DocHRL Optimizes Document Classification with Hierarchical Reinforcement Learning.

Traditional document classification systems often apply a uniform processing pipeline to all documents, leading to inefficiencies where simple documents are over-processed and complex ones lack sufficient scrutiny. Researchers have introduced DocHRL, a novel framework leveraging hierarchical reinforcement learning to address this. This system intelligently adapts its classification strategy on a per-document basis. DocHRL operates with a two-tiered decision-making process. A high-level policy determines the general approach, such as employing vision classifiers, large language models, optical character recognition, or involving human review. Subsequently, a lower-level policy selects the specific model or tool within that chosen option. The framework is trained to minimize total expected costs, encompassing inference expenses, the financial impact of misclassifications, and human labeling costs. Evaluated on the RVL-CDIP benchmark, DocHRL achieved a high macro F1 score of 0.973 across 16 document classes. Crucially, it significantly reduced the average per-document processing cost to 2.74 normalized units, demonstrating substantial operational efficiency gains compared to fixed, standalone classifiers. This approach highlights the potential of cost-aware reinforcement learning to enhance both accuracy and resource utilization in document understanding.