Research On Link Prediction Method Of Knowledge Graph Based On Neural Network

Knowledge Graph(KG)represents the real world in the form of a graph,which can deal with and exhibit multi-source heterogeneous data and their complex relationships.It is one of the key technologies of the Cognitive Intelligence driven by data and knowledge,and has been widely utilized in various fields such as e-commerce and medical care.However,due to the diversity and variability of the entities themselves,and the knowledge obtained by automatic or semi-automatic knowledge extraction techniques is usually incomplete,as a result,the knowledge in the KGs may either have some errors or missing entities or relationships,which will greatly affect its effective connection and application with downstream AI tasks.To this end,scholars and researchers put forward the task of KG link prediction,which aims to predict unknown information from the known one,complete the KG,and promote its application in downstream tasks.At present,by using use convolutional neural network or graph attention network,link prediction methods based on neural network fully mine information and obtain excellent prediction performance.In spite of this,there are still some challenges.On the one hand,most neural network models focus on entity representation learning,while ignoring the learning of relation representation,resulting in information loss and lower link prediction performance of the model.On the other hand,when learning the interaction information of entities and relations,convolutional neural networks only pay attention to part of the interaction information,and do not learn the hidden information inside a triple abundantly,so that it cannot represent entities and relations adequately.To solve the above two problems,this thesis makes the graph attention network better by introducing a parameter generator(PG),and advances the convolutional neural network by applying a self-attention mechanism,and then to improve the link prediction effect.The main contents of this thesis is as follows:(1)A graph attention network model with PG of relationship embedding is proposed to implement link prediction.Inspired by the idea of the parameter generator,the model takes the updated entity embeddings as the input and the network parameters as the output of the PG,which will subsequently be adopted to learn the vector representation of relations,which will realize the update of the relationship embedding,improve the quality of the relationship embedding,and boost the effect of link prediction.Experiments on the WN18 RR and FB15k-237 public datasets demonstrate that the proposed model has optimal performance than the baseline models.On FB15k-237,the average increase of MRR index is about 21.0%,and the average increase of Hits@10 index is about 17.4%.On WN18 RR,the average increase of MRR index is about 6.4%,and the average increase of Hits@10 index is about 9.1%.(2)This thesis makes up a convolutional neural network model with self-attention mechanism for link prediction.It is accounted that there is also extremely abundant information between different dimensions of entity and relation embeddings that is of great assistance to link prediction.Accordingly,for the sake of learning the internal information of triples,the self-attention mechanism is proposed to extract information from the triple matrix,fuse the information of different dimensions of the triplet,which can capture the rich entity and relationship interaction information inside the triplet,thereby improving the performance of the model.The proposed model is tested independently on the WN18 RR and FB15k-237 public datasets,and it is also used as the decoder of the model in(1).The two experimental results on datasets illustrate the effectiveness of the method.For example,in the two-stage experiment,the MRR index is increased by about 15.3% on average on FB15k-237,and the Hits@10 index is increased by about 10.4% on average.On WN18 RR,the average growth of MRR index is about 9.2%,and the average growth of Hits@10 index is about 6.0%,which proves that the model can fully mine the rich interactive information between entity and relation embeddings and reduce information loss.(3)Finally,this thesis also implements an intelligent question answering(QA)system based on a medical knowledge graph.A Chinese medical knowledge graph is constructed,and the link prediction model mentioned above is used to reason about unknown medical knowledge.On this basis,medical intelligent question answering and entity and relationship query functions will also be implemented.