Research Of Entity-connectivity-difference-aware And Target-oriented Attention Network

As the cornerstone of the development of artificial intelligence in the current society,knowledge graphs play a very important role.With the development of data scale and automation technology,the scale of knowledge graph has grown pretty fast,but it is still far from the ideal goal of knowledge-complete.In order to achieve this goal,the link prediction task came into being,which aims to infer the missing facts through the existing knowledge in the knowledge graph,so as to perfect and complement the knowledge graph.Researchers have done a lot of research on this task.In the past two years,the method based on representation learning has attracted wide attention from scholars at home and abroad because of its high efficiency and strong ability.After investigation,we found that although the knowledge graph embedding model has achieved good results,most of the existing models ignore the connectivity differences of the entity itself and are relatively rough in the implementation of the attention strategy,resulting in the network's ability to aggregate neighborhood information and the efficiency has relatively large limitations,so the model's predictive ability is poor.To solve these problems mentioned above,we made improvements on the basis of Endto-End Structure-Aware Convolutional Networks for Knowledge Base Completion(SACN),and proposed Entity-Connectivity-Difference-Aware and Target-Oriented Attention Network for Link Prediction(ETAN).ETAN divides the entities in the knowledge graph into three categories by statistically analyzing the distribution of entity connectivity: remote entities,ordinary entities,and over-saturated entities.For remote entities,we use rules for neighborhood completion,and for over-saturated entities,we perform neighborhood sorting and truncation based on attention to improve the neighborhood structure of entities.In addition,by modeling the interaction between neighborhood entities and neighborhood relations,it provides more effective information for the embedding learning of the central entity.Finally,we integrate the prediction links of the link prediction task into the neighborhood interaction to achieve a complete target-oriented attention strategy,which provides a powerful support for the network to efficiently and highly aggregate neighborhood information.This thesis conducts comparative experiments on the standard FB15K-237 and WN18 RR data sets,and evaluates the model capability with the common evaluation tool of the link prediction task.Compared with the basic model SACN,ETAN has greatly improved indicators such as MRR and Hit@N and has reached the high level in the existing knowledge graph embedding models for link prediction.In addition,we also conducted ablation experiments for each core component of our model,and further analyzed and verified the impact and importance of each component on the model's capability.