Using Twitter to Detect Polling Place Issues on U.S. Election Days

Abstract

In this article we analyze whether Twitter can be used to detect barriers to voting at polling places. We use 20,322 tweets geolocated to U.S. states that match a series of keywords on the 2010, 2012, 2014, 2016, and 2018 general election days. We fine-tune BERTweet, a pre-trained language model, using a training set of 6,365 tweets labeled as issues or non-issues. We develop a model with an accuracy of 96.9% and a recall of 72.2%, and another model with an accuracy of 90.5% and a recall of 93.5%, far exceeding the performance of baseline models. Based on these results, we argue that these BERTweet-based models are promising methods for detecting polling place issues on U.S. election days. We suggest that outputs from these models can be used to supplement existing voter protection efforts and to research the impact of policies, demographics, and other variables on voting access.

Keywords: natural language processing, election administration, voting, voter suppression, machine learning, twitter