Topic Extraction from Eminent Domain Adjudications

Abstract Legal documents are widely reputed to be long and complex, making the manual identification of key topics time-consuming and resource intensive. However, due to the recent advent of large language models (LLMs), topic modeling has become more accessible and easier. This study aims to efficiently extract topics from eminent domain adjudications of the Land Tribunal in South Korea by using two popular topic extraction models: Latent Dirichlet Allocation (LDA) and Bidirectional Encoder Representations from Transformers Topic (BERTopic). We evaluate the topics identified by LDA and BERTopic using both quantitative (topic coherence metric) and qualitative (domain-specific knowledge) methods. The qualitative approach yielded a more reliable assessment than the quantitative one, revealing that BERTopic performed best when applied to English adjudication text. This model extracted eight meaningful topics, including land compensation, farming and fishing compensation, appeals for exclusion, and relocation issues. For legal information professionals, the findings demonstrate how advanced topic modeling can streamline the retrieval and analysis of legal texts, enabling efficient access to precedents and ultimately supporting informed decision-making in specialized domains.