Research On Long-tail Relation Extraction In Cyberspace Security Knowledge Graph

The cyberspace security knowledge graph can mine the relation of heterogeneous data in massive data and help maintain cyber security.In the real world,the distribution of the relation classes used to build such a knowledge graph is long-tailed and hard to be extracted by machine learning techniques which require a large amount of training data.Though some works have made progress,the performances of these works are still poor.Layer-enhanced knowledge aggregation network,named Le KAN,is presented to extract the relations between two annotated entities from texts.Le KAN can boost the performance of the data-poor classes by transferring knowledge from the outside knowledge graph and different branches to the long-tail relation.First,layer-enhanced relational tree construction and aggregation algorithm is used to build the layer-enhanced relational tree.Second,translation and Graph SAGE embedding methods are used to form the relation representation for the layer-enhanced hierarchical relational tree.Finally,under the guidance of the layer-enhanced knowledge-aggregation attention mechanism,Le KAN can extract the relations more efficiently.To verify the effectiveness and interpretability of Le KAN,experiments on a publiclyavailable long-tail benchmark NYT-10 are conducted.The experimental results demonstrate that Le KAN outperforms other baseline models in terms of overall relation extraction performance,and archives better performance in terms of long-tail relation extraction evaluation.The rationality of modules in Le KAN is proved via the ablation study.The interpretability of transferring relational knowledge is verified via class embedding visualization.