Construction of Diabetes Knowledge Graph Based on Deep Learning

Abstract

To integrate medical data which is scattered over the internet, natural language processing (NLP) is widely used in medical text mining. BERT (Bidirectional Encoder Representations from Transformers) is outstanding among many other representation models and vector representation based on Bert pre-training language model can help the target task learn more semantic information. The knowledge graph intuitively reveals the relationship between entities and helps explore deeper semantic connections between entities. There are three important parts in the construction of a knowledge graph, including entity extraction, relation extraction, and graph generation. Based on these methods this paper proposes a Bert-based named entities identification model Bert-BiLSTM-CRF and it is outperforming the established methods. In the relation extraction part, use the BERT-Softmax to improve the semantic expression and its F1-value increased by 12 percent compared with the traditional entity relation extraction model. Based on the above redefined the entities of diabetes and their relationships to enrich the semantics of the knowledge graph. Finally, the Neo4j graph database was used to realize the visualization of the diabetes knowledge map.