InnoVAE: Generative AI for Mapping Patents and Firm Innovation

Abstract

We propose a generative AI approach (InnoVAE) to map unstructured patent text into an interpretable, spatial representation of firms' innovative activities. InnoVAE learns a vector representation of patents and places them within a "disentangled" space to facilitate managerial intuition and action. After validating the internal consistency of our approach, we apply it to three decades of AI patents to show that it can be used to construct interpretable measures, at scale, that characterize firms' AI-based IP portfolios. We demonstrate three use cases including (1) generating technology landscapes that inform businesses about their competitive positions, (2) engineering new, intuitive features from unstructured text that facilitate analysis of patent activity, and (3) augmenting patent applications to mitigate the risk of patent rejection. Our approach demonstrates the potential of generative AI methods to make actionable the vast quantities of text stored in unstructured corporate databases.

Keywords: generative AI, patents, machine learning, interpretability, economics of innovation, managerial decision support, ChatGPT, large language models