To facilitate the knowledge reuse in engineering design, several dataset approaches have been proposed and applied by designers. This paper builds a patent-based knowledge graph, patent-KG, to represent the knowledge facts in patents for engineering design. The arising patent-KG approach proposes a new unsupervised mechanism to extract the knowledge facts in patent, by searching the attention graph in language models. This method avoids using expensive labelled data in supervised learning or listing complex syntactic rules in rule-based extraction. The extracted entities are compared with other benchmarks and shows a higher coverage of engineering words. The extracted relationships are also compared with other benchmarks, and the result shows meaningful advantages.

Patent-KG: A Patent Knowledge Graph for Engineering Design