Development planning documents in the potato industry provide essential guidance for production, technology upgrading, and industrial transformation, yet their hierarchical and context-dependent structures make automatic classification a challenging task. This study proposes a Policy Coherence Loss (PCL) to enhance the logical consistency of text classification models applied to development planning documents. PCL introduces a coherence-based regularization term that encourages semantically related sentences within the same policy role to remain close in the representation space while maintaining separation across distinct roles. Experiments on an annotated corpus of potato industry planning texts show that integrating PCL consistently improves classification accuracy and F1-score across multiple backbone models, confirming its effectiveness in capturing the semantic and logical coherence embedded in agricultural development planning.
Full publication URL
