Deep Voting: Predicting Votes from Bill Texts using Neural Networks

• In Event: Text as Data and Lawmaking

Abstract

We develop a deep learning neural network model of legislative voting that utilizes bill text as well as other sources of exogenous information to predict both vote margins and individual votes on Congressional bills. Our model is 89% accurate at predicting individual votes and is able to predict forward in time across Congresses with the same party leadership. The model derives most of its leverage from predicting non-unanimous votes by minority-party members, and we use node weight clustering to reveal the most predictive text features and how they vary by party and vote. We then develop a novel multi-layer neural network model that scales both legislators and bills in a two-dimensional space, seeing only a small accuracy cost from the spatial bottleneck. This spatial output appears to work better than Nominate for both in-sample and out-of-sample prediction, and suggests new substantive interpretations of the first and second dimensions.

Authors

• Nicholas Beauchamp, Northeastern University

• Alexander Herzog, Clemson University