Development planning texts in agriculture are layered, technical, and rhetorically varied, which makes reliable role classification nontrivial yet essential for policy analytics. In the potato sector, separating strategic goals, implementation measures, support conditions, and evaluation statements enables quantitative assessment of priorities and execution readiness for governance. This work introduces the Policy-Aware Optimal Transport (PAOT) Loss, which frames learning as an optimal-transport alignment between sentence representations and ontology-guided role prototypes. PAOT integrates lexical supervision with prototype-based semantic grounding, counterfactual invariance to trigger words, and document-position priors within a single cost function, avoiding manual weighting and preserving interpretability. Experiments across convolutional and transformer backbones show consistent gains over cross-entropy in accuracy, precision, and F1, alongside reduced confusion among semantically adjacent roles. A full ablation over the semantic, invariance, and structural components confirms their complementary effects, and transport plans offer transparent diagnostics of sentence–role alignment. Beyond model performance, PAOT provides a computable lens for governance in the potato industry by supporting reproducible analysis of development plans, clarifying how long-term aims are connected to concrete actions and enabling conditions, and facilitating evidence-based monitoring and accountability in sectoral public administration.
Full publication URL
