Implementing Text-Based AIs to Investigate and Measure Private Sector Software Innovation

Overview

We study the landscape of product innovation in the computer software sector, leveraging publicly available opportunity data news articles obtained from Dow Jones, a business news and data provider. We implement a series of Bidirectional Encoder Representations from Transformers (BERT) neural networks, a sophisticated natural language processing method, for a number of tasks. Our work developed a BERT classification model to identify news articles describing innovation broadly, making use of a training set of 600 manually labeled articles and demonstrating an accuracy rate over 96%. This model was then applied to 1 year’s worth of news articles about the computer software industry to predict which articles describe innovation. We applied a different BERT algorithm to this set of predicted innovation articles for the purposes of named entity recognition, which was used here to extract the company and new product names mentioned in these predicted innovation-describing articles.