Knowledge Graph Link Prediction Based On Model Embedding And Rule Inference

Link prediction is one of the important directions and difficulties in knowledge mapping research.The link prediction method based on model embedding almost completely depends on triple data,which is easy to operate,but because good embedding depends on rich data,it is difficult to learn a good representation of sparse entities.The link prediction method based on logical rules is highly accurate and interpretable,but the huge search space in the process of logical reasoning makes its computational burden high.Therefore,this paper combines the advantages of the above two methods and proposes a knowledge map link prediction algorithm based on the combination of model embedding and rule reasoning.The main research work of the paper includes:(1)To solve the problem that model-based embedding methods rely on the richness of triple data and rule-based reasoning algorithms have high computational burden,this paper proposes a rule induction and model embedding iteration(RIMEI)algorithm,which includes the credibility calculation stage and the model training stage.Before reaching the specified number of iterations,the credibility calculation phase calculates the credibility of the quasi triplet,and then predicts a new quasi triplet.In the training phase of the model,trusted triples are added to the knowledge map,trained together with known triples,and modified the vector learning of entities and relationships.Through this iterative process,the knowledge contained in the logical rules can be better transferred to the learned embedding.(2)Aiming at the problem that the triplet inferred by rules is not necessarily correct,and the existing methods cannot deal with the uncertainty of logical rules,this paper proposes an Embedded Model Combined with Markov Logical Network(EMCMLN).Based on RIMEI,the algorithm uses Markov logic network to model the distribution of all possible triples,learn the joint distribution of these triples,and use variational EM algorithm for effective training to improve the accuracy of link prediction.(3)Build a TCM ancient book link prediction platform with entity prediction and relationship prediction functions,apply the RIMEI and EMCMLN algorithms proposed above to the actual TCM ancient book data,and effectively improve the efficiency of data mining for TCM ancient books.