An embedding inversion approach to interpretation of patent vacancy

This study presents an approach to identifying emerging technology opportunities by extracting patent vacancies and concretizing their meaning in textual form. Patent abstracts are mapped into a high-dimensional vector space using a text embedding model, then reduced to a two-dimensional map using an autoencoder. Density estimation is applied to these coordinates to identify hotspots and define vacant cells as patent vacancies. The two-dimensional coordinates of these patent vacancies are then converted back into high-dimensional embedding vectors using the decoder of a trained autoencoder. Finally, the embedding inversion model converts the embedding vectors into text describing the technology overview. For validation, 7,413 patents related to solar cell technology registered in the last ten years as of 2023 were collected. The first eight years of patent data were used to extract vacancies and generate technical text. Consequently, patents exhibiting a resemblance to the generated text were observed to emerge in the subsequent two years, thereby substantiating the innovative potential of our approach.