Research On Algorithms Of Knowledge Graph Completion Based On Multi-source Information Representation Learning

Knowledge Graph contains a large number of triples such as(entity1,relation,entity2).It provides structured data that can be understood by computers for many artificial intelligence applications.Knowledge Graphs store vast amounts of knowledge,but they are still not complete.Knowledge Graph Completion aims to solve the data sparseness problem that exists in the Knowledge Graph and improve its internal perfection.From the existing algorithms,there are mainly two kinds of information that can be used for Knowledge Graph Completion.The first is the existing triplet data in the Knowledge Graphs.The second is multi-source information such as text and images outside the triple.If this information is integrated manually,it will not be able to meet the needs of the complementary work of an increasingly large-scale Knowledge Graph.In recent years,the development of knowledge representation learning has greatly enhanced the efficiency of Knowledge Graph Completion.Knowledge representation learning aims to map the entities and relationships in the Knowledge Graph to a low-dimensional vector space.In this way,we can easily calculate the semantic similarity and efficiently use the above-mentioned two kinds of information to complete the Knowledge Graphs.The completion algorithm based on knowledge representation learning has made great progress,but there are still the following problems:(1)The quality of multi-source information is mixed,and there is no effective way to extract useful information from it;(2)Multi-sources information is abundant but not fully utilized;(3)The heterogeneity of entity1 and entity2 in the triple is not considered.In response to the above issues,the main work of this thesis are as follows:1).In entity representation learning,the entity's text neighborhood and structure neighborhood information are introduced.The text neighborhood of an entity refers to a collection of other entities that are frequently co-occurrence in the text,which can capture the useful information in the text more precisely.The structure neighborhood of an entity refers to a collection of other entities connected to it in the Knowledge Graph.We consider these two kinds of neighborhood information at the same time when modeling the entity,so that the learned entity representation is more expressive.2).When the entity in the Knowledge Graph is lonely and there are few entities that appear together with it,the model based on the entity neighborhood cannot achieve the desired effect.Therefore,we purposely introduce the description information of the entity and use CNN as an encoder.Then we propose two ways of describing information and triple joint learning:(i)Cross-fusion training of entities based on descriptive information and triple-based representations.For the part of the triple,we use different matrices to map the two entities and specify that the mapping matrix must be low-rank.(ii)The description information is integrated into the triples in a mapping manner.This method is less complex and can better represent triplets in non-one-to-one relationships.In addition,the models constructed by the above two methods all can be used in a zero-shot learning scenario,that is,generating a representation of an entity that has not appeared in the Knowledge Graph.In this way,the description information of the entity is better utilized.We performed experiments on Knowledge Graph Completion,such as link prediction and triple classification,on benchmark datasets such as FB15 k and WN18.Experiments show that compared with the existing methods,the model presented in this thesis can better complement the work of Knowledge Graph Completion.