Language Models in Sustainability Governance: An AI-driven Policy Monitoring Framework for Hydrogen

Text data has been a longstanding key source for political science research. It provides an instructive and explanatory lens for understanding policy dynamics and governance evolution, for examining policy discourse and narratives, for analyzing policy development and governance behavior, and most importantly, for policy impact assessment. The significant role of text data in political research, along with the complexity and difficulty in numerically digesting unstructured data in natural languages, has motivated the use of natural language processing (NLP) to extract meaningful insights from large-scale datasets. NLP enables the analysis and structuring of unstructured textual data, allowing for a systematic examination of large collections of policy documents that would otherwise be difficult to assess manually. The emergence of large language models (LLMs) marked a new phase in computational political science research. By combining contextual understanding with generative capabilities, LLMs extend NLP toward more nuanced interpretation and synthesis of political texts. In this work, we explored how NLP and LLMs can supplement the policy analysis and impact assessment in an automated manner. For an AI-driven policy monitoring framework, we developed a tool that uses regulatory documents as input and conducts comprehensive policy analysis and impact assessment using retrieval-augmented generation (RAG), topic modeling through Latent Dirichlet Allocation (LDA), and semantic analysis. First, we developed a chatbot based on an RAG model, enabling transparent policy review and analysis. Next, a dashboard was developed, providing visual and textual insights through topic modeling and semantic analysis on policy instruments and binding targets. In general, our work illustrates the potential of artificial intelligence, particularly NLP and LLMs, for policy monitoring, moving beyond traditional approaches in political science, toward automated understanding of policy design, coherence, instruments, and cross-country differences in a way that manual reading or word search cannot. To show the applicability of the developed framework, we used the case of hydrogen, considering the rapid development of the global political hydrogen landscape. National hydrogen strategies serve as both planning and signaling regulatory frameworks, aiming to reflect each country’s objectives, priorities, institutional settings, and governance mechanisms. For this analysis, a global corpus of 67 national hydrogen strategies has been assembled, encompassing 15 languages and comprising 101 policy documents. However, the scale, heterogeneity, and complexity of these documents make it extremely difficult to obtain a clear picture of how countries design their strategies, where they differ or converge, and how policy priorities evolve. Traditional qualitative reading or keyword-based approaches cannot capture these patterns, nor can they handle multilingual corpora or provide systematic cross-country comparisons. This creates a pressing need for computational tools that can treat structure, identify policy instruments, and reveal temporal trends across large regulatory collections. Against this background, example outputs of our framework include: (i) identifying the evolution of policy focus areas, such as import/export, infrastructure, production, and R&D, and tracking their growth within more detailed sub-topics (e.g., economic viability, safety, environmental impacts); (ii) mapping policy goal horizons across focus categories; and (iii) deriving sentiment profiles to assess the tone and direction of policy commitments.