Defense of Daniela Flores

Abstract

Sustainability has gained much attention in recent years when we have started to see the effects of climate change and environmental damage, and action needs to be taken to keep inhabiting planet Earth. In this thesis, we explore using state-of-the-art Transformer-based language models to develop a Sustainable Development Goals (SDGs) classifier for academic articles. This model could lead academic institutions to measure their contribution to Sustainability and promote collaboration between interdisciplinary researchers to tackle current world challenges using knowledge from different fields. We propose a fine-tuned RoBERTa model that reaches an f1-score of 73\%. We also studied two Explainable Artificial Intelligence techniques to better understand the model predictions, Attention Mechanism and Integrated Gradients. Finally, we conducted a user study to discover the best explanation method between the two using text visualization techniques. We concluded that Attention Mechanism visualizations better help the users understand model predictions, even when said predictions are erroneous.