Screening ideas in the early stages of technology development: A word2vec and convolutional neural network approach

Abstract

Previous patent-based methods for assessing the value of technological ideas face challenges in screening ideas in the early stages of technology development because they require information available at the time or after a patent is granted. Given that the technical descriptions of ideas are usually available in the early stages, we propose an analytical framework for screening ideas by associating the technical descriptions of ideas implied in patents with the number of patent forward citations as a proxy for the technological value of the ideas. Accordingly, word2vec is used to examine the semantic relationships among words and construct matrices representing the technical content of ideas implied in patents. A convolutional neural network is used to model the nonlinear relationships between the matrices and the number of patent forward citations. Once trained, the proposed analytical framework can screen early-stage ideas using only the technical descriptions of the ideas. We explore the varying performance of our framework across different analysis contexts and discuss the research implications for theory and practice. A case study covering 35,376 patents in pharmaceutical technology confirms that the proposed analytical framework identifies most ideas with little technological value and outperforms existing models in terms of accuracy and reliability.