Words Speak As Loudly As Actions: Deep Learning Methods For Stance-Based Ideal Points From Congressional Speeches

Abstract

This thesis presents a framework for analyzing text data to extract ideological positions, with a particular application to estimating legislators’ ideal points from their speeches. Traditional vote-based methods such as DW-NOMINATE capture limited information from binary roll-call votes, while existing text-based methods rely solely on the lexical co-occurrence of words, leaving a gap in leveraging the rich semantic content of legislative discourse. To address this, we introduce the first theoretical formulation for stance-based ideal points and develop computational methods for their estimation directly from the Congressional Record using language models and contextual embeddings. Our contributions include: (1) a comprehensive new dataset linking congressional speeches to the bills they discuss; (2) BARTExpan, a novel topic taxonomy generation method that discovers topics from speeches and organizes them in a hierarchical taxonomy of increasing specificity; (3) StrideStance, a state-of-the-art zero-shot stance detection model that identifies legislators’ stances on these topics; and (4) an integrated framework that combines all of these methodological developments to estimate new stance-based ideal points that reveal patterns not captured in voting data. Applied to the 117th Congress (2021-2023), we validate our method against existing measures, and show that it captures fiscal and economic issues along the primary dimension as well as contemporary social flashpoints on issues such as abortion and civil rights along secondary dimensions. Furthermore, we use these ideal points to illustrate the strategic behavior of swing-district representatives who moderate their speeches while ultimately voting with the party line, as well as to examine the political positions of non-voting delegates. Beyond Congress, our framework provides a broadly applicable methodology for estimating ideological positions across a diversity of textual domains, including judicial opinions and news media.