关键词: 关键词: 关系抽取, 原子特征, 特征组合, 异构图, 图注意力网络

Abstract: Abstract: Relation extraction aims to identify predefined semantic relationships between two entities within a sentence. Traditional graph neural network-based relation extraction methods generally use dependency trees to construct a graphical representation of the sentence. However, the graph structure constructed by the dependency tree has limited expressive power and is unable to fully capture the rich syntactic structure information of the target entity. To address this issue, a feature combination-based heterogeneous graph attention network method for relation extraction is proposed. The method first extracts atomic features from the sentence and then obtains composite features by combining these atomic features. Subsequently, the composite features and relation labels are represented as two types of nodes to construct a "feature-relation bipartite graph." Finally, a graph attention network dynamically updates the nodes to perform relation extraction. The approach effectively utilizes the composite features and the syntactic structural information in the sentence, enhancing the performance of relation extraction. Experiments conducted on the ACE05 English dataset and the SemEval-2010 task 8 dataset achieved F1 scores of 84.11% and 90.67%, respectively, demonstrating the effectiveness of the method.

Key words: Keywords: Relation extraction, Atomic feature, Feature combination, Heterogeneous graph, Graph attention networks