Cities all over the world are undergoing a significant shift towards civic technologies. Local governments are deploying these technologies to scale up civic participation and transform decision-making: they allow a large number of citizens to generate proposals for new public policies. However, not all the resources of public institutions for these processes scale in the same way. The large number of proposals resulting from the increase in citizen participation triggered by civic technologies becomes a major challenge for public servants, who are responsible for their evaluation and conversion into public policies.
To investigate this phenomenon, we present a study of Decidim Barcelona, the digital platform for citizen participation launched by the local government in 2016. In particular, we analyse its first urban planning process (Pla dโActuaciรณ Municipal, hereinafter PAM) that established the actions to be implemented by the local government. The process started with 1,300 official proposals authored by the local government. Then, 9,560 proposals were added by citizens over three months. All proposals were manually reviewed by public servants considering the votes and comments they received, the contributions by civic organizations, as well as the face-to-face meetings in which proposals were presented and discussed. Accepted proposals were then manually grouped by topical similarity to define the actions of PAM to be implemented by the local government.
In this paper, we audit this review process through the analysis of text similarity between actions and their corresponding proposals. We propose the use of word embeddings to numerically represent the content as real-valued vectors. This allows quantifying the similarity between actions and proposals with distance measures. The results show that proposals written by the local government were linguistically closer to the action description than those written by citizens. In fact, we find noteworthy lexical differences between proposals authored by the local government and the ones posted by citizens. While the former uses an administrative vocabulary, the latter are usually characterized by terms related to specific urban issues such as mobility. We then assess the manual process of reviewing and grouping proposals into actions, implementing a distance method that automatically finds the closest action for each proposal, both accepted and rejected. This computational task confirms that accepted proposals and actions have similar content. The analysis also reveals the popularity of proposals was not considered to design the content of actions.
Our approach leads to a new avenue on auditing participatory democratic processes mediated by civic technologies, which are essential to understand how decisions are made. Given the large amount of information generated on these massive participatory processes, both public servants and citizens can leverage these computational methods for the evaluation and conversion process of proposals into public policies and find relevant information, respectively. However, we raise concerns over the risks and limitations of using linguistic models in civic participation, such as techno-solutionism or biases, which suggests that further empirical work is needed to critically examine the opportunities and pitfalls of algorithmic approaches for participatory democracy.