Research On Relation Exntraction Technology Graph Construction

Knowledge graph means data stored in the form of graphs,and has been widely used in many fields.Knowledge graph with graph matching algorithms can provide theoretical support for medical diagnosis,public opinion discovery,and event reasoning.The knowledge triplet<entity-relation-entity>is the most basic unit of knowledge in the knowledge graph.It is very important to efficiently and quickly extracts the relations between entities from the free text of cyber threat intelligence,for building a cyber security knowledge graph.This paper has conducted in-depth research about the relation extraction algorithm for cyber security knowledge graph construction.The current distantly supervised relation extraction algorithms have great defects in sentence feature extraction and noise reduction ability between sentence bags,which also leads to the poor performance of many relation extraction algorithms.Based on the attention mechanism,this paper has improved the existing relation extraction algorithm in sentence feature extraction and sentence bag noise reduction.The main work content and results of this article are as follows:(1)The long short-term memory networks(LSTM)and convolutional neural networks used in traditional relations extraction methods cannot capture the features of longer distances in the sentence,so the extraction of the features of the sentence is insufficient,which also affects the effect of subsequent classification of the relation expressed by the sentence.In response to this problem,this paper proposes a relation extraction neural network model based on hierarchical attention mechanism,using self-attention mechanism for feature capturing between words,and sentence-level soft attention mechanism to reduce dimension of sentence feature.Compared with previous models,our model can better capture sentence features and improve the effect of subsequent sentence relation classification.Good results have been achieved on the public dataset NYT-10,the P@100,P@200,and P@300,AUC indicators have increased by 4.8%,4.9%,2.3%,and 1.1%,respectively.(2)In the process of automatic labelling of corpus,the distant supervision method introduces a large number of incorrectly labelled training corpus because of the assumption is too broad.This paper proposes a relation similarity attention mechanism aligned with relation vectors.Compared with the previous models,our model can better capture the similarity between each sentence and all relation vectors,and better extract the relation in the sentence under distant supervision.On the standard dataset NYT-10,compared with the previous work the indicators of P@100,P@200,P@300 and AUC have increased by 2.5%,2.4%,2.8% and 2.0% respectively.