Research And Implementation Of Knowledge Graph Entity Typing Based On Meta-learning

The prosperity and development of the internet over the past decades has spawned many large-scale knowledge bases.These knowledge bases usually contain ontology views knowledge graph composed of abstract concepts and instance views knowledge graph composed of entities.Both views contain a large number of triples in the form of(head entity/type,relation,tail entity/type),cross-view links from instance to ontology view are also included to indicate the type of each entity.There are often some problems related to cross-view links,such as incompleteness,only a small number of entities have type information.Another example is the long-tail distribution problem,which means that the number of entity types is large at low frequency and small at high frequency.The existence of these problems often leads to the loss of many important knowledge in the knowledge graphs,which affects its effectiveness in downstream task.Knowledge graph entity typing,which aims to predict the missing type of each entity,has attracted great attention in recent years.However,previous studies assumed that each type has a large number of corresponding entities,ignored the long-tail problem in type information,and failed to make fully use of the structural information in the two views.To solve the above problems,this thesis designs and implements a knowledge graph entity typing framework based on meta-learning,which can solve the entity typing problem under the few-shot scenario.The specific contents include the following three parts:(1)Research and implement a knowledge graph representation method based on graph neural network.Inspired by the factorization machine algorithm,this thesis proposes a new graph neural network algorithm for multi-relational data,which can explicitly model the interactive information among relations with different types.(2)Construct training rules based on contrastive learning.In order to make fully use of the labeled samples under the fewshot scenario,this thesis proposes training methods based on contrastive learning,and implements two different contrastive learning rules for intertypes and intra-types.(3)Construct the knowledge graph entity typing framework based on meta-learning methods.The ontology view knowledge graph is constructed in the form of tree,which contains a large amount of hierarchical information.In this thesis,three information extraction rules are constructed according to hierarchical information for the meta-training stage,and in the meta-testing stage,the model can fast adapt to new tasks.This framework can solve the few-shot problem of entity typing based on the Model-Agnostic Meta-Learning algorithm under few-shot scenario.Experimental results show that the knowledge graph entity typing method based on meta-learning proposed in this thesis can effectively extract type related information from the knowledge graph,so as to help the problem of entity typing under few-shot scenario and improve the accuracy of entity typing.