Research On Model Of Hyper-relational Knowledge Graph For Link Prediction

Nowadays,the Knowledge Graph(KG)has also got into a rapid development stage.However,the issue of data integrity in the KG restricts research and development in many fields.The binary-relational knowledge graph is more common in the KG but is limited by less information,it can only rely on the triple structure to predict missing entities and relations.The hyper-relational knowledge graph is another form of the KG,which introduces much supplementary information based on the binary-relational knowledge graph and can effectively improve prediction accuracy.Therefore,the hyper-relational knowledge graph is of great significance in solving the data integrity problem in the KG.The existing hyper-relational link prediction models only consider the overall perspective when dealing with additional information(hyper-relation or supplementary information)and calculate the score function by combining the additional information and the main triple.These methods ignore the inherent characteristics of entity and relation in hyper-relational data.Even worse,traditional methods ignore the similarity between the main triple and additional information when aggregating additional information.These methods may result in combining additional information with low predictive correlation with the main triple.However,the contribution of these additional information to the prediction results is tiny.In addition,the additional entity and relation embedding often exist independently and lack correlation,which leads to poor prediction performance of the model.To solve the above problems,the following work content has been performed in this thesis:(1)LGHAE: Local and Global Hyper-relation Aggregation Embedding for Link Prediction is proposed.The LGHAE can capture the semantic features of hyper-relational data from local and global perspectives.To fully utilize local and global features,the Hyper-Interact E is designed to predict missing entities.(2)HIAE: Hyper-Relational Interaction Aware Embedding for Link Prediction is proposed.The HIAE utilizes attention networks to calculate the semantic similarity between the main triple and additional information.To solve the problem of low interactivity between additional information,a new Hyper-relational interaction aggregation method is designed.In addition,the HIAE converts the hyper-relational link prediction task into the binary-relational link prediction task,which is suitable for multiple binary-relational link prediction models.Finally,experimental comparisons are conducted with baselines on public datasets such as Wiki People,JF17 K,and WD50 K,using commonly used evaluation indicators in the field of link prediction-Mean Reciprocal Rank and Hit Rate.Through analysis of experimental results,it is shown that both LGHAE and HIAE models achieved better performance.