Keywords: Depolarization, Deliberative Democracy, Topic Modeling, Sentiment Analysis, Partisan Divides, Consensus-Building

This study aims to examine how deliberations contribute to depolarization across partisan lines by analyzing speech acts from a deliberative forum using natural language processing (NLP) to identify themes and sentiments. Entering the 2024 Presidential Campaign Season, a group of citizen organizers planned and executed a deliberative citizens forum on electoral systems reform with a nationally representative sample via NORC’s AmeriSpeak panel.
The deliberative forum yielded depolarization on contentious agendas such as online voter registration, restoring voting rights for convicted felons, and auditing random samples of ballots (Fishkin & Diamond, 2023). Given the political divisions in American electoral systems exemplified by the January 6th Insurrection, these depolarizing outcomes of deliberation underscore its importance in counteracting political division. However, previous literature has seldom explored the content and mechanisms of participant exchanges in deliberation or how these exchanges build common ground. Thus, this research will analyze participants’ speech acts with a cluster of topics and sentiments and examine how much of the topics and sentiments effectively reduce differences.
In June 2023, 582 citizens across the US joined an AI-moderated deliberation platform to deliberate nineteen agenda items (69 proposals) on electoral system reforms in four sessions. Participants were randomly assigned to groups upon logging into the AI-moderated platform. This research will analyze 3,126 out of 15,480 speech acts in the session dedicated to discussing voter access and election administration, which lasted about an hour. In this session, the participants deliberated on proposals that included access to voter registration online and restoration of voting rights for convicted felons. Speech acts with fewer than four words, typically brief agreements, were excluded. Citizens’ opinions were recorded twice – before and after discussions – using an 11-point Likert scale (0, ‘extremely oppose,’ to 10, ‘extremely support’), with a ‘don’t know’ option.
This research will use a Python package, BERTopic, to identify hierarchical clusters of topics related to the agenda and see differences in topics across partisanship. Also, the DLATK package will be used to analyze sentiment, emotional tone, and levels of agreement across partisan lines. Thus, this research will investigate how problematic the current election systems are to Democrats, Republicans, and Independents and how much they agree on the needed changes.
This research examines how structured deliberations reduce partisan divides by analyzing speech acts with NLP to uncover the shared topics and sentiments that emerge during dialogue. It highlights the importance of deliberative forums in addressing and overcoming today’s increasingly polarized political landscape.

Citation: Fishkin, J. & Diamond, L. (2023, August 21). Can deliberation cure our divisions about democracy? Boston Globe, https://www.bostonglobe.com/2023/08/21/opinion/2024-elections-partisanship-democracy-common-ground/