Misinformation surrounding climate change (CC) proliferates across the internet at such rapid speeds and in such large quantities that human fact-checkers are unable to feasibly verify the veracity of most online CC-related information. While automatic fact-checking algorithms can supplement human factchecking efforts, existing models suffer from a lack of domainspecific training data to robustly fact-check CC information. To address this problem, we tailor an existing automatic factchecking system to the CC domain by introducing the popular semi-supervised training method, Unsupervised Data Augmentation (UDA), into our system's pipeline, allowing us to leverage large amounts of unlabeled CC-related claims. We evaluate our fact-checking model on the CC fact-checking dataset CLIMATE-FEVER, yielding a state-of-the-art (SotA) F1 score of 0.7182, improving upon the previously reported SotA F1 score of 0.3285.

PDF