Building Knowledge Graphs in Heliophysics and Astrophysics.

We propose a method to build a fully connected knowledge graph in the scientific domains of heliophysics and astrophysics, using word embeddings from BERT which are adaptively fine-tuned to these domains. We extract the scientific concepts automatically by a keyword extractor. The graph nodes representing these concepts are connected and weighed based on the cosine similarities computed from their fine-tuned embeddings. Our method is able to capture various meaningful scientific connections, and it incorporates the possibility to enable knowledge discovery.