Content area
Countries regularly publish strategies for economic development, healthcare, transportation, and other areas of importance for improving governance and the lives of their citizens. As Artificial Intelligence (AI) becomes an increasingly important technology, countries and entities within them are formulating and publishing AI strategy documents. This dissertation introduces novel machine learning-based methodologies that leverage representation learning to extract and compare the values and priorities in these documents. Central to this research is a novel framework that treats AI strategies as narratives and the latent values and priorities in them as themes or topics to be extracted from text through computation. A novel Latent Dirichlet Allocation (LDA)-based method, STAR-LDA, is developed to enhance the stability and robustness of topic extraction. The nearest-neighbor graph construct is then leveraged to capture the connectivity of the topic representation space revealed by STAR-LDA. Finally, large language models are incorporated in creative ways to both semantically enhance the topic representation space as well as aid in validating it, better connecting knowledge produced by humans and machines. With varied applications across national and sub-national AI policies, ethical AI, and health-equity, this dissertation advances our understanding of global AI governance, bridges computational and policy research, and offers rigorous methods for analyzing complex policy landscapes.