ENHANCING KNOWLEDGE GRAPH COMPLETION USING TEXTUAL CONTENT

Abstract

SeveralKnowledge graphs likeDBpedai, Freebase, Wordnet and others are far fromcomplete. Thus, Knowledge Graph Completionisataskwhich has a main objective to completethese graphs withmissingknowledge. Everyknowledge graph is a set of triples like “AlgierscapitalOfAlgeria” where the first is the subjectentity, the second is the relation and the last is the objectentity. The principal tasks are the linkprediction and the relation classification where the former predict the relation betweentwogivenentities and the later classifies given triples withtrue or false. Approachesthat use only the observed triples cangive best results but theyfail in case of unseenentitiesbecause the predictionmodels have trainedwithonlyexisting triples. Therefore, new directions have been proposed to solvethisproblem and the main isusing an externalresourceliketext, becausethislater have rich contents.