Staying on the democratic script? A deep learning analysis of the speechmaking of U.S. presidents

Abstract

Dynamic agenda representation assumes a linkage between the policy emphases prescribed by various democratic inputs (electoral promises and public opinion polls) and policy agendas ranging from the media to executive orders. An extrapolation of this idea would propose that, in the U.S. context, policy emphasis in major programmatic messages such as State of the Union addresses would be followed by the president's day‐to‐day communication. We investigate this congruence with a new database of presidential speeches that, for the first time, offers a deep learning‐enhanced sentence‐level policy topic coding of various forms of the speeches U.S. presidents made from Truman to Trump (for a total count of 16,523 speeches divided into nearly 2 million individual sentences). Using this database, we demonstrate that presidents' occasional, day‐to‐day remarks strongly correlate with the annual policy messages—in this sense, presidents are staying on the democratic script.