Research On Embedding Technology For Temporal Knowledge Graph

With the continuous increase of data knowledge,a large amount of temporary knowledge data emerges with complex temporal characteristics,and the conventional static knowledge graph can no longer meet the urgent demand for modeling knowledge with temporal characteristics.Therefore,the temporal knowledge graph came into being.The temporal knowledge graph takes temporal information as the constraint of facts and highlights the relevant timeliness of facts,and this description is more in line with the pattern of real-world events.In the temporal knowledge graph,for entities with multiple answers to the same relationship,precise querying can be achieved based on temporal conditions,avoiding the 1-toN phenomenon.However,the missing information generated in the fact inclusion process leads to the reduction of valid facts and the decrease of answer accuracy.On the other hand,there is a strong correlation among temporal information in the temporal knowledge graph,and it is impossible to effectively distinguish between entities,relationships,and temporal information if they are embedded in the same dimension.Based on the above reasons,the research on embedding techniques for temporal knowledge graphs is carried out,and its main research contents are as follows.A review study is done for the existing static embedding models and temporal embedding models,and the main problems of the current stage knowledge graph embedding models are identified by analyzing the main solution methods of each type of model.Meanwhile,the problems that need to be solved are identified for the characteristics of temporal knowledge graph data.A temporal knowledge graph embedding model(Spherical Temporal Knowledge Graph Embedding,STKE)in a spherical coordinate system is proposed for the problem of not being able to distinguish entities,relations,and time well in the temporal knowledge graph.The model embeds facts with temporal information constraints into a spherical coordinate system and treats entities as spatial vectors from the origin in a spherical coordinate space,with each entity containing the radial part,the azimuth part,and the polar part.In this case,the radial part aims to calculate the modal length of each entity vector.The azimuth part is used to distinguish entities with the same modal length,and the polar part is intended to measure the shift in temporal information.The above-mentioned model STKE is improved by using the attention-enhanced mechanism(Compositive Graph Attention Network,Comp GAT).The powerful learning ability and expressiveness of graph neural networks are utilized to capture the relationship between target entities and neighbors,and the information is aggregated into each target entity to enhance the embedding representation of target entities.Meanwhile,the attention-enhanced mechanism is utilized to assign attention coefficients to neighbors.Finally,an attentionenhanced temporal knowledge graph embedding model(Compositive Spherical Temporal Knowledge Graph Embedding,Comp STKE)is obtained.The embedding ability of the models is evaluated by four typical temporal datasets.Using the link prediction task as the benchmark,STKE and Comp STKE slightly outperform the benchmark model in different aspects by comparing with the mainstream static knowledge graph embedding model and the temporal knowledge graph embedding model.In addition,the effectiveness of different components of this model is analyzed by ablation experiments to prove the importance of each module of this model and to verify the feasibility of this scheme.