Benjamin Guinaudeau and Simon Roth (University of Konstanz)
Abstract: Estimating ideological position has always been challenging for political scientists. The technical progress of the last decades -digitalization, computational improvements- opened new opportunities to measure ideological position. While classical approaches relied on surveys, new strategies relying on other sources of data such as social networks or text have been explored. This paper develops a strategy to estimate the ideological position of representatives on the basis of their floor speeches. One big challenge when scaling from text is to ensure that the captured dimension matches the left-right axis. Existing models deal differently with this issue: wordscore requires the identification of two documents meant to represent the extreme points of the dimension, wordfish assumes the principal component maps the left-right and wordshoal require to filter the document to build a homogenous corpus structured around the dimension to measure. We propose another approach, which relies on a two-stage model. First, we train a neural network to predict the party of each speaker. Thus, we obtain for each document a vector of probabilities, which informs on the affiliation to each party. In other words, we represent the speeches in a low-dimensional space, where each dimension corresponds to one party. In the second stage, we further reduce the dimensions to obtain one linear scale, expected to map the left-right dimension. To test the measure, we use each floor speech pronounced in the German, British, Spanish and French Parliaments in the last two decades (~2000000 Speeches). The speech-level ideological point estimation allows representing each representative as a distribution of ideological position, which offers more flexibility than single-point estimation. After validating the measure using more classical measure of political ideology, we use it to substantially investigate how the ideological position of representatives responds to electoral incentives.