSN：1008-973X
年,卷(期)：2024.58(5)
  • 28