Research On Knowledge Graph Completion Based On Relation Path

The knowledge graph is a valuable resource for building applications such as intelligent question answering and recommendation systems.The knowledge graph is actually a collection of fact triples,each triple includes a head entity,a relationship,and a tail entity.At present,many large-scale knowledge graphs are relatively sparse.The incompleteness of the knowledge graph affects the performance of their downstream applications.The task of knowledge graph completion is to add new triples based on the existing information in the knowledge graph.The knowledge graph completion methods based on knowledge graph embedding utilize the structural information of triples,and have achieved good results in recent years.However,these methods only pay attention to the direct relationship between entities,and ignores the existence of relationship paths in the knowledge graph.The path-based knowledge graph completion methods infer the relationship between entities according to the relationship paths.The early path-based method PRA uses each path between entities as atomic features,and then trains a binary classifier to infer the relationship between entities.This method is difficult to handle large knowledge graphs.Subsequent Path-RNN and Single-Model models use recurrent neural networks to represent the semantic information of relational paths as vectors,which improves the scalability of the model.This paper improves the shortcomings of the Single-Model model.The specific work is as follows:(1)The RNN used by the Single-Model model cannot solve the long-term dependency problem,so it cannot generate better vector representation when dealing with long relationship paths.In addition,RNN has the problem of gradient vanishing and gradient explosion,which makes the model difficult to train.So we use LSTM network instead of RNN.(2)Single-Model uses the type of entity to replace the entity in the relationship path,which will cause the loss of path information.We use the text auxiliary information provided in the data set,combined with the structural information of the triples in the knowledge graph,to pre-train the embedding representations of entities and relationships,and use them in the vector generation of the relation path.(3)The Single-Model model does not take into account that the amount of information provided by each path between entity pairs is different for a specific prediction relationship.Based on this,we introduce attention mechanism to assign different weights to each path.The information of the paths is fused to obtain a vector representation of an entity pair,and then this vector is used to infer the relationship between the entity pairs.This paper evaluates the proposed model on the validation data set.The experiment shows that our model has better performance than the benchmark model on the triple classification and link prediction task.