摘要
针对基于句法依赖树和图卷积神经网络(GCN)的关系抽取(RE)任务中,由于句子中词与词之间的依赖连接被赋予相同权重所引入的噪声问题,提出一种基于多特征驱动图注意卷积网络(MFDA-GCN)的RE模型。该模型充分利用了句子的依赖类型、词性、相对实体位置等多种特征信息,通过引入注意力机制计算句法依赖树中不同连接的重要程度,再将多种特征信息动态地融入句子的词向量表示中。最后,根据词之间依赖连接的重要程度更有效地引导词信息传递,优化整个句子的词向量表示,进一步提高RE性能。实验结果表明,相较于其他基线模型,基于MFDA-GCN的RE模型具有更强的远距离词依赖捕获能力,且该模型在数据集SemEval-2010Task8和ACE2005EN上的F1值分别达到90.39%和79.86%。
Syntactic dependency tree contains rich syntactic information and is often applied in Relation Extraction (RE) tasks together with Graph Convolutional Network (GCN). However, standard GCNs treat every connection in the syntactic dependency tree equally, which introduces noise. To address this issue, Multiple Features Driven Attentive Graph Convolutional Network (MFDA-GCN) was proposed. Firstly, various feature information of sentence such as dependency type, part of speech and relative entity position were fully utilized;by introducing attention mechanism, the importance of different connections in the syntactic dependency tree was calculated. Then, the word vector representation of sentence was enhanced in a dynamic way by using multiple features. Finally, based on the importance of dependency connections between words, the transmission of word information could be better guided and new word vector representation of sentence was obtained for relation extraction. Experimental results show that compared to other baseline models, MFDA-GCN model has a stronger ability to capture long-distance word dependencies, and the F1 scores of this model on the SemEval-2010 Task 8 and ACE2005EN datasets reach 90. 39% and 79. 86%, respectively.
作者
李航程
钟勇
LI Hangcheng;ZHONG Yong(Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610213,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《计算机应用》
CSCD
北大核心
2024年第S01期24-28,共5页
journal of Computer Applications
基金
四川省科技成果转移转化平台项目(2020ZHCG0002)。
关键词
关系抽取
图卷积神经网络
句法依赖树
注意力机制
自然语言处理
Relation Extraction(RE)
Graph Convolutional Network(GCN)
syntactic dependency tree
attention mechanism
Natural Language Processing(NLP)
