Research On Knowledge Representation Learning Method Based On Multi-source Information

Knowledge graph stores structured data in the form of triple(entity 1,relation,entity 2),which has become the foundation of many intelligent applications.A large number of knowledge graphs,such as Freebase,DBpedia and YAGO,have been successfully constructed and applied to many scenarios such as relation extraction,information retrieval,question answering system,entity linking and so on.Effective representation of knowledge graph is a key to the whole process of knowledge graph construction and application.The traditional representation scheme represented by the one-hot representation assumes that all objects are independent and irrelevant,which leads to the waste of a large amount of information and the failure to make full use of the semantic information of the object and the problem of sparse data.Therefore,it is impossible to effectively represent triples.As the same time,knowledge graphs need to be constantly enriched due to the explosive growth of knowledge,and there are still a lot of knowledge that need to be complemented.If the special graph algorithm is designed to carry out semantic calculation and relational reasoning for every entity and relation,it will have the shortcoming of computing expensive,poor transportability and it will be hard to do large-scale operations.However,by representing entities and relations as low-dimensional dense real-valued vectors,the representation learning of knowledge graph can efficiently calculate the semantic relation between entities and relations,and then conduct relational reasoning,achieve heterogeneous information fusion,and improve computing efficiency.However,there are still many challenges in knowledge representation.First of all,most representation learning methods only explain triples from the perspective of structure,ignoring diversified multi-source information,lacking of effective ways to extract complementary information and lack of effective fusion methods.Secondly,since most methods only learn entities and relations from a global perspective through latent feature,they cannot provide accurate semantic representation in some scenarios and effectively represent complex relations.At the same time,most of the distributed representations learned by existing methods can embody the basic semantic and structural information of entities,but the representations cannot express higher-level information.In view of the above problems,this paper improves the representation learning method on the basis of the existing work.The main work is as follows:(1)A knowledge representation learning method based on discriminative path is proposed,which consists of latent feature learning model and graph feature learning model.The former extracts the semantic information contained in the multi-step relational paths,and learns the triple representation from the global perspective based on the semantic similarity hypothesis,the latter extracts the graph features of entities and relations from the local perspective based on graph pattern,and integrates it with the former module by treating the graph features as prior information.DPTransE makes full use ofthe advantages of the two models and integrates the characteristics of the two modules.The path clustering algorithm improves the confidence level of the path features and solves the problem of data sparsity to a certain extent.Experiments show that DPTransE can improve the quality of representation learning and verify the effectiveness of the method.(2)A joint representation learning method CEKE based on structure and entity category information is proposed.This method makes full use of entity category information and proposes a joint learning framework.The closed loop of the learning process is constructed by combining the representation learning based on knowledge graph structure with the distributed representation learning of entity categories.Knowledge representation can be optimized by learning the distributed representation of entity categories,so that the representation of entities and relationships can not only contain semantic information and structural information,but also embody higher-level category information.The experimental results show that CEKE has made remarkable progress in link prediction and triple classification tasks,which further illustrates the effectiveness of the method.