Temporal Knowledge Graph Inference Based On Copy-Generation Model And Its Application

Knowledge graphs contain a large amount of human knowledge and have a wide application in tasks such as search engines,self-questioning,machine translation,and recommendation systems.Knowledge graphs usually contain dynamic facts in the temporal dimension,which model the dynamic relationships or interactions of entities along a temporal line.Recent research has started to incorporate temporal information into knowledge graphs,i.e.,temporal knowledge graphs.Since such temporal knowledge graphs often suffer from incompleteness,it is important to develop time-aware inference models for temporal knowledge graphs that help infer the missing temporal facts in such graphs.Recently,temporal knowledge graph inference has become a hot topic with a wide range of applications in Social Network Analysis,Event Prediction,Recommender Systems,Intent Recognition,and other fields.Although facts on the temporal knowledge graph are usually in constant change,it is noteworthy that many of them can recur in history.In this thesis,we conduct a study on the inference of temporal knowledge graphs using the public dataset of temporal knowledge graphs.A novel temporal knowledge graph inference method based on a timeaware copy-generation model is proposed to address the problem that the temporal knowledge graph embedding method focuses only on calculating the latent representations of each timestamp knowledge graph separately leading to the inability to capture the long-term dependencies of facts in serial timestamp.In this paper,we propose CyGNet(Temporal Copy-Generation Network)by combining the idea of implementing the copy mechanism in natural language generation.Extensive experimental results demonstrate that CyGNet improves the prediction accuracy and the ability to predict future events.For the activity data features of naval and air ship reconnaissance aircraft,etc.,CyGNet is applied to dynamic target intent inference.First,the military activity knowledge atlas data already constructed based on a project is randomly cut into training data and test data,and entity linking is performed by linking a large number of entities with different statements of the same target or the same region in the training data and test data,and the graph neural network model CompGCN is used to learn the semantic vector representation of each node in the quaternion to obtain the pre-trained node vector representation by the trained semantic vectors to initialize the CyGNet quadruplet model.The CyGNet has a hits@10 value of 91.81% for dynamic target intent inference,which is a practical guide for military activities.However,the performance of the algorithm may be hindered by the problem of having too unbalanced probabilities of historical repeated events in the Dataset.In future work,we plans to attempt to learn logical inference between events of different time slices by pre-training the temporal knowledge graph of globally salient entities and events.The proposed algorithmic techniques will also be used to help understand dynamic target intent inference for military activities.