The beliefs communicated and not communicated. Discourse networks of urban sustainability governance in Zürich over time.

The UN Agenda 2030 and its associated sustainable development goals (SDGs) put an emphasis on the role of cities for achieving a sustainable future for all. Implementing the SDGs in cities is a complex governance challenge, touching on issues ranging from gender equality over biodiversity to climate protection. This means that implementing the SDGs requires cross-scale governance involving solutions in many policy areas and all societal sectors. Actor networks play a key role for successful cross-scale governance in complex urban sustainability governance settings. These networks, also called governance networks, consist of organizations from all societal sectors, including civil society, the private sector, government and scientific institutions. Issues in urban sustainability governance are also often discursively contested. Actors in urban governance networks constantly need to make choices on how to position themselves in relation to various issues in public discourse. They may want raise the profile of certain issues and downplay others (agenda-setting) and are sometimes forced, sometimes interested in publicly taking a stance in relation to them. Doing so, they create discourse networks as the interplay between policy issues, organizational stances towards them and vis-à-vis each other, which provide an essential backdrop to material policymaking and governance activity. Crucially, following theory on beliefs in policy systems in the Advocacy Coalition Framework (ACF), we propose that actors can be seen as communicating their beliefs in the total of their discursive activity in discourse networks. In a highly cross-sector governance setting such as in urban sustainability, this crucially entails a) policy beliefs about issues within a given sector, b) deep core beliefs in the coherence of stances across issues and sectors and c) for many actors also beliefs about the nature of issue interdependencies across sectors. Thus, we propose a deeply relational approach to identifying beliefs beyond the explicitly communicated by exploring stable and less stable patterns within the activity of actors in discourse networks over time. We test our approach for the empirical example of urban sustainability governance in Zürich, mapping activity patterns with discourse networks of actors to different types of beliefs. Traditionally, discourse network analysis has often relied on small datasets of high quality annotations of documents. In contrast (but complementary), we make use of an approach based on automated text analysis of large volumes of newspaper articles (about 1 Mio articles over 12 years). This enables us to cover a larger temporal, thematic and sectoral scope in our analysis. Our processing pipeline uses supervised machine learning to first categorize text at the paragraph level, assigning paragraphs to issues in key urban SDG implementation areas. In a second step, we use a named entity recognition classifier to identify organizational actors occurring in these paragraphs. In a third step, we build on few-shot learning approaches to stance detection to classify actor stances toward issues and toward each other. The final output is a time-stamped bipartite actor-issue graph on stances of actors toward governance issues. We explore this graph using clustering algorithms to detect patterns of related actor activity.