| Knowledge graph is an important basic technology on information service applications such as intelligent question answer and semantic search,and the more comprehensive the corresponding facts covered by the knowledge graph,the better able to provide the desired answers.Although there have been many large-scale knowledge graphs,they are still far from completeness,e.g.,in Freebase,30% person entities miss information about their parents.With traditional knowledge graphs representation which based on network structure,specific algorithms have to be designed to make inferences from knowledge graphs,which suffer from computational efficiency.Representation learning of knowledge graphs can the problem that the traditional knowledge representation cannot effectively measure and use the semantic relationship between entities,and improve the ability of knowledge and reasoning of knowledge graphs.TransE,which is the representative model of knowledge graph representation learning,is simple but effective,however,it could apply well to modeling one-one relations rather than complex relations in knowledge graphs,and it only used the triple information in the knowledge graph,and these are many text information for entities in knowledge graphs,which cannot be well used in this method.To solve these problems,we present a text-embodied translation embedding method with binary credibility vector for knowledge graph representation learning.In our method,we can apply well to modeling complex relations in knowledge graphs,and we use textual information related to entities in knowledge graph.Supported by 863 Project “The Key Technologies of Humanoid Intelligent Knowledge Understanding and Reasoning oriented to Primary Education”(2015AA015403),in order to make better use of the constructed knowledge graph,we studied the representation learning of knowledge graphs.Our work is mainly divided into those following parts:(1)According to the relations of reflexive,one-to-many,many-to-one,or many-to-many mapping,the knowledge graph embedding model--TransE cannot distinguish entities which have same semantics in one aspect.We constructed a binary credibility vector,which make different relations focus on different attribute information of entities.Thus,we propose a translating embedding method with binary credibility vector model,which can improve the distinguishing ability.In our model,we constructed a binary credibility vector for each relation,and entities have multiple aspects,different relations focus on different aspects of entities base on the binary credibility vector of the relations,we updated the credibility vector using the method similar to the pheromone in ant colony algorithm.Experimental results show that our model achieves better performances under lower model complexity.(2)In addition to the triple information in the knowledge graphs,the unstructured text contains rich semantic information,which can enrich the embedding ability of the representation learning model.In this thesis,we proposed a text-embodied knowledge representation learning with binary credibility vector based on the text information of entities.In this model,the representation of an entity is divided into two parts: a representation based on triple information and the other based on text information.We used a five-layer deep convolutional neural network to extract semantics of text information of entities as the textual-based representations of entities.Finally,the structured representation and the textual representation of entities are embedded in the score function of our model,so that these two representations can be fused.The main contribution of this thesis are that in view of the two shortcomings of TransE,which are difficult to deal with complex relations and use only triplet information in the knowledge graph.We proposed a text-embodied knowledge representation learning with binary credibility vector model,which can provide a certain theory and technical support for the application of the knowledge graph. |
PDF Full Text Request