Rule Learning And Reasoning On Uncertain Knowledge Graph

In recent years,knowledge graphs have played an increasingly important role in many knowledge-driven artificial intelligence applications.However,with the increasing scale of knowledge graphs,the way of manually constructing graphs has been difficult to meet the demand.Many automatic graph construction steps,such as entity linking and relationship extraction,have been introduced into the graph construction process.Compared with manual labeling,the correctness of these methods is difficult to guarantee,which makes the data in graphs uncertain.Uncertain knowledge graph describes this uncertainty by adding confidence to triple to represent the probability that a fact is true.Rule learning and reasoning methods play an irreplaceable role in many applications because of their strong interpretability.However,the current rule learning and reasoning methods are usually designed for the deterministic knowledge graph,without considering noise and uncertainty,which leads to the low quality of the rules learned in the uncertain knowledge graph,and can’t complete the reasoning of the graph well.In order to solve the above problems,this thesis proposes a novel uncertain knowledge graph rule learning and reasoning method.This method is based on uncertain knowledge graph representation learning.Firstly,the structural semantic information,confidence information and relational logic information in uncertain knowledge graph are embedded into the representation vector.Then,in the process of rule learning,the representation vector is used to assist rule generation,so as to reduce the interference of uncertainty on rule learning.In rule reasoning,the application range of rules is described by learning the variable feature vector of rules,so that the matching degree between rules and instances is also taken into account in rule reasoning.The main contributions of this thesis are as follows:(1)An uncertain knowledge graph representation learning and rule coding evaluation method is proposed.Through this method,the representation vector of entity and relation can embed the structural semantic information,confidence information and logical properties of relation of uncertain knowledge graph at the same time,which can be used to support the subsequent rule learning and reasoning process.In addition,this thesis also proposes a rule coding and evaluation method based on attention mechanism,which expresses the semantics of vector coding rules through attention weight and relation,and scores the quality of rules to guide rule learning.(2)An iterative uncertain knowledge graph rule learning and reasoning method is proposed.The method mainly consists of three modules: firstly,the rule generator generates candidate logic rules based on the representation vector of uncertain knowledge graph? Then,the rule variable learner learns the variable feature vector of the rule by sampling and induction,which is used to describe the applicable scope of the rule? Finally,the rule reasoner makes use of the rule and the variable feature vector of the rule to complete the inference of the graph,and selects high-quality rules from the inference completion for iterative training of the rule generator.(3)The proposed method is implemented and verified by experiments on four uncertain knowledge graph data sets.Experimental results show that the proposed method is superior to the existing related methods on four data sets.To study the learning and reasoning method of uncertain knowledge graph rules is to explore the logical rule mining and reasoning completion in the graph with noise and confidence description.From the practical level,this research can effectively support many applications that require high interpretability and accuracy,such as military decision-making,medical research and development,etc.From the theoretical level,this study is also a useful exploration of symbol-neural joint reasoning method.