Lightning Bugs or Fireflies? A Machine-Learning Approach to U.S. Dialects and Presidential Elections

Abstract

The United States is a heterogeneous nation, containing diverse regional sub-populations with dif- ferent dialects and political stances. Research has already been conducted about different dialects of the United States, but there existed a gap in research on how dialects influence voting pattern. In this study, we investigate whether someone’s state and vote can be predicted based on their dialect, how states relate into regions based on dialect and voting, and whether dialect and past voting data can be used to predict presidential elections. Using Linear Discriminant Analysis modelling techniques on data from the Harvard Dialect Survey, we find that the dialect data of a survey re- spondent can be predict the respondent’s state with over 80% accuracy and census region with over 93% accuracy. By correlating state dialect and correlating state 1996-2016 election data, we devise a 7-region state classification system that shows that states speak and vote in blocs. Further applying Linear Discriminant Analysis, we show that dialect and past voting behavior can predict the current political leanings of over 96% of states and predict over 94% of swing states. We apply LDA models to predict a close 2020 Republican victory that could easily swing Democrat. Our analysis and modeling techniques could prove useful for voice recognition technology to analyze the geographic location of a speaker and improve targeted advertising based on dialect. Also, our dialect and voting models could help predict future elections, and our analysis could help political organizations target their campaign trails and political rhetoric.