Research On The Construction Of Knowledge Graph Of Segmentation Domain Under The Guidance Of Small-scale Knowledge Base

The knowledge graph is essentially an efficient and interpretable way of knowledge organization.Through the transfer of relationships between multiple knowledge units,more factual knowledge can be reflected.Some traditional methods of knowledge organization,such as taxonomy,subject vocabulary,etc.have not yet reached the semantic level of knowledge,and it is difficult to show multiple relationships between knowledge units.The emergence of knowledge maps has just solved this problem.Nowadays,as a new method for describing concepts,examples and their relationships in the objective world,knowledge graph technology has attracted widespread attention.The rational use of knowledge graph can effectively expand the depth and breadth of knowledge exploration.So,how to construct a domain knowledge map at low cost is an urgent problem that needs to be solved.The effective construction of knowledge map mainly involves two elements of entity and entity relationship.In more subdivided domain scenarios,the main reason that restricts the existing generalization of entity recognition and entity relationship recognition technology is the lack of reliable tagging corpus.Correspondingly,taking the existing small-scale knowledge base in the subdivided domain as a guide,and fully extracting and utilizing the semantic information in it reasonably,it can help the automatic construction of the domain knowledge map to a certain extent.Taking the field of "cardiovascular diseases" as an example,this article starts with the exploration of two key elements of building a knowledge map of subdivided domains,namely,entity identification in subdivided domains and entity relationship identification in subdivided domains.In the entity recognition scheme,the existing domain knowledge(encyclopedia,vocabulary)is used to assist small-scale domain corpus annotation,reduce labor costs,and use several neural network models such as RNN,LSTM,LSTM-CRF,Bi LSTM-CRF and The BERT pre-trained model was trained separately on the sequence labeling model,and finally achieved a good automatic recognition of the entity.The F1 value of the Bi LSTM-CRF model reached more than 85%,and the F1 of the BERT pre-trained model's recognition result after finetune reached 88.35% The recognition accuracy of two types of entities,"organ" and "diagnostic method" is more than 90%.In the relationship recognition scheme,a scheme that combines domain meta-knowledge(ie relationship constraints)and word embedding vector analogy is proposed.The small-scale knowledge base of a small-scale knowledge base formulates the entity relationship constraints of a specific segment,and calculates the word embedding vectors of the domain entities based on the corresponding domain background corpus.The word embedding analogy learning of a small number of easily accessible entity relationships in the knowledge base generates different types Entity relationship initial classifier.And test the performance of the classifier in simulated real scenes,and then use active learning The effect of the classifier on the weaker relational layer is modified.The experimental results show that the recognition effect of the initial trainer can be effectively improved through active learning.