Construction And Application Of Domain Knowledge Graph By Joint Extraction Of Entities And Relations

With the rapid development of Knowledge Graph(KG)techniques,domain KG has become the focus of academia and industry.Compared with the methods of traditional domain data management,domain KG is a structured semantic knowledge base to facilitate the access of relations between knowledge,and thus provide data for domain applications,like intelligent question and answer,search engines and decision support.Therefore,it is significant to construct domain KG efficiently for effective management and intuitive display of domain knowledge.Compared to the large-scale knowledge of the general KG,domain KG is less applicable and more focuses on the accuracy and depth of knowledge,so the construction still suffers from a lack of labeled corpus.Therefore,it is critical to construct domain KG with as few labeled corpus as possible.In addition,the joint extraction of entities and relations is a key part of the domain KG construction,and existing models are unable to effectively extract entities and overlapping relations between entities in the domain text data.Therefore,how to effectively extract the overlapping relations between entities in domain text data is another key problem to be solved.In this case,we propose a method to jointly extract entities and relations by incorporating active learning to construct the domain KG upon triples,and make empirical studies on the domain of Dulong minority.This thesis includes the following aspects:1.To address the problem of the lack of labeled corpus in domain area,we propose a method of sampling data to be labeled based on active learning.The samples to be labeled are selected by evaluating their values and similarity.Experimental results show that the method can reduce the cost of labeling effectively.2.To address the problem of the overlapping relations in domain text data,we propose a model to jointly extract entities and relations based on BERT-Bi GRU*-CRF from domain data.Besides,experimental results on the NYT public dataset and the real dataset of ethnic minorities show that our method can identify entities and relationships more accurately than other models.3.Taking the domain knowledge of Dulong culture as an example,we design and develop a system of KG construction and semantic question answering for Dulong culture by the method of constructing domain KG proposed in this thesis.KG construction,KG visualization query and question answering of Dulong culture are implemented in the system.