Research On Knowledge Graph Completion Based On Convolutional Neural Network

Knowledge graph is a semantic network composed of nodes and edges.It usually uses triple(head entity,relationship,tail entity)to represent knowledge,and provides the ability to describe the objective world in a structured way.It is widely used in search and intelligent assistants.Although the existing knowledge graphs contain a large amount of knowledge,their completeness need to be improved.The knowledge graph completion is aimed at completing the existing knowledge graphs,improving their completeness,solving the problem of data sparsity,and then conducting knowledge reasoning and mining implicit knowledge.The existing knowledge graph completion models usually adopt the method of knowledge representation learning,and use the triple structure information in the knowledge graph.Then they embed the entity and relationship into the low-dimensional dense vector space to obtain the vector representation.And the knowledge graph completion is conducted.In addition to the triple information,knowledge graphs also contain rich multi-source information,such as text and images.According to the used information and the scoring function,the existing knowledge graph completion models can be divided into translation distance models,semantic matching models and models combining multi-source information.The translation distance models are simple in structure and easy to expand,but have limited ability to fit complex semantic relations between entities and relationships.The semantic matching models have strong fitting ability but high complexity and low scalability.The models combining multi-source information are relatively simple on fusing multi-source information,with low utilization for target information.Most of the existing models are difficult in balancing the fitting ability and scalability on the complex semantic relationship between entities and relationships.And this thesis proposes a knowledge graph completion model for the triple structure,namely CNN-KG model.This model uses the powerful nonlinear representation ability and feature extraction ability of the convolutional neural network to extract the implicit features between entities and relationships.This model could efficiently use triple structure information,improve the complex semantic relationship between entities and relationships and better balance the fitting ability and scalability.When the existing models utilize text information such as entity description,the fusing strategies are relatively simple and the information is utilized in a low degree.Therefore,this thesis proposes a model that combines triple structure and entity description information,namely CNN-DKG model.This model firstly uses the convolutional neural network to encode the entity description information to obtain the entity representation.Then it uses the TransE model to obtain the knowledge representation based on the triple structure.Then it fuses the two kinds of representation,and extracts its characteristics by convolutional neural network.It also defines the relevant scoring function to evaluate the possibility of the establishment of the triple.In this thesis,the training target is optimized by Adam on two famous datasets,WN18 RR and FB15k-237.Three tasks,entity prediction,relationship prediction and triple classification are adopted to evaluate the training results of the proposed models.The experimental results show that CNN-KG and CNN-DKG models are superior to TransE and other models in most of the evaluating indicators such as mean rank,mean reciprocal rank,top ten hit rate and classification accuracy.