A Study On Systems And Applications Of Scalable Knowledge Graph Serving

These years have witnessed the great progress on Artificial Intelligence(AI)tech-niques.Knowledge graph techniques,as one of the main field in AI,are also experi-enced a rapid development.Due to the simplicity and flexibility of this representation,a large number of knowledge repositories are emerging for managing the knowledge from many fields.The size of the available knowledge built by human experts or extracted from large text corpora reaches an unprecedented scale.These knowledge bases,along with other open data sets,are further interlinked with each other,yielding tens of bil-lions of facts.Despite the flexibility of knowledge graph,how to serve big knowledge graphs in real-time is a challenging task.In this dissertation,the research work on this topic,ranging from systems to applications,will be presented as follows:1.This dissertation will first introduce the existing graph systems which are pro-posed to manage and serve general graphs and knowledge graphs that scale to tens of billions of nodes.Graph data are inherently hard to manage and process due to the heterogeneous structures and rich connections,the problems become even more serious when the graph data are large in sizes.As a special kind of graph,knowledge graph introduces more challenges for serving efficiently.This disser-tation will first demonstrate the challenges for massive graph data and knowledge graphs,and then this dissertation summarizes the design principles for graph sys-tems and RDF stores.2.This dissertation will present our strongly-typed RDF store named Stylus for serv-ing massive knowledge data.This dissertation argues that a strongly-typed stor-age scheme is the key to serve massive knowledge data efficiently.However,it is non-trivial to design a type system for knowledge graphs due to the flexibility of its representation.Stylus exploits a strongly-typed storage scheme to enable extremely efficient management of massive knowledge data.The defined types are derived from served data,meanwhile,a dedicated structure is proposed for data modeling.On top of this storage scheme which is equipped with a subgraph based query processing strategy,Stylus is able to serve large-scale knowledge graphs in nearly real time.3.For the task of relation finding in knowledge graph,semantic navigation meth-ods are proposed to find the paths connecting the given entities in an efficient and effective way.These methods leverage distributed representations of knowledge graphs to guide the navigation for short and meaningful paths.This disserta-tion exploits machine learning techniques to reduce the complexity of the search space.A few semantic navigation methods are proposed for this task.This disser-tation first introduces a similarity-based navigation method which is straightfor-ward.Then,a neurally-guided navigation method is proposed which is based on a pre-trained value net to improve the performance.To tackle the issue of meeting dead ends of these navigation methods,a method based on tailored Monte Carlo tree search is further proposed.In experiments,these methods are proved to be both efficient and effective.