Application of NLP Analysis on Innofund Grant Applications to Predict Innovation Outcomes

Abstract

The development of new technologies is typically a capital-intensive process, requiring significant investments before producing any commercially viable technology. Hence, many nations have enacted national programmes to provide funding and grants to SMEs (Small Medium Enterprises) with potential to advance industrial technology. China's Innofund is one such national programme, attracting many applications due to its equity-free nature. The process of reviewing grant applications is a labour-intensive and potentially arbitrary process that often requires domain-specific knowledge from human reviewers. Hence, the adoption of Natural Language Processing (NLP) techniques may help to identify and shortlist promising applicants to improve the effectiveness and fairness of the review process. This study applies modern NLP methods to Innofund grant applications to predict the applicant company�s innovation outcomes. Transformer-based text analysis was applied to the Market Competition and Business Model sections of Innofund grant applications, with the aim of developing a method that exceeds the predictive power of human evaluators. The domain-adapted, stacked inference model is found to be the most effective. Specific choice of base transformer does not matter, and two years training data suffices. Additional observations suggest that analysing a firm's competition in conjunction with their business model provides the most predictive power. Implementation Software and Hardware: Ubuntu Linux, Hugging Face Transformers, BERT Transformers and variants, XGBoost Keywords: Natural Language Processing, Business Grant Review, Competition Strategy, Business Model, Patent Filing