Revealing a Country’s Government Discourse Through BERT-based Topic Modeling in the US Presidential Speeches

This paper analyzes presidential speeches from the United States, delivered by the current president, Joe Biden, from his inauguration until April 3, 2024, using BERT-based topic modeling techniques. Since the role and actions of the United States affect global affairs, inferring latent topics from these speeches are important for making informed decisions and also useful to researchers. We provide experiments combining different embedding methods and clustering algorithms for dimensionality reduction. Our experiment results demonstrated that using DistilBERT embeddings followed by HDBSCAN clustering and UMAP dimensionality reduction consistently produced meaningful and interpretable topics, and this paper introduces the UMASS and NPMI metrics for evaluating topic coherence, with the best values of these metrics among the other schemes investigated. Some of the dominant themes conveyed in the American president’s state of the union speech that were successfully revealed in this study include international cooperation, alongside significant attention to environmental, community, health, and economic issues.