Unsupervised NLP for quantifying sentiment deviation and framing in global climate discourse

Abstract

Climate change news coverage varies by region and media organization. These differences show in the way news is framed, the themes chosen, and communication priorities. This study suggests a scalable and unsupervised natural language processing (NLP) framework to measure sentiment changes and framing patterns in global climate discussions across major news outlets. The proposed system uses Zero-Shot BART-large-MNLI as the main sentiment classification model, while VADER serves as a comparative baseline during model validation. Key themes are identified using BERTopic. We detect changes in climate narratives over time with the Pruned Exact Linear Time (PELT) changepoint detection algorithm. We introduce a dual-baseline framework to calculate relative sentiment changes. This combines regional consensus with a scientific baseline. It allows comparison of framing patterns without hiding systematic regional differences. The BART-MNLI model was tested against a manually annotated sample of articles. It achieved an accuracy of 98.0%, an F1-score of 0.98, and Cohen’s κ of 0.96, showing strong agreement with human annotations. It significantly outperformed lexicon-based sentiment analysis for professional news text. We also assessed statistical robustness with the Kruskal-Wallis H-test, bootstrap confidence intervals, and PELT sensitivity analysis. The framework reveals regional sentiment differences, outlet-level framing patterns, thematic trends, and notable shifts in response to major climate events. It offers a clear, reproducible, and flexible method for analyzing global climate communication.