Knowledge graph enhanced citation recommendation model for patent examiners

Yonghe Lu, Xinyu Tong, Xin Xiong, Hou Zhu

Abstract: In the face of a growing volume of patents, patent examiners grapple with prolonged examination cycles, prompting the need for more effective citation recommendations. To address this, we introduce the patent knowledge graph embedded in Bert (PK-Bert) model. This innovation combines a patent knowledge graph with semantic information in an advanced Transformer framework, outperforming conventional common-sense knowledge graph embedding. PK-Bert exhibits substantial improvements, boosting the recall of accurate citation recommendations by 2.15% over the benchmark model Bert and 1.25% over K-Bert with CnDBpedia. Ablation experiments highlight the significance of knowledge graph elements, with the inventor proving most influential, followed by the IPC number and assignee. At the same time, publication time and title information have a minor impact. Moreover, PK-Bert excels when trained with earlier data and evaluated for patents issued post-November 2023. Our study not only advances patent examiner recommendations but also presents an efficient integration method for knowledge graph-enhanced semantic patent characterization.