Disaster preparedness plans must not only be logistically sound but also institutionally feasible. Institutions are critical to assure both public acceptance of disaster response strategies and their alignment with the existing regulations, both of which improve the effectiveness of response and eventually reduce the economic costs of disasters.

A significant problem within institutional frameworks is the imprecise allocation of tasks and division of responsibilities in policy and planning documents, leading to poor disaster preparedness. For instance, during flood disaster management, overlapping or even conflicting procedures between federal, state, and local agencies can cause significant delays and inefficiencies. Or, legal disputes over land use and water management rights can prevent the timely construction of flood defenses or delay evacuation orders. These institutional voids and conflicts create bottlenecks that hinder timely and effective disaster response, escalating economic costs due to redundant efforts or misallocation of resources.

LEARNS aims to facilitate the identification and analysis of institutional discrepancies in disaster management policies and plans. This form of discrepancy analysis is currently done manually, which is cumbersome, inaccurate, and even infeasible with a larger number of formal documents. The LEARNS project will therefore automate this process with document coding and analysis, which will make the analysis more efficient, consistent, and scalable.  In order to achieve this, the project team will:

1. Analyze the requirements for integrating Institutional Grammar and network analysis with the latest innovations in large language models (LLMs);
2. Select and benchmark the most suitable LLM (open-pretrained-foundation and instruct models, e.g., Llama, Claude, and BERT);
3. Optimize and prompt-engineer the selected model, and evaluate its performance against standard data to achieve an acceptable level of consistency and accuracy.

To demonstrate the efficacy and impact of our approach, we will conduct a complete analysis of disaster management documents in Chennai, India. Chennai has experienced major urban flooding events in recent years, prompting million-dollar international project investments in disaster resilience. Chennai is selected for its complex governance and multi-actor environment and the abundance of strategy documents, presenting the perfect opportunity to demonstrate the ability to identify and resolve ambiguities in disaster management policies.

The resulting open-source LLM model from the LEARNS project will enable a scalable analysis of the specific legal, administrative, and operational barriers to successful disaster management, which can be further used to analyze local cases of disaster management (e.g., Limburg) and the COVID-19 pandemic. Ultimately, the model will strengthen our disaster response, making it more timely and effective while reducing economic costs through more efficient use of resources.

Collaboration & funding

The LEARNS cross-fertilizes with the DISINT project.

The project is funded by the Pandemic and Disaster Preparedness Center (PDPC), EsCHER and the SCBH networks.