AI Framework Improves Regulation-Driven Fine-Grained Classification

A new framework has been developed to address the complex challenge of regulation-driven fine-grained classification, a task critical for areas like customs tariff classification and export control. Unlike standard text classification, these tasks require adherence to explicit regulatory hierarchies, rule-defined boundaries, and specific conditions, where semantic similarity alone is insufficient. Existing methods, including flat classifiers and retrieval-augmented LLMs, often fail to jointly enforce hierarchical validity, rule consistency, and fine-grained boundary reasoning. The proposed constraint-aware hierarchical search framework tackles this by converting regulatory documents into a searchable tree structure. It then retrieves only valid local candidate nodes and leverages structured regulatory fields alongside evidence snippets to guide each decision step. This approach ensures that the assigned label follows a valid path within the regulatory hierarchy and is supported by auditable evidence. Tested on four benchmark datasets derived from real-world regulation-intensive scenarios, the method achieved superior accuracy, particularly in cases involving fine-grained neighboring categories and rule-based boundary conditions, while also providing interpretable decision paths.