Measuring climate salience across different issues in political text with supervised machine learning

This paper aims to contribute to the panel's discussion on climate policy linkages by introducing a measure of the association between climate change and various issues in party communication. Literature in international relations and public policy has explored the intersection of climate change with other policy issues using concepts such as climate integration, mainstreaming, or climatisation. However, a research gap exists in understanding how political parties link climate change to different issues in their communication. It is surprising, considering that a crucial aspect of political actors' responses to climate change involves reframing other issues. For instance, while nuclear energy has long been a contentious topic in France, politicians have only recently started to use climate change as justification for their stance on nuclear energy. This communication is likely strategic, but studying it requires measures that are currently unavailable. After reviewing existing approaches to measure issues and climate change in political text, this paper develops a measure to examine how parties associate climate change with different issues in their communication. Party literature has a longstanding tradition of examining how political actors compete by emphasizing different issues. Scholars have extensively relied on manual content analysis to measure the salience of policy issues in manifestoes within important projects such as the Comparative Manifesto Project and the Comparative Agendas Project. While manual content analysis has been the primary method for measuring the frequency of these issue categories in political text, advances in NLP are increasingly used to automate this process. However, relying on mutually exclusive categories, these classifications are unable to capture the complexity of cross-sectoral issues like climate change, as communication on this aspect falls into different categories. Party research on climate change has then led to the development of new climate measures, relying on manual content analysis to detect political text referring to climate change. These measures have not been automated yet, limiting their scope. In this paper, I build on this methodological literature and advances in NLP to develop a measure of the association between climate change and other issues in political text. Firstly, I train supervised learning classifiers to detect whether a given political text is about climate change or not. This first contribution enriches existing measures by providing a new method to measure climate change content in political text, which I further compare with a dictionary approach. Secondly, I apply on the same texts existing models trained to classify political text in the 21 different policy issues of the Comparative Agendas Project. For each document of a corpora, I then obtain a probability of the document to be about climate change and a probability of the document to be about each of the 21 issues. This enables measuring the diffusion of climate change references across various issues and assessing how climate change is reconfiguring them by quantifying their "climatization" in party communication. Lastly, I discuss how these measures might vary across different types of political text, languages, and the challenges of using such approach to measure climate policy linkages.