融合语义和句法图神经网络的实体关系联合抽取

Fusion of Semantic and Syntactic Graph Convolutional Networks for Joint Entity and Relation Extraction

实体关系抽取任务是信息抽取的核心任务,它对于有效地从爆炸性增长的数据中提取出关键性的信息有着不可替代的作用,也是构建大规模知识图谱的基础任务,因此研究实体关系抽取对各种自然语言处理任务具有重要意义。尽管现有的基于深度学习方法的实体关系抽取已经有了很成熟的理论和较好的性能,但依然还存在着误差累积、实体冗余、交互缺失、三元组重叠等问题。语义信息和句法信息对自然语言处理任务都具有重要作用,为了充分利用这些信息以解决上述提到的问题,提出了一种融合语义和句法图神经网络的二元标记实体关系联合抽取模型FSSRel(Fusion of Semantic and Syntactic Graph Convolutional Networks Binary Tagging Framework for Relation triple extraction)。该模型分为三个阶段进行:第一阶段,对三元组主体的开始结束位置进行预测标记;第二阶段,分别通过语义图神经网络和句法图神经网络提取语义特征和句法特征,并将其融合进编码向量;第三阶段,对语句的每种关系的客体位置进行预测标记,完成最终三元组的提取。实验结果表明,在NYT数据集和WebNLG数据集上,该模型的F1值较基线模型分别提升了2.5%和1.6%,并且在拥有重叠三元组和多三元组等问题的复杂数据上也有良好的表现。

Entity and relation extraction task is the core task of information extraction.It plays an irreplaceable role in effectively extracting key information from explosive growth data,and is also the basic task of building a large-scale knowledge graph.Therefore,the research on entity relationship extraction task is of great significance for various natural language processing(NLP)tasks.Although the existing entity and relation extraction based on deep learning method has a very mature theory and good performance,there are still some problems,such as error accumulation,entity redundancy,lack of interaction,entity and relation overlap.Semantic information and syntactic information play an important role in NLP tasks.In order to make full use of them to solve the above problems,a fusion of semantic and syntactic graph convolutional networks binary tagging framework for relation triple extraction(FSSRel)is proposed.The model is divided into three stages.In the first stage,the start and end positions of the triple body are predicted.In the second stage,semantic features and syntactic features are extracted by semantic graph neural network and syntactic graph neural network respectively,and fused into the coding vector.In the third stage,the object position of each relation of the statement is predicted and marked to complete the extraction of the final triple.Experimental results show that the F1 value of the model increases by 2.5%and 1.6%respectively compared with the baseline model on the NYT dataset and the WebNLG dataset,and it also performs well on complex data with overlapping triples and multiple triples.