Uncovering Multi-Level Governance and Policy Idea Transfer in Energy Policy Using Topic Modelling on Large Policy Corpuses

This paper uses a natural language processing method called dynamic topic modelling to map the transfer of ideas between policy levels. The paper presents a case study on energy policy in the European Union, Finland and the Helsinki metropolitan area, using policy documents spanning the past 25 years. The documents include laws and regulations, but also scenarios, roadmaps and administrative documents, and constitute a massive corpus of circa 2500 documents and tens of thousands of pages. The corpus has been collected with web scraping methods.

The paper uses dynamic topic modelling to map the structure of this corpus and analyze trends over time at the three governance levels. Topic models are a family of machine learning methods that map word co-occurrence in documents to find word probability distributions that are, ideally, interpretable as topics to a human reader. Each document is a selection of words drawn from a mixture of topics. For example, the words “carbon” and “capture” might occur with a high probability in a topic, and that topic might then be interpreted as discussing carbon capture and storage technologies. Dynamic topic models also allow for the evolution of the word distribution, so that the prevalence of “greenhouse effect” might be overtaken with “climate change” as the vocabulary evolves.

This paper looks at the evolution of the topic structure and the words within topics in the three policy contexts, and evaluates whether the emergence of new issues and ideas happens first at the international, national or the local level, and whether the three levels are similar in the topics that are discussed and what vocabulary is used to discuss them. The analysis opens a novel vantage point into the relationships between formal and informal institutions and their development.