Scalable Hybrid Framework for Domain Patent Knowledge Graph Construction Using Deep Learning and Graph Databases

In the context of technological innovation in rare earth materials, the construction of a rare earth patent knowledge graph remains a research field requiring further exploration, particularly regarding the mining and utilization of unstructured patent data during the graph construction process. This study proposes a novel method for constructing a rare earth patent knowledge graph based on an improved information extraction model, SpERT-LG. By integrating low-rank matrix and graph neural network techniques, the method enables the extraction of knowledge from unstructured patent data. The ontology construction is achieved using a dual strategy combining “top-down” and “bottom-up” approaches, ultimately realizing efficient construction and relationship mapping of the rare earth patent knowledge graph. Experimental results demonstrate a 3.4 task with the enhanced model. The constructed graph comprises 669,220 nodes and 2,815,109 edges, providing strong support for exploring the evolution of rare earth patent technology directions.