Research On Completion Algorithm Of Temporal Knowledge Graphs Based On Representation Learning

Because the knowledge graphs is inherently incomplete.The sparse knowledge graphs is not only a defect for the graphs itself,but also severely restricts the subsequent calculation and reasoning on the graphs.Therefore,we need to complete the existing graphs.The scalability and computational efficiency of the completion method based on knowledge representation learning are more in line with the needs of the current task,so more research is focused on this kind of completion algorithm.The model representing learning needs to create an embedding vector for each entity and relationship.Compared with the method of graph structure,the method based on sequence learning can reduce the model by more than 50%,and the computational efficiency has been greatly improved.In the past,the complement of temporal knowledge graphs based on sequence learning only focused on the relationship between the structures of quads,and did not focus on the internal logic.In addition,according to the knowledge graphs data set ICEWS,more than 80% of the events occurred more than once in the 24 years from 1995 to 2019.The existence of this regularity highlights that we should use the facts we have mastered to reasonably complete the knowledge graphs.Based on the above reasons,this paper designs a temporal knowledge graphs completion model based on representation learning.The main work and innovations are as follows:(1)In the representation learning of quaternion,the traditional sequence model adopts the joint coding mechanism of time and relationship,ignoring the interaction between timestamp and entities and relationships,and introduces the combination of attention mechanism and cyclic neural network into the inner of its own quaternion to obtain the semantic vector fused with context representation,making semantic understanding more sufficient.(2)A sequence completion model based on Encoder-Decoder is proposed.Compared with the traditional seq2 seq model,we choose the structure applicable to the domain of knowledge graphs,which reduces the model volume on the premise of complete functions of each part.This lightweight model reflects high application value when facing the huge data volume of knowledge graphs.In addition,based on the characteristics of easy expansion of the model,this paper proposes the idea of completing the spatiotemporal knowledge graphs.(3)A model combining historical knowledge frequency is proposed,and knowledge frequency is introduced into the completion task for the first time.At the same time of completing the knowledge graphs,the existing knowledge can be used to predict the regular large-scale crisis events that may occur in the future.This work has great practical significance.Early prediction of the crisis can be deployed in time to deal with the crisis and minimize social losses.The model carries out the ablation and contrast experiments of the knowledge map completion task on the YAGO,WIKI,GDELT and ICEWS18 data sets respectively,which confirms that the enhancement of knowledge representation and the use of historical knowledge frequency can improve the level of completion ability.