Comparative Climate Change Sentiment Analysis Using Machine Learning

This paper investigates how sentiments towards climate change are expressed in YouTube comments and how they vary over time across major USA and European news channels. A supervised sentiment classifier is developed that combines a RoBERTa backbone adapted for socialmedia text with a convolutional neural network (CNN) head discovered through multitask neural architecture search (MTNAS). The model is tuned on four public <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{X}$</tex> and YouTube datasets and then refined using domain-adaptive pre-training (DAPT) on a large corpus of general YouTube comments with parameter-efficient adapters. The final classifier attains 89.73 % validation accuracy on a three-class YouTube benchmark (positive, neutral, negative) while remaining operationally lightweight. It is applied to top-level comments posted between 2010 and 2025 under climate-related videos from eight USA and eight European mainstream news channels. Negative comments form the largest share across the full period, with sentiments stabilising from around 2017 onwards. USA channels exhibit a higher overall share of negativity than European channels, driven by a few very large outlets that contribute a disproportionate volume of comments.