Critical Algorithm Studies On Large-scale Knowledge Graph Completion

In recent years,with the booming of the Semantic Web data,knowledge graphs based on the framework of graph structure have attracted widespread attention of both current academy and industry.A knowledge graph usually contains substantial structure information,stores tens of millions of facts,and covers massive practical entities and relations.It provides the reliable sources of information and the underlying supports for many technologies of artificial intelligences,intelligentize a large number of real-life applications.However,the coverage of such knowledge graph is still far from complete compared with real-world knowledge.With the goal of knowledge graph completion,it is an effective way of extracting structure information from unstructured and semi-structured data by natural language processing.However,the process of extraction needs many heuristic rules made by much domain expert knowledge and many participations of human being to make sure of extraction quality.At the same time,artificial intelligence,especially the field of machine learning,gets rapid developments.Researchers expect to exploit the machine learning models to achieve the goal of knowledge reasoning and fusion based on the existed facts.It has become one of the hottest research areas in natural language processing,and it is also the same theme for our work.In this dissertation,we introduce two mainly kinds of knowledge graph completion algorithm in detail: path ranking algorithm based on graph feature and knowledge representation learning based on latent feature.Path ranking algorithm can use the random walks to find common relation paths as explicit features.Knowledge representation learning,based on the framework of representation learning,can better tackle the problems of inherent semantic information in knowledge graph and the scale needs.Building upon these works,we propose a new relation path embedding model.Our model incorporates the rich semantics of relation paths into the framework of knowledge representation learning.Based on the semantic similarity between relations and reliable relation paths,we extend the relation-specific type constraints to novel path-specific type constraints.The Path projection enables entities to learn their low-dimensional representations in different latent spaces simultaneously;the path constraints can better distinguish similar embeddings in the latent space,which improve the discriminability of our method.With the goal of achieving the better algorithm performances,both of these innovations can be incorporated into many knowledge representation learning models seamlessly.The proposed model is evaluated on three common datasets of two benchmark tasks of link prediction and triple classification.The results of experiments demonstrate our method outperforms all baselines on both tasks.They indicate that our model is capable of catching the semantics of relation paths which is significant for knowledge representation learning.