Evaluating Environment & Climate Truthfulness in Social Media using Deep Learning & Large Language Models (LLMs)
Awarded Best Dissertation in Cohort, this MSc project explores the detection of climate and environmental misinformation on social media using a comparative framework of traditional natural language processing techniques, deep learning, and Large Language Models (LLMs). Leveraging a web-scraped dataset from PolitiFact, the study highlights the superiority of CNNs trained on ordinal truthfulness data, with accuracy boosted from 80.1% to 84.0% through GPT-4o-driven feature augmentation. While LLMs enhanced contextual understanding and sentiment analysis, their time complexity posed practical limitations. The project contributes novel insights into model performance trade-offs, evaluation metrics tailored to ordinal classification, and the practical integration of LLMs for misinformation mitigation in climate discourse.