This study analyzes congressional floor speeches from 1994 to 2024 to assess climate-related claims using a comprehensive dataset. We employed a Congressional Record scraper to gather 2,515,806 paragraphs, classifying 110,837 as climate-related with the ClimateBERT model. A rigorous annotation process was undertaken for 2,151 paragraphs to establish a performance benchmark, revealing a 79.8% inter-coder agreement. Statistical analysis included Bayesian mixed effects models, capturing demographic and political covariates such as party affiliation and fossil fuel campaign contributions. The study introduced an advanced machine learning framework, enhancing classification through in-context learning and dynamic few-shot methods, resulting in a highly efficient CARDS model. This model delivered exceptional performance akin to leading models but at significantly reduced costs. The research illustrates the intricate relationships between political speech, economic factors, and climate change discourse, contributing valuable insights into congressional responses to climate issues.
Source link
