Paul Selzner

Patent classifications are central to technology and innovation research, providing structured yet detailed technological information. However, the hierarchical design of widely used systems such as IPC and CPC imposes three long-standing constraints: limited horizontal comparability within domains, the separation of technologies into siloed domains, and inconsistencies in classification across patent office jurisdictions. The contribution of this paper is threefold: Firstly, we address these constraints conceptually and methodologically by developing a machine learning method, based on Meta’s Llama 3, that assigns Japanese F-terms - granular, attribute-level descriptors - to patents worldwide, and simultaneously generates high-dimensional vector representations that map all F-terms into a unified technological space. Secondly, we operationalize this capability by releasing both the trained classifier and the complete vector dataset, enabling researchers and practitioners to apply F-term–based methods beyond Japan, to conduct consistent cross-national analyses, and to measure technological distances across the entire innovation landscape. Thirdly, we outline a research agenda that leverages these new resources to advance the study of technological change, innovation systems, and cross-domain analysis. While the classifier’s performance can be further enhanced, this work provides a new empirical basis for studying technological change and innovation with greater granularity, cross-domain comparability, and consistency across jurisdictions.