Data Driven Knowledge Graph Completion

Knowledge Graph(denoted as KG)has good structural features and is easy to support computing and reasoning.Now KG has been a vital source of knowledge for AI applications,which can provide knowledge reserve for Search Engine or Question and answer system.The typical KG cases include FreeBase and WordNet,etc.As the underneath core technology of the AI applications,KG needs to provide sufficient and reliable knowledge for the upper level.So,the richness of the KG content determines the value of AI applications.Even though the current KG is rich in content and large in volume,it's far from completion,still needed to be completed,that is KG completion.KG completion aims to perform link prediction between entities,which leads us to discover new effective facts.There is a lot of potential knowledge in text data from Internet,which is an important supplement for KG.At the same time,extracting structured information(we call it the external information)from the text data of Internet to make KG completion,can not only reflect the latest changes of the current knowledge,but also establish the relevant connections with the existing knowledge from the KG,which can help us get more knowledge.It is not to be ignored that the text data from the news media of Internet is usually not accurate and reliable.How to get effective information from the text data and use it to fuse with the existing knowledge of the KG is the key of KG Completion task and also an important challenge.In recent years,with the introduction of the TransE,Representation Learning plays a key role of KG completion.Based on Representation Learning,the main work of this thesis is divided into the following aspects:(1)Introduce a method of extracting structured information from text data and its relevant model.(2)Several typical knowledge representation models are introduced and reviewed.(3)Present an effective method of using text data from Internet to make KG completion: First of all,we extract structured information from the text data and save it in the form of triple.Second of all,we use the existing knowledge from KG to build the KG's knowledge representation model.Finally,based on the knowledge representation model,we evaluate the rationality of the extracted triples,and add the effective information to the KG.Taking the typical Freebase Knowledge Graph as an example,this thesis makes the KG completion based on the different knowledge representation models and gives a experiment to have a validity test of this method with the index of Precision,Recall and F1-Measure.The experiment results show that compared with the traditional method,we do have a significant F1 value growth on our method combining external information.It proves that the idea of this thesis is effective and feasible.