Detecting innovative companies via their website

Abstract

Producing an overview of innovative companies in a country is a challenging task. Traditionally, this is done by sending a questionnaire to a sample of companies. This approach, however, usually only focuses on large companies. We therefore investigated an alternative approach: determining if a company is innovative by studying the text on its website. For this task a model was developed based on the texts of the websites of companies included in the Community Innovation Survey of the Netherlands. The latter is a survey carried out every two years that focusses on the detection of innovative companies with 10 or more working persons. We found that the text-based model developed was able to reproduce the result from the Community Innovation Survey and was also able to detect innovative companies with less than 10 employees, such as startups. Model stability, model bias, the minimal number of words extracted from a website and companies without a website were found to be important issues in producing high quality results. How these issues were dealt with and the findings on the number of innovative companies with large and small numbers of employees are discussed in the paper.

Improving the Output Quality of Official Statistics Based on Machine Learning Algorithms

“…Many of the projects presented by Beck et al (2018) that use machine learning for classification, were intended to measure variables that could only be measured by employing machine learning algorithms. A typical example is identifying small innovative companies (too small to be included in traditional surveys), using data from their websites (Daas and Van der Doef 2020). Therefore, we aim for the best of Breiman's two cultures, complementing statistical quality with information quality.…”

“…Although this type of statistical output is rather uncomplicated, we stress that it occurs in a wide variety of applications, also outside official statistics. Examples include counting solar panels (Curier et al 2018), estimating the relative occurrence of small innovative companies (Daas and Van der Doef 2020), measuring deforestation and other applications of land cover mapping (Costa et al 2018), and predicting election outcomes based on sentiment analysis (O'Connor et al 2010). Moreover, results for estimating the base rate generalise to more complicated statistical output.…”

New perspectives for data‐supported foresight: The hybrid AI‐expert approach

“…By adding the possibility to analyze also unstructured data without a predefined search realm, however, the advancements in big data and AI create unprecedented opportunities to develop data‐supported and meaningful insights in future technology and innovation trends and their potential societal impact. That is, due to the rapid development of computing capacity, algorithms to analyze patterns in unstructured data sources—such as natural language processing (NLP), topic‐modeling such as latent Dirichlet allocation (LDA), and deep learning methods based on artificial neural networks—are increasingly available (Daas & van der Doef, 2020; LeCun et al, 2015; Mühlroth & Grottke, 2020; Porter, 2019). Recently, a number of proposals have been made on how to mobilize this potential in foresight, especially for identifying early signals of emerging changes (e.g., Krigsholm & Riekkinen, 2019; Lee & Park, 2018; Mühlroth & Grottke, 2018; 2020).…”

The need for timely official statistics. The COVID-19 pandemic as a driver for innovation

“…Another example is detecting small innovative companies using text data from their websites [2]. Statistics Netherlands sends out surveys to collect information on innovation in companies but these do not include the smaller innovative companies.…”