Policy documents often embody complex reasoning that shapes governance outcomes. Participatory commitments and managerial control coexist within the same text, but the tensions between them are rarely articulated. Existing computational methods struggle to express the frame-mediated relations that characterize these tensions, where one argument narrows or instrumentalizes another rather than outright rejecting it.
We introduce Apaf, a hybrid LLM-symbolic pipeline that operationalizes critical discourse analysis as a quantitative bipolar argumentation framework for policy texts. Arguments are classified into deliberative or managerial frames, and four frame-mediated relation subtypes (agency reduction, agenda shift, instrumental support, and normative support) are generated through deterministic rules applied to LLM-extracted features.
Additionally, we release a novel dataset comprising 100 sub-documents of disaster-risk-reduction policies from the USA, UK, Canada, and Australia. Our findings demonstrate that the resulting argument graphs are accurate, interpretable, and stable across jurisdictions.
Blogger's Review: This study breaks the limitations of traditional methods in policy analysis by introducing a hybrid framework, offering a fresh perspective on the intricate relationships within policy texts. The precise argument graphs allow policymakers to better identify and address potential contradictions and challenges in governance.