Learning Inclusion Axioms Via Representation Learning

Ontology axiom learning is an important task in ontology engineering,which mainly constructs logical inclusion axioms from knowledge graph.Recently,how to use representation learning to learn ontology inclusion axioms from large-scale knowledge graphs is a big challenge.On the one hand,the scale of knowledge graphs are getting bigger and bigger which traditional construction methods cannot cope with.On the other hand,existing representation models hardly learn the logical relationship between predicates.Therefore,studying the representation learning model that can represent the logical inclusion relationship between predicates is the key to the problem.Firstly,this paper designs a new representation learning model named Set E,which gives a unified continuous low-dimensional dense vector representation to the unary predicate type and the binary predicate relationship in the knowledge graph.Specifically,the type as a unary predicate is encoded into a set of entities,the type is represented as the boundary of the set;the relation as a binary predicate is encoded into a set of pairs of entities,and the relation is expressed as the boundary of the set.This representation preserves the inclusion relationship between predicates and can be used to learn ontology inclusion axioms.Secondly,this paper proposes a method based on linear programming,which uses the predicate representation trained by Set E to learn the inclusion relationship between predicates.That is,the score function of the subpredicate is regarded as the feasible domain to find the minimum value of the suppredicate score function,and then the axiom is selected according to the minimum value and the threshold.Finally,this paper designed kinds of experiments to compare the performance of the Set E model-based method with traditional methods.Experiments show that the proposed method can obtain high-quality inclusion axioms on the real data set DBpedia,and can learn considerable axioms when the knowledge graph is incomplete.The whole process of learning axioms has certain interpretability,including the representation of predicates and the process of problem reduction.And this paper provides a feasible method for large-scale ontology automation construction.