Contrastive Approaches to Climate Change Narrative Classification

Abstract

Climate change is a topic surrounded by many different narratives, either giving ex- planations to the problem and describing solutions, or denying one or more aspects of it. In recent years, there has been development in natural language processing (NLP) towards computational models for classifying texts by narrative or misinformation claim regarding climate change, following a hierarchical taxonomy of labels. However, training models using such taxonomies enforce predictions confined to the predefined taxonomy. Ideally, we would like to approach the task in a less supervised way, letting a model generalize to predict outside of the taxonomy it has been trained on, since the climate change discourse changes through time and between domains. This thesis aims at finding models whose representations are useful across and beyond taxonomies. As a mean for this, we use the contrastive learning approach as seen in Sentence-BERT, and investigate how it can be used for multiclass and multi-label climate change narrative classification by re-framing the task as a clustering task. We compare contrastive models to classifiers built on BERT, including cross-dataset and taxonomy performance. The results show that both classifiers and contrastive models along with clustering can be effectively used for accurate predictions on the task they are trained on. Contrastive models furthermore have the advantage of being able to predict on several label granularities and to provide interpretable document similarity scores. However, when evaluating on a dataset with a different taxonomy altogether, we find that both classifiers and contrastive models perform poorly.