Data-augmented Knowledge Graph Embedding With Filtered Rules

Knowledge graphs are now widely used in personalized recommendation systems,intelligent question and answer systems,etc.Since the completeness of knowledge graphs greatly affects our subsequent use of knowledge graphs,and the factual information in knowledge graphs is often incomplete,and some obvious relationships between entities do not exist in the original knowledge graphs,we need to complement the knowledge graphs.Among the methods for knowledge graph complementation,representation learning methods occupy the mainstream position,and among the representation learning methods,translation model is one of the representative models,and the representation learning methods based on translation model are highly scalable and can effectively evaluate the complex semantic information and relationships among entities.The richness of data is crucial for representation learning models,but there are often a large number of sparse entities in existing knowledge graphs,and the connections between entities are not strong,which leads to the inability of representation learning models to accurately complement the triples with sparse entities.Traditional approaches alleviate this problem by combining a rule-based learning approach with a representation learning approach,but require that the representation learning model needs to satisfy the linear map assumption,which is not satisfied by the translation model.To solve this problem,we propose to use a bottom-up rule learning algorithm to combine with a translation model to enhance the representational power of the translation model,i.e.,by using a rule learning algorithm to generate augmented data to enhance the translation model.When we use this bottom-up rule learning algorithm,the confidence level of the generated rules is defined differently.When we infer triples,most of the triples generated by high confidence rules already exist in the original knowledge graph,and only a few of the new triples are valid triples,and when we perform data augmentation,we need to ensure the number of augmented data.This problem makes it impossible to simply specify a rule confidence threshold to filter rules and generate new triples.To solve this problem,we propose a rule filtering method based on rule confidence.Also,since the existence of sparse entities and sparse relations in the knowledge graph causes the representation learning model to not learn well enough for triples containing sparse entities or sparse relations,we propose another approach that focuses on enhancing triples containing sparse entities or sparse relations.The main contributions of this paper are as follows:(1)To enhance the accuracy of link prediction of the translation model,a bottomup rule learning algorithm is proposed to generate high quality rules and an intermediate module is used to infer reliable triples to enhance the expressiveness of the translation model.(2)Given the definition of rule confidence in the bottom-up rule learning algorithm and the requirement of data quantity for data augmentation,in order to achieve a balance between the quality and quantity of new triples generated,we improve the traditional method by not simply specifying a confidence threshold to filter rules as in the traditional method,but by traversing rules in descending order according to their confidence.(3)Sparse relations also exist in the knowledge graph,and the existence of sparse relations makes the traditional model for link prediction of sparse relations unsatisfactory.We enhance the triples containing sparse relations on the basis of the existing methods focusing on enhancing the triples containing sparse entities at the same time.In this paper,we apply the two improved methods proposed in this paper on the WN18,FB15 k,WN18RR,FB15k-237 datasets and multiple models,and compare them with multiple models without these two methods.The results show that these two methods outperform the original models overall under the three types of link prediction evaluation metrics,and enhance the expression of the translation models to some extent.