Research On N-ary Relational Facts For Link Prediction

Different from the link prediction on traditional binary relationships,the link prediction on N-ary relational facts or hyper-relational facts is the prediction of complex relationship facts containing multiple entities with one or more relations,which can directly learn the complex facts within the knowledge graph,reducing information loss and knowledge ambiguity,and then improve the accuracy of prediction results.However,the existing methods suffer from inadequate understanding of multivariate relational facts and insufficient information utilization,which reduces the accuracy of prediction results and the practicality of the methods.For this reason,this paper conducts research in two directions:representation of N-ary relational fact and its mining and utilization of factual information.The main research contents are as follows.1.A link prediction model is proposed based on a heterogeneous graph representation of N-ary relational facts with a graph transformer network.Aiming at the problem of ignoring fact heterogeneity,the thesis proposes a heterogeneous graph representation of N-ary relational facts.In addition to the intra-interaction in relation-entity pairs and the inter-interaction between primary triple and auxiliary relation-entity pairs,it adds the inter-interaction between auxiliary relation-entity pairs to enhance the representation of N-ary relational facts.Then,the thesis introduces a graph transformer neural network,which effectively extracts heterogeneous interactions within relationship facts and predicts both entities and relations.The model conducts comparison experiments on the commonly used public dataset JF17K with advanced prediction models based on a set of relation-entity pairs representation,such as t Na LP~+.The results show it improves on MRR metrics with mixed-relation prediction by 0.52% and mixed-entity prediction by 1.36%.2.An enhanced graph transformer prediction model based on lightweight aggregation is proposed.Aiming at the problem of information loss on graph transformer deep layers,which is caused by variable lengths of relational facts,the thesis proposes a lightweight aggregation method based on cross-channel graph maximum pooling.Based on the previous link prediction model of the graph transformer network,a graph maximum pooling is used to aggregate key features on each layer after aligning all network layers by channel,which achieves the information complement on the output layer.The model conducts comparison experiments on dataset JF17K with the advanced models based on primary triple and auxiliary relation-entity pairs representation,such as HINGE.The results show it improves on MRR metrics with primary-relation prediction by 1.37% and primary-entity prediction by 13.88%.3.An N-ary relational link prediction system for public figure information query is constructed.This paper demonstrates the practicality of this approach by designing and building a system with entity prediction,relation prediction,data management and historical query functions in conjunction with the demand for public figure information query in social network scenarios.