Apaf: Hybrid LLM-Symbolic Framework Reveals Hidden Tensions in Disaster Policy

Policy documents often contain implicit tensions between participatory commitments and managerial control, but existing computational approaches lack the structure to make these conflicts explicit. End-to-end LLM summarization produces fluent text that domain experts cannot easily inspect or contest. Vasileiou and Derendiaeva address this gap with Apaf, a hybrid framework that combines LLM-powered feature extraction with symbolic reasoning rules. The system first classifies arguments into deliberative (participatory) or managerial frames, then identifies four subtypes of frame-mediated relations: agency reduction, agenda shift, instrumental support, and normative support. These relations capture how one argument narrows or instrumentalizes another without outright rejection.

The researchers created a novel dataset of 100 sub-documents from disaster-risk-reduction policies across the USA, UK, Canada, and Australia. Experimental results show that Apaf produces argument graphs that are not only accurate but also interpretable by domain experts and stable across different jurisdictions. This hybrid approach bridges the gap between black-box LLM summaries and the structured, contestable analysis needed for real policy evaluation. By making implicit tensions explicit, Apaf empowers policymakers and analysts to scrutinize the reasoning behind governance documents with greater confidence and precision.