Research On Hyponymic Relations Recongnition Based On Representation Learning

For various purposes,knowledge bases in different fields are constructed.However,due to different needs and different ways of constructing knowledge bases,knowledge in knowledge bases is loosely organized and multi-source heterogeneous.Therefore,it is significant to construct ontology,a tool that can describe d omain concepts structurally,for knowledge reuse and information extraction.Because domain ontology construction is a relatively complex project,which requires a large amount of human participation,this paper aims to use automated methods to help recognize hyponymy relations in the ontology construction process to save human resources.As information is growing,new entities without class labels will be added to the existing knowledge base,while for new knowledge bases,the categories of entities are unknown.Recognizing hyponymy relations is such a problem that how to divide the unknown-classed entities.This paper mainly deals with the problem of recognizing hyponymy relations in the medical field.Therefore,the required knowledge is extracted from the medical knowledge base SemMedDB,and uses the corresponding medical domain ontology UMLS to verify the identified hyponymy relations.The key to recognizing hyponymy relations is how to learn the characteristics of entities effectively.We can find that knowledge can be represented in the form of multi-relational graph.In this paper,the entities' feature vectors are obtained by using the heterogeneous graph embedding algorithm,using the knowledge graph representation learning algorithm and directly using the relationships as entities' features to construct vectors.After learning entities' feature vectors,we classify and cluster the entities.The results of classification and clustering can be used as guidance to divide unknown-classed entities and assist the recognition of hyponymy relations.Visualization results of entities' feature vectors,which is learned by knowledge graph representation learning algorithm,shows that in vector space,the feature vectors of the same kind of entities are close,and there are obvious boundaries between different types of entities.Meanwhile,the feature vectors,which are concatenated by knowledge graph representation learning vectors and the vectors using relationship as features,can class entities effectively and classification accuracy can reach 71.2% when there are 122 classes.Experiments show that feature vectors learned by knowledge graph completion models and relations as entities' features can act as entities' features to recognize hyponymy relations.