Climate-NLI: A Model for Natural Language Inference and Zero-Shot Classification on Climate-Related Text

Abstract

Climate change is one of the most significant challenges of our era, necessitating innovative solutions across multiple fields. Advancements in NLP offer a promising pathway, particularly through the development of generalized models applicable to various tasks. Despite recent progress, specialized NLP models excel in individual tasks but require substantial domain-specific training data and fail to generalize well to new scenarios. This paper introduces the Climate-NLI, an approach that utilizes NLI models to create a versatile NLP model that can be used for fact-checking and text classification on climate-related text. Experiment results on 10 climate-related datasets show that our proposed model obtained comparable results to the models that have been fine-tuned on task-specific datasets. Our model improves adaptability to new classes by adding training samples without full retraining but struggles with certain classes due to limited related samples and similar but distinct concepts.