Automated Institutional Grammar Coding of Flood Governance Texts: A Bilingual English–Korean Transformer-Based Approach
Description
Water resources planning and management often require coordinated action among multiple agencies working under different rules, plans, and procedures. These coordinated actions are structured by a governance architecture embodied in governing texts that assign responsibilities and specify procedures across agencies. Understanding this governance architecture is important for clarifying roles, identifying coordination gaps, and reducing the risk of operational failures. In practice, however, the large volume of governing texts, their distribution across jurisdictions, and the need to work with texts written in different languages make manual review difficult to sustain. This study develops and evaluates an automated approach for coding flood governance texts in English and Korean. Using a BERT based token classification model aligned with the Institutional Grammar, we identify the core components of rule statements, including actors, actions, obligations, affected objects, and applicable conditions. The model was trained and tested on 1,186 manually coded statements from 15 flood governance documents in two jurisdictions: reservoir governance in California and urban drainage and flood governance in South Korea. The approach achieved 82% token-level accuracy for English texts and 95% for Korean texts. These results show that automated analysis can make large collections of governing texts more tractable for planners and managers. To illustrate use of the coded output, we assemble the coded statements into institutional networks and compare rules across agencies and governance levels in the two regions, identifying areas where coordination is unclear, overlapping, or missing. Results from English and Korean texts also suggest that the approach can support cross national comparison of flood governance.