Research On Domain Knowledge Graph Completion Technology Based On Meta-learning

With the continuous study of knowledge description,knowledge graph has become a hot topic in the field of artificial intelligence.As an organizational form of industry knowledge,domain knowledge graphs have high demands on specialization and accuracy of data,while due to the errors in the construction process,the knowledge in the existing domain knowledge graphs is often incomplete.To provide more accurate data support to the upper applications,knowledge graphs need to be completed,but domain knowledge graph often contains knowledge triples with low occurrence frequency,and traditional completion methods could not achieve beneficial results on such triples.There is a need to develop a knowledge graph completion method that can not only capture complex semantic information,but also perform well under the few-shot conditions.Based on the research of meta-learning method and knowledge representation learning method,this thesis aims to build a domain knowledge graph completion system,and studies the domain knowledge graph completion method in the case of few-shot condition.In summary,main contributions of the study are:(1)A presentation learning task construction under the meta-learning framework is developed.A meta-learning-based knowledge graph completion algorithm is proposed,which can obtain more accurate distributed knowledge representation under the condition of few-shot triples.The metarelation vector generated by the algorithm can adapt to the new task quickly and reduce the training cost.(2)A single step descent optimization method and the first-order optimization method for the learning and training process of meta-knowledge representation are proposed,which realize the improvement of computing efficiency when the algorithm is applied.(3)A domain knowledge graph completion system is designed and implemented,which can perform distributed representation generation and missing entity recommendation for domain knowledge graph.