Ideological Feature Extraction: A General NLP-Based Methodolgy

• In Event: Political Scaling

Abstract

We propose a text-based framework for policy stance and ideology feature extraction to produce real-time measures of positions in the political space. The key insight is to apply intricate topic modeling techniques to reduce very large text data into topics and sentiment orientations. This has the effect of extracting both the issues that politicians are discussing on social media, as well as their position around those issues. From here, we can project this data in a interpretable two dimensional space using the eigenvalue decomposition on the portfolio of positions that social media users have taken on issues at a given time. Because we have information on which topic-sentiment features drive variation in this two dimensional space, this framework means we can apply standard interpretations of the space as being driven by key features in the data. With this methodology, we can uncover that in March and April 2020 politicians initially de-polarized in their public stances on COVID prevention policies, before re-polarizing by the June 2020. We also show that state legislatures are generally less polarized in their online speech, with evidence from the Texas and California State legislatures. Finally, to demonstrate the general nature of the framework, we recover national and state journalists' political positioning in this same space.

Our approach to ideology feature extraction involves three methodological steps. First we obtain and preprocess the social media data. Second we use Joint Sentiment-Topic (JST) modeling (a type of topic modeling) to extract topics from the social media data. Third we use principal-components analysis (PCA) on the JST results to reduce the dimensionality of the topic space.

Authors

• Daniel Ebanks, Harvard University

• R. Michael Alvarez, California Institute of Technology

• Betsy Sinclair, Washington University in St. Louis