Type-Augmented Link Prediction Based On Bayesian Formula

Nowadays,artificial intelligence has developed into a national strategy,and knowledge graph help artificial intelligence systems better understand the world.Because knowledge graph contains rich information,knowledge graph has been widely used in many downstream natural language processing applications,including question answering,machine reading,dialogue system and semantic search.Entity missing is an important problem in knowledge graph,which may lead to incomplete or inaccurate knowledge in knowledge graph,thus affecting the application effect of knowledge graph and may lead to reasoning errors in knowledge graph.In order to solve the above problems,many studies focus on link prediction tasks.The task attempts to infer the missing links of a given knowledge graph and predict all possible candidate entities.Most of the existing work is proposed by maximizing the possibility of observed instance-level triples.There is not much research on type information,and there is still much room for improvement.Aiming at the problem of insufficient application of type information,a new link prediction method,called Type-Augmented Link Prediction(TALP),is proposed in this paper.TALP is divided into two parts : prior model and likelihood model.Firstly,TALP proposes a prior model,which encodes the type information into a prior probability according to the hierarchy,and sets a training matrix in the prior model,so that the type is specific under different relationships.Secondly,TALP improves the type enhancement of the Trans E model in the likelihood model part,embeds the multiple semantics of the entity and feeds the semantic information back to the prior model,which makes up for the difficulty of multi-semantic processing of the entity in the Trans E model to a certain extent,thus improving the prediction results of the overall model,so that the likelihood model can be better combined with the prior model for prediction.In addition,the likelihood model part of TALP also uses the convolutionbased Interact E model and the graph attention model GAT.Finally,TALP calculates the results of the prior model and the likelihood model according to Bayesian rules to obtain the final link prediction results.The experimental results show that the proposed TALP method is significantly better than the method using the likelihood model alone on the FB15k-237,YAGO26K-906,DB111K-174 and other data sets.Especially on the DB111K-174 data set,all kinds of likelihood models spliced according to the TALP method have more than 6 % improvement.This shows that the TALP method can make full use of type information and make up for the shortcomings of the likelihood model ignoring type information.At the same time,the experimental results also show that the TALP method is significantly better than the current optimal model using type information.And it can improve the representation ability and generalization ability of the model,so as to achieve better performance in the knowledge graph link prediction task.