Abstract

Misinformation and disinformation on the internet present a significant challenge in the context of climate change debate. The dissemination of false or misleading information can hinder public understanding and impede efforts to combat the growing issue of climate change. While social media platforms have implemented automatic fact-checking algorithms, existing models lack domain-specific training to effectively verify climate change-related information. As a remedy, a new fact-checking dataset is proposed that combines data from CLIMATE-FEVER with web-scraped information, resulting in a comprehensive dataset comprising 8,115 annotated claim-evidence pairs. The improved dataset is used to fine-tune a variety of pre-trained transformers for climate claim verification tasks. The best model, RoBERTa, achieved an accuracy of 0.7288 and F1-score of 0.7229, improving upon previously reported state-of-the-art (SoTA) F1-score of 0.7182.

Please note that any unauthorized copying or distribution is strictly prohibited. All rights reserved.