Research On Temporal Knowledge Graph Completion Algorithm Based On Knowledge Representation Learning

In recent years,knowledge graph has attracted more and more attention from academia and industry and has become one of the basic technologies of artificial intelligence research.It represents a large number of events in the real world as structured facts.It is an effective method to store massive data and is applied in many tasks.However,data sparsity is a common problem in knowledge graph.In order to improve the completion of knowledge graph,many knowledge representation learning algorithms have emerged in recent years to infer new facts in knowledge graph and this task is also called knowledge graph completion.However,the existing researches mainly focus on static knowledge graph and ignore the information about when the facts happened.Actually,time is an important property that most facts have,and different events or actions can cause triples to change.Therefore,the research on temporal knowledge graph becomes more and more important.In recent years,some temporal knowledge graph completion models show improved results by further integrating the temporal information of facts.Due to the instability and sparsity of data,as well as the complex temporal dependence,the completion of temporal knowledge graph is a challenging problem.The key point is that how to effectively utilize time information or how to model the dynamic evolution characteristics of facts.In view of these problems,the thesis carries on the research along two aspects,and the main work is as follows:Firstly,in order to better capture the evolution process of entities over time,the thesis proposes a new embedding algorithm for temporal knowledge graph based on quaternion rotation,which defines the evolution process of entities over time as temporal rotation transformation in quaternion space.Potential interdependencies can be captured better by Hamilton product in quaternion space than by Hermitian inner product in complex space,so the learning process is more efficient and expressive.Experimental results on benchmark datasets show that this model achieves relatively excellent performance.Secondly,most of the existing work on temporal knowledge graph completion is an extension of static knowledge graph completion,which fails to make full use of the multihop structure and the temporal dependence of the temporal knowledge graph.In order to make full use of these two characteristics to improve the performance of the model,the thesis proposes a joint encoder based on attention mechanism,which is made up of structural attention encoder and temporal attention encoder.In order to model the evolution characteristics of temporal knowledge graph in many facets,the thesis also introduces soft gate to adjust the importance differences of different evolutionary facets.Experimental results show that the performance of this model is better than the existing mainstream methods by making full use of the structural characteristics of temporal knowledge graph.