Constructing biomedical domain-specific knowledge graph with minimum supervision

Domain-specific knowledge graph is an effective way to represent complex domain knowledge in a structured format and has shown great success in real-world applications. Most existing work on knowledge graph construction and completion shares several limitations in that sufficient external resources such as large-scale knowledge graphs and concept ontologies are required as the starting point. However, such extensive domain-specific labeling is highly time-consuming and requires special expertise, especially in biomedical domains. Therefore, knowledge extraction from unstructured contexts with minimum supervision is crucial in biomedical fields. In this paper, we propose a versatile approach for knowledge graph construction with minimum supervision based on unstructured biomedical domain-specific contexts including the steps of entity recognition, unsupervised entity and relation embedding, latent relation generation via clustering, relation refinement and relation assignment to assign cluster-level labels. The experimental results based on 24,687 unstructured biomedical science abstracts show that the proposed framework can effectively extract 16,192 structured facts with high precision. Moreover, we demonstrate that the constructed knowledge graph is a sufficient resource for the task of knowledge graph completion and new knowledge inference from unseen contexts.