Scientific Knowledge Combination in Networks: New Perspectives on Analyzing Knowledge Absorption and Integration

Recombinant innovation is considered a significant driver in generating new ideas, and it has been evidenced to have a higher rate of occurrence in scientific papers. Therefore, modeling and measuring the combination of scientific knowledge in articles has garnered widespread research interest. This paper aims to provide a new perspective to understand and measure the absorption and integration of scientific ideas and insights by leveraging knowledge networks. The references and content of the articles function as input for knowledge absorption and output for knowledge integration, respectively, in which the content refers to the substance or core elements found within the articles. These knowledge elements are extracted using KeyBERT, fused and consolidated with string fuzzy match and embedding-based semantic similarity provided by SciBERT, and labeled as supplied knowledge elements, absorbed knowledge elements, and generated knowledge elements. Knowledge networks are then constructed using the extracted elements and the cooccurrence of elements. Three types of metrics are developed to measure the structure and properties of knowledge networks, including descriptive statistics of nodes, degrees, edges, and components, network global structure metrics, and knowledge proximity calculated using document embedding. We finally use the key publications of the Nobel prize in physics to perform an empirical study.