Ontology Construction Based On Open World Assumption

With the development of information technologies,especially Artificial Intelligence(AI)technologies,large-scale knowledge graphs(KGs)are published on the World Wide Web(WWW).This provides support for many AI applications,such as knowledge interlinking,intelligent question answering,semantic searching.In general,a KG can be logically divided into a data layer and a schematic layer.The former is used to describe factual knowledge in the real world and the latter consists of axioms and is used to organize knowledge in the data layer.The schematic knowledge is often called ontological knowledge.Due to the lack of schema information in the semi-structured data of encyclopedia websites,there is a lack of ontology in some KGs.This lead to a series of applications can not be implemented on KGs.For example,for a KG without ontological knowledge,the logic reasoning engine can not automatically detect the logic errors and can not reason with the implied facts.To build ontologies,existing ontology construction approaches use statistical methods to obtain candidate axioms from factual knowledge.Such approaches assume that the data not in the KG is taken as a negative example based on the closed world assumption.These methods will introduce a lot of false negative examples which greatly affects the quality of constructed ontology.To overcome these problems,this paper proposes the approach of building ontology based on the open world assumption,which is mainly divided into three parts: facts completion,axioms mining and the axioms completion.In this paper,we firstly use a KG embedding model to complete the data layer.By completing facts of the data layer,axioms mining algorithm can reduce biased estimation of the confidence.At present,the KG embedding model has achieved some success in the fact completion.However,most KG embedding approaches ignore the local structured information.How to introduce the structural information in the KG embedding model and improve the data completion of complex relationship are the important problem.In addition,when building ontology,statistical methods introduce a lot of noise in building negative examples,which affects the quality of ontology.The challenge we are facing is how to obtain high-quality negative examples and design a novel axioms mining algorithms that conform to open world assumption.Finally,the ontology constructed by axiomatic mining is still incomplete.Traditional KG embedding models often neglect the logical attributes of relationships,such as transitivity and symmetry,which makes such models unable to complete the axioms in the schematic layer.Another challenge is how to design a KG embedding model to complement the ontology.To better accomplish the above challenges,this paper mainly conducts the following researches:1)With respect to fact completion,a new method based on KG embedding model named TCE is proposed.In this model,TCE defines a triple context of each triple.The triples context contains the local structureinformation,which enables TCE to effectively handle complex relationships.At the same time,TCE pro-poses a new score function,which can calculate the probability of the triples by given the correspondingtriple context.The experimental results show that TCE can effectively complete factual knowledge afteradding the context.2)With respect to axioms mining,a novel mining algorithm named SIFS(Schema Induction From Incom-plete Semantic Data)is proposed to automatically obtain disjointness and subclass axioms.This frame-work firstly obtains probabilistic type assertions by exploiting a type inference algorithm.A mining ap-proach based on association rule mining is proposed to learn high-quality schema information.To addressthe incompleteness problem of semantic data,the mining model introduces new support and confidencefor pruning false axioms.3)With respect to axioms completion,a novel ontology representation model named CosE is proposed toautomatically complete ontologies.This model projects each concept into two semantic spaces.One isan angle-based semantic space that is utilized to preserve transitivity or symmetry of an axiom.The otheris a translation-based semantic space that is utilized to measure the confidence score of an axiom.Weconduct extensive experiments on link prediction on benchmark datasets like YAGO and FMA ontologies.Experimental results indicate that our method outperforms state-of-the-art methods.