Research And Implementation Of Temporal Knowledge Graph Completion And Forecasting Algorithms

This paper mainly studies the temporal knowledge graph(TKG)completion and forecasting algorithms and the application of the completion algorithm in temporal question answering.The task of temporal knowledge graph completion is to complete the missing facts of the graph,while the task of forecasting is to predict the future events of the graph.The task of temporal question answering is to find the answer of temporal natural language query by using the completion algorithm from temporal knowledge graph.At present,there are many problems in the research of temporal knowledge graph:(1)For the task of temporal knowledge graph completion,the existing models only take time as a parameter,ignore the corelations in time evolution,and don’t consider how to fill the newly emerged timestamp,so it is not capable of online learning.(2)For the task of temporal knowledge graph forecasting,the existing work is independent completion task or forecasting task,without considering the relationship between the two tasks;(3)For the task of temporal knowledge graph question answering task,the difficulty lies in finding the implied target time of complex temporal questions.The existing models still have sufficient space to improve the performance for complex questions.In view of the above problems,the main research work of this paper is as follows:(1)The temporal knowledge graph completion algorithm RTFE based on state transition is proposed:the temporal knowledge graph is regarded as a Markov chain.Based on RTFE,the rich static knowledge graph embedding methods are transferred to temporal knowledge graph embedding,which bridges the gap between static graphs and temporal graphs.Besides,RTFE further enhances the effect of the existing temporal knowledge graph embedding models.In the open data set,RTFE improved the MRR metric(mean reciprocal rank)by 6.2%over the previous optimal model(2)The event forecasting algorithm CompTF based on temporal knowledge graph completion is proposed:a training algorithm of"completing the temporal knowledge graph first and then predicting the future".The search space is reduced according to the periodicity of events,and the discriminative completion model is directly used to complete missing events.Then,the completed graph is used to train the forecasting model to improve the prediction ability on the uncomplete graph and the complete graph.(3)To explore the further application of temporal knowledge graph completion algorithm in temporal question answering,the complex question answering algorithm Time-trans over temporal knowledge graph based on time transformation is proposed.To solve the challenge of finding the target time of complex temporal,a neural network is used to transform the possible potential time to the target time.Then the temporal knowledge graph completion model is directly used to answer the transformed questions,which not only improves the accuracy but also simplifies the solving process of temporal questions.Time-trans improves the accuracy by 13.1%compared with the previous optimal model on the complex questions of open data setFinally,this work supports the research of‘Knowledge graph representation model combining structural and temporal information’ in the national Natural Science Foundation of China(NSFC)project’Research on knowledge graph fusion and knowledge graph completion of Graph Neural Network for medical knowledge graph’.