Vinicius Woloszyn, Joseph Kobti, Vera Schmitt - Proceedings of the 13th Annual Meeting of the Forum for Information Retrieval Evaluation
Companies frequently make claims about positive impacts on the environment, but these claims are not always valid. In this sense, “greenwashing” is used when a company states a false claim about its products and practices being environmentally friendly. It is an important issue and has attracted the attention of policymakers to strengthen consumer law and new companies with the mission of helping brands and consumers to become more eco-friendly. However, manual screening of websites and social networks for sustainable claims (green claims) is time-consuming. Automatic detection of green claims is an underexplored problem from a computer science perspective, and thus, we present the design, training and evaluation of different approaches in this study. Our experiments reveal that although pre-trained models present high performance, they also show sensibility to adversarial attacks, such as character-swap-based methods, which are common in social networks. In order to understand the applicability in a real-world scenario, we also evaluated its generalization performance, which showed a notable performance across different domains.