Exploring Polarizing Political Discourse Among U.S. Congressional Members with Large Language Models

Abstract

In July 2024, a shooting incident involving President Trump drew widespread public attention, highlighting the need for a deeper understanding of political rhetoric among U.S. Congressional members. This study analyzes discourse patterns on Twitter using large language models (LLMs), specifically ChatGPT and bidirectional encoder representations from transformers (BERT), to explore underlying factors that may contribute to polarizing political language. By collecting and preprocessing Twitter data, we initially labeled 20,000 tweets using ChatGPT and then utilized the BERT-large model to classify the remaining 980,000 tweets. The analysis identified party affiliation and geographic region as significant factors influencing political rhetoric. Republican lawmakers exhibited a higher prevalence of polarizing language, while New Jersey recorded the highest rate among the states. Newly elected Congressional members also tended to adopt more provocative language, potentially as a strategy to engage with their voter base or distinguish themselves in a competitive political environment. Temporal analysis revealed spikes in polarizing rhetoric corresponding to events such as discussions on the new fiscal year budget. This study offers insights into the dynamics of political discourse, providing a foundation for promoting constructive dialogue and fostering institutional resilience.